reactos/win32ss/gdi/ntgdi/gdipool.c

370 lines
11 KiB
C
Raw Normal View History

/*
* PROJECT: ReactOS win32 kernel mode subsystem
* LICENSE: GPL - See COPYING in the top level directory
* FILE: win32ss/gdi/ntgdi/gdipool.c
* PURPOSE: Static size allocator for user mode object attributes
* PROGRAMMERS: Timo Kreuzer
*/
/* INCLUDES ******************************************************************/
#include <win32k.h>
#define NDEBUG
#include <debug.h>
typedef struct _GDI_POOL_SECTION
{
LIST_ENTRY leInUseLink;
LIST_ENTRY leReadyLink;
PVOID pvBaseAddress;
ULONG ulCommitBitmap;
ULONG cAllocCount;
RTL_BITMAP bitmap;
ULONG aulBits[1];
} GDI_POOL_SECTION, *PGDI_POOL_SECTION;
typedef struct _GDI_POOL
{
ULONG ulTag;
ULONG cjAllocSize;
ULONG cjSectionSize; // 32 * cjAllocSize, rounded up to pages
ULONG cSlotsPerSection;
ULONG cEmptySections;
EX_PUSH_LOCK pushlock; // For pool growth
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
#if DBG_ENABLE_EVENT_LOGGING
SLIST_HEADER slhLog;
#endif
LIST_ENTRY leInUseList;
LIST_ENTRY leEmptyList;
LIST_ENTRY leReadyList;
} GDI_POOL;
#define GDI_POOL_ALLOCATION_GRANULARITY 64 * 1024
static
PGDI_POOL_SECTION
GdiPoolAllocateSection(PGDI_POOL pPool)
{
PGDI_POOL_SECTION pSection;
PVOID pvBaseAddress;
SIZE_T cjSize;
NTSTATUS status;
/* Allocate a section object */
cjSize = sizeof(GDI_POOL_SECTION) + pPool->cSlotsPerSection / sizeof(ULONG);
pSection = EngAllocMem(0, cjSize, pPool->ulTag);
if (!pSection)
{
return NULL;
}
/* Reserve user mode memory */
cjSize = GDI_POOL_ALLOCATION_GRANULARITY;
pvBaseAddress = NULL;
status = ZwAllocateVirtualMemory(NtCurrentProcess(),
&pvBaseAddress,
0,
&cjSize,
MEM_RESERVE,
PAGE_READWRITE);
if (!NT_SUCCESS(status))
{
EngFreeMem(pSection);
return NULL;
}
/* Initialize the section */
pSection->pvBaseAddress = pvBaseAddress;
pSection->ulCommitBitmap = 0;
pSection->cAllocCount = 0;
RtlInitializeBitMap(&pSection->bitmap,
pSection->aulBits,
pPool->cSlotsPerSection);
RtlClearAllBits(&pSection->bitmap);
/* Return the section */
return pSection;
}
static
VOID
GdiPoolDeleteSection(PGDI_POOL pPool, PGDI_POOL_SECTION pSection)
{
NTSTATUS status;
SIZE_T cjSize = 0;
/* Should not have any allocations */
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
if (pSection->cAllocCount != 0)
{
DPRINT1("There are %lu allocations left, section=%p, pool=%p\n",
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
pSection->cAllocCount, pSection, pPool);
DBG_DUMP_EVENT_LIST(&pPool->slhLog);
ASSERT(FALSE);
}
/* Release the virtual memory */
status = ZwFreeVirtualMemory(NtCurrentProcess(),
&pSection->pvBaseAddress,
&cjSize,
MEM_RELEASE);
ASSERT(NT_SUCCESS(status));
/* Free the section object */
EngFreeMem(pSection);
}
PVOID
NTAPI
GdiPoolAllocate(
PGDI_POOL pPool)
{
PGDI_POOL_SECTION pSection;
ULONG ulIndex, cjOffset, ulPageBit;
PLIST_ENTRY ple;
PVOID pvAlloc, pvBaseAddress;
SIZE_T cjSize;
NTSTATUS status;
/* Disable APCs and acquire the pool lock */
KeEnterCriticalRegion();
ExAcquirePushLockExclusive(&pPool->pushlock);
/* Check if we have a ready section */
if (!IsListEmpty(&pPool->leReadyList))
{
/* Get a free section */
ple = pPool->leReadyList.Flink;
pSection = CONTAINING_RECORD(ple, GDI_POOL_SECTION, leReadyLink);
if (pSection->cAllocCount >= pPool->cSlotsPerSection)
{
DPRINT1("pSection->cAllocCount=%lu, pPool->cSlotsPerSection=%lu\n",
pSection->cAllocCount, pPool->cSlotsPerSection);
DBG_DUMP_EVENT_LIST(&pPool->slhLog);
ASSERT(FALSE);
}
ASSERT(pSection->cAllocCount < pPool->cSlotsPerSection);
}
else
{
/* No, check if we have something on the empty list */
if (!IsListEmpty(&pPool->leEmptyList))
{
/* Yes, remove it from the empty list */
ple = RemoveHeadList(&pPool->leEmptyList);
pSection = CONTAINING_RECORD(ple, GDI_POOL_SECTION, leInUseLink);
pPool->cEmptySections--;
ASSERT(pSection->cAllocCount == 0);
}
else
{
/* No, allocate a new section */
pSection = GdiPoolAllocateSection(pPool);
if (!pSection)
{
DPRINT1("Couldn't allocate a section\n");
pvAlloc = NULL;
goto done;
