- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
/*
|
|
|
|
* PROJECT: ReactOS Kernel
|
|
|
|
* LICENSE: BSD - See COPYING.ARM in the top level directory
|
|
|
|
* FILE: ntoskrnl/mm/ARM3/mdlsup.c
|
|
|
|
* PURPOSE: ARM Memory Manager Memory Descriptor List (MDL) Management
|
|
|
|
* PROGRAMMERS: ReactOS Portable Systems Group
|
|
|
|
*/
|
|
|
|
|
|
|
|
/* INCLUDES *******************************************************************/
|
|
|
|
|
|
|
|
#include <ntoskrnl.h>
|
|
|
|
#define NDEBUG
|
|
|
|
#include <debug.h>
|
|
|
|
|
|
|
|
#define MODULE_INVOLVED_IN_ARM3
|
2014-11-10 16:26:55 +00:00
|
|
|
#include <mm/ARM3/miarm.h>
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* GLOBALS ********************************************************************/
|
2010-12-26 15:23:03 +00:00
|
|
|
|
2010-04-20 22:47:51 +00:00
|
|
|
BOOLEAN MmTrackPtes;
|
|
|
|
BOOLEAN MmTrackLockedPages;
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
SIZE_T MmSystemLockPagesCount;
|
2010-04-20 22:47:51 +00:00
|
|
|
|
2018-10-01 11:47:06 +00:00
|
|
|
ULONG MiCacheOverride[MiNotMapped + 1];
|
|
|
|
|
2017-06-18 11:18:41 +00:00
|
|
|
/* INTERNAL FUNCTIONS *********************************************************/
|
|
|
|
static
|
|
|
|
PVOID
|
|
|
|
NTAPI
|
|
|
|
MiMapLockedPagesInUserSpace(
|
|
|
|
_In_ PMDL Mdl,
|
|
|
|
_In_ PVOID StartVa,
|
|
|
|
_In_ MEMORY_CACHING_TYPE CacheType,
|
|
|
|
_In_opt_ PVOID BaseAddress)
|
|
|
|
{
|
|
|
|
NTSTATUS Status;
|
|
|
|
PEPROCESS Process = PsGetCurrentProcess();
|
|
|
|
PETHREAD Thread = PsGetCurrentThread();
|
|
|
|
TABLE_SEARCH_RESULT Result;
|
|
|
|
MI_PFN_CACHE_ATTRIBUTE CacheAttribute;
|
2018-10-01 11:47:06 +00:00
|
|
|
MI_PFN_CACHE_ATTRIBUTE EffectiveCacheAttribute;
|
2017-06-18 11:18:41 +00:00
|
|
|
BOOLEAN IsIoMapping;
|
|
|
|
KIRQL OldIrql;
|
|
|
|
ULONG_PTR StartingVa;
|
|
|
|
ULONG_PTR EndingVa;
|
|
|
|
PMMADDRESS_NODE Parent;
|
|
|
|
PMMVAD_LONG Vad;
|
|
|
|
ULONG NumberOfPages;
|
|
|
|
PMMPTE PointerPte;
|
|
|
|
PMMPDE PointerPde;
|
|
|
|
MMPTE TempPte;
|
|
|
|
PPFN_NUMBER MdlPages;
|
|
|
|
PMMPFN Pfn1;
|
|
|
|
PMMPFN Pfn2;
|
|
|
|
BOOLEAN AddressSpaceLocked = FALSE;
|
|
|
|
|
|
|
|
PAGED_CODE();
|
|
|
|
|
|
|
|
DPRINT("MiMapLockedPagesInUserSpace(%p, %p, 0x%x, %p)\n",
|
|
|
|
Mdl, StartVa, CacheType, BaseAddress);
|
|
|
|
|
|
|
|
NumberOfPages = ADDRESS_AND_SIZE_TO_SPAN_PAGES(StartVa,
|
|
|
|
MmGetMdlByteCount(Mdl));
|
|
|
|
MdlPages = MmGetMdlPfnArray(Mdl);
|
|
|
|
|
|
|
|
ASSERT(CacheType <= MmWriteCombined);
|
|
|
|
|
|
|
|
IsIoMapping = (Mdl->MdlFlags & MDL_IO_SPACE) != 0;
|
|
|
|
CacheAttribute = MiPlatformCacheAttributes[IsIoMapping][CacheType];
|
|
|
|
|
|
|
|
/* Large pages are always cached, make sure we're not asking for those */
|
|
|
|
if (CacheAttribute != MiCached)
|
|
|
|
{
|
|
|
|
DPRINT1("FIXME: Need to check for large pages\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Allocate a VAD for our mapped region */
|
|
|
|
Vad = ExAllocatePoolWithTag(NonPagedPool, sizeof(MMVAD_LONG), 'ldaV');
|
|
|
|
if (Vad == NULL)
|
|
|
|
{
|
|
|
|
Status = STATUS_INSUFFICIENT_RESOURCES;
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Initialize PhysicalMemory VAD */
|
|
|
|
RtlZeroMemory(Vad, sizeof(*Vad));
|
|
|
|
Vad->u2.VadFlags2.LongVad = 1;
|
|
|
|
Vad->u.VadFlags.VadType = VadDevicePhysicalMemory;
|
|
|
|
Vad->u.VadFlags.Protection = MM_READWRITE;
|
|
|
|
Vad->u.VadFlags.PrivateMemory = 1;
|
|
|
|
|
|
|
|
/* Did the caller specify an address? */
|
|
|
|
if (BaseAddress == NULL)
|
|
|
|
{
|
|
|
|
/* We get to pick the address */
|
|
|
|
MmLockAddressSpace(&Process->Vm);
|
|
|
|
AddressSpaceLocked = TRUE;
|
|
|
|
if (Process->VmDeleted)
|
|
|
|
{
|
|
|
|
Status = STATUS_PROCESS_IS_TERMINATING;
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
|
|
|
|
Result = MiFindEmptyAddressRangeInTree(NumberOfPages << PAGE_SHIFT,
|
|
|
|
MM_VIRTMEM_GRANULARITY,
|
|
|
|
&Process->VadRoot,
|
|
|
|
&Parent,
|
|
|
|
&StartingVa);
|
|
|
|
if (Result == TableFoundNode)
|
|
|
|
{
|
|
|
|
Status = STATUS_NO_MEMORY;
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
EndingVa = StartingVa + NumberOfPages * PAGE_SIZE - 1;
|
|
|
|
BaseAddress = (PVOID)StartingVa;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* Caller specified a base address */
|
|
|
|
StartingVa = (ULONG_PTR)BaseAddress;
|
|
|
|
EndingVa = StartingVa + NumberOfPages * PAGE_SIZE - 1;
|
|
|
|
|
|
|
|
/* Make sure it's valid */
|
|
|
|
if (BYTE_OFFSET(StartingVa) != 0 ||
|
|
|
|
EndingVa <= StartingVa ||
|
|
|
|
EndingVa > (ULONG_PTR)MM_HIGHEST_VAD_ADDRESS)
|
|
|
|
{
|
|
|
|
Status = STATUS_INVALID_ADDRESS;
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
|
|
|
|
MmLockAddressSpace(&Process->Vm);
|
|
|
|
AddressSpaceLocked = TRUE;
|
|
|
|
if (Process->VmDeleted)
|
|
|
|
{
|
|
|
|
Status = STATUS_PROCESS_IS_TERMINATING;
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Check if it's already in use */
|
|
|
|
Result = MiCheckForConflictingNode(StartingVa >> PAGE_SHIFT,
|
|
|
|
EndingVa >> PAGE_SHIFT,
|
|
|
|
&Process->VadRoot,
|
|
|
|
&Parent);
|
|
|
|
if (Result == TableFoundNode)
|
|
|
|
{
|
|
|
|
Status = STATUS_CONFLICTING_ADDRESSES;
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Vad->StartingVpn = StartingVa >> PAGE_SHIFT;
|
|
|
|
Vad->EndingVpn = EndingVa >> PAGE_SHIFT;
|
|
|
|
|
|
|
|
MiLockProcessWorkingSetUnsafe(Process, Thread);
|
|
|
|
|
|
|
|
ASSERT(Vad->EndingVpn >= Vad->StartingVpn);
|
|
|
|
|
|
|
|
MiInsertVad((PMMVAD)Vad, &Process->VadRoot);
|
|
|
|
|
|
|
|
/* Check if this is uncached */
|
|
|
|
if (CacheAttribute != MiCached)
|
|
|
|
{
|
|
|
|
/* Flush all caches */
|
|
|
|
KeFlushEntireTb(TRUE, TRUE);
|
|
|
|
KeInvalidateAllCaches();
|
|
|
|
}
|
|
|
|
|
|
|
|
PointerPte = MiAddressToPte(BaseAddress);
|
|
|
|
while (NumberOfPages != 0 &&
|
|
|
|
*MdlPages != LIST_HEAD)
|
|
|
|
{
|
|
|
|
PointerPde = MiPteToPde(PointerPte);
|
|
|
|
MiMakePdeExistAndMakeValid(PointerPde, Process, MM_NOIRQL);
|
|
|
|
ASSERT(PointerPte->u.Hard.Valid == 0);
|
|
|
|
|
|
|
|
/* Add a PDE reference for each page */
|
|
|
|
MiIncrementPageTableReferences(BaseAddress);
|
|
|
|
|
|
|
|
/* Set up our basic user PTE */
|
|
|
|
MI_MAKE_HARDWARE_PTE_USER(&TempPte,
|
|
|
|
PointerPte,
|
|
|
|
MM_READWRITE,
|
|
|
|
*MdlPages);
|
|
|
|
|
2018-10-01 11:47:06 +00:00
|
|
|
EffectiveCacheAttribute = CacheAttribute;
|
|
|
|
|
|
|
|
/* We need to respect the PFN's caching information in some cases */
|
2017-06-18 11:18:41 +00:00
|
|
|
Pfn2 = MiGetPfnEntry(*MdlPages);
|
|
|
|
if (Pfn2 != NULL)
|
|
|
|
{
|
|
|
|
ASSERT(Pfn2->u3.e2.ReferenceCount != 0);
|
|
|
|
|
2018-10-01 11:47:06 +00:00
|
|
|
switch (Pfn2->u3.e1.CacheAttribute)
|
2017-06-18 11:18:41 +00:00
|
|
|
{
|
2018-10-01 11:47:06 +00:00
|
|
|
case MiNonCached:
|
|
|
|
if (CacheAttribute != MiNonCached)
|
|
|
|
{
|
|
|
|
MiCacheOverride[1]++;
|
|
|
|
EffectiveCacheAttribute = MiNonCached;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case MiCached:
|
|
|
|
if (CacheAttribute != MiCached)
|
|
|
|
{
|
|
|
|
MiCacheOverride[0]++;
|
|
|
|
EffectiveCacheAttribute = MiCached;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case MiWriteCombined:
|
|
|
|
if (CacheAttribute != MiWriteCombined)
|
|
|
|
{
|
|
|
|
MiCacheOverride[2]++;
|
|
|
|
