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
16 package PublicInbox::DS;
19 use POSIX qw(WNOHANG);
21 use Fcntl qw(SEEK_SET :DEFAULT);
22 use Time::HiRes qw(clock_gettime CLOCK_MONOTONIC);
23 use parent qw(Exporter);
24 our @EXPORT_OK = qw(now msg_more);
27 use Scalar::Util qw(blessed);
29 use PublicInbox::Syscall qw(:epoll);
30 use PublicInbox::Tmpfile;
32 use fields ('sock', # underlying socket
33 'rbuf', # scalarref, usually undef
34 'wbuf', # arrayref of coderefs or GLOB refs
35 'wbuf_off', # offset into first element of wbuf to start writing at
38 use Errno qw(EAGAIN EINVAL);
39 use Carp qw(confess carp);
41 my $nextq; # queue for next_tick
42 my $WaitPids; # list of [ pid, callback, callback_arg ]
43 my $later_queue; # callbacks
44 my $EXPMAP; # fd -> [ idle_time, $self ]
45 our $EXPTIME = 180; # 3 minutes
46 my ($later_timer, $reap_timer, $exp_timer);
48 %DescriptorMap, # fd (num) -> PublicInbox::DS object
49 $Epoll, # Global epoll fd (or DSKQXS ref)
50 $_io, # IO::Handle for Epoll
51 @ToClose, # sockets to close when event loop is done
53 $PostLoopCallback, # subref to call at the end of each loop, if defined (global)
55 $LoopTimeout, # timeout of event loop in milliseconds
56 $DoneInit, # if we've done the one-time module init yet
63 #####################################################################
64 ### C L A S S M E T H O D S
65 #####################################################################
67 =head2 C<< CLASS->Reset() >>
78 $reap_timer = $later_timer = $exp_timer = undef;
80 $LoopTimeout = -1; # no timeout by default
83 $PostLoopCallback = undef;
86 $_io = undef; # closes real $Epoll FD
87 $Epoll = undef; # may call DSKQXS::DESTROY
89 *EventLoop = *FirstTimeEventLoop;
92 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
94 Set the loop timeout for the event loop to some value in milliseconds.
96 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
101 return $LoopTimeout = $_[1] + 0;
104 =head2 C<< PublicInbox::DS::add_timer( $seconds, $coderef ) >>
106 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
107 are not guaranteed to fire at the exact time you ask for.
111 my ($secs, $coderef) = @_;
113 my $fire_time = now() + $secs;
115 my $timer = [$fire_time, $coderef];
117 if (!@Timers || $fire_time >= $Timers[-1][0]) {
118 push @Timers, $timer;
122 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
123 # but it was compared against calendar queues, heaps, naive push/sort,
124 # and a bunch of other versions, and found to be fastest with a large
125 # variety of datasets.)
126 for (my $i = 0; $i < @Timers; $i++) {
127 if ($Timers[$i][0] > $fire_time) {
128 splice(@Timers, $i, 0, $timer);
133 die "Shouldn't get here.";
136 # keeping this around in case we support other FD types for now,
137 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
138 sub set_cloexec ($) {
141 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
142 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
143 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
151 if (PublicInbox::Syscall::epoll_defined()) {
152 $Epoll = epoll_create();
153 set_cloexec($Epoll) if (defined($Epoll) && $Epoll >= 0);
156 for (qw(DSKQXS DSPoll)) {
157 $cls = "PublicInbox::$_";
158 last if eval "require $cls";
160 $cls->import(qw(epoll_ctl epoll_wait));
163 *EventLoop = *EpollEventLoop;
