Exiting a process calls exit() which isn’t thread-safe

[DemiMarie]

On Unix-like platforms, std::process::exit calls libc::exit(). However, this can lead to undefined behavior in a multithreaded process.

In my case, this showed up as use-after-free crashes when running the testsuite for RPM Oxide. librpm registers an atexit() handler to clean up global resources, but this causes some of the resources to be freed while other threads are still using them. Other projects have had similar issues, such as the Rust bindings for RocksDB.

The best answer I know of is to call quick_exit instead, which only calls functions registered with at_quick_exit. Such functions should be safe to be called while other threads are running.

[jyn514]

Hmm, it seems really weird to me for exit() not to call libc::exit. It seems like the issue is libraries shouldn't be registering exit handler that aren't thread safe?

I've seen a similar issue before: sfackler/rust-openssl#1293

[jyn514]

I'm going to move this from T-libs-impl to T-libs because not calling libc::exit is user-facing.

The best answer I know of is to call quick_exit instead, which only calls functions registered with at_quick_exit. Such functions should be safe to be called while other threads are running.

I don't see how would quick_exit be different from exit if you registered a non-thread-safe exit handler (by at_quick_exit). You should either use _Exit or not registering one at all.

[hameerabbasi]

Removing I-prioritize and adding I-nominate as discussed on Zulip.

@rustbot labels -I-prioritize +I-nominate

[DemiMarie]

The best answer I know of is to call quick_exit instead, which only calls functions registered with at_quick_exit. Such functions should be safe to be called while other threads are running.

I don't see how would quick_exit be different from exit if you registered a non-thread-safe exit handler (by at_quick_exit). You should either use _Exit or not registering one at all.

The assumption (from an old C++ proposal I read) is that doing so is considered a bug on the part of the caller of at_quick_exit.
In particular, quick_exit does not call C++ static destructors for this reason.