EffectiveCacheAttribute = MiWriteCombined;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
/* We don't support AWE magic (MiNotMapped) */
|
|
|
|
DPRINT1("FIXME: MiNotMapped is not supported\n");
|
|
|
|
ASSERT(FALSE);
|
|
|
|
break;
|
2017-06-18 11:18:41 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Configure caching */
|
2018-10-01 11:47:06 +00:00
|
|
|
switch (EffectiveCacheAttribute)
|
2017-06-18 11:18:41 +00:00
|
|
|
{
|
|
|
|
case MiNonCached:
|
|
|
|
MI_PAGE_DISABLE_CACHE(&TempPte);
|
|
|
|
MI_PAGE_WRITE_THROUGH(&TempPte);
|
|
|
|
break;
|
|
|
|
case MiCached:
|
|
|
|
break;
|
|
|
|
case MiWriteCombined:
|
|
|
|
MI_PAGE_DISABLE_CACHE(&TempPte);
|
|
|
|
MI_PAGE_WRITE_COMBINED(&TempPte);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
ASSERT(FALSE);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Make the page valid */
|
|
|
|
MI_WRITE_VALID_PTE(PointerPte, TempPte);
|
|
|
|
|
|
|
|
/* Acquire a share count */
|
|
|
|
Pfn1 = MI_PFN_ELEMENT(PointerPde->u.Hard.PageFrameNumber);
|
2017-11-21 22:33:42 +00:00
|
|
|
OldIrql = MiAcquirePfnLock();
|
2017-06-18 11:18:41 +00:00
|
|
|
Pfn1->u2.ShareCount++;
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
2017-06-18 11:18:41 +00:00
|
|
|
|
|
|
|
/* Next page */
|
|
|
|
MdlPages++;
|
|
|
|
PointerPte++;
|
|
|
|
NumberOfPages--;
|
|
|
|
BaseAddress = (PVOID)((ULONG_PTR)BaseAddress + PAGE_SIZE);
|
|
|
|
}
|
|
|
|
|
|
|
|
MiUnlockProcessWorkingSetUnsafe(Process, Thread);
|
|
|
|
ASSERT(AddressSpaceLocked);
|
|
|
|
MmUnlockAddressSpace(&Process->Vm);
|
|
|
|
|
|
|
|
ASSERT(StartingVa != 0);
|
|
|
|
return (PVOID)((ULONG_PTR)StartingVa + MmGetMdlByteOffset(Mdl));
|
|
|
|
|
|
|
|
Error:
|
|
|
|
if (AddressSpaceLocked)
|
|
|
|
{
|
|
|
|
MmUnlockAddressSpace(&Process->Vm);
|
|
|
|
}
|
|
|
|
if (Vad != NULL)
|
|
|
|
{
|
|
|
|
ExFreePoolWithTag(Vad, 'ldaV');
|
|
|
|
}
|
|
|
|
ExRaiseStatus(Status);
|
|
|
|
}
|
|
|
|
|
|
|
|
static
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MiUnmapLockedPagesInUserSpace(
|
|
|
|
_In_ PVOID BaseAddress,
|
|
|
|
_In_ PMDL Mdl)
|
|
|
|
{
|
|
|
|
PEPROCESS Process = PsGetCurrentProcess();
|
|
|
|
PETHREAD Thread = PsGetCurrentThread();
|
|
|
|
PMMVAD Vad;
|
|
|
|
PMMPTE PointerPte;
|
|
|
|
PMMPDE PointerPde;
|
|
|
|
KIRQL OldIrql;
|
|
|
|
ULONG NumberOfPages;
|
|
|
|
PPFN_NUMBER MdlPages;
|
|
|
|
PFN_NUMBER PageTablePage;
|
|
|
|
|
|
|
|
DPRINT("MiUnmapLockedPagesInUserSpace(%p, %p)\n", BaseAddress, Mdl);
|
|
|
|
|
|
|
|
NumberOfPages = ADDRESS_AND_SIZE_TO_SPAN_PAGES(MmGetMdlVirtualAddress(Mdl),
|
|
|
|
MmGetMdlByteCount(Mdl));
|
|
|
|
ASSERT(NumberOfPages != 0);
|
|
|
|
MdlPages = MmGetMdlPfnArray(Mdl);
|
|
|
|
|
|
|
|
/* Find the VAD */
|
|
|
|
MmLockAddressSpace(&Process->Vm);
|
|
|
|
Vad = MiLocateAddress(BaseAddress);
|
|
|
|
if (!Vad ||
|
|
|
|
Vad->u.VadFlags.VadType != VadDevicePhysicalMemory)
|
|
|
|
{
|
|
|
|
DPRINT1("MiUnmapLockedPagesInUserSpace invalid for %p\n", BaseAddress);
|
|
|
|
MmUnlockAddressSpace(&Process->Vm);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
MiLockProcessWorkingSetUnsafe(Process, Thread);
|
|
|
|
|
|
|
|
/* Remove it from the process VAD tree */
|
|
|
|
ASSERT(Process->VadRoot.NumberGenericTableElements >= 1);
|
|
|
|
MiRemoveNode((PMMADDRESS_NODE)Vad, &Process->VadRoot);
|
|
|
|
|
|
|
|
/* MiRemoveNode should have removed us if we were the hint */
|
|
|
|
ASSERT(Process->VadRoot.NodeHint != Vad);
|
|
|
|
|
|
|
|
PointerPte = MiAddressToPte(BaseAddress);
|
2017-11-21 22:33:42 +00:00
|
|
|
OldIrql = MiAcquirePfnLock();
|
2017-06-18 11:18:41 +00:00
|
|
|
while (NumberOfPages != 0 &&
|
|
|
|
*MdlPages != LIST_HEAD)
|
|
|
|
{
|
|
|
|
ASSERT(MiAddressToPte(PointerPte)->u.Hard.Valid == 1);
|
|
|
|
ASSERT(PointerPte->u.Hard.Valid == 1);
|
|
|
|
|
|
|
|
/* Dereference the page */
|
|
|
|
MiDecrementPageTableReferences(BaseAddress);
|
|
|
|
|
|
|
|
/* Invalidate it */
|
|
|
|
MI_ERASE_PTE(PointerPte);
|
|
|
|
|
|
|
|
/* We invalidated this PTE, so dereference the PDE */
|
|
|
|
PointerPde = MiAddressToPde(BaseAddress);
|
|
|
|
PageTablePage = PointerPde->u.Hard.PageFrameNumber;
|
|
|
|
MiDecrementShareCount(MiGetPfnEntry(PageTablePage), PageTablePage);
|
|
|
|
|
|
|
|
/* Next page */
|
|
|
|
PointerPte++;
|
|
|
|
NumberOfPages--;
|
|
|
|
BaseAddress = (PVOID)((ULONG_PTR)BaseAddress + PAGE_SIZE);
|
|
|
|
MdlPages++;
|
|
|
|
|
|
|
|
/* Moving to a new PDE? */
|
|
|
|
if (PointerPde != MiAddressToPde(BaseAddress))
|
|
|
|
{
|
|
|
|
/* See if we should delete it */
|
|
|
|
KeFlushProcessTb();
|
|
|
|
PointerPde = MiPteToPde(PointerPte - 1);
|
|
|
|
ASSERT(PointerPde->u.Hard.Valid == 1);
|
|
|
|
if (MiQueryPageTableReferences(BaseAddress) == 0)
|
|
|
|
{
|
|
|
|
ASSERT(PointerPde->u.Long != 0);
|
|
|
|
MiDeletePte(PointerPde,
|
|
|
|
MiPteToAddress(PointerPde),
|
|
|
|
Process,
|
|
|
|
NULL);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
KeFlushProcessTb();
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
2017-06-18 11:18:41 +00:00
|
|
|
MiUnlockProcessWorkingSetUnsafe(Process, Thread);
|
|
|
|
MmUnlockAddressSpace(&Process->Vm);
|
|
|
|
ExFreePoolWithTag(Vad, 'ldaV');
|
|
|
|
}
|
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
/* PUBLIC FUNCTIONS ***********************************************************/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
PMDL
|
|
|
|
NTAPI
|
|
|
|
MmCreateMdl(IN PMDL Mdl,
|
|
|
|
IN PVOID Base,
|
2010-07-15 23:56:53 +00:00
|
|
|
IN SIZE_T Length)
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
{
|
2010-07-15 23:56:53 +00:00
|
|
|
SIZE_T Size;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if we don't have an MDL built
|
|
|
|
//
|
|
|
|
if (!Mdl)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Calculate the size we'll need and allocate the MDL
|
|
|
|
//
|
|
|
|
Size = MmSizeOfMdl(Base, Length);
|
|
|
|
Mdl = ExAllocatePoolWithTag(NonPagedPool, Size, TAG_MDL);
|
|
|
|
if (!Mdl) return NULL;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Initialize it
|
|
|
|
//
|
|
|
|
MmInitializeMdl(Mdl, Base, Length);
|
|
|
|
return Mdl;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
2010-07-15 23:56:53 +00:00
|
|
|
SIZE_T
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
NTAPI
|
|
|
|
MmSizeOfMdl(IN PVOID Base,
|
2010-07-15 23:56:53 +00:00
|
|
|
IN SIZE_T Length)
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
{
|
|
|
|
//
|
|
|
|
// Return the MDL size
|
|
|
|
//
|
|
|
|
return sizeof(MDL) +
|
|
|
|
(ADDRESS_AND_SIZE_TO_SPAN_PAGES(Base, Length) * sizeof(PFN_NUMBER));
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmBuildMdlForNonPagedPool(IN PMDL Mdl)
|
|
|
|
{
|
|
|
|
PPFN_NUMBER MdlPages, EndPage;
|
|
|
|
PFN_NUMBER Pfn, PageCount;
|
|
|
|
PVOID Base;
|
|
|
|
PMMPTE PointerPte;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity checks
|
|
|
|
//
|
|
|
|
ASSERT(Mdl->ByteCount != 0);
|
|
|
|
ASSERT((Mdl->MdlFlags & (MDL_PAGES_LOCKED |
|
|
|
|
MDL_MAPPED_TO_SYSTEM_VA |
|
|
|
|
MDL_SOURCE_IS_NONPAGED_POOL |
|
2010-12-26 15:23:03 +00:00
|
|
|
MDL_PARTIAL)) == 0);
|
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// We know the MDL isn't associated to a process now
|
|
|
|
//
|
|
|
|
Mdl->Process = NULL;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get page and VA information
|
|
|
|
//
|
|
|
|
MdlPages = (PPFN_NUMBER)(Mdl + 1);
