1 # This library is free software; you can redistribute it and/or modify
2 # it under the same terms as Perl itself.
4 # This license differs from the rest of public-inbox
6 # This is a fork of the unmaintained Danga::Socket (1.61) with
7 # significant changes. See Documentation/technical/ds.txt in our
10 # Do not expect this to be a stable API like Danga::Socket,
11 # but it will evolve to suite our needs and to take advantage of
12 # newer Linux and *BSD features.
13 # Bugs encountered were reported to bug-Danga-Socket@rt.cpan.org,
14 # fixed in Danga::Socket 1.62 and visible at:
15 # https://rt.cpan.org/Public/Dist/Display.html?Name=Danga-Socket
18 # sock: underlying socket
19 # rbuf: scalarref, usually undef
20 # wbuf: arrayref of coderefs or tmpio (autovivified))
21 # (tmpio = [ GLOB, offset, [ length ] ])
22 package PublicInbox::DS;
25 use parent qw(Exporter);
26 use bytes qw(length substr); # FIXME(?): needed for PublicInbox::NNTP
27 use POSIX qw(WNOHANG sigprocmask SIG_SETMASK);
28 use Fcntl qw(SEEK_SET :DEFAULT O_APPEND);
29 use Time::HiRes qw(clock_gettime CLOCK_MONOTONIC);
30 use Scalar::Util qw(blessed);
31 use PublicInbox::Syscall qw(:epoll);
32 use PublicInbox::Tmpfile;
33 use Errno qw(EAGAIN EINVAL ECHILD EINTR);
34 use Carp qw(carp croak);
35 our @EXPORT_OK = qw(now msg_more awaitpid add_timer add_uniq_timer);
38 my $nextq; # queue for next_tick
39 my $AWAIT_PIDS; # pid => [ $callback, @args ]
41 my $ToClose; # sockets to close when event loop is done
43 %DescriptorMap, # fd (num) -> PublicInbox::DS object
44 $Epoll, # Global epoll fd (or DSKQXS ref)
45 $ep_io, # IO::Handle for Epoll
47 $PostLoopCallback, # subref to call at the end of each loop, if defined (global)
49 $LoopTimeout, # timeout of event loop in milliseconds
57 #####################################################################
58 ### C L A S S M E T H O D S
59 #####################################################################
61 =head2 C<< CLASS->Reset() >>
68 $in_loop = undef; # first in case DESTROY callbacks use this
72 $PostLoopCallback = undef;
74 # we may be iterating inside one of these on our stack
75 my @q = delete @Stack{keys %Stack};
76 for my $q (@q) { @$q = () }
77 $AWAIT_PIDS = $nextq = $ToClose = undef;
78 $ep_io = undef; # closes real $Epoll FD
79 $Epoll = undef; # may call DSKQXS::DESTROY
80 } while (@Timers || keys(%Stack) || $nextq || $AWAIT_PIDS ||
81 $ToClose || keys(%DescriptorMap) ||
82 $PostLoopCallback || keys(%UniqTimer));
85 $LoopTimeout = -1; # no timeout by default
88 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
90 Set the loop timeout for the event loop to some value in milliseconds.
92 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
96 sub SetLoopTimeout { $LoopTimeout = $_[1] + 0 }
98 sub _add_named_timer {
99 my ($name, $secs, $coderef, @args) = @_;
100 my $fire_time = now() + $secs;
101 my $timer = [$fire_time, $name, $coderef, @args];
103 if (!@Timers || $fire_time >= $Timers[-1][0]) {
104 push @Timers, $timer;
108 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
109 # but it was compared against calendar queues, heaps, naive push/sort,
110 # and a bunch of other versions, and found to be fastest with a large
111 # variety of datasets.)