166 =head2 C<< CLASS->EventLoop() >>
168 Start processing IO events. In most daemon programs this never exits. See
169 C<PostLoopCallback> below for how to exit the loop.
172 sub FirstTimeEventLoop {
180 sub now () { clock_gettime(CLOCK_MONOTONIC) }
186 # we avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak:
187 # https://rt.perl.org/Public/Bug/Display.html?id=114340
196 # runs timers and returns milliseconds for next one, or next event loop
200 return ((@$nextq || @ToClose) ? 0 : $LoopTimeout) unless @Timers;
205 while (@Timers && $Timers[0][0] <= $now) {
206 my $to_run = shift(@Timers);
207 $to_run->[1]->($now) if $to_run->[1];
210 # timers may enqueue into nextq:
211 return 0 if (@$nextq || @ToClose);
213 return $LoopTimeout unless @Timers;
215 # convert time to an even number of milliseconds, adding 1
216 # extra, otherwise floating point fun can occur and we'll
217 # call RunTimers like 20-30 times, each returning a timeout
218 # of 0.0000212 seconds
219 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
221 # -1 is an infinite timeout, so prefer a real timeout
222 return $timeout if $LoopTimeout == -1;
224 # otherwise pick the lower of our regular timeout and time until
226 return $LoopTimeout if $LoopTimeout < $timeout;
230 # We can't use waitpid(-1) safely here since it can hit ``, system(),
231 # and other things. So we scan the $WaitPids list, which is hopefully
237 foreach my $ary (@$tmp) {
238 my ($pid, $cb, $arg) = @$ary;
239 my $ret = waitpid($pid, WNOHANG);
241 push @$WaitPids, $ary;
243 eval { $cb->($arg, $pid) };
247 # we may not be donea, and we may miss our
248 $reap_timer = add_timer(1, \&reap_pids);
252 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
253 sub enqueue_reap ($) { push @$nextq, \&reap_pids };
255 sub in_loop () { $in_loop }
262 my $timeout = RunTimers();
264 # get up to 1000 events
265 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
266 for ($i=0; $i<$evcount; $i++) {
267 # it's possible epoll_wait returned many events, including some at the end
268 # that ones in the front triggered unregister-interest actions. if we
269 # can't find the %sock entry, it's because we're no longer interested
271 $DescriptorMap{$events[$i]->[0]}->event_step;
273 } while (PostEventLoop());
277 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
279 Sets post loop callback function. Pass a subref and it will be
280 called every time the event loop finishes.
282 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
285 The callback function will be passed two parameters: \%DescriptorMap
288 sub SetPostLoopCallback {
289 my ($class, $ref) = @_;
292 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
295 # Internal function: run the post-event callback, send read events
296 # for pushed-back data, and close pending connections. returns 1
297 # if event loop should continue, or 0 to shut it all down.
299 # now we can close sockets that wanted to close during our event processing.
300 # (we didn't want to close them during the loop, as we didn't want fd numbers
301 # being reused and confused during the event loop)
302 delete($DescriptorMap{fileno($_)}) for @ToClose;
303 @ToClose = (); # let refcounting drop everything all at once
305 # by default we keep running, unless a postloop callback (either per-object
306 # or global) cancels it
307 my $keep_running = 1;
309 # now we're at the very end, call callback if defined
310 if (defined $PostLoopCallback) {
311 $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
314 return $keep_running;
317 #####################################################################
318 ### PublicInbox::DS-the-object code
319 #####################################################################
321 =head2 OBJECT METHODS
323 =head2 C<< CLASS->new( $socket ) >>
325 Create a new PublicInbox::DS subclass object for the given I<socket> which will
326 react to events on it during the C<EventLoop>.
328 This is normally (always?) called from your subclass via:
330 $class->SUPER::new($socket);
334 my ($self, $sock, $ev) = @_;
335 $self = fields::new($self) unless ref $self;
337 $self->{sock} = $sock;
338 my $fd = fileno($sock);
342 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
343 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
344 $ev &= ~EPOLLEXCLUSIVE;
347 die "couldn't add epoll watch for $fd: $!\n";
349 confess("DescriptorMap{$fd} defined ($DescriptorMap{$fd})")
350 if defined($DescriptorMap{$fd});
352 $DescriptorMap{$fd} = $self;
356 #####################################################################
357 ### I N S T A N C E M E T H O D S
358 #####################################################################
360 sub requeue ($) { push @$nextq, $_[0] }
362 =head2 C<< $obj->close >>
369 my $sock = delete $self->{sock} or return;
371 # we need to flush our write buffer, as there may
372 # be self-referential closures (sub { $client->close })
373 # preventing the object from being destroyed
374 delete $self->{wbuf};
376 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
377 # notifications about it
378 my $fd = fileno($sock);
379 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
380 confess("EPOLL_CTL_DEL: $!");
382 # we explicitly don't delete from DescriptorMap here until we
383 # actually close the socket, as we might be in the middle of
384 # processing an epoll_wait/etc that returned hundreds of fds, one
385 # of which is not yet processed and is what we're closing. if we
386 # keep it in DescriptorMap, then the event harnesses can just
387 # looked at $pob->{sock} == undef and ignore it. but if it's an
388 # un-accounted for fd, then it (understandably) freak out a bit
389 # and emit warnings, thinking their state got off.