}
}
/* Insert it into the in-use and ready list */
InsertHeadList(&pPool->leInUseList, &pSection->leInUseLink);
InsertHeadList(&pPool->leReadyList, &pSection->leReadyLink);
}
/* Find and set a single bit */
ulIndex = RtlFindClearBitsAndSet(&pSection->bitmap, 1, 0);
ASSERT(ulIndex != MAXULONG);
/* Calculate the allocation address */
cjOffset = ulIndex * pPool->cjAllocSize;
pvAlloc = (PVOID)((ULONG_PTR)pSection->pvBaseAddress + cjOffset);
/* Check if memory is committed */
ulPageBit = 1 << (cjOffset / PAGE_SIZE);
ulPageBit |= 1 << ((cjOffset + pPool->cjAllocSize - 1) / PAGE_SIZE);
if ((pSection->ulCommitBitmap & ulPageBit) != ulPageBit)
{
/* Commit the pages */
pvBaseAddress = PAGE_ALIGN(pvAlloc);
cjSize = ADDRESS_AND_SIZE_TO_SPAN_PAGES(pvAlloc, pPool->cjAllocSize) * PAGE_SIZE;
status = ZwAllocateVirtualMemory(NtCurrentProcess(),
&pvBaseAddress,
0,
&cjSize,
MEM_COMMIT,
PAGE_READWRITE);
if (!NT_SUCCESS(status))
{
pvAlloc = NULL;
goto done;
}
pSection->ulCommitBitmap |= ulPageBit;
}
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
/* Increase alloc count */
pSection->cAllocCount++;
ASSERT(RtlNumberOfSetBits(&pSection->bitmap) == pSection->cAllocCount);
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
DBG_LOGEVENT(&pPool->slhLog, EVENT_ALLOCATE, pvAlloc);
/* Check if section is now busy */
if (pSection->cAllocCount == pPool->cSlotsPerSection)
{
/* Remove the section from the ready list */
RemoveEntryList(&pSection->leReadyLink);
}
done:
/* Release the pool lock and enable APCs */
ExReleasePushLockExclusive(&pPool->pushlock);
KeLeaveCriticalRegion();
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
DPRINT("GdiPoolallocate: %p\n", pvAlloc);
return pvAlloc;
}
VOID
NTAPI
GdiPoolFree(
PGDI_POOL pPool,
PVOID pvAlloc)
{
PLIST_ENTRY ple;
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
PGDI_POOL_SECTION pSection = NULL;
ULONG_PTR cjOffset;
ULONG ulIndex;
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
DPRINT("GdiPoolFree: %p\n", pvAlloc);
/* Disable APCs and acquire the pool lock */
KeEnterCriticalRegion();
ExAcquirePushLockExclusive(&pPool->pushlock);
/* Loop all used sections */
for (ple = pPool->leInUseList.Flink;
ple != &pPool->leInUseList;
ple = ple->Flink)
{
/* Get the pointer to the section */
pSection = CONTAINING_RECORD(ple, GDI_POOL_SECTION, leInUseLink);
/* Calculate offset */
cjOffset = (ULONG_PTR)pvAlloc - (ULONG_PTR)pSection->pvBaseAddress;
/* Check if the allocation is from this section */
if (cjOffset < pPool->cjSectionSize)
{
/* Calculate the index of the allocation */
ulIndex = cjOffset / pPool->cjAllocSize;
/* Mark it as free */
ASSERT(RtlTestBit(&pSection->bitmap, ulIndex) == TRUE);
RtlClearBit(&pSection->bitmap, ulIndex);
/* Decrease allocation count */
pSection->cAllocCount--;
ASSERT(RtlNumberOfSetBits(&pSection->bitmap) == pSection->cAllocCount);
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
DBG_LOGEVENT(&pPool->slhLog, EVENT_FREE, pvAlloc);
/* Check if the section got valid now */
if (pSection->cAllocCount == pPool->cSlotsPerSection - 1)
{
/* Insert it into the ready list */
InsertTailList(&pPool->leReadyList, &pSection->leReadyLink);
}
/* Check if it got empty now */
else if (pSection->cAllocCount == 0)
{
/* Remove the section from the lists */
RemoveEntryList(&pSection->leInUseLink);
RemoveEntryList(&pSection->leReadyLink);
if (pPool->cEmptySections >= 1)
{
/* Delete the section */
GdiPoolDeleteSection(pPool, pSection);
}
else
{
/* Insert it into the empty list */
InsertHeadList(&pPool->leEmptyList, &pSection->leInUseLink);
pPool->cEmptySections++;
}
}
goto done;
}
}
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
DbgPrint("failed to free. pvAlloc=%p, base=%p, size=%lx\n",
pvAlloc, pSection ? pSection->pvBaseAddress : NULL, pPool->cjSectionSize);
ASSERT(FALSE);
// KeBugCheck()
done:
/* Release the pool lock and enable APCs */
ExReleasePushLockExclusive(&pPool->pushlock);
KeLeaveCriticalRegion();
}
PGDI_POOL
NTAPI
GdiPoolCreate(
ULONG cjAllocSize,
