Deadlock in loader lock (CoreCLR) from background GC in XAML apps

[Sergio0694]

Description

We're hitting random hangs in the Microsoft Store in debug builds on CoreCLR.

Context from @hoyosjs:

"This is a loader lock deadlock
[...]

• Thread 3 tries to do a GC from the added pressure, which tries to start the background GC thread.
• BG thread (and a ton of others) fail to start - they are stuck in LdrpInitializeThread on LdrpAcquireLoaderLock
• They can't start bc thread 16 has the lock (LdrShutdownThread) and it's doing TLS destruction. Part of it must happen in COOP GC. Transitioning to coop means it waits for the original GC...

Crash dump (MSFT only): loader_hang_dmp.zip

Reproduction Steps

It just happens when using the app for a while.

Configuration

• .NET 9.0.202

[jkotas]

Duplicate of #110350 . cc @VSadov

[jkotas]

@Sergio0694 Could you please validate that it is fixed in 9.0.5?

(backport PR #113055)

[Sergio0694]

Oh that's nice! I can try, yes, though it might be slightly tricky as I can't repro this myself.
Just to confirm, is this only affecting CoreCLR, or does this affect Native AOT as well?

[VSadov]

Yes, this looks like #110350 . The fix has been backported to 9.0, but may not be available till the next servicing release.
As Jan suggested - it would be very helpful if you could try with 9.0.5 bits (i.e. #113055)

[jkotas]

Just to confirm, is this only affecting CoreCLR, or does this affect Native AOT as well?

It does not affect Native AOT. The fix was to switch regular CoreCLR to how Native AOT works.

[Sergio0694]

Perfect, thank you for confirming! At least we know this only affects our dev inner loop then 🙂