391 # defer closing the actual socket until the event loop is done
392 # processing this round of events. (otherwise we might reuse fds)
393 push @ToClose, $sock;
398 # portable, non-thread-safe sendfile emulation (no pread, yet)
399 sub psendfile ($$$) {
400 my ($sock, $fh, $off) = @_;
402 seek($fh, $$off, SEEK_SET) or return;
403 defined(my $to_write = read($fh, my $buf, 16384)) or return;
405 while ($to_write > 0) {
406 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
410 return if $written == 0;
418 sub epbit ($$) { # (sock, default)
419 ref($_[0]) eq 'IO::Socket::SSL' ? PublicInbox::TLS::epollbit() : $_[1];
422 # returns 1 if done, 0 if incomplete
423 sub flush_write ($) {
425 my $wbuf = $self->{wbuf} or return 1;
426 my $sock = $self->{sock};
429 while (my $bref = $wbuf->[0]) {
430 if (ref($bref) ne 'CODE') {
431 my $off = delete($self->{wbuf_off}) // 0;
433 my $w = psendfile($sock, $bref, \$off);
439 } elsif ($! == EAGAIN) {
440 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
441 $self->{wbuf_off} = $off;
447 } else { #($ref eq 'CODE') {
449 my $before = scalar(@$wbuf);
452 # bref may be enqueueing more CODE to call (see accept_tls_step)
453 return 0 if (scalar(@$wbuf) > $before);
457 delete $self->{wbuf};
462 my ($self, $rbuf) = @_;
463 if ($$rbuf eq '') { # who knows how long till we can read again
464 delete $self->{rbuf};
466 $self->{rbuf} = $rbuf;
470 sub do_read ($$$;$) {
471 my ($self, $rbuf, $len, $off) = @_;
472 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
473 return ($r == 0 ? $self->close : $r) if defined $r;
474 # common for clients to break connections without warning,
475 # would be too noisy to log here:
477 epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT);
478 rbuf_idle($self, $rbuf);
485 # drop the socket if we hit unrecoverable errors on our system which
486 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
493 # n.b.: use ->write/->read for this buffer to allow compatibility with
494 # PerlIO::mmap or PerlIO::scalar if needed
496 my ($self, $bref, $off) = @_;
497 my $fh = tmpfile('wbuf', $self->{sock}, 1) or
498 return drop($self, "tmpfile $!");
500 my $len = bytes::length($$bref) - $off;
501 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
505 =head2 C<< $obj->write( $data ) >>
507 Write the specified data to the underlying handle. I<data> may be scalar,
508 scalar ref, code ref (to run when there).
