Eamon Daly wrote:
and even wrapping an eval/ALRM around it fails to the tune
of 50%.

About alarm()...


On Win32, alarm() is considered unsupported (man perlport).
On Unices, alarm is affected by the signal handling API.
The first standard one is unsafe (race condition, need to
set the signal handler in the handler code, etc).
The second one (POSIX) is safer and well-defined (sigaction).
Try replacing this

> local $SIG{ALRM} = sub { die "alarm timeout" };

with something like this

sigaction SIGALRM,
   new POSIX::SigAction( sub { die "alarm timeout" } )
or die "Error setting SIGALRM handler: $!\n";

The signal handling model in Perl itself has changed from
unsafe to safer (Perl 5.8). But the safer one can take
longuer (no signal handling between Perl's internal "op-code").
I think there's a env var to force Perl using old model,
starting with Perl 5.8.1 (cant remember its name).

See man perlipc(1) for more details and comments
about timeout, sigaction, model, gethostbyname, etc

  eval {
      local $SIG{ALRM} = sub { die "alarm timeout" };
      alarm($timeout);
      $res = $ua->get($np_key_url);
      alarm(0);
  };
  alarm(0);

What's the purpose of last alarm(0) ? Just becoming paranoid ? ;-) A more classical line replacing that call would be like if ($@ and $@ !~ /alarm timeout/) { die "oops time out"}

A possible problem is that there is only one ALRM signal
handler at a time, by process. So using several handlers
can lead to strange behaviour. Ex if a "eval/alarm" is used
around LWP calls and LWP also uses a "eval/alarm".
I guess LWP catches ALRM to set a 500 return code.

> Does this sound like an issue with LWP, IO:Socket, perl,

It's an issue with DNS timeout in IO::Socket::INET, apparently.
I saw this when trying to reach down servers or inexistant
ones. Perhaps the system dns lookup uses its own signal handler,
-- I remember 'strace' showed me a EINTR in the last syscalls.
Gisle's idea to use 'strace -tt' can be more informative to you.

LWP::UserAgent uses IO::Socket::INET, and the timeout handling
code is different following what versions you use.
With LWP 5.69, Perl 5.8.0 (Linux 2.4.18) the whole code is commented
out, and the workaround seems to use your own signal handling around
LWP calls (sigaction prefered). Seems to work for me.
With LWP 5.43, Perl 5.005_03 (Linux 2.2.16) the code is there, and
(dns) timeout cant be avoided easyly

All this seems quite complex, but hope this helps. It's not always
easy to put together relevant informations in a 'big picture'.

Christophe



Reply via email to