and in some cases, return an adequate value. In some places however, I add an ASSERT(FALSE); before keeping the while (TRUE); in critical regions.
(Only x86, as well as in r58110). More ASSERTS (in debug mode) or BSODs may appear, instead of having hung threads.
Part 1-bis/2
svn path=/trunk/; revision=58111
- Fix several locking issues to better fit with RosMm
- Page out is critical as we use it now (shutdown) so force address space locking
- Fix an address space lock leak for VM regions
- Quiet some debugging and enable/change some other
- Still some lingering issues in page out exist which seem to be related to zefklop's PDE ref counting stuff and a weird one where we fault on the PFN database
- Besides the aforementioned issues, NewCC is surprisingly solid. I was able to boot to 3rd stage without a problem and run some apps.
svn path=/trunk/; revision=55895
A reintegration checkpoint for the NewCC branch, brought to you by Team NewCC.
Differences with current ReactOS trunk:
* A new memory area type, MEMORY_AREA_CACHE, is added, which represents a mapped region of a file. In NEWCC mode, user sections are MEMORY_AREA_CACHE type as well, and obey new semantics. In non-NEWCC mode, they aren't used.
* A way of claiming a page entry for a specific thread's work is added. Placing the special SWAPENTRY value MM_WAIT_ENTRY in a page table, or in a section page table should indicate that memory management code is intended to wait for another thread to make some status change before checking the state of the page entry again. In code that uses this convention, a return value of STATUS_SUCCESS + 1 is used to indicate that the caller should use the MiWaitForPageEvent macro to wait until somebody has change the state of a wait entry before checking again. This is a lighter weight mechanism than PAGEOPs.
* A way of asking the caller to perform some blocking operation without locks held is provided. This replaces some spaghettified code in which locks are repeatedly taken and broken by code that performs various blocking operations. Using this mechanism, it is possible to do a small amount of non-blocking work, fill in a request, then return STATUS_MORE_PROCESSING_REQUIRED to request that locks be dropped and the blocking operation be carried out. A MM_REQUIRED_RESOURCES structure is provided to consumers of this contract to use to accumulate state across many blocking operations. Several functions wrapping blocking operations are provided in ntoskrnl/cache/reqtools.c.
* Image section pages are no longer direct mapped. This is done to simplify consolidation of ownership of pages under the data section system. At a later time, it may be possible to make data pages directly available to image sections for the same file. This is likely the only direct performance impact this code makes on non-NEWCC mode.
RMAPs:
* A new type of RMAP entry is introduced, distinguished by RMAP_IS_SEGMENT(Address) of the rmap entry. This kind of entry contains a pointer to a section page table node in the Process pointer, which in turn links back to the MM_SECTION_SEGMENT it belongs to. Therefore, a page belonging only to a segment (that is, a segment page that isn't mapped) can exist and be evicted using the normal page eviction mechanism in balance.c. Each of the rmap function has been modified to deal with segment rmaps.
* The low 8 bits of the Address field in a segment rmap denote the entry number in the generic table node pointed to by Process that points to the page the rmap belongs to. By combining them, you can determine the file offset the page belongs to.
* In NEWCC mode, MmSharePageEntry/UnsharePageEntry are not used, and instead the page reference count is used to keep track of the number of mappings of a page, allowing the last reference expiring to allow the page to be recycled without much intervention. These are still used in non-NEWCC mode. One change has been made, the count fields have been narrowed by 1 bit to make room for a dirty bit in SSE entries, needed when a page is present but unmapped.
Section page tables:
* The section page tables are now implemented using RtlGenericTables. This enables a fairly compact representation of section page tables without having the existence of a section object imply 4k of fake PDEs. In addition, each node in the generic table has a wide file offset that is a multiple of 256 pages, or 1 megabyte total. Besides needing wide file offsets, the only other visible change caused by the switch to generic tables for section page tables is the need to lock the section segment before interacting with the section page table.
Eviction:
* Page eviction in cache sections is accomplished by MmpPageOutPhysicalAddress. In the case of a shared page, it tries to remove all mappings of the indicated page. If this process fails at any point, the page will simply be drawn back into the target address spaces. After succeeding at this, if TRUE has been accumulated into the page's dirty bit in the section page table, it is written back, and then permanently removed.
NewCC mode:
* NEWCC mode is introduced, which rewrites the file cache to a set of cache stripes actively mapped, along with unmapped section data.
* NewCC is more authentic in its interpretation of the external interface to the windows cache than the current cache manager, implementing each of the cache manager functions according to the documented interface with no preconceived ideas about how anything should be implemented internally. Cache stripes are implemented on top of section objects, using the same memory manager paths, and therefore economizing code and complexity. This replaces a rather complicated system in which pages can be owned by the cache manager and the memory manager simultaneously and they must cooperate in a fairly sophisticated way to manage them. Since they're quite interdependent in the current code, modifying either is very difficult. In NEWCC, they have a clear division of labor and thus can be worked on independently.
* Several third party filesystems that use the kernel Cc interface work properly using NEWCC, including matt wu's ext3 driver.
* In contrast with code that tries to make CcInitializeCacheMap and CcUninitializeCacheMap into a pair that supports reference counting, NEWCC lazily initializes the shared and private cache maps as needed and uses the presence of a PrivateCacheMap on at least one file pointing to the SharedCacheMap as an indication that the FILE_OBJECT reference in the SharedCacheMap should still be held. When the last PrivateCacheMap is discarded, that's the appropriate time to tear down caching for a specific file, as the SharedCacheMap data is allowed to be saved and reused. We honor this by making the SharedCacheMap into a depot for keeping track of the PrivateCacheMap objects associated with views of a file.
svn path=/trunk/; revision=55833
The cache manager rewrite I started years ago has finally appeared in
ReactOS' trunk and although at this point it's not quite perfectly
integrated, it's enough to boot up the bootcd or livecd. To check out
the more mature original, check out arty-newcc-reactos, branch
arty-newcc on bitbucket.org . Amine Khaldi encouraged me quite a bit
to not give up on it, and was able to reach out and be an advocate
when i really wasn't able to. Others agree that the time has come to
begin removing the old cache manager. I expect the remaining problems
in the version going to trunk will be taken care of relatively
quickly.
The motivation for this effort lies in the particularly hairy
relationship between ReactOS' cache manager and data sections. This
code completely removes page sharing between cache manager and section
and reimagines cache manager as being a facility layered on the memory
manager, not really caring about individual pages, but simply managing
data section objects where caching might occur.
It took me about 2 years to do the first pass of this rewrite and most
of this year to fix some lingering issues, properly implement demand
paging in ReactOS (code which didn't come with this patch in a
recognizable form), and finish getting the PrivateCacheMap and
SharedCacheMap relationship correct.
Currently, the new ntoskrnl/cache directory contains an own
implementation of data file sections. After things have settled down,
we can begin to deprecate and remove the parts of ReactOS' section
implementation that depend on a close relationship with cache
manager. Eventually, I think that the extra code added to
ntoskrnl/cache/section will be removed and ReactOS' own sections will
replace the use of the special MM_CACHE_SECTION_SEGMENT in the cache
path.
Note also, that this makes all cache manager (and new section parts)
use wide file offsets. If my section code were to take over other
parts of the ReactOS memory manager, they would also benefit from
these improvements.
I invite anyone who wants to to peek at this code and fix whatever
bugs can be found.
svn path=/trunk/; revision=49423