|
|
|
|
Base = Mdl->StartVa;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Set the system address and now get the page count
|
|
|
|
//
|
|
|
|
Mdl->MappedSystemVa = (PVOID)((ULONG_PTR)Base + Mdl->ByteOffset);
|
|
|
|
PageCount = ADDRESS_AND_SIZE_TO_SPAN_PAGES(Mdl->MappedSystemVa,
|
|
|
|
Mdl->ByteCount);
|
|
|
|
ASSERT(PageCount != 0);
|
|
|
|
EndPage = MdlPages + PageCount;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Loop the PTEs
|
|
|
|
//
|
|
|
|
PointerPte = MiAddressToPte(Base);
|
|
|
|
do
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Write the PFN
|
|
|
|
//
|
|
|
|
Pfn = PFN_FROM_PTE(PointerPte++);
|
|
|
|
*MdlPages++ = Pfn;
|
|
|
|
} while (MdlPages < EndPage);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Set the nonpaged pool flag
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags |= MDL_SOURCE_IS_NONPAGED_POOL;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if this is an I/O mapping
|
|
|
|
//
|
2010-06-06 14:15:18 +00:00
|
|
|
if (!MiGetPfnEntry(Pfn)) Mdl->MdlFlags |= MDL_IO_SPACE;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
PMDL
|
|
|
|
NTAPI
|
|
|
|
MmAllocatePagesForMdl(IN PHYSICAL_ADDRESS LowAddress,
|
|
|
|
IN PHYSICAL_ADDRESS HighAddress,
|
|
|
|
IN PHYSICAL_ADDRESS SkipBytes,
|
|
|
|
IN SIZE_T TotalBytes)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Call the internal routine
|
|
|
|
//
|
|
|
|
return MiAllocatePagesForMdl(LowAddress,
|
|
|
|
HighAddress,
|
|
|
|
SkipBytes,
|
|
|
|
TotalBytes,
|
|
|
|
MiNotMapped,
|
|
|
|
0);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
PMDL
|
|
|
|
NTAPI
|
|
|
|
MmAllocatePagesForMdlEx(IN PHYSICAL_ADDRESS LowAddress,
|
|
|
|
IN PHYSICAL_ADDRESS HighAddress,
|
|
|
|
IN PHYSICAL_ADDRESS SkipBytes,
|
|
|
|
IN SIZE_T TotalBytes,
|
|
|
|
IN MEMORY_CACHING_TYPE CacheType,
|
|
|
|
IN ULONG Flags)
|
|
|
|
{
|
|
|
|
MI_PFN_CACHE_ATTRIBUTE CacheAttribute;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check for invalid cache type
|
|
|
|
//
|
|
|
|
if (CacheType > MmWriteCombined)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Normalize to default
|
|
|
|
//
|
|
|
|
CacheAttribute = MiNotMapped;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Conver to internal caching attribute
|
|
|
|
//
|
|
|
|
CacheAttribute = MiPlatformCacheAttributes[FALSE][CacheType];
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Only these flags are allowed
|
|
|
|
//
|
|
|
|
if (Flags & ~(MM_DONT_ZERO_ALLOCATION | MM_ALLOCATE_FROM_LOCAL_NODE_ONLY))
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Silently fail
|
|
|
|
//
|
|
|
|
return NULL;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Call the internal routine
|
|
|
|
//
|
|
|
|
return MiAllocatePagesForMdl(LowAddress,
|
|
|
|
HighAddress,
|
|
|
|
SkipBytes,
|
|
|
|
TotalBytes,
|
|
|
|
CacheAttribute,
|
|
|
|
Flags);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmFreePagesFromMdl(IN PMDL Mdl)
|
|
|
|
{
|
|
|
|
PVOID Base;
|
|
|
|
PPFN_NUMBER Pages;
|
|
|
|
LONG NumberOfPages;
|
|
|
|
PMMPFN Pfn1;
|
2009-09-30 18:24:00 +00:00
|
|
|
KIRQL OldIrql;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
DPRINT("Freeing MDL: %p\n", Mdl);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity checks
|
|
|
|
//
|
|
|
|
ASSERT(KeGetCurrentIrql() <= APC_LEVEL);
|
|
|
|
ASSERT((Mdl->MdlFlags & MDL_IO_SPACE) == 0);
|
|
|
|
ASSERT(((ULONG_PTR)Mdl->StartVa & (PAGE_SIZE - 1)) == 0);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get address and page information
|
|
|
|
//
|
|
|
|
Base = (PVOID)((ULONG_PTR)Mdl->StartVa + Mdl->ByteOffset);
|
|
|
|
NumberOfPages = ADDRESS_AND_SIZE_TO_SPAN_PAGES(Base, Mdl->ByteCount);
|
2009-09-30 18:24:00 +00:00
|
|
|
|
|
|
|
//
|
|
|
|
// Acquire PFN lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
OldIrql = MiAcquirePfnLock();
|
2009-09-30 18:24:00 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Loop all the MDL pages
|
|
|
|
//
|
|
|
|
Pages = (PPFN_NUMBER)(Mdl + 1);
|
|
|
|
do
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Reached the last page
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
if (*Pages == LIST_HEAD) break;
|
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the page entry
|
|
|
|
//
|
|
|
|
Pfn1 = MiGetPfnEntry(*Pages);
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
ASSERT(Pfn1);
|
|
|
|
ASSERT(Pfn1->u2.ShareCount == 1);
|
|
|
|
ASSERT(MI_IS_PFN_DELETED(Pfn1) == TRUE);
|
2010-12-26 15:23:03 +00:00
|
|
|
if (Pfn1->u4.PteFrame != 0x1FFEDCB)
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
{
|
|
|
|
/* Corrupted PFN entry or invalid free */
|
|
|
|
KeBugCheckEx(MEMORY_MANAGEMENT, 0x1236, (ULONG_PTR)Mdl, (ULONG_PTR)Pages, *Pages);
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Clear it
|
|
|
|
//
|
|
|
|
Pfn1->u3.e1.StartOfAllocation = 0;
|
|
|
|
Pfn1->u3.e1.EndOfAllocation = 0;
|
2015-06-26 10:40:49 +00:00
|
|
|
Pfn1->u3.e1.PageLocation = StandbyPageList;
|
2010-11-24 14:39:23 +00:00
|
|
|
Pfn1->u2.ShareCount = 0;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Dereference it
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
ASSERT(Pfn1->u3.e2.ReferenceCount != 0);
|
|
|
|
if (Pfn1->u3.e2.ReferenceCount != 1)
|
|
|
|
{
|
|
|
|
/* Just take off one reference */
|
|
|
|
InterlockedDecrement16((PSHORT)&Pfn1->u3.e2.ReferenceCount);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* We'll be nuking the whole page */
|
|
|
|
MiDecrementReferenceCount(Pfn1, *Pages);
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Clear this page and move on
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
*Pages++ = LIST_HEAD;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
} while (--NumberOfPages != 0);
|
2009-09-30 18:24:00 +00:00
|
|
|
|
|
|
|
//
|
|
|
|
// Release the lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
2009-09-30 18:24:00 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Remove the pages locked flag
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags &= ~MDL_PAGES_LOCKED;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
PVOID
|
|
|
|
NTAPI
|
|
|
|
MmMapLockedPagesSpecifyCache(IN PMDL Mdl,
|
|
|
|
IN KPROCESSOR_MODE AccessMode,
|
|
|
|
IN MEMORY_CACHING_TYPE CacheType,
|
|
|
|
IN PVOID BaseAddress,
|
|
|
|
IN ULONG BugCheckOnFailure,
|
2020-09-24 19:30:52 +00:00
|
|
|
IN ULONG Priority) // MM_PAGE_PRIORITY
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
{
|
|
|
|
PVOID Base;
|
|
|
|
PPFN_NUMBER MdlPages, LastPage;
|
[HAL/NDK]
- Make Vector parameter in HalEnableSystemInterrupt, HalDisableSystemInterrupt and HalBeginSystemInterrupt an ULONG, not an UCHAR
[NDK]
- 64bit fixes for HANDLE_TABLE, KPROCESS, SECTION_IMAGE_INFORMATION, MMADDRESS_LIST, MMVAD_FLAGS, MMVAD, MMVAD_LONG, MMVAD_SHORT, MEMORY_DESCRIPTOR, MEMORY_ALLOCATION_DESCRIPTOR, LdrVerifyMappedImageMatchesChecksum
- KDPC_DATA::DpcQueueDepth is signed on amd64, unsigned on x86
[NTOSKRNL]
- Fix hundreds of MSVC and amd64 warnings
- add a pragma message to FstubFixupEfiPartition, since it looks broken
- Move portable Ke constants from <arch>/cpu.c to krnlinit.c
- Fixed a bug in amd64 KiGeneralProtectionFaultHandler
svn path=/trunk/; revision=53734
2011-09-18 13:11:45 +00:00
|
|
|
PFN_COUNT PageCount;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
BOOLEAN IsIoMapping;
