Re: [PATCH 2/2] stdlib: Make abort AS-safe (BZ 26275)

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 01/08/23 05:26, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> +void
>> +__abort_lock_lock (void)
>> +{
>> +  __libc_lock_lock (lock);
>> +}
> 
> I think __abort_lock_lock and __abort_lock_unlock should take a signal
> mask argument, so that they are more like a async-signal-safe mutex.

That was my first approach, but I changed it due __spawnix usage.  The
lock is really only required on clone call, so we can release it after
clone returns.  However we need to only umask signals after the execve
call in helper process.

We can move the abort unlock later, it should be ok as well (it would
add a slight more latency on highly multithead programs that spawns
a lot of thread and try to abort(), but it should be ok). 

> 
>> +/* Cause an abnormal program termination with core-dump.  */
>> +_Noreturn void
>> +abort (void)
>> +{
>> +  raise (SIGABRT);
>> +
>> +  /* There is a SIGABRT handler installed and it returned, or SIGABRT was
>> +     blocked or ignored.  In this case use a AS-safe lock to prevent sigaction
>> +     to change the signal disposition, reinstall the handle to abort the
>> +     process, and raise the signal again.  */
>> +  internal_signal_block_all (NULL);
>> +  __libc_lock_lock (lock);
> 
> This could call __abort_lock_lock.

Ack.