Re: [PATCH] libgomp: fix hang on fatal error

[Jakub Jelinek <jakub@redhat.com>]

On Fri, Oct 14, 2022 at 06:04:06PM +0100, Andrew Stubbs wrote:
> This patch fixes a problem in which fatal errors inside mutex-locked regions
> (i.e. basically anything in the plugin) will cause it to hang up trying to
> take the lock to clean everything up.
> 
> Using abort() instead of exit(1) bypasses the atexit handlers and solves the
> problem.
> 
> OK for mainline?
> 
> Andrew

> libgomp: fix hang on fatal error
> 
> Don't try to clean up if a fatal error occurs in libgomp.  Typically the
> cleanup is not reentrant so we end up hung on a lock.
> 
> libgomp/ChangeLog:
> 
> 	* error.c (gomp_vfatal): Use abort instead of exit.
> 
> diff --git a/libgomp/error.c b/libgomp/error.c
> index 50ed85eedb1..25548c14a82 100644
> --- a/libgomp/error.c
> +++ b/libgomp/error.c
> @@ -77,7 +77,7 @@ void
>  gomp_vfatal (const char *fmt, va_list list)
>  {
>    gomp_verror (fmt, list);
> -  exit (EXIT_FAILURE);
> +  abort ();
>  }
>  
>  void

I don't like this, abort has quite different user visible behavior
from exit, e.g. the former often dumps core.

I believe in most places libgomp handles this by releasing locks before
calling gomp_{,v}fatal:
      gomp_mutex_unlock (&register_lock);
      gomp_fatal ("Out of memory allocating %lu bytes", (unsigned long) size);

      gomp_mutex_unlock (&devicep->lock);
      gomp_fatal ("Copying of %s object [%p..%p) to %s object [%p..%p) failed",
                  src, srcaddr, srcaddr + size, dst, dstaddr, dstaddr + size);

etc.
I could live with a gomp_fatal/gomp_vfatal alternative that would
use _exit/_Exit (but not sure if it is supported on all targets where
libgomp is) for uses where releasing locks is for whatever reason
not an option.

	Jakub