|
|
|
|
MI_PFN_CACHE_ATTRIBUTE CacheAttribute;
|
|
|
|
PMMPTE PointerPte;
|
|
|
|
MMPTE TempPte;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity check
|
|
|
|
//
|
|
|
|
ASSERT(Mdl->ByteCount != 0);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the base
|
|
|
|
//
|
|
|
|
Base = (PVOID)((ULONG_PTR)Mdl->StartVa + Mdl->ByteOffset);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Handle kernel case first
|
|
|
|
//
|
|
|
|
if (AccessMode == KernelMode)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Get the list of pages and count
|
|
|
|
//
|
|
|
|
MdlPages = (PPFN_NUMBER)(Mdl + 1);
|
|
|
|
PageCount = ADDRESS_AND_SIZE_TO_SPAN_PAGES(Base, Mdl->ByteCount);
|
|
|
|
LastPage = MdlPages + PageCount;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity checks
|
|
|
|
//
|
|
|
|
ASSERT((Mdl->MdlFlags & (MDL_MAPPED_TO_SYSTEM_VA |
|
|
|
|
MDL_SOURCE_IS_NONPAGED_POOL |
|
|
|
|
MDL_PARTIAL_HAS_BEEN_MAPPED)) == 0);
|
|
|
|
ASSERT((Mdl->MdlFlags & (MDL_PAGES_LOCKED | MDL_PARTIAL)) != 0);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the correct cache type
|
|
|
|
//
|
2009-07-01 09:59:47 +00:00
|
|
|
IsIoMapping = (Mdl->MdlFlags & MDL_IO_SPACE) != 0;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
CacheAttribute = MiPlatformCacheAttributes[IsIoMapping][CacheType];
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Reserve the PTEs
|
|
|
|
//
|
|
|
|
PointerPte = MiReserveSystemPtes(PageCount, SystemPteSpace);
|
|
|
|
if (!PointerPte)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// If it can fail, return NULL
|
|
|
|
//
|
|
|
|
if (Mdl->MdlFlags & MDL_MAPPING_CAN_FAIL) return NULL;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Should we bugcheck?
|
|
|
|
//
|
|
|
|
if (!BugCheckOnFailure) return NULL;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Yes, crash the system
|
|
|
|
//
|
|
|
|
KeBugCheckEx(NO_MORE_SYSTEM_PTES, 0, PageCount, 0, 0);
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the mapped address
|
|
|
|
//
|
|
|
|
Base = (PVOID)((ULONG_PTR)MiPteToAddress(PointerPte) + Mdl->ByteOffset);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the template
|
|
|
|
//
|
2010-02-10 13:56:54 +00:00
|
|
|
TempPte = ValidKernelPte;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
switch (CacheAttribute)
|
|
|
|
{
|
|
|
|
case MiNonCached:
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Disable caching
|
|
|
|
//
|
2009-07-15 17:53:45 +00:00
|
|
|
MI_PAGE_DISABLE_CACHE(&TempPte);
|
|
|
|
MI_PAGE_WRITE_THROUGH(&TempPte);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
break;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
case MiWriteCombined:
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Enable write combining
|
|
|
|
//
|
2009-07-15 17:53:45 +00:00
|
|
|
MI_PAGE_DISABLE_CACHE(&TempPte);
|
|
|
|
MI_PAGE_WRITE_COMBINED(&TempPte);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
break;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
default:
|
|
|
|
//
|
|
|
|
// Nothing to do
|
|
|
|
//
|
|
|
|
break;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Loop all PTEs
|
|
|
|
//
|
|
|
|
do
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// We're done here
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
if (*MdlPages == LIST_HEAD) break;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Write the PTE
|
|
|
|
//
|
|
|
|
TempPte.u.Hard.PageFrameNumber = *MdlPages;
|
2010-06-06 18:45:46 +00:00
|
|
|
MI_WRITE_VALID_PTE(PointerPte++, TempPte);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
} while (++MdlPages < LastPage);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Mark it as mapped
|
|
|
|
//
|
|
|
|
ASSERT((Mdl->MdlFlags & MDL_MAPPED_TO_SYSTEM_VA) == 0);
|
|
|
|
Mdl->MappedSystemVa = Base;
|
|
|
|
Mdl->MdlFlags |= MDL_MAPPED_TO_SYSTEM_VA;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if it was partial
|
|
|
|
//
|
|
|
|
if (Mdl->MdlFlags & MDL_PARTIAL)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Write the appropriate flag here too
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags |= MDL_PARTIAL_HAS_BEEN_MAPPED;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Return the mapped address
|
|
|
|
//
|
|
|
|
return Base;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
2017-06-18 11:18:41 +00:00
|
|
|
return MiMapLockedPagesInUserSpace(Mdl, Base, CacheType, BaseAddress);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
PVOID
|
|
|
|
NTAPI
|
|
|
|
MmMapLockedPages(IN PMDL Mdl,
|
|
|
|
IN KPROCESSOR_MODE AccessMode)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Call the extended version
|
|
|
|
//
|
|
|
|
return MmMapLockedPagesSpecifyCache(Mdl,
|
|
|
|
AccessMode,
|
|
|
|
MmCached,
|
|
|
|
NULL,
|
|
|
|
TRUE,
|
|
|
|
HighPagePriority);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmUnmapLockedPages(IN PVOID BaseAddress,
|
|
|
|
IN PMDL Mdl)
|
|
|
|
{
|
|
|
|
PVOID Base;
|
[HAL/NDK]
- Make Vector parameter in HalEnableSystemInterrupt, HalDisableSystemInterrupt and HalBeginSystemInterrupt an ULONG, not an UCHAR
[NDK]
- 64bit fixes for HANDLE_TABLE, KPROCESS, SECTION_IMAGE_INFORMATION, MMADDRESS_LIST, MMVAD_FLAGS, MMVAD, MMVAD_LONG, MMVAD_SHORT, MEMORY_DESCRIPTOR, MEMORY_ALLOCATION_DESCRIPTOR, LdrVerifyMappedImageMatchesChecksum
- KDPC_DATA::DpcQueueDepth is signed on amd64, unsigned on x86
[NTOSKRNL]
- Fix hundreds of MSVC and amd64 warnings
- add a pragma message to FstubFixupEfiPartition, since it looks broken
- Move portable Ke constants from <arch>/cpu.c to krnlinit.c
- Fixed a bug in amd64 KiGeneralProtectionFaultHandler
svn path=/trunk/; revision=53734
2011-09-18 13:11:45 +00:00
|
|
|
PFN_COUNT PageCount, ExtraPageCount;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
PPFN_NUMBER MdlPages;
|
|
|
|
PMMPTE PointerPte;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity check
|
|
|
|
//
|
|
|
|
ASSERT(Mdl->ByteCount != 0);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if this is a kernel request
|
|
|
|
//
|
|
|
|
if (BaseAddress > MM_HIGHEST_USER_ADDRESS)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Get base and count information
|
|
|
|
//
|
|
|
|
Base = (PVOID)((ULONG_PTR)Mdl->StartVa + Mdl->ByteOffset);
|
|
|
|
PageCount = ADDRESS_AND_SIZE_TO_SPAN_PAGES(Base, Mdl->ByteCount);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity checks
|
|
|
|
//
|
|
|
|
ASSERT((Mdl->MdlFlags & MDL_PARENT_MAPPED_SYSTEM_VA) == 0);
|
|
|
|
ASSERT(PageCount != 0);
|
|
|
|
ASSERT(Mdl->MdlFlags & MDL_MAPPED_TO_SYSTEM_VA);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the PTE
|
|
|
|
//
|
|
|
|
PointerPte = MiAddressToPte(BaseAddress);
|
2009-10-20 16:47:01 +00:00
|
|
|
|
|
|
|
//
|
|
|
|
// This should be a resident system PTE
|
|
|
|
//
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
ASSERT(PointerPte >= MmSystemPtesStart[SystemPteSpace]);
|
|
|
|
ASSERT(PointerPte <= MmSystemPtesEnd[SystemPteSpace]);
|
2009-10-20 16:47:01 +00:00
|
|
|
ASSERT(PointerPte->u.Hard.Valid == 1);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if the caller wants us to free advanced pages
|
|
|
|
//
|
|
|
|
if (Mdl->MdlFlags & MDL_FREE_EXTRA_PTES)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Get the MDL page array
|
|
|
|
//
|
[HAL/NDK]
