Re: USE_SIGCHLD

In other words, it eliminates that up-to-1-second wait for child  
processes that always annoyed you.  Without resorting to a faster,  
less efficient polling interval. :)

Yuval: Reviewing the code, it looks like the SIGCHLD handler may be  
triggered multiple times before the EN_SCPOLL event could be  
dispatched.  That would enqueue redundant EN_SCPOLL events, and it  
could be marginally unoptimal.

Would it be feasible to disable the SIG$_[0] handler from the handler  
and then re-enable it just before the children are reaped?  Something  
like this:

# In POE::Loop::PerlSignals:
sub _loop_signal_handler_chld {
   if (TRACE_SIGNALS) {
     POE::Kernel::_warn "<sg> Enqueuing CHLD-like SIG$_[0] event";
   }

   $poe_kernel->_data_sig_enqueue_poll_event();
   $SIG{$_[0]} = sub { }; # temporarily ignore
}

# In POE::Resource::Signals:
sub _data_sig_handle_poll_event {
   my $self = shift;

   if ( USE_SIGCHLD ) {
     # Turn on the handler just before polling.  Leaves a window
     # open for redundancy, but avoids a gap where races could
     # lose children.
     $SIG{CHLD} = $SIG{CLD} = \&_loop_signal_handler_chld;
     $polling_for_signals = undef;
   }

   ... gather child process state ...;
}

-- 
Rocco Caputo - [email protected]


On Jun 6, 2008, at 06:33, Yuval Kogman wrote:

> Hello POE hackers
>
> POE 1.0001 had a patch that conditionally enabled $SIG{CHLD} support
> for immediate reaping of child processes (as opposed to POE's once
> per second polling).
>
> Instead of polling at an interval, the same polling code is
> immediatley queued after SIGCHLD arrives, so that on the next loop
> cycle wait() is called, and SIGCHLD is delivered. This means that
> system() etc are still supposed to continue working inside POE code,
> because the SIGCHLD handler never calls wait() itself, and thus
> doesn't interfere with Perl's wait() call.
>
> Anyway, this code tries to be compatible, but is probably subtly
> different, so please run
>
> 	env POE_USE_SIGCHLD=1 make test
>
> on your code and tell me if it breaks before I forget what I did.
>
> kthx, xoxo
> nothingmuch
>
> -- 
>  Yuval Kogman <[email protected]>
> http://nothingmuch.woobling.org  0xEBD27418
>