509 Returns 1 if writes all went through, or 0 if there are writes in queue. If
510 it returns 1, caller should stop waiting for 'writable' events)
514 my ($self, $data) = @_;
516 # nobody should be writing to closed sockets, but caller code can
517 # do two writes within an event, have the first fail and
518 # disconnect the other side (whose destructor then closes the
519 # calling object, but it's still in a method), and then the
520 # now-dead object does its second write. that is this case. we
521 # just lie and say it worked. it'll be dead soon and won't be
523 my $sock = $self->{sock} or return 1;
525 my $bref = $ref ? $data : \$data;
526 my $wbuf = $self->{wbuf};
527 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
528 if ($ref eq 'CODE') {
531 my $last = $wbuf->[-1];
532 if (ref($last) eq 'GLOB') { # append to tmp file buffer
533 $last->print($$bref) or return drop($self, "print: $!");
535 my $tmpio = tmpio($self, $bref, 0) or return 0;
540 } elsif ($ref eq 'CODE') {
544 my $to_write = bytes::length($$bref);
545 my $written = syswrite($sock, $$bref, $to_write);
547 if (defined $written) {
548 return 1 if $written == $to_write;
549 requeue($self); # runs: event_step -> flush_write
550 } elsif ($! == EAGAIN) {
551 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
557 # deal with EAGAIN or partial write:
558 my $tmpio = tmpio($self, $bref, $written) or return 0;
560 # wbuf may be an empty array if we're being called inside
561 # ->flush_write via CODE bref:
562 push @{$self->{wbuf} ||= []}, $tmpio;
567 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
571 my $sock = $self->{sock} or return 1;
572 my $wbuf = $self->{wbuf};
574 if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) &&
575 ref($sock) ne 'IO::Socket::SSL') {
576 my $n = send($sock, $_[1], MSG_MORE);
578 my $nlen = bytes::length($_[1]) - $n;
579 return 1 if $nlen == 0; # all done!
580 # queue up the unwritten substring:
581 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
582 $self->{wbuf} //= $wbuf //= [];
584 epwait($sock, EPOLLOUT|EPOLLONESHOT);
589 # don't redispatch into NNTPdeflate::write
590 PublicInbox::DS::write($self, \($_[1]));
594 my ($sock, $ev) = @_;
595 epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and
596 confess("EPOLL_CTL_MOD $!");
599 # return true if complete, false if incomplete (or failure)
600 sub accept_tls_step ($) {
602 my $sock = $self->{sock} or return;
603 return 1 if $sock->accept_SSL;
604 return $self->close if $! != EAGAIN;
605 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
606 unshift @{$self->{wbuf} ||= []}, \&accept_tls_step;
610 # return true if complete, false if incomplete (or failure)
611 sub shutdn_tls_step ($) {
613 my $sock = $self->{sock} or return;
614 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
615 return $self->close if $! != EAGAIN;
616 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
617 unshift @{$self->{wbuf} ||= []}, \&shutdn_tls_step;
621 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
622 # or fork w/o exec, so no inadvertant socket sharing
625 my $sock = $self->{sock} or return;
626 if (ref($sock) eq 'IO::Socket::SSL') {
627 shutdn_tls_step($self);
633 # must be called with eval, PublicInbox::DS may not be loaded (see t/qspawn.t)
635 my ($pid, $cb, $arg) = @_;
637 push @$WaitPids, [ $pid, $cb, $arg ];
639 # We could've just missed our SIGCHLD, cover it, here:
640 requeue(\&reap_pids);
642 die "Not in EventLoop\n";
647 my $run = $later_queue;
648 $later_timer = undef;
655 push @$later_queue, $cb;
656 $later_timer //= add_timer(60, \&_run_later);
662 my $old = $now - $exp;
664 while (my ($fd, $v) = each %$EXPMAP) {
665 my ($idle_time, $ds_obj) = @$v;
666 if ($idle_time < $old) {
667 if (!$ds_obj->shutdn) {
675 $exp_timer = scalar(keys %new) ? later(\&expire_old) : undef;
678 sub update_idle_time {
680 my $sock = $self->{sock} or return;
681 $EXPMAP->{fileno($sock)} = [ now(), $self ];
682 $exp_timer //= later(\&expire_old);
686 my ($self, $now) = @_;
687 my $sock = $self->{sock} or return;
688 my $ary = $EXPMAP->{fileno($sock)} or return;
689 my $exp_at = $ary->[0] + $EXPTIME;
695 =head1 AUTHORS (Danga::Socket)
697 Brad Fitzpatrick <brad@danga.com> - author
699 Michael Granger <ged@danga.com> - docs, testing
701 Mark Smith <junior@danga.com> - contributor, heavy user, testing
703 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits