bug#28211: Stack marking issue in multi-threaded code, 2020 edition

["Ludovic Courtès" <ludo@gnu.org>]

Hi!

Andy Wingo <wingo@igalia.com> skribis:

> On Thu 12 Mar 2020 22:59, Ludovic Courtès <ludo@gnu.org> writes:
>
>> I think I’ve found another race condition involving stack marking, as a
>> followup to <https://issues.guix.gnu.org/issue/28211> (this time on
>> 3.0.1+, but the code is almost the same.)
>>
>> ‘abort_to_prompt’ does this:
>>
>>   fp = vp->stack_top - fp_offset;
>>   sp = vp->stack_top - sp_offset;
>>
>>   /* Continuation gets nargs+1 values: the one more is for the cont.  */
>>   sp = sp - nargs - 1;
>>
>>   /* Shuffle abort arguments down to the prompt continuation.  We have
>>      to be jumping to an older part of the stack.  */
>>   if (sp < vp->sp)
>>     abort ();
>>   sp[nargs].as_scm = cont;
>>   while (nargs--)
>>     sp[nargs] = vp->sp[nargs];
>>
>>   /* Restore VM regs */
>>   vp->fp = fp;
>>   vp->sp = sp;
>>   vp->ip = vra;
>>
>>
>> What if ‘scm_i_vm_mark_stack’ walks the stack right before the ‘vp->fp’
>> assignment?  It can determine that one of the just-assigned ‘sp[nargs]’
>> is a dead slot, and thus set it to SCM_UNSPECIFIED.
>
> I think you're right here.
>
> Given that the most-recently-pushed frame is marked conservatively, I
> think it would be sufficient to reset vp->fp before shuffling stack
> args; that would make it so that the frame includes the values to
> shuffle, their target locations, and probably some other crap in
> between.  Given that marking the crap is harmless, I think that would be
> enough.  WDYT?

Sounds good.  Following our discussion on IRC, I pushed what you
proposed as 89edd1bc2dcff50fb05c3598a846d6b51b172f7c.  \o/

I confirmed with and without rr that it no longer triggers the dreaded
crash.

BTW, pro tip: to run ./meta/guile under rr, I do:

  sed -i libguile/guile \
      -e 's/exec /exec rr record -n --syscall-buffer-sig=SIGUSR1 /g'

where ‘-n’ disables stack switching.

> In a more perfect world, initiating GC should tell threads to reach a
> safepoint and mark their own stacks -- preserves thread locality and
> prevents this class of bug.  But given that libgc uses signals to stop
> threads, we have to be less precise.

Yup, agreed.

Thanks,
Ludo’.