112 for (my $i = 0; $i < @Timers; $i++) {
113 if ($Timers[$i][0] > $fire_time) {
114 splice(@Timers, $i, 0, $timer);
118 die "Shouldn't get here.";
121 sub add_timer { _add_named_timer(undef, @_) }
123 sub add_uniq_timer { # ($name, $secs, $coderef, @args) = @_;
124 $UniqTimer{$_[0]} //= _add_named_timer(@_);
127 # caller sets return value to $Epoll
129 if (defined $PublicInbox::Syscall::SYS_epoll_create) {
130 my $fd = epoll_create();
131 die "epoll_create: $!" if $fd < 0;
132 open($ep_io, '+<&=', $fd) or return;
133 fcntl($ep_io, F_SETFD, FD_CLOEXEC);
137 for (qw(DSKQXS DSPoll)) {
138 $cls = "PublicInbox::$_";
139 last if eval "require $cls";
141 $cls->import(qw(epoll_ctl epoll_wait));
146 sub now () { clock_gettime(CLOCK_MONOTONIC) }
149 my $q = $nextq or return;
151 $Stack{cur_runq} = $q;
153 # avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak:
154 # https://rt.perl.org/Public/Bug/Display.html?id=114340
161 delete $Stack{cur_runq};
164 # runs timers and returns milliseconds for next one, or next event loop
168 return (($nextq || $ToClose) ? 0 : $LoopTimeout) unless @Timers;
173 while (@Timers && $Timers[0][0] <= $now) {
174 my $to_run = shift(@Timers);
175 delete $UniqTimer{$to_run->[1] // ''};
176 $to_run->[2]->(@$to_run[3..$#$to_run]);
179 # timers may enqueue into nextq:
180 return 0 if ($nextq || $ToClose);
182 return $LoopTimeout unless @Timers;
184 # convert time to an even number of milliseconds, adding 1
185 # extra, otherwise floating point fun can occur and we'll
186 # call RunTimers like 20-30 times, each returning a timeout
187 # of 0.0000212 seconds
188 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
190 # -1 is an infinite timeout, so prefer a real timeout
191 ($LoopTimeout < 0 || $LoopTimeout >= $timeout) ? $timeout : $LoopTimeout
194 sub sig_setmask { sigprocmask(SIG_SETMASK, @_) or die "sigprocmask: $!" }
196 sub block_signals () {
197 my $oldset = POSIX::SigSet->new;
198 my $newset = POSIX::SigSet->new;
199 $newset->fillset or die "fillset: $!";
200 sig_setmask($newset, $oldset);
205 my ($pid, @cb_args) = @_;
206 my $cb = shift @cb_args or return;
207 eval { $cb->($pid, @cb_args) };
208 warn "E: awaitpid($pid): $@" if $@;
211 # This relies on our Perl process is single-threaded, or at least
212 # no threads are spawning and waiting on processes (``, system(), etc...)
213 # Threads are officially discouraged by the Perl5 team, and I expect
214 # that to remain the case.
217 my $oldset = block_signals();
219 my $pid = waitpid(-1, WNOHANG) // last;
221 if (defined(my $cb_args = delete $AWAIT_PIDS->{$pid})) {
222 await_cb($pid, @$cb_args) if $cb_args;
224 warn "W: reaped unknown PID=$pid: \$?=$?\n";
227 sig_setmask($oldset);
230 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
231 sub enqueue_reap () { $reap_armed //= requeue(\&reap_pids) }
233 sub in_loop () { $in_loop }
235 # Internal function: run the post-event callback, send read events
236 # for pushed-back data, and close pending connections. returns 1
237 # if event loop should continue, or 0 to shut it all down.
238 sub PostEventLoop () {
239 # now we can close sockets that wanted to close during our event
240 # processing. (we didn't want to close them during the loop, as we
241 # didn't want fd numbers being reused and confused during the event
243 if (my $close_now = $ToClose) {
244 $ToClose = undef; # will be autovivified on push
245 @$close_now = map { fileno($_) } @$close_now;
247 # ->DESTROY methods may populate ToClose
248 delete @DescriptorMap{@$close_now};
251 # by default we keep running, unless a postloop callback cancels it
252 $PostLoopCallback ? $PostLoopCallback->(\%DescriptorMap) : 1;
255 # Start processing IO events. In most daemon programs this never exits. See
256 # C<PostLoopCallback> for how to exit the loop.
257 sub event_loop (;$$) {
258 my ($sig, $oldset) = @_;
259 $Epoll //= _InitPoller();
260 require PublicInbox::Sigfd if $sig;
261 my $sigfd = PublicInbox::Sigfd->new($sig, 1) if $sig;
262 local @SIG{keys %$sig} = values(%$sig) if $sig && !$sigfd;
263 local $SIG{PIPE} = 'IGNORE';
264 if (!$sigfd && $sig) {
265 # wake up every second to accept signals if we don't
266 # have signalfd or IO::KQueue:
267 sig_setmask($oldset);
268 PublicInbox::DS->SetLoopTimeout(1000);
270 $_[0] = $sigfd = $sig = undef; # $_[0] == sig
274 my $timeout = RunTimers();
276 # get up to 1000 events
277 epoll_wait($Epoll, 1000, $timeout, \@events);
278 for my $fd (@events) {
279 # it's possible epoll_wait returned many events,
280 # including some at the end that ones in the front
281 # triggered unregister-interest actions. if we can't
282 # find the %sock entry, it's because we're no longer
283 # interested in that event.