ULONG ulTag)
{
PGDI_POOL pPool;
/* Allocate a pool object */
pPool = EngAllocMem(0, sizeof(GDI_POOL), 'lopG');
if (!pPool) return NULL;
/* Initialize the object */
ExInitializePushLock(&pPool->pushlock);
InitializeListHead(&pPool->leInUseList);
InitializeListHead(&pPool->leReadyList);
InitializeListHead(&pPool->leEmptyList);
pPool->cEmptySections = 0;
pPool->cjAllocSize = cjAllocSize;
pPool->ulTag = ulTag;
pPool->cjSectionSize = GDI_POOL_ALLOCATION_GRANULARITY;
pPool->cSlotsPerSection = pPool->cjSectionSize / cjAllocSize;
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
DBG_INITLOG(&pPool->slhLog);
return pPool;
}
VOID
NTAPI
GdiPoolDestroy(PGDI_POOL pPool)
{
PGDI_POOL_SECTION pSection;
PLIST_ENTRY ple;
/* Loop all empty sections, removing them */
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
while (!IsListEmpty(&pPool->leEmptyList))
{
/* Delete the section */
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
ple = RemoveHeadList(&pPool->leEmptyList);
pSection = CONTAINING_RECORD(ple, GDI_POOL_SECTION, leInUseLink);
GdiPoolDeleteSection(pPool, pSection);
}
/* Loop all ready sections, removing them */
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
while (!IsListEmpty(&pPool->leInUseList))
{
/* Delete the section */
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
ple = RemoveHeadList(&pPool->leInUseList);
pSection = CONTAINING_RECORD(ple, GDI_POOL_SECTION, leInUseLink);
GdiPoolDeleteSection(pPool, pSection);
}
[WIN32K] Rewrite of the GDI handle manager - The old handle manager used a completely retarded spinlock in combination with KeDelayExecutionThread() for both exclusive and shared locks. This is probably the most uneffective algorithm possible. It was also duplicating code everywhere and it was a overall mess It is now replaced with a lock-free reference counter for shared locks and a pushlock for exclusive locks. -> Better performance and scalability. - Allocate user mode object attributes from the new gdi pool. This way, we don't need any caching, since the pool serves as a cache. Its also much faster and uses much less memory. - Allow object allocations of different size, instead of fixed size from a table. This way a single allocation can take care of actual needs. - Allow allcoating objects without a handle and insert them into the handle table later - Properly synchronize the process GDIHandleCount. Now gdiview and taskmanager show the correct number of gdi handles. - Implement a new event tracking system, that is capable of tracking all reverences and locks of objects and pool allocations to help track possible leaks - Make sure that all objects of a process are deleted in cleanup - Make sure all usermode memory allocations are freed, when cleaning up the process pool. - Make sure that each object type is using the correct type of lock (either shared or exclusive, not a mixture) - Fix some object / reference leaks - Lots of inferface improvements - Use global variables for certain things instead of members in the mapped gdi handle table - Make IntSysCreateRectpRgn create a region without a handle - Fix detection od source and mask use in GreStretchBltMask - Use GDIOBJ_bLockMultipleObjects in NtGdiCombineRegion to avoid possible deadlocks - Fix NtGdiAbortPath to reset DCPATH_ACTIVE flag in the dc and only bail out on error, instead of always - Replace DC_AllocateDcAttr and DC_AllocDcAttr with DC_bAllocDcAttr using the new user mode pool - Remove DCU_SyncDcAttrtoUser and DCU_SynchDcAttrtoUser. Those functions were unused and didn't do anything useful anyway, - Replace IntGdiSetDCOwnerEx and DC_SetOwnership with GreSetDCOwner, remove unused NoSetBrush parameter - Replace GDIOBJ_bValidateHandle and IsObjectDead with GreIsHandleValid - Chage GDIOBJ_bLockMultipleObjects: pass object type, return a BOOL, whether all objects could be locked, cleanup on failure svn path=/trunk/; revision=51470
2011-04-28 08:26:46 +00:00
DBG_CLEANUP_EVENT_LIST(&pPool->slhLog);
EngFreeMem(pPool);
}