Re: 31.0.50; Emacs scratch/igc 6682d0e6c96 Crash on Linux, wayland

[Helmut Eller <eller.helmut@gmail.com>]

On Wed, Sep 04 2024, Pip Cet wrote:
>> #0  0x00000000006dfe09 in record_backtrace.lto_priv ()
>> #1  0x00000000006e03df in deliver_profiler_signal ()
>> #2  <signal handler called>
>> #3  0x00000000007dea30 in sigHandle ()
>> #4  <signal handler called>
>
> That looks like the SIGPROF hit at a very bad time, after sigHandle was
> called but before it had time to lock the arena, which would have
> stopped the profiler handler from accessing memory.

If sigHandle hasn't acquired the lock, then accessing memory by the
profiler should trigger another SIGSEGV and a nested call to sigHandle
should then be able to acquire the lock.

Looking at the code of ArenaAccess, it seems that there are two locks
involved: a global lock and an arena lock.  And igc_busy_p only checks
the arena lock.  Hmm.

> Helmut, do you have any ideas?

I think it would work, if we could check the global lock instead for the
arena lock.  Unfortunately, MPS offers no public function for that.

> Maybe we'll have to add SIGPROF to the
> SIGSEGV sa_mask after all and explicitly unblock it in the handler after
> setting a "just count this" flag?

That would probably also work.