285 # guard stack-not-refcounted w/ Carp + @DB::args
286 my $obj = $DescriptorMap{$fd};
289 } while (PostEventLoop());
292 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
294 Sets post loop callback function. Pass a subref and it will be
295 called every time the event loop finishes.
297 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
300 The callback function will be passed two parameters: \%DescriptorMap
303 sub SetPostLoopCallback {
304 my ($class, $ref) = @_;
307 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
310 #####################################################################
311 ### PublicInbox::DS-the-object code
312 #####################################################################
314 =head2 OBJECT METHODS
316 =head2 C<< CLASS->new( $socket ) >>
318 Create a new PublicInbox::DS subclass object for the given I<socket> which will
319 react to events on it during the C<event_loop>.
321 This is normally (always?) called from your subclass via:
323 $class->SUPER::new($socket);
327 my ($self, $sock, $ev) = @_;
328 $self->{sock} = $sock;
329 my $fd = fileno($sock);
331 $Epoll //= _InitPoller();
333 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
334 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
335 $ev &= ~EPOLLEXCLUSIVE;
338 die "EPOLL_CTL_ADD $self/$sock/$fd: $!";
340 croak("FD:$fd in use by $DescriptorMap{$fd} (for $self/$sock)")
341 if defined($DescriptorMap{$fd});
343 $DescriptorMap{$fd} = $self;
346 # for IMAP, NNTP, and POP3 which greet clients upon connect
348 my ($self, $sock) = @_;
351 if ($sock->can('accept_SSL') && !$sock->accept_SSL) {
352 return CORE::close($sock) if $! != EAGAIN;
353 $ev = PublicInbox::TLS::epollbit() or return CORE::close($sock);
354 $wbuf = [ \&accept_tls_step, $self->can('do_greet')];
356 new($self, $sock, $ev | EPOLLONESHOT);
358 $self->{wbuf} = $wbuf;
365 #####################################################################
366 ### I N S T A N C E M E T H O D S
367 #####################################################################
369 sub requeue ($) { push @$nextq, $_[0] } # autovivifies
371 =head2 C<< $obj->close >>
378 my $sock = delete $self->{sock} or return;
380 # we need to flush our write buffer, as there may
381 # be self-referential closures (sub { $client->close })
382 # preventing the object from being destroyed
383 delete $self->{wbuf};
385 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
386 # notifications about it
387 my $fd = fileno($sock);
388 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
389 croak("EPOLL_CTL_DEL($self/$sock): $!");
391 # we explicitly don't delete from DescriptorMap here until we
392 # actually close the socket, as we might be in the middle of
393 # processing an epoll_wait/etc that returned hundreds of fds, one
394 # of which is not yet processed and is what we're closing. if we
395 # keep it in DescriptorMap, then the event harnesses can just
396 # looked at $pob->{sock} == undef and ignore it. but if it's an
397 # un-accounted for fd, then it (understandably) freak out a bit
398 # and emit warnings, thinking their state got off.
