decommit unusable page(s) in medium heap blocks (bug 5822302) #840
Conversation
leirocks
force-pushed
the
mediumdecommit
branch
from
6225048 to
5d6803d
Compare
leirocks
force-pushed
the
mediumdecommit
branch
from
5d6803d to
2f52f5d
Compare
|
The behavior on free is a little concerning in the case where commit failed. Is it possible to just call PartialDecommit on the decommitted page to update the PageAllocator state and then call release on the entire set of pages? Release doesn't always just decommit, it can also call VirtualFree in some cases, so just decommitting if the commit fails might not be enough. |
|
Is this tied to #742? |
|
yes |
leirocks
force-pushed
the
mediumdecommit
branch
2 times, most recently
from
655fe7d to
0a9e42f
Compare
lib/Common/Memory/HeapInfo.cpp
Outdated
| return ((MediumAllocationBlockAttributes::PageCount*AutoSystemInfo::PageSize) % sizeCat) / AutoSystemInfo::PageSize; | ||
| } | ||
| /* static */ | ||
| void MediumAllocationBlockAttributes::ProtectUnusablePages(HeapBlock* heapBlock) |
There was a problem hiding this comment.
I'm not sure why this is here and not in the HeapBlock- GetUnusablePageCount can be argued as getting an attribute of the block, this is definitely not an attribute. Also, nitpicky but please fix spacing of usage of *- not sure why its the only operator that you use with no spaces
in medium heap blocks, if the object size is bigger than 1 page, it can cause the whole last 1~3 page(s) never been allocated. when such case is hit, decommit those pages to save memory as well as capture corruption. when returning the pages back to page allocator, we should commit those pages again for reuse, in case of OOM here, just decommit all pages in the heap block and let the page allocator to manage the decommitted pages. in rescan code, asserting that we never scan those unallocatable pages
it's hard to make the partial decommit withou refactoring page allocator, especially in background releasing pages, there will have race issues. changing to protect with NOACCESS for now. In the future after we implemented the capability of partially commit pages in pageAllocator we can change to use that, because that can save some memories.
…, do specialization on heapblock itself
|
|
|
need to merge to master, see #985 |
4 participants
in medium heap blocks, if the object size is bigger than 1 page, it can cause the whole last 1~3 page(s) never been allocated. when such case is hit, decommit those pages to save memory as well as capture corruption. when returning the pages back to page allocator, we should commit those pages again for reuse, in case of OOM here, just decommit all pages in the heap block and let the page allocator to manage the decommitted pages.
in rescan code, asserting that we never scan those unallocatable pages
Fixes #742