bug#21424: Bug#685919: guile-1.6: SIGALRM signal handler does not get called when guile blocks on I/O

[Andy Wingo <wingo@pobox.com>]

It does seem clear that we have a bug here; evidently signals do not
wake up the reader, neither with readline nor without.

Andy

On Sun 06 Sep 2015 19:18, Rob Browning <rlb@defaultvalue.org> writes:

> [If possible, please preserve the -forwarded address in any replies.]
>
> Reference: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=765497
>
> Panu Kalliokoski <panu.kalliokoski@gmail.com> writes:
>
>> While playing with guile on my system, I discovered a weird anomaly
>> which I could not reproduce on other systems running guile.  If I
>> install a signal handler for SIGALRM, it won't get called while guile is
>> making an I/O system call.  To demonstrate:
>>
>> [atehwa@karaihin ~/proj/psyk]$ guile
>> guile> (alarm 2)
>> 0
>> guile> Herätyskello
>> [atehwa@karaihin ~/proj/psyk]$ guile
>> guile> (sigaction SIGALRM (lambda (x) (display "now!") (newline)))
>> (0 . 335544320)
>> guile> (alarm 2)
>> 0
>> guile> now a lot more than two seconds has passed, while I wrote this
>> now!
>> <unnamed port>: In expression now:
>> <unnamed port>: Unbound variable: now
>> ABORT: (unbound-variable)
>> [...]
>>
>> As you can see, the signal handler gets called as soon as guile returns
>> from read(2), already before calling (eval).
>>
>> I can't get to understand what causes this on my system, because another
>> Debian system with exact same versions of guile-1.6, libc6 and
>> libguile-ltdl-1 seems to work fine, and interrupts the read(2) call with
>> the signal handler.
>
> This appears to still be the case with at least Debian's 2.0.11+1-10
> package, and setting the handler to something that doesn't perform IO
> has the same effect (i.e. no alarm until you hit return):
>
>   (sigaction SIGALRM (lambda (x) (exit 1)))
>
> Thanks