400 # defer closing the actual socket until the event loop is done
401 # processing this round of events. (otherwise we might reuse fds)
402 push @$ToClose, $sock; # autovivifies $ToClose
407 # portable, non-thread-safe sendfile emulation (no pread, yet)
408 sub send_tmpio ($$) {
409 my ($sock, $tmpio) = @_;
411 sysseek($tmpio->[0], $tmpio->[1], SEEK_SET) or return;
412 my $n = $tmpio->[2] // 65536;
413 $n = 65536 if $n > 65536;
414 defined(my $to_write = sysread($tmpio->[0], my $buf, $n)) or return;
416 while ($to_write > 0) {
417 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
421 return if $written == 0;
425 $tmpio->[1] += $written; # offset
426 $tmpio->[2] -= $written if defined($tmpio->[2]); # length
430 sub epbit ($$) { # (sock, default)
431 $_[0]->can('stop_SSL') ? PublicInbox::TLS::epollbit() : $_[1];
434 # returns 1 if done, 0 if incomplete
435 sub flush_write ($) {
437 my $sock = $self->{sock} or return;
438 my $wbuf = $self->{wbuf} or return 1;
441 while (my $bref = $wbuf->[0]) {
442 if (ref($bref) ne 'CODE') {
444 my $w = send_tmpio($sock, $bref); # bref is tmpio
450 } elsif ($! == EAGAIN) {
451 my $ev = epbit($sock, EPOLLOUT) or return $self->close;
452 epwait($sock, $ev | EPOLLONESHOT);
458 } else { #(ref($bref) eq 'CODE') {
460 my $before = scalar(@$wbuf);
463 # bref may be enqueueing more CODE to call (see accept_tls_step)
464 return 0 if (scalar(@$wbuf) > $before);
468 delete $self->{wbuf};
473 my ($self, $rbuf) = @_;
474 if ($$rbuf eq '') { # who knows how long till we can read again
475 delete $self->{rbuf};
477 $self->{rbuf} = $rbuf;
481 sub do_read ($$$;$) {
482 my ($self, $rbuf, $len, $off) = @_;
483 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
484 return ($r == 0 ? $self->close : $r) if defined $r;
485 # common for clients to break connections without warning,
486 # would be too noisy to log here:
488 my $ev = epbit($sock, EPOLLIN) or return $self->close;
489 epwait($sock, $ev | EPOLLONESHOT);
490 rbuf_idle($self, $rbuf);
497 # drop the socket if we hit unrecoverable errors on our system which
498 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
506 my ($self, $bref, $off) = @_;
507 my $fh = tmpfile('wbuf', $self->{sock}, O_APPEND) or
508 return drop($self, "tmpfile $!");
510 my $len = length($$bref) - $off;
511 my $n = syswrite($fh, $$bref, $len, $off) //
512 return drop($self, "write ($len): $!");
513 $n == $len or return drop($self, "wrote $n < $len bytes");
514 [ $fh, 0 ] # [1] = offset, [2] = length, not set by us
517 =head2 C<< $obj->write( $data ) >>
519 Write the specified data to the underlying handle. I<data> may be scalar,
520 scalar ref, code ref (to run when there).
521 Returns 1 if writes all went through, or 0 if there are writes in queue. If
522 it returns 1, caller should stop waiting for 'writable' events)
526 my ($self, $data) = @_;
528 # nobody should be writing to closed sockets, but caller code can
529 # do two writes within an event, have the first fail and
530 # disconnect the other side (whose destructor then closes the
531 # calling object, but it's still in a method), and then the
532 # now-dead object does its second write. that is this case. we
533 # just lie and say it worked. it'll be dead soon and won't be
535 my $sock = $self->{sock} or return 1;
537 my $bref = $ref ? $data : \$data;
538 my $wbuf = $self->{wbuf};
539 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
540 if ($ref eq 'CODE') {
543 my $tmpio = $wbuf->[-1];
544 if ($tmpio && !defined($tmpio->[2])) { # append to tmp file buffer
545 $tmpio->[0]->print($$bref) or return drop($self, "print: $!");
547 my $tmpio = tmpio($self, $bref, 0) or return 0;
552 } elsif ($ref eq 'CODE') {
556 my $to_write = length($$bref);
557 my $written = syswrite($sock, $$bref, $to_write);
559 if (defined $written) {
560 return 1 if $written == $to_write;
561 requeue($self); # runs: event_step -> flush_write
562 } elsif ($! == EAGAIN) {
563 my $ev = epbit($sock, EPOLLOUT) or return $self->close;
564 epwait($sock, $ev | EPOLLONESHOT);
570 # deal with EAGAIN or partial write:
571 my $tmpio = tmpio($self, $bref, $written) or return 0;
573 # wbuf may be an empty array if we're being called inside
574 # ->flush_write via CODE bref:
575 push @{$self->{wbuf}}, $tmpio; # autovivifies
580 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
584 my $sock = $self->{sock} or return 1;
585 my $wbuf = $self->{wbuf};
587 if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) &&
588 !$sock->can('stop_SSL')) {
589 my $n = send($sock, $_[1], MSG_MORE);
591 my $nlen = length($_[1]) - $n;
592 return 1 if $nlen == 0; # all done!
