[x86 Windows] GC stackwalking crash with R2R PInvoke frames

[ViktorHofer]

Configuration: netcoreapp5.0-Windows_NT-Release-x86-CoreCLR_release-Windows.81.Amd64.Open

console.4a227872.log:
https://helix.dot.net/api/2019-06-17/jobs/7b2af8b3-f1bf-4cd8-912b-47f19883cf58/workitems/System.Xml.Xsl.XslTransformApi.Tests/files/console.4a227872.log

dotnet.exe.1760.dmp:
https://helix.dot.net/api/2019-06-17/jobs/7b2af8b3-f1bf-4cd8-912b-47f19883cf58/workitems/System.Xml.Xsl.XslTransformApi.Tests/files/dotnet.exe.1760.dmp

https://dev.azure.com/dnceng/public/_build/results?buildId=508623&view=ms.vss-test-web.build-test-results-tab&runId=16140738&paneView=debug&resultId=180629

cc @jkotas @buyaa-n @krwq

[jkotas]

We crash during GC while stackwalking GetFullPathNameW:

08 00dfc9e0 712d180b coreclr!WKS::GCHeap::WaitUntilGCComplete+0x24 [F:\workspace\_work\1\s\src\coreclr\src\gc\gcee.cpp @ 292] 
09 00dfca24 7134bcfe coreclr!Thread::RareDisablePreemptiveGC+0xc9 [F:\workspace\_work\1\s\src\coreclr\src\vm\threadsuspend.cpp @ 2530] 
0a 00dfcaa0 70d68b6b coreclr!JIT_PInvokeEndRarePath+0x2e [F:\workspace\_work\1\s\src\coreclr\src\vm\jithelpers.cpp @ 4923] 
0b 00dfcaf0 099055f7 Interop+Kernel32.GetFullPathNameW(Char ByRef, UInt32, Char ByRef, IntPtr)+0x57
0c 00dfcb08 09907cc2 System_Private_CoreLib!System.IO.PathHelper.GetFullPathName(System.ReadOnlySpan`1<Char>, System.Text.ValueStringBuilder ByRef)+0x989fbb47 [/_/src/libraries/System.Private.CoreLib/src/System/IO/PathHelper.Windows.cs @ 91] 

[jkotas]

#31659 hit it as well. That one crashes under:

08 0a66e4c0 73a5bcfe coreclr!Thread::RareDisablePreemptiveGC+0xc9 [F:\workspace\_work\1\s\src\coreclr\src\vm\threadsuspend.cpp @ 2530] 
09 0a66e53c 7347a038 coreclr!JIT_PInvokeEndRarePath+0x2e [F:\workspace\_work\1\s\src\coreclr\src\vm\jithelpers.cpp @ 4923] 
0a 0a66e584 0af36ddd Interop+Ole32.CoCreateGuid(System.Guid ByRef)+0x38
0b 0a66e5a4 0a698408 System_Private_CoreLib!System.Guid.NewGuid()+0x97a54e5d [/_/src/libraries/System.Private.CoreLib/src/System/Guid.Windows.cs @ 23] 
0c 0a66e658 0a6983ad CoreFx_Private_TestUtilities!System.IO.FileCleanupTestBase..ctor()+0x40
0d 0a66e664 0b2f0c8d System_IO_FileSystem_Tests!System.IO.Tests.FileSystemTest..ctor()+0x1d

[jkotas]

@fadimounir This is most likely caused by #26834. I am going to revert it. Sorry...

[jkotas]

Revert in #31811

[fadimounir]

@jkotas Ok thanks. I'll take a look at it. Quite interesting... I've been running a couple of instances of gcstress tests with the fix for a couple of days, and none of them crashed

[jkotas]

This may be the type of issue that gc stress is not good at catching. Let me know if you need to help with a repro. I think running Guid.NewGuid and some allocations, on multiple threads, with release build, and tiered compilation disabled should reproduce it reliably.

[fadimounir]

Let me know if you need to help with a repro

@jkotas I tried to repro this with a multi-threaded test that calls Guid.NewGuid() and does a bunch of allocations, and GC.Collect calls in parallel, but couldn't get it to fail (tested with the pinvoke changes that you reverted). Tried both release and checked builds, with tiered compilation disabled as you suggested to keep using the R2R code.
I'll leave the test running for a while longer, maybe it would get luck and fail...

I think at this point, it would be more useful to investigated a dump file but looks like the one mentioned in this issue is no longer available.

Any suggestions? Create a PR with the changes and keep running CI until it fails?

[jkotas]

I am not able to reproduce it either. Yes, create PR and try to rerun x86 legs on it a few times. If you are not able to hit it after a day of trying, commit it and hopefully we are going to get a dump eventually.

[jkotas]

Duplicate of #34279