I expect that this will break tracing tests. Tracing depends on C runtime shutdown notification to close the session:

Also, the same fundamental problem exists in regular coreclr as well.

It seems that the only way to fix the problem and keep the cleanup at exit untouched would be to link to ucrt dynamically. The cleanup then occurs in DllMain of the library on DLL_PROCESS_DETACH.

I meant the low level CRT cleanup, not the atexit stuff.

Tracing depends on C runtime shutdown notification to close the session

Playing the devil's advocate, you can always make that a at_quick_exit.

Ah, damn, that's C++ and not in ucrt, so that would not help.

Ah, damn, that's C++ and not in ucrt, so that would not help.

There is no C++ in the mix. Neither atexit or at_quick_exit are directly exposed by the ucrt.dll. They are statically linked stubs that use CRT global state for .exes and module-local state for .dlls. For example:

• Exiting via quick_exit from main still runs atexit handlers registered in other .dlls in the process, but it does not run at_quick_exit handlers registered in other .dlls in the process
• Exiting via quick_exit from .dll with statically linked CRT (the "global" state is module-local with static linking) runs at_quick_exit handlers registered in the given .dll, but it does not run atexit nor at_quick_exit handlers registered by the .exe

/azp run runtime-nativeaot-outerloop

/azp run runtime-nativeaot-outerloop

It looks like we do not have any diagnostics tests that would be sensitive for improperly closed eventpipe sessions.

I think we should also update #109006 (comment) to register the callbacks using at_quick_exit .

Looks good to me otherwise.

link to ucrt dynamically

Why don't we do it currently?

Why don't we do it currently?

We do not dynamically link to debug CRT to save us from troubles with deploying it to CI machines.

We do dynamically link to release CRT (ie the shipping bits link to release CRT).

Thanks, I remember trying to link some static libraries with dynamic CRT linkage and getting errors related to static/dynamic CRT mismatch. But that was years ago and I will try it again.