- Make Vector parameter in HalEnableSystemInterrupt, HalDisableSystemInterrupt and HalBeginSystemInterrupt an ULONG, not an UCHAR
[NDK]
- 64bit fixes for HANDLE_TABLE, KPROCESS, SECTION_IMAGE_INFORMATION, MMADDRESS_LIST, MMVAD_FLAGS, MMVAD, MMVAD_LONG, MMVAD_SHORT, MEMORY_DESCRIPTOR, MEMORY_ALLOCATION_DESCRIPTOR, LdrVerifyMappedImageMatchesChecksum
- KDPC_DATA::DpcQueueDepth is signed on amd64, unsigned on x86
[NTOSKRNL]
- Fix hundreds of MSVC and amd64 warnings
- add a pragma message to FstubFixupEfiPartition, since it looks broken
- Move portable Ke constants from <arch>/cpu.c to krnlinit.c
- Fixed a bug in amd64 KiGeneralProtectionFaultHandler
svn path=/trunk/; revision=53734
2011-09-18 13:11:45 +00:00
|
|
|
MdlPages = MmGetMdlPfnArray(Mdl);
|
|
|
|
|
|
|
|
/* Number of extra pages stored after the PFN array */
|
|
|
|
ExtraPageCount = (PFN_COUNT)*(MdlPages + PageCount);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Do the math
|
|
|
|
//
|
[HAL/NDK]
- Make Vector parameter in HalEnableSystemInterrupt, HalDisableSystemInterrupt and HalBeginSystemInterrupt an ULONG, not an UCHAR
[NDK]
- 64bit fixes for HANDLE_TABLE, KPROCESS, SECTION_IMAGE_INFORMATION, MMADDRESS_LIST, MMVAD_FLAGS, MMVAD, MMVAD_LONG, MMVAD_SHORT, MEMORY_DESCRIPTOR, MEMORY_ALLOCATION_DESCRIPTOR, LdrVerifyMappedImageMatchesChecksum
- KDPC_DATA::DpcQueueDepth is signed on amd64, unsigned on x86
[NTOSKRNL]
- Fix hundreds of MSVC and amd64 warnings
- add a pragma message to FstubFixupEfiPartition, since it looks broken
- Move portable Ke constants from <arch>/cpu.c to krnlinit.c
- Fixed a bug in amd64 KiGeneralProtectionFaultHandler
svn path=/trunk/; revision=53734
2011-09-18 13:11:45 +00:00
|
|
|
PageCount += ExtraPageCount;
|
|
|
|
PointerPte -= ExtraPageCount;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
ASSERT(PointerPte >= MmSystemPtesStart[SystemPteSpace]);
|
|
|
|
ASSERT(PointerPte <= MmSystemPtesEnd[SystemPteSpace]);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the new base address
|
|
|
|
//
|
|
|
|
BaseAddress = (PVOID)((ULONG_PTR)BaseAddress -
|
[HAL/NDK]
- Make Vector parameter in HalEnableSystemInterrupt, HalDisableSystemInterrupt and HalBeginSystemInterrupt an ULONG, not an UCHAR
[NDK]
- 64bit fixes for HANDLE_TABLE, KPROCESS, SECTION_IMAGE_INFORMATION, MMADDRESS_LIST, MMVAD_FLAGS, MMVAD, MMVAD_LONG, MMVAD_SHORT, MEMORY_DESCRIPTOR, MEMORY_ALLOCATION_DESCRIPTOR, LdrVerifyMappedImageMatchesChecksum
- KDPC_DATA::DpcQueueDepth is signed on amd64, unsigned on x86
[NTOSKRNL]
- Fix hundreds of MSVC and amd64 warnings
- add a pragma message to FstubFixupEfiPartition, since it looks broken
- Move portable Ke constants from <arch>/cpu.c to krnlinit.c
- Fixed a bug in amd64 KiGeneralProtectionFaultHandler
svn path=/trunk/; revision=53734
2011-09-18 13:11:45 +00:00
|
|
|
(ExtraPageCount << PAGE_SHIFT));
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Remove flags
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags &= ~(MDL_MAPPED_TO_SYSTEM_VA |
|
|
|
|
MDL_PARTIAL_HAS_BEEN_MAPPED |
|
|
|
|
MDL_FREE_EXTRA_PTES);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Release the system PTEs
|
|
|
|
//
|
|
|
|
MiReleaseSystemPtes(PointerPte, PageCount, SystemPteSpace);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2017-06-18 11:18:41 +00:00
|
|
|
MiUnmapLockedPagesInUserSpace(BaseAddress, Mdl);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmProbeAndLockPages(IN PMDL Mdl,
|
|
|
|
IN KPROCESSOR_MODE AccessMode,
|
|
|
|
IN LOCK_OPERATION Operation)
|
|
|
|
{
|
2010-07-15 22:50:12 +00:00
|
|
|
PPFN_NUMBER MdlPages;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
PVOID Base, Address, LastAddress, StartAddress;
|
|
|
|
ULONG LockPages, TotalPages;
|
|
|
|
NTSTATUS Status = STATUS_SUCCESS;
|
|
|
|
PEPROCESS CurrentProcess;
|
|
|
|
NTSTATUS ProbeStatus;
|
2010-02-09 22:56:21 +00:00
|
|
|
PMMPTE PointerPte, LastPte;
|
|
|
|
PMMPDE PointerPde;
|
2012-02-04 23:08:20 +00:00
|
|
|
#if (_MI_PAGING_LEVELS >= 3)
|
|
|
|
PMMPDE PointerPpe;
|
|
|
|
#endif
|
|
|
|
#if (_MI_PAGING_LEVELS == 4)
|
|
|
|
PMMPDE PointerPxe;
|
|
|
|
#endif
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
PFN_NUMBER PageFrameIndex;
|
|
|
|
BOOLEAN UsePfnLock;
|
|
|
|
KIRQL OldIrql;
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
PMMPFN Pfn1;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
DPRINT("Probing MDL: %p\n", Mdl);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity checks
|
|
|
|
//
|
|
|
|
ASSERT(Mdl->ByteCount != 0);
|
|
|
|
ASSERT(((ULONG)Mdl->ByteOffset & ~(PAGE_SIZE - 1)) == 0);
|
|
|
|
ASSERT(((ULONG_PTR)Mdl->StartVa & (PAGE_SIZE - 1)) == 0);
|
|
|
|
ASSERT((Mdl->MdlFlags & (MDL_PAGES_LOCKED |
|
|
|
|
MDL_MAPPED_TO_SYSTEM_VA |
|
|
|
|
MDL_SOURCE_IS_NONPAGED_POOL |
|
|
|
|
MDL_PARTIAL |
|
|
|
|
MDL_IO_SPACE)) == 0);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get page and base information
|
|
|
|
//
|
|
|
|
MdlPages = (PPFN_NUMBER)(Mdl + 1);
|
2011-09-11 11:47:28 +00:00
|
|
|
Base = Mdl->StartVa;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the addresses and how many pages we span (and need to lock)
|
|
|
|
//
|
|
|
|
Address = (PVOID)((ULONG_PTR)Base + Mdl->ByteOffset);
|
|
|
|
LastAddress = (PVOID)((ULONG_PTR)Address + Mdl->ByteCount);
|
|
|
|
LockPages = ADDRESS_AND_SIZE_TO_SPAN_PAGES(Address, Mdl->ByteCount);
|
|
|
|
ASSERT(LockPages != 0);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Block invalid access */
|
|
|
|
if ((AccessMode != KernelMode) &&
|
|
|
|
((LastAddress > (PVOID)MM_USER_PROBE_ADDRESS) || (Address >= LastAddress)))
|
|
|
|
{
|
|
|
|
/* Caller should be in SEH, raise the error */
|
|
|
|
*MdlPages = LIST_HEAD;
|
|
|
|
ExRaiseStatus(STATUS_ACCESS_VIOLATION);
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
// Get the process
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
if (Address <= MM_HIGHEST_USER_ADDRESS)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Get the process
|
|
|
|
//
|
|
|
|
CurrentProcess = PsGetCurrentProcess();
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// No process
|
|
|
|
//
|
|
|
|
CurrentProcess = NULL;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Save the number of pages we'll have to lock, and the start address
|
|
|
|
//
|
|
|
|
TotalPages = LockPages;
|
|
|
|
StartAddress = Address;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Large pages not supported */
|
|
|
|
ASSERT(!MI_IS_PHYSICAL_ADDRESS(Address));
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Now probe them
|
|
|
|
//
|
|
|
|
ProbeStatus = STATUS_SUCCESS;
|
|
|
|
_SEH2_TRY
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Enter probe loop
|
|
|
|
//
|
|
|
|
do
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Assume failure
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
*MdlPages = LIST_HEAD;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Read
|
|
|
|
//
|
|
|
|
*(volatile CHAR*)Address;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if this is write access (only probe for user-mode)
|
|
|
|
//
|
|
|
|
if ((Operation != IoReadAccess) &&
|
|
|
|
(Address <= MM_HIGHEST_USER_ADDRESS))
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Probe for write too
|
|
|
|
//
|
|
|
|
ProbeForWriteChar(Address);
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Next address...
