Front page | perl.perl5.porters |
Postings from January 2001
Raphael Manfredi wrote:
> The problem is to know the semantics we wish to guarantee when one
> does this:
>
> $SIG{ALRM} = 'IGNORE';
>
> eval { &foo };
> print "back from foo()\n";
> alarm(2);
> select(undef, undef, undef, 4);
>
> sub TIMEOUT {
> die "TIMEOUT";
> }
>
> sub foo {
> $SIG{ALRM} = 'TIMEOUT';
> alarm(2);
> select(undef, undef, undef, 4);
> }
>
> For me (Linux 2.4.0, perl 5.005_03), it yields:
>
> back from foo()
> TIMEOUT at /home/ram/tmp/alarm line 9
>
> Is this what is expected?
>
> I wonder where sigsetjmp() comes into play here, because what is
> happening shows that the signal mask is not restored, and that ALRM
> continues to be trapped even though we came back through a siglongjmp()?
Goodness knows what is happening here - tracing the program shows
Perl_sighandler being called when the first alarm call expires, but it
doesn't then seem to call the script-level signal handler. I suspect
that because the alarm is caught inside an eval, perl is mistakenly
siglongjmp-ing out of the eval *before* calling the interrupt handler.
The subsequent alarm/select pair then works as expected because it
hasn't been called inside an eval block, and the siglongjmp doesn't
therefore happen. Here is the evidence:
(/tmp/a:14) alarm
=> IV(0)
(/tmp/a:14) nextstate
=>
(/tmp/a:15) undef
=> SV_UNDEF
(/tmp/a:15) undef
=> SV_UNDEF SV_UNDEF
(/tmp/a:15) undef
=> SV_UNDEF SV_UNDEF SV_UNDEF
(/tmp/a:15) const(IV(4))
=> SV_UNDEF SV_UNDEF SV_UNDEF IV(4)
(/tmp/a:15) sselect
(/tmp/a:15) ENTER scope 6 at perl.c:1608
(/tmp/a:15) ENTER scope 7 at pp_hot.c:2303
Entering block 0, type SUB
=> PV("ALRM"\0)
(/tmp/a:15) nextstate
=>
(/tmp/a:9) pushmark
=> *
(/tmp/a:9) const(PV("TIMEOUT"\0))
=> * PV("TIMEOUT"\0)
(/tmp/a:9) die
Unwinding block 0, type SUB
(/tmp/a:9) (Found eval #1)
Unwinding block 2, type SUB
Leaving block 1, type EVAL
(/tmp/a:3) LEAVE scope 3 at pp_ctl.c:1385
=>
(/tmp/a:3) nextstate
=>
(/tmp/a:4) pushmark
=> *
(/tmp/a:4) const(PV("back from foo()\n"\0))
=> * PV("back from foo()\n"\0)
(/tmp/a:4) print
=> SV_YES
There is so much brokenness and platform-dependent behaviour in the area
of perl signal handling that I don't think a switch from sigsetjmp to
setjmp is going to make any discernable difference.
Alan Burlison
Thread Previous