Re: seccomp causes pthread_join() to hang

[Rich Felker <dalias@aerifal.cx>]

On Wed, Jun 26, 2019 at 12:18:05AM +0100, Radostin Stoyanov wrote:
> Hello,
> 
> In the test suite of CRIU [1] we have noticed an interesting bug
> which is caused by commit 8f11e6127fe93093f81a52b15bb1537edc3fc8af
> ("track all live threads in an AS-safe, fully-consistent linked
> list") [2].
> 
> When seccomp is used in a multithreaded application it may cause
> pthread_join() to hang.
> 
> This is a minimal application to reproduce the issue:
> 
> 
> #include <errno.h>
> #include <seccomp.h>
> #include <stdio.h>
> #include <stdlib.h>
> #include <string.h>
> #include <pthread.h>
> #include <unistd.h>
> 
> static void *fn()
> {
>     scmp_filter_ctx ctx = seccomp_init(SCMP_ACT_KILL);
>     if (!ctx) {
>         perror("seccomp_init");
>         goto err;
>     }
> 
>     if (seccomp_load(ctx) < 0) {
>         perror("seccomp_load");
>         goto err;
>     }
> 
>     /* This should cause SIG_KILL */
>     getpid();
> err:
>     return (void *)1;
> }
> 
> int main()
> {
>     pthread_t t1;
> 
>     if (pthread_create(&t1, NULL, fn, NULL)) {
>         perror("pthread_create");
>         return -1;
>     }
> 
>     if (pthread_join(t1, NULL)) {
>         perror("pthread_join");
>         return -1;
>     }
> 
>     return 0;
> }
> 
> 
> Expected behaviour: Thread t1 should receive SIG_KILL and the main
> thread should return 0.
> Actual behaviour: pthread_join() hangs.
> Reproducibility: Always
> Regression: Yes
> 
> 
> This bug can be reproduced with Alpine 3.10 ($ docker run -it
> alpine:3.10 sh).

A fundamental property of the pthread API, and part of why threads are
a much better primitive than processes for lots of purposes, is that
threads are not killable; only whole processes are. Any configuration
that results in a thread being terminated out from under the process
has all sorts of extremely dangerous conditions with memory/locks
being left in inconsistent state, tid reuse while the application
thinks the old thread is still alive, etc., and fundamentally can't be
supported. What you're seeing is exposure of a serious existing
problem with this seccomp usage, not a regression.

Rich