593 # queue up the unwritten substring:
594 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
595 push @{$self->{wbuf}}, $tmpio; # autovivifies
596 epwait($sock, EPOLLOUT|EPOLLONESHOT);
601 # don't redispatch into NNTPdeflate::write
602 PublicInbox::DS::write($self, \($_[1]));
606 my ($sock, $ev) = @_;
607 epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and
608 croak("EPOLL_CTL_MOD($sock): $!");
611 # return true if complete, false if incomplete (or failure)
612 sub accept_tls_step ($) {
614 my $sock = $self->{sock} or return;
615 return 1 if $sock->accept_SSL;
616 return $self->close if $! != EAGAIN;
617 my $ev = PublicInbox::TLS::epollbit() or return $self->close;
618 epwait($sock, $ev | EPOLLONESHOT);
619 unshift(@{$self->{wbuf}}, \&accept_tls_step); # autovivifies
623 # return true if complete, false if incomplete (or failure)
624 sub shutdn_tls_step ($) {
626 my $sock = $self->{sock} or return;
627 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
628 return $self->close if $! != EAGAIN;
629 my $ev = PublicInbox::TLS::epollbit() or return $self->close;
630 epwait($sock, $ev | EPOLLONESHOT);
631 unshift(@{$self->{wbuf}}, \&shutdn_tls_step); # autovivifies
635 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
636 # or fork w/o exec, so no inadvertent socket sharing
639 my $sock = $self->{sock} or return;
640 if ($sock->can('stop_SSL')) {
641 shutdn_tls_step($self);
647 sub dflush {} # overridden by DSdeflate
648 sub compressed {} # overridden by DSdeflate
649 sub long_response_done {} # overridden by Net::NNTP
653 # wbuf is unset or empty, here; {long} may add to it
654 my ($fd, $cb, $t0, @args) = @{$self->{long_cb}};
655 my $more = eval { $cb->($self, @args) };
656 if ($@ || !$self->{sock}) { # something bad happened...
657 delete $self->{long_cb};
658 my $elapsed = now() - $t0;
659 $@ and warn("$@ during long response[$fd] - ",
660 sprintf('%0.6f', $elapsed),"\n");
661 $self->out(" deferred[$fd] aborted - %0.6f", $elapsed);
663 } elsif ($more) { # $self->{wbuf}:
664 # control passed to ibx_async_cat if $more == \undef
665 requeue_once($self) if !ref($more);
667 delete $self->{long_cb};
668 $self->long_response_done;
669 my $elapsed = now() - $t0;
670 my $fd = fileno($self->{sock});
671 $self->out(" deferred[$fd] done - %0.6f", $elapsed);
672 my $wbuf = $self->{wbuf}; # do NOT autovivify
673 requeue($self) unless $wbuf && @$wbuf;
679 # COMPRESS users all share the same DEFLATE context.
680 # Flush it here to ensure clients don't see each other's data
683 # no recursion, schedule another call ASAP,
684 # but only after all pending writes are done.
685 # autovivify wbuf. wbuf may be populated by $cb,
686 # no need to rearm if so: (push returns new size of array)
687 $self->requeue if push(@{$self->{wbuf}}, \&long_step) == 1;
690 sub long_response ($$;@) {
691 my ($self, $cb, @args) = @_; # cb returns true if more, false if done
692 my $sock = $self->{sock} or return;
693 # make sure we disable reading during a long response,
694 # clients should not be sending us stuff and making us do more
695 # work while we are stream a response to them
696 $self->{long_cb} = [ fileno($sock), $cb, now(), @args ];
697 long_step($self); # kick off!
702 my ($pid, @cb_args) = @_; # @cb_args = ($cb, @args), $cb may be undef
703 $AWAIT_PIDS->{$pid} = \@cb_args if @cb_args;
704 # provide synchronous API
705 if (defined(wantarray) || (!$in_loop && !@cb_args)) {
708 $ret = waitpid($pid, 0) // -2;
710 my $cb_args = delete $AWAIT_PIDS->{$pid};
711 @cb_args = @$cb_args if !@cb_args && $cb_args;
712 await_cb($pid, @cb_args);
714 goto again if $! == EINTR;
715 carp "waitpid($pid): $!";
716 delete $AWAIT_PIDS->{$pid};
719 } elsif ($in_loop) { # We could've just missed our SIGCHLD, cover it, here:
726 =head1 AUTHORS (Danga::Socket)
728 Brad Fitzpatrick <brad@danga.com> - author
730 Michael Granger <ged@danga.com> - docs, testing
732 Mark Smith <junior@danga.com> - contributor, heavy user, testing
734 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits