Support detached threads still running when the main thread exits ("thread leaks")

[RalfJung]

With #1284 we can run multi-threaded programs in Miri. However, right now we error if, when the main thread exits, there are any other threads still left. We should properly support that case.

For that, we need to:

• not complain about memory reachable from those threads as leaking.
• figure out what the semantics are wrt. TLS dtors. I assume those do not run for threads that are still around (after all the thread might be working on that piece of TLS right now, so a dtor would find it in an inconsistent state), but we should confirm that and leave appropriate comments.
• decide what to do about non-detached threads still running when the main thread exits. I am inclined to call those threads a leak, but maybe if they were spawned by a detached thread it's okay? IMO we should report a leak when e.g. the main thread spawns a thread and then neither detaches nor joins it.

Cc @vakaras

[oli-obk]

I always found it surprising that it was fine to have threads still exist when the main thead exits. The other threads don't get shut down time at all, they are just killed. I wonder if we should at least warn about it.

[RalfJung]

Some libraries might have "background threads" to do stuff, that does not seem unreasonable to me -- I doubt rayon's thread pool or crossbeam's epoch cleanup thread will be joined ever, and indeed I don't think there is a good way to join them.

[RalfJung]

In other words, "detaching" a thread IMO is an explicit marker that it is entirely okay for this thread to be suddenly killed when the program is done.

I am fine warning about this for non-detached threads, see item 3 above. Detaching then serves as explicit opt-out from that warning.

[oli-obk]

Detaching is solely a library thing. Dropping a JoinHandle is a nop afaict, there's no operation that guarantees that a thread has been detached, the user may at any point get a thread id from the ether and then decide to join the thread.

[RalfJung]

Dropping a JoinHandle is a nop afaict

Nope, docs explicitly say

A JoinHandle detaches the associated thread when it is dropped, which means that there is no longer any handle to thread and no way to join on it.

there's no operation that guarantees that a thread has been detached, the user may at any point get a thread id from the ether and then decide to join the thread.

It is UB to join on a detached thread, and @vakaras patch detects that UB.

[oli-obk]

oh, interesting. In that case it may be impossible to do a leak check without running those threads to completion after the main thread exits. Since they may never terminate, I'm not sure how we can achieve that though.

The reason I think it's impossible is that the threads may hold integer addresses of memory that they could free, but we can't tell whether that's the case.

[RalfJung]

The reason I think it's impossible is that the threads may hold integer addresses of memory that they could free, but we can't tell whether that's the case.

That's already a problem (#1318), so I don't think it should affect our decision here.

[oli-obk]

so... we treat everything that the thread refers to with provenance pointers the same way we treat them in statics? That seems reasonable to me, maybe with a flag to disable it specifically when ppl want to debug leaking from threads that they didn't clean up properly (e.g. when accidentally detaching them)

[RalfJung]

we treat everything that the thread refers to with provenance pointers the same way we treat them in statics? That seems reasonable to me

We have agreement then. :)

maybe with a flag to disable it specifically when ppl want to debug leaking from threads that they didn't clean up properly (e.g. when accidentally detaching them)

In that case I'd just say we have a flag to warn about detached threads still sitting around when the main thread exits -- there is little point in accepting that silently, but then complaining about all its stack memory leaking.

[oli-obk]

true! That's much better. Oh, we could even just warn when detaching, to show the user where it's happening 😄

[jonhoo]

Following on from jonhoo/bus#29 (comment), I believe that Rust currently does run destructors for anything in thread-local storage. It does this by using various hooks that the OS provides, such as __cxa_thread_atexit_impl on Linux/glibc.

To follow-on from your comment there Ralf, I believe that function is called when each thread exits. There is also atexit, which I believe is only called after all the threads have terminated, though I'm not entirely sure. You may need someone from rustc to confirm that part.

[RalfJung]

I made an experiment and it confirmed what I suspected -- destructors of still-running detached threads are not run.

When the process exits, I don't think the still-running threads are considered to "exit" it any clean sense. It's more like thread cancellation -- they get abruptly stopped without any cleanup.