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
Address = PAGE_ALIGN((ULONG_PTR)Address + PAGE_SIZE);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Next page...
|
|
|
|
//
|
|
|
|
LockPages--;
|
|
|
|
MdlPages++;
|
|
|
|
} while (Address < LastAddress);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Reset back to the original page
|
|
|
|
//
|
|
|
|
ASSERT(LockPages == 0);
|
|
|
|
MdlPages = (PPFN_NUMBER)(Mdl + 1);
|
|
|
|
}
|
|
|
|
_SEH2_EXCEPT(EXCEPTION_EXECUTE_HANDLER)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Oops :(
|
|
|
|
//
|
|
|
|
ProbeStatus = _SEH2_GetExceptionCode();
|
|
|
|
}
|
|
|
|
_SEH2_END;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// So how did that go?
|
|
|
|
//
|
|
|
|
if (ProbeStatus != STATUS_SUCCESS)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Fail
|
|
|
|
//
|
|
|
|
DPRINT1("MDL PROBE FAILED!\n");
|
|
|
|
Mdl->Process = NULL;
|
|
|
|
ExRaiseStatus(ProbeStatus);
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the PTE and PDE
|
|
|
|
//
|
|
|
|
PointerPte = MiAddressToPte(StartAddress);
|
|
|
|
PointerPde = MiAddressToPde(StartAddress);
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
#if (_MI_PAGING_LEVELS >= 3)
|
2012-02-04 23:08:20 +00:00
|
|
|
PointerPpe = MiAddressToPpe(StartAddress);
|
|
|
|
#endif
|
|
|
|
#if (_MI_PAGING_LEVELS == 4)
|
|
|
|
PointerPxe = MiAddressToPxe(StartAddress);
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
#endif
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity check
|
|
|
|
//
|
2009-10-20 16:47:01 +00:00
|
|
|
ASSERT(MdlPages == (PPFN_NUMBER)(Mdl + 1));
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check what kind of operation this is
|
|
|
|
//
|
|
|
|
if (Operation != IoReadAccess)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Set the write flag
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags |= MDL_WRITE_OPERATION;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Remove the write flag
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags &= ~(MDL_WRITE_OPERATION);
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Mark the MDL as locked *now*
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags |= MDL_PAGES_LOCKED;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if this came from kernel mode
|
|
|
|
//
|
2012-02-04 23:08:20 +00:00
|
|
|
if (Base > MM_HIGHEST_USER_ADDRESS)
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
{
|
|
|
|
//
|
|
|
|
// We should not have a process
|
|
|
|
//
|
|
|
|
ASSERT(CurrentProcess == NULL);
|
|
|
|
Mdl->Process = NULL;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// In kernel mode, we don't need to check for write access
|
|
|
|
//
|
|
|
|
Operation = IoReadAccess;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Use the PFN lock
|
|
|
|
//
|
|
|
|
UsePfnLock = TRUE;
|
2017-11-21 22:33:42 +00:00
|
|
|
OldIrql = MiAcquirePfnLock();
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Sanity checks
|
|
|
|
//
|
|
|
|
ASSERT(TotalPages != 0);
|
|
|
|
ASSERT(CurrentProcess == PsGetCurrentProcess());
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Track locked pages
|
|
|
|
//
|
|
|
|
InterlockedExchangeAddSizeT(&CurrentProcess->NumberOfLockedPages,
|
|
|
|
TotalPages);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Save the process
|
|
|
|
//
|
|
|
|
Mdl->Process = CurrentProcess;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Lock the process working set */
|
|
|
|
MiLockProcessWorkingSet(CurrentProcess, PsGetCurrentThread());
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
UsePfnLock = FALSE;
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
OldIrql = MM_NOIRQL;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the last PTE
|
|
|
|
//
|
|
|
|
LastPte = MiAddressToPte((PVOID)((ULONG_PTR)LastAddress - 1));
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Loop the pages
|
|
|
|
//
|
|
|
|
do
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Assume failure and check for non-mapped pages
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
*MdlPages = LIST_HEAD;
|
2012-02-04 23:08:20 +00:00
|
|
|
while (
|
|
|
|
#if (_MI_PAGING_LEVELS == 4)
|
|
|
|
(PointerPxe->u.Hard.Valid == 0) ||
|
|
|
|
#endif
|
2010-07-26 21:45:42 +00:00
|
|
|
#if (_MI_PAGING_LEVELS >= 3)
|
2012-02-04 23:08:20 +00:00
|
|
|
(PointerPpe->u.Hard.Valid == 0) ||
|
2010-07-26 21:45:42 +00:00
|
|
|
#endif
|
2012-02-04 23:08:20 +00:00
|
|
|
(PointerPde->u.Hard.Valid == 0) ||
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
(PointerPte->u.Hard.Valid == 0))
|
|
|
|
{
|
|
|
|
//
|
2011-09-11 11:47:28 +00:00
|
|
|
// What kind of lock were we using?
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
if (UsePfnLock)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Release PFN lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Release process working set */
|
|
|
|
MiUnlockProcessWorkingSet(CurrentProcess, PsGetCurrentThread());
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Access the page
|
|
|
|
//
|
|
|
|
Address = MiPteToAddress(PointerPte);
|
2011-10-26 14:38:57 +00:00
|
|
|
|
|
|
|
//HACK: Pass a placeholder TrapInformation so the fault handler knows we're unlocked
|
2018-01-01 14:24:05 +00:00
|
|
|
Status = MmAccessFault(FALSE, Address, KernelMode, (PVOID)(ULONG_PTR)0xBADBADA3BADBADA3ULL);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
if (!NT_SUCCESS(Status))
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Fail
|
|
|
|
//
|
|
|
|
DPRINT1("Access fault failed\n");
|
|
|
|
goto Cleanup;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
2011-09-11 11:47:28 +00:00
|
|
|
// What lock should we use?
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
if (UsePfnLock)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Grab the PFN lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
OldIrql = MiAcquirePfnLock();
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Lock the process working set */
|
|
|
|
MiLockProcessWorkingSet(CurrentProcess, PsGetCurrentThread());
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if this was a write or modify
|
|
|
|
//
|
|
|
|
if (Operation != IoReadAccess)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Check if the PTE is not writable
|
|
|
|
//
|
2009-07-15 17:53:45 +00:00
|
|
|
if (MI_IS_PAGE_WRITEABLE(PointerPte) == FALSE)
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
{
|
|
|
|
//
|
|
|
|
// Check if it's copy on write
|
|
|
|
//
|
2009-07-15 17:53:45 +00:00
|
|
|
if (MI_IS_PAGE_COPY_ON_WRITE(PointerPte))
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
{
|
|
|
|
//
|
|
|
|
// Get the base address and allow a change for user-mode
|
|
|
|
//
|
|
|
|
Address = MiPteToAddress(PointerPte);
|
|
|
|
if (Address <= MM_HIGHEST_USER_ADDRESS)
|
|
|
|
{
|
|
|
|
//
|
2011-09-11 11:47:28 +00:00
|
|
|
// What kind of lock were we using?
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
if (UsePfnLock)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Release PFN lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Release process working set */
|
|
|
|
MiUnlockProcessWorkingSet(CurrentProcess, PsGetCurrentThread());
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Access the page
|
|
|
|
//
|
2011-10-26 14:38:57 +00:00
|
|
|
|
|
|
|
//HACK: Pass a placeholder TrapInformation so the fault handler knows we're unlocked
|
2018-01-01 14:24:05 +00:00
|
|
|
Status = MmAccessFault(TRUE, Address, KernelMode, (PVOID)(ULONG_PTR)0xBADBADA3BADBADA3ULL);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
if (!NT_SUCCESS(Status))
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Fail
|
|
|
|
//
|
|
|
|
DPRINT1("Access fault failed\n");
|
|
|
|
goto Cleanup;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Re-acquire the lock
|
|
|
|
//
|
|
|
|
if (UsePfnLock)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Grab the PFN lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
OldIrql = MiAcquirePfnLock();
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Lock the process working set */
|
|
|
|
MiLockProcessWorkingSet(CurrentProcess, PsGetCurrentThread());
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Start over
|
|
|
|
//
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Fail, since we won't allow this
|
|
|
|
//
|
|
|
|
Status = STATUS_ACCESS_VIOLATION;
|
|
|
|
goto CleanupWithLock;
|
|
|
|
}
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Grab the PFN
|
|
|
|
//
|
|
|
|
PageFrameIndex = PFN_FROM_PTE(PointerPte);
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
Pfn1 = MiGetPfnEntry(PageFrameIndex);
|
|
|
|
if (Pfn1)
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
{
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Either this is for kernel-mode, or the working set is held */
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
ASSERT((CurrentProcess == NULL) || (UsePfnLock == FALSE));
|
2010-12-26 15:23:03 +00:00
|
|
|
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* No Physical VADs supported yet */
|
|
|
|
if (CurrentProcess) ASSERT(CurrentProcess->PhysicalVadRoot == NULL);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* This address should already exist and be fully valid */
|
2012-08-03 11:34:35 +00:00
|
|
|
MiReferenceProbedPageAndBumpLockCount(Pfn1);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// For I/O addresses, just remember this
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags |= MDL_IO_SPACE;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Write the page and move on
|
|
|
|
//
|
|
|
|
*MdlPages++ = PageFrameIndex;
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
PointerPte++;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Check if we're on a PDE boundary */
|
2012-02-04 23:08:20 +00:00
|
|
|
if (MiIsPteOnPdeBoundary(PointerPte)) PointerPde++;
|
|
|
|
#if (_MI_PAGING_LEVELS >= 3)
|
|
|
|
if (MiIsPteOnPpeBoundary(PointerPte)) PointerPpe++;
|
|
|
|
#endif
|
|
|
|
#if (_MI_PAGING_LEVELS == 4)
|
|
|
|
if (MiIsPteOnPxeBoundary(PointerPte)) PointerPxe++;
|
|
|
|
#endif
|
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
} while (PointerPte <= LastPte);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
2011-09-11 11:47:28 +00:00
|
|
|
// What kind of lock were we using?
