microsoft / WSL

Issues found on WSL
https://docs.microsoft.com/windows/wsl
MIT License
17.48k stars 822 forks source link

Crash on TLS destructor call after libd3d12core.so has already been unloaded #12171

Open kasper93 opened 1 month ago

kasper93 commented 1 month ago

Windows Version

Microsoft Windows [Version 10.0.26100.2033]

WSL Version

2.3.24.0

Are you using WSL 1 or WSL 2?

Kernel Version

5.15.153.1-microsoft-standard-WSL2

Distro Version

2024.3

Other Software

OpenGL / mpv

Repro Steps

  1. mpv --no-config <video>
  2. Close with q key

Expected Behavior

Doesn't crash on exit.

Actual Behavior

* thread #42, name = 'vo', stop reason = signal SIGSEGV: address not mapped to object (fault address: 0x7ffec043ab10)
  * frame #0: 0x00007ffec043ab10
    frame #1: 0x00007ffff3269801 libc.so.6`__GI___nptl_deallocate_tsd [inlined] __GI___nptl_deallocate_tsd at nptl_deallocate_tsd.c:73:29
    frame #2: 0x00007ffff3269786 libc.so.6`__GI___nptl_deallocate_tsd at nptl_deallocate_tsd.c:22:1
    frame #3: 0x00007ffff326c52f libc.so.6`start_thread(arg=<unavailable>) at pthread_create.c:455:3
    frame #4: 0x00007ffff32e72b8 libc.so.6`__clone3 at clone3.S:78

After comparing the memory map before closing, the address we are trying to call 0x00007ffec043ab10 is inside /usr/lib/wsl/lib/libd3d12core.so. However, by the time libc tries to clean it up, libd3d12core.so has already been unloaded. As expected, adding a dummy dlopen("/usr/lib/wsl/lib/libd3d12core.so", RTLD_LAZY | RTLD_GLOBAL) to prevent it from being unloaded works around the crash.

I didn't look into exactly how libd3d12core.so is loaded/unloaded. It is likely handled somewhere in the OpenGL driver. However, it's clear that the pthread_key_delete() call is missing before it gets unloaded. And when libc tries to clean outstanding keys it's already too late.

Diagnostic Logs

To make bot happy after #12125. I can send actual crash dumps if needed, but they are too big for GitHub attachment. Generally the issue is trivially reproducible, so not much more digging into my logs is needed. WslLogs-2024-10-15_20-20-11.zip

github-actions[bot] commented 1 month ago
Diagnostic information ``` .wslconfig found Custom kernel command line found: 'mitigations=off cgroup_no_v1=all' Detected appx version: 2.3.24.0 ```