[Arm64/Ubuntu] GCStress = 0x4/0x8/0xC/0xF Regression

[sdmaclea]

I am seeing ~10K new GCStress failures on a231c3a. Problem did not exist on 5d023eb

Looking at the logs, I would guess this was introduced by dotnet/coreclr#16384

I would guess this is not Arm64 or Ubuntu specific.

@jkotas @janvorli @mlabiuk
@dotnet/jit-contrib @dotnet/arm64-contrib

[mlabiuk]

Could you provide logs and how to reproduce test? Link to CI build?

[sdmaclea]

This is my log from
JIT/CodeGenBringUpTests/Add1/Add1.sh

corerun Add1.exe

Assert failure(PID 12683 [0x0000318b], Thread: 12683 [0x318b]): 
!CREATE_CHECK_STRING(pMT && pMT->Validate())
    File: src/vm/object.cpp Line: 733
    Image: bin/tests/pr1/Tests/coreoverlay/corerun

You should be able to reproduce locally by setting one of:
export COMPlus_GCStress=F
export COMPlus_GCStress=C
export COMPlus_GCStress=8
export COMPlus_GCStress=4

The assertion above typically indicates that the GC either:

• Moved an object without finding and updating all the references to the object
• Deleted an object because it failed to find all the references to the object

@Maoni0 @swgillespie FYI

[sdmaclea]

You should be able to trigger an x64 gcstress=0xc CI run with something like
@dotnet-bot test Ubuntu x64 Checked gcstress0xc

[Maoni0]

@RussKeldorph, could someone from your team take a look at these GCStress failures from jit tests?

[BruceForstall]

@Maoni0 It's almost certainly a regression due to VM change dotnet/coreclr#16384, not a JIT change, according to @sdmaclea 's investigation. Someone could verify that.

[janvorli]

Let me try to create a PR reverting that change and run the gc stress in the CI. I doubt it is that change though.

[janvorli]

Ah, I can see there is a bug that I've missed during the PR. The signalContextRecord is left uninitialized, no wonder it causes the crashes. @mlabiuk I'm going to send out a PR with a fix.

[janvorli]

Oh, I was just seeing things. It seems actually correct. Sorry for the false alarm.

[janvorli]

So there actually is a bug, but more subtle. I'll create a PR with a fix.

[sdmaclea]

FWIW I did run gcStress0xc on a231c3a~1 and it was working well.

[janvorli]

I've merged the fix in. It took more than 24 hours to run the CI gcstress0xc job.