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
if (UsePfnLock)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Release PFN lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Release process working set */
|
|
|
|
MiUnlockProcessWorkingSet(CurrentProcess, PsGetCurrentThread());
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity check
|
|
|
|
//
|
|
|
|
ASSERT((Mdl->MdlFlags & MDL_DESCRIBES_AWE) == 0);
|
|
|
|
return;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
CleanupWithLock:
|
|
|
|
//
|
|
|
|
// This is the failure path
|
|
|
|
//
|
|
|
|
ASSERT(!NT_SUCCESS(Status));
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
2011-09-11 11:47:28 +00:00
|
|
|
// What kind of lock were we using?
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
if (UsePfnLock)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Release PFN lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Release process working set */
|
|
|
|
MiUnlockProcessWorkingSet(CurrentProcess, PsGetCurrentThread());
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
Cleanup:
|
|
|
|
//
|
|
|
|
// Pages must be locked so MmUnlock can work
|
|
|
|
//
|
|
|
|
ASSERT(Mdl->MdlFlags & MDL_PAGES_LOCKED);
|
|
|
|
MmUnlockPages(Mdl);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Raise the error
|
|
|
|
//
|
|
|
|
ExRaiseStatus(Status);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @implemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmUnlockPages(IN PMDL Mdl)
|
|
|
|
{
|
|
|
|
PPFN_NUMBER MdlPages, LastPage;
|
|
|
|
PEPROCESS Process;
|
|
|
|
PVOID Base;
|
|
|
|
ULONG Flags, PageCount;
|
|
|
|
KIRQL OldIrql;
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
PMMPFN Pfn1;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
DPRINT("Unlocking MDL: %p\n", Mdl);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Sanity checks
|
|
|
|
//
|
|
|
|
ASSERT((Mdl->MdlFlags & MDL_PAGES_LOCKED) != 0);
|
|
|
|
ASSERT((Mdl->MdlFlags & MDL_SOURCE_IS_NONPAGED_POOL) == 0);
|
|
|
|
ASSERT((Mdl->MdlFlags & MDL_PARTIAL) == 0);
|
|
|
|
ASSERT(Mdl->ByteCount != 0);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the process associated and capture the flags which are volatile
|
|
|
|
//
|
|
|
|
Process = Mdl->Process;
|
|
|
|
Flags = Mdl->MdlFlags;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Automagically undo any calls to MmGetSystemAddressForMdl's for this MDL
|
|
|
|
//
|
|
|
|
if (Mdl->MdlFlags & MDL_MAPPED_TO_SYSTEM_VA)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Unmap the pages from system space
|
|
|
|
//
|
|
|
|
MmUnmapLockedPages(Mdl->MappedSystemVa, Mdl);
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Get the page count
|
|
|
|
//
|
|
|
|
MdlPages = (PPFN_NUMBER)(Mdl + 1);
|
|
|
|
Base = (PVOID)((ULONG_PTR)Mdl->StartVa + Mdl->ByteOffset);
|
|
|
|
PageCount = ADDRESS_AND_SIZE_TO_SPAN_PAGES(Base, Mdl->ByteCount);
|
|
|
|
ASSERT(PageCount != 0);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// We don't support AWE
|
|
|
|
//
|
2010-12-26 15:23:03 +00:00
|
|
|
if (Flags & MDL_DESCRIBES_AWE) ASSERT(FALSE);
|
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if the buffer is mapped I/O space
|
|
|
|
//
|
|
|
|
if (Flags & MDL_IO_SPACE)
|
2010-12-26 15:23:03 +00:00
|
|
|
{
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Acquire PFN lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
OldIrql = MiAcquirePfnLock();
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Loop every page
|
|
|
|
//
|
|
|
|
LastPage = MdlPages + PageCount;
|
|
|
|
do
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Last page, break out
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
if (*MdlPages == LIST_HEAD) break;
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if this page is in the PFN database
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
Pfn1 = MiGetPfnEntry(*MdlPages);
|
2012-08-03 11:34:35 +00:00
|
|
|
if (Pfn1) MiDereferencePfnAndDropLockCount(Pfn1);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
} while (++MdlPages < LastPage);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Release the lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if we have a process
|
|
|
|
//
|
|
|
|
if (Process)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Handle the accounting of locked pages
|
|
|
|
//
|
|
|
|
ASSERT(Process->NumberOfLockedPages > 0);
|
|
|
|
InterlockedExchangeAddSizeT(&Process->NumberOfLockedPages,
|
[HAL/NDK]
- Make Vector parameter in HalEnableSystemInterrupt, HalDisableSystemInterrupt and HalBeginSystemInterrupt an ULONG, not an UCHAR
[NDK]
- 64bit fixes for HANDLE_TABLE, KPROCESS, SECTION_IMAGE_INFORMATION, MMADDRESS_LIST, MMVAD_FLAGS, MMVAD, MMVAD_LONG, MMVAD_SHORT, MEMORY_DESCRIPTOR, MEMORY_ALLOCATION_DESCRIPTOR, LdrVerifyMappedImageMatchesChecksum
- KDPC_DATA::DpcQueueDepth is signed on amd64, unsigned on x86
[NTOSKRNL]
- Fix hundreds of MSVC and amd64 warnings
- add a pragma message to FstubFixupEfiPartition, since it looks broken
- Move portable Ke constants from <arch>/cpu.c to krnlinit.c
- Fixed a bug in amd64 KiGeneralProtectionFaultHandler
svn path=/trunk/; revision=53734
2011-09-18 13:11:45 +00:00
|
|
|
-(LONG_PTR)PageCount);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// We're done
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags &= ~MDL_IO_SPACE;
|
|
|
|
Mdl->MdlFlags &= ~MDL_PAGES_LOCKED;
|
|
|
|
return;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Check if we have a process
|
|
|
|
//
|
|
|
|
if (Process)
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Handle the accounting of locked pages
|
|
|
|
//
|
|
|
|
ASSERT(Process->NumberOfLockedPages > 0);
|
|
|
|
InterlockedExchangeAddSizeT(&Process->NumberOfLockedPages,
|
[HAL/NDK]
- Make Vector parameter in HalEnableSystemInterrupt, HalDisableSystemInterrupt and HalBeginSystemInterrupt an ULONG, not an UCHAR
[NDK]
- 64bit fixes for HANDLE_TABLE, KPROCESS, SECTION_IMAGE_INFORMATION, MMADDRESS_LIST, MMVAD_FLAGS, MMVAD, MMVAD_LONG, MMVAD_SHORT, MEMORY_DESCRIPTOR, MEMORY_ALLOCATION_DESCRIPTOR, LdrVerifyMappedImageMatchesChecksum
- KDPC_DATA::DpcQueueDepth is signed on amd64, unsigned on x86
[NTOSKRNL]
- Fix hundreds of MSVC and amd64 warnings
- add a pragma message to FstubFixupEfiPartition, since it looks broken
- Move portable Ke constants from <arch>/cpu.c to krnlinit.c
- Fixed a bug in amd64 KiGeneralProtectionFaultHandler
svn path=/trunk/; revision=53734
2011-09-18 13:11:45 +00:00
|
|
|
-(LONG_PTR)PageCount);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Loop every page
|
|
|
|
//
|
|
|
|
LastPage = MdlPages + PageCount;
|
|
|
|
do
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// Last page reached
|
|
|
|
//
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
if (*MdlPages == LIST_HEAD)
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
{
|
|
|
|
//
|
|
|
|
// Were there no pages at all?
|
|
|
|
//
|
|
|
|
if (MdlPages == (PPFN_NUMBER)(Mdl + 1))
|
|
|
|
{
|
|
|
|
//
|
|
|
|
// We're already done
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags &= ~MDL_PAGES_LOCKED;
|
|
|
|
return;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Otherwise, stop here
|
|
|
|
//
|
|
|
|
LastPage = MdlPages;
|
|
|
|
break;
|
|
|
|
}
|
2010-12-26 15:23:03 +00:00
|
|
|
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Save the PFN entry instead for the secondary loop */
|
|
|
|
*MdlPages = (PFN_NUMBER)MiGetPfnEntry(*MdlPages);
|
2011-09-11 11:47:28 +00:00
|
|
|
ASSERT(*MdlPages != 0);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
} while (++MdlPages < LastPage);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Reset pointer
|
|
|
|
//
|
|
|
|
MdlPages = (PPFN_NUMBER)(Mdl + 1);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Now grab the PFN lock for the actual unlock and dereference
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
OldIrql = MiAcquirePfnLock();
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
do
|
|
|
|
{
|
[NTOS]: Implement MiDecrementReferenceCount and rewrite large parts of the ProbeAndLock/Unlock MDL API to fully use ARM3 APIs, dropping MmReference/DereferencePage behind.
