[all-commits] [llvm/llvm-project] 0d5b51: [sanitizer][asan][win] Only unmap unneeded shadow ...

alvinhochun via All-commits all-commits at lists.llvm.org
Tue May 2 09:23:33 PDT 2023


  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 0d5b51e0ac885b6119e788cf6f0a6836c3908afc
      https://github.com/llvm/llvm-project/commit/0d5b51e0ac885b6119e788cf6f0a6836c3908afc
  Author: Alvin Wong <alvin at alvinhc.com>
  Date:   2023-05-03 (Wed, 03 May 2023)

  Changed paths:
    M compiler-rt/lib/asan/asan_win.cpp

  Log Message:
  -----------
  [sanitizer][asan][win] Only unmap unneeded shadow memory on x86_64

D21942 / 1128db8fe1c13800ebc77206efc50d0a219b8750 added support for
committing shadow memory on demand on Win 64-bit. The reason it is not
enabled on 32-bit wasn't clear but the page table overhead on Windows 7
may be a contributing factor.

In `AsanMapUnmapCallback::OnUnmap`, `FlushUnneededASanShadowMemory` is
called to release shadow memory. It calls `ReleaseMemoryPagesToOS`,
which had been a no-op on Windows, until D95892 /
81b1d3da094c54ffd75e05c8d4683792edf17f4c in which it was changed to
unmap full pages that the memory region covers. This was done on both
32-bit and 64-bit.

AddressSanitizerInterface.GetHeapSizeTest appears to fail on i686
targets as a side effect of this. This test allocates and frees a huge
chunk of memory which causes shadow memory to be unmapped immediately.
When the test allocates the chunk of memory a second time, asan tries to
reuse the same shadow memory region, but because the shadow memory has
now been unmapped, it causes an access violation and crashes the test.

x86_64 is not affected, because the code that handles commiting shadow
memory on demand also handles this situation, allowing the test to work
without crashing.

Therefore, this patch changes `FlushUnneededASanShadowMemory` on Windows
to only release/unmap the shadow memory on x86_64 to stop this from
happening on i686.

Differential Revision: https://reviews.llvm.org/D149025




More information about the All-commits mailing list