Bug #102403 for IO-Async-SSL: acceptor wrapper fails to account for EAGAIN

Fri Feb 27 12:01:51 2015 DAKKAR [...] cpan.org - Ticket created

CC:	TEAM [...] cpan.org
Subject:	IO::Async::Listener shouldn't consider EAGAIN as a success

I wrote this: http://paste.scsys.co.uk/468224 When I connect to it, one child process responds, the others call on_ssl_error with "not a socket". Tom diagnosed that the other children get EAGAIN on their 'accept' call, and proposed this patch: http://paste.scsys.co.uk/468228 It looks like only the second hunk is needed: when getting EAGAIN or EWOULDBLOCK, consider it a failure (we didn't get a socket, after all).

Fri Mar 13 13:11:47 2015 leonerd-cpan [...] leonerd.org.uk - Correspondence added

That paste has expired :( If anyone still has it to hand could they attach it to the bug? -- Paul Evans

Fri Mar 13 13:11:48 2015 The RT System itself - Status changed from 'new' to 'open'

Tue Mar 17 06:15:37 2015 DAKKAR [...] cpan.org - Correspondence added

On Fri Mar 13 13:11:47 2015, PEVANS wrote: Show quoted text

> That paste has expired :( > > If anyone still has it to hand could they attach it to the bug?

Yeah, sorry about that. The attached patch should be identical to the expired one. I applied it to my installation just as Tom typed it.

Subject:

listener-team.patch

--- lib/IO/Async/Listener.pm 2015-02-15 14:43:51.000000000 +0000 +++ /home/dakkar/.perlbrew/libs/perl-5.20.1@nap/lib/perl5/IO/Async/Listener.pm 2015-02-27 16:58:03.191958253 +0000 @@ -295,7 +295,8 @@ $on_done->( $self, $result ); })->on_fail( sub { my ( $message, undef, $socket, $dollarbang ) = @_; - $self->maybe_invoke_event( on_accept_error => $socket, $dollarbang ); + $self->maybe_invoke_event( on_accept_error => $socket, $dollarbang ) + #unless ($dollarbang||0) == EAGAIN or ($dollarbang||0) == EWOULDBLOCK; }); # Caller is not going to keep hold of the Future, so we have to ensure it @@ -321,7 +322,7 @@ } } elsif( $! == EAGAIN or $! == EWOULDBLOCK ) { - return Future->new->done; + return Future->new->fail( "Missed it", accept => $listen_sock, $! ); } else { return Future->new->fail( "Cannot accept() - $!", accept => $listen_sock, $! );

Sat Apr 04 08:47:15 2015 leonerd-cpan [...] leonerd.org.uk - Correspondence added

On Fri Feb 27 12:01:51 2015, DAKKAR wrote: Show quoted text

> I wrote this: http://paste.scsys.co.uk/468224 > > When I connect to it, one child process responds, the others call > on_ssl_error with "not a socket". Tom diagnosed that the other > children get EAGAIN on their 'accept' call, and proposed this patch: > http://paste.scsys.co.uk/468228 > > It looks like only the second hunk is needed: when getting EAGAIN or > EWOULDBLOCK, consider it a failure (we didn't get a socket, after > all).

Ah. Looking again at this I think the problem is in fact in IO::Async::SSL, whose accept-wrapper code fails to take account of the fact that done with an empty result means EAGAIN; it's hard otherwise to account for the non-success but non-failure of EAGAIN-like errors. -- Paul Evans

Sat Apr 04 08:48:17 2015 leonerd-cpan [...] leonerd.org.uk - Queue changed from IO-Async to IO-Async-SSL

Sat Apr 04 08:48:53 2015 leonerd-cpan [...] leonerd.org.uk - Subject changed from 'IO::Async::Listener shouldn't consider EAGAIN as a success' to 'acceptor wrapper fails to account for EAGAIN'

Sat Apr 04 08:48:54 2015 leonerd-cpan [...] leonerd.org.uk - Broken in 0.14 added

Sat Apr 04 09:36:26 2015 leonerd-cpan [...] leonerd.org.uk - Correspondence added

Patched -- Paul Evans

Subject:

rt102403.patch

=== modified file 'lib/IO/Async/SSL.pm' --- lib/IO/Async/SSL.pm 2015-04-04 13:31:09 +0000 +++ lib/IO/Async/SSL.pm 2015-04-04 13:36:04 +0000 @@ -440,6 +440,8 @@ $listener->$cleartext_acceptor( $listen_sock )->then( sub { my ( $socket ) = @_; + return Future->done() unless $socket; # EAGAIN + $stream->configure( handle => $socket ) if $stream; $loop->SSL_upgrade(

Sat Apr 04 09:36:33 2015 leonerd-cpan [...] leonerd.org.uk - Status changed from 'open' to 'patched'

Fri Apr 17 13:17:15 2015 leonerd-cpan [...] leonerd.org.uk - Status changed from 'patched' to 'resolved'

Fri Apr 17 13:17:21 2015 leonerd-cpan [...] leonerd.org.uk - Fixed in 0.15 added