bug#14640: SA_RESTART prevents execution of signal handlers

[Andy Wingo]

On Tue 21 Jun 2016 09:48, address@hidden (Ludovic Courtès) writes:

> Andy Wingo <address@hidden> skribis:
>
>> On Mon 17 Jun 2013 15:54, address@hidden (Ludovic Courtès) writes:
>>
>>> When using SA_RESTART, signal handlers are never executed, as in this
>>> example (checked on 2.0.9+):
>>>
>>> (sigaction SIGALRM
>>>   (lambda (signum)
>>>     (pk 'sig signum))
>>>   SA_RESTART)
>>> (alarm 3)
>>> (pk 'char (read-char))
>>>
>>> Presumably this is because the read(2) syscall is automatically
>>> restarted, leaving no chance for the handler async to run.
>>
>> Thinking about this a bit -- since we always handle signals
>> asynchronously and have no intention of handling them synchronously,
>> then we just have to document this behavior.  Done in e877e1b:
>
> I think it’s problematic though.  With the current design, signal
> delivery is unreliable (with or without SA_RESTART; what we observe with
> SA_RESTART occurs similarly if you make a syscall right after queuing,
> but not running, an async.)

Can you expect any kind of reasonable behavior with SA_RESTART?  I think
not.

> The more I think about it, the more I think a different approach is
> needed.  On GNU/Linux, signalfd(2) may be part of the solution.

We already do the equivalent of signalfd(), with our self-pipe trick.
And an fd doesn't help you if the syscall has no associated fd.  Signals
are just a problem.  I agree we can do better though :)

If you are just concerned about read and write, I think the right thing
is non-blocking fd's, and making the C read/write waiters also add the
signal FD to their poll set.  WDYT?

Andy