[NTOS]: Fix many MDL API bugs: correctly check for I/O pages, use LIST_HEAD instead of -1, track system-wide locked pages, use the process working set lock instead of the address space lock, add check for cross-ring MDL mappings, and make some small optimizations.
[NTOS]: Make some more fixes in MmAllocatePagesForMdl, MmFreeMdlPages to make the PFN entries more "correct".
[NTOS]: Had a little breakthrough: instead of complicating our lives and hiding certain ReactOS-Mm fields inside legitimate ARM3/MMPFN fields, differentiate between "legacy" (RosMm) and ARM3 pages. The legacy allocator (MmAllocPage/MmRequestPageMemoryConsumer) will use the non-paged pool to allocate a MMROSPFN add-on (8 bytes), in which the RMAP list head and SWAPENTRY are stored. When a legacy "free" is done, this data is deleted. Additionally, we can now tell apart between ARM3 and RosMm pages, so appropriate ASSERTs have been added to make sure the two never cross paths (which should safely let us use all the PFN fields now and implement working sets, etc...). I don't know why I didn't think of this sooner.
svn path=/trunk/; revision=49201
2010-10-19 04:30:48 +00:00
|
|
|
/* Get the current entry and reference count */
|
2011-09-11 11:47:28 +00:00
|
|
|
Pfn1 = (PMMPFN)*MdlPages;
|
2012-08-03 11:34:35 +00:00
|
|
|
MiDereferencePfnAndDropLockCount(Pfn1);
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
} while (++MdlPages < LastPage);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// Release the lock
|
|
|
|
//
|
2017-11-21 22:33:42 +00:00
|
|
|
MiReleasePfnLock(OldIrql);
|
2010-12-26 15:23:03 +00:00
|
|
|
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
//
|
|
|
|
// We're done
|
|
|
|
//
|
|
|
|
Mdl->MdlFlags &= ~MDL_PAGES_LOCKED;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @unimplemented
|
|
|
|
*/
|
|
|
|
NTSTATUS
|
|
|
|
NTAPI
|
|
|
|
MmAdvanceMdl(IN PMDL Mdl,
|
|
|
|
IN ULONG NumberOfBytes)
|
|
|
|
{
|
2011-09-11 11:47:28 +00:00
|
|
|
UNIMPLEMENTED;
|
|
|
|
return STATUS_NOT_IMPLEMENTED;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @unimplemented
|
|
|
|
*/
|
|
|
|
PVOID
|
|
|
|
NTAPI
|
|
|
|
MmMapLockedPagesWithReservedMapping(IN PVOID MappingAddress,
|
|
|
|
IN ULONG PoolTag,
|
|
|
|
IN PMDL MemoryDescriptorList,
|
|
|
|
IN MEMORY_CACHING_TYPE CacheType)
|
|
|
|
{
|
2011-09-11 11:47:28 +00:00
|
|
|
UNIMPLEMENTED;
|
|
|
|
return 0;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @unimplemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmUnmapReservedMapping(IN PVOID BaseAddress,
|
|
|
|
IN ULONG PoolTag,
|
|
|
|
IN PMDL MemoryDescriptorList)
|
|
|
|
{
|
2011-09-11 11:47:28 +00:00
|
|
|
UNIMPLEMENTED;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @unimplemented
|
|
|
|
*/
|
|
|
|
NTSTATUS
|
|
|
|
NTAPI
|
|
|
|
MmPrefetchPages(IN ULONG NumberOfLists,
|
|
|
|
IN PREAD_LIST *ReadLists)
|
|
|
|
{
|
2011-09-11 11:47:28 +00:00
|
|
|
UNIMPLEMENTED;
|
|
|
|
return STATUS_NOT_IMPLEMENTED;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @unimplemented
|
|
|
|
*/
|
|
|
|
NTSTATUS
|
|
|
|
NTAPI
|
|
|
|
MmProtectMdlSystemAddress(IN PMDL MemoryDescriptorList,
|
|
|
|
IN ULONG NewProtect)
|
|
|
|
{
|
2011-09-11 11:47:28 +00:00
|
|
|
UNIMPLEMENTED;
|
|
|
|
return STATUS_NOT_IMPLEMENTED;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @unimplemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmProbeAndLockProcessPages(IN OUT PMDL MemoryDescriptorList,
|
|
|
|
IN PEPROCESS Process,
|
|
|
|
IN KPROCESSOR_MODE AccessMode,
|
|
|
|
IN LOCK_OPERATION Operation)
|
|
|
|
{
|
2011-09-11 11:47:28 +00:00
|
|
|
UNIMPLEMENTED;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @unimplemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmProbeAndLockSelectedPages(IN OUT PMDL MemoryDescriptorList,
|
|
|
|
IN LARGE_INTEGER PageList[],
|
|
|
|
IN KPROCESSOR_MODE AccessMode,
|
|
|
|
IN LOCK_OPERATION Operation)
|
|
|
|
{
|
2011-09-11 11:47:28 +00:00
|
|
|
UNIMPLEMENTED;
|
- Major rewrite of Memory Descriptor List (MDL) implementation (moving it towards using System PTEs).
- MmCreateMdl, MmSizeOfMdl: No Change.
- MmBuildMdlForNonPagedPool: Do not use MmGetPfnForProcess, just normal PMMPTE manipulation.
- This seems to cause issues in certain scenarios, because in ReactOS, nonpaged pool, a resident and guaranteed resources, does not always have its PDEs mapped!
- By calling MmGetPfnForProcess, this wound up in the annals of ReactOS mm code, which lazy-remapped the PDE. We detected this issue specifically in the cache manager, and fixed it there. It should not appear anywhere else.
- MmAllocatePagesForMdl, MmAllocatePagesForMdlEx, MmFreePagesFromMdl:
- The *Ex function is now implemented.
- Allocating pages now uses MiAllocatePagesForMdl, which is based on the older MmAllocPagesSpecifyRange.
- The code is cleaner, better commented, and better handles partial MDLs.
- Cache flags are still ignored (so the Ex functionality isn't really there).
- MmMapLockedPages, MmMapLockedPagesSpecifyCache, MmUnmapLockedPages:
- These functions now use System PTEs for the mappings, instead of the hacked-up "MDL Mapping Space".
- This frees up 256MB of Kernel Virtual Address Space.
- Takes advantage of all System PTE functionality.
- Once again, optimizations in the System PTE code will be felt here.
- For user-space mappings however, the old code is still kept and used.
- MiMapLockedPagesInUserSpace and MiUnMapLockedPagesInUserSpace are now in virtual.c and provide this.
- MmProbeAndLockPages, MmUnlockPages:
- The pages are actually probed now, in SEH. This did not seem to happen before (did someone misread the function's name?)
- Probe for write is only done for write access to user pages (as documented).
- We do not probe/check for write access for kernel requests (force Operation to be IoReadAccess).
- Proper locking is used now: Address Space lock for user mappings, PFN lock for kernel mappings.
- Faulting in pages (to make them available before locking) is now done outside the address space/PFN lock.
- You don't want to be holding a spinlock/mutex while doing disk I/O!
- For write/modify access, if the PTE is not writable, fail the request since the PTE protection overrides.
- However, if the PTE is writable but also copy on write, then we'll fault the page in for write access, which is a legitimate operation for certain user-mode scenarios.
- The old version always provided the CopyOnWrite behavior, even for non-CopyOnWrite pages!
- Reference and lock every valid page that has a PFN entry (non-I/O Pages).
- The older code did not seem to lock pages that had to be faulted in (weren't already valid).
- Cleanup the cleanup code (no pun intended). Because we now mark the pages as locked early-on, and because of changes in MmUnlockPages, we can simply use MmUnlockPages in case of error, since it will be able to fully back-out and references/locks that we did.
- Previous code attempted to do this on its own, in a pretty inconsistent manner, which would leave page leaks (both in references and lock count).
- In MmUnlockPages, not as many changes, but we now:
- Still make sure that an I/O Mapping MDL doesn't have valid PFN database pages (non-I/O).
- An MDL can cover pages that are both I/O mapped and RAM mapped, so we have to unlock/dereference the latter instead of skipping them as the old code did.
- Use the PFN lock when checking pages and unlocking/dereferencing them.
- Overall, non-complete MDLs are now marked by having a -1 PFN, and the MDL code has been updated to early-break out of page-scanning loops and/or ignore such pages, which can happen in a sparse MDL.
- Implementation has been tested on VMWare and QEMU for a variety of tasks and was found to be reliable and stable.
svn path=/trunk/; revision=41707
2009-06-30 08:29:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* @unimplemented
|
|
|
|
*/
|
|
|
|
VOID
|
|
|
|
NTAPI
|
|
|
|
MmMapMemoryDumpMdl(IN PMDL Mdl)
|
|
|
|
{
|
|
|
|
UNIMPLEMENTED;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* EOF */
|