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 (autovivified)
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 $wait_pids; # 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);
47 my $ToClose; # sockets to close when event loop is done
49 %DescriptorMap, # fd (num) -> PublicInbox::DS object
50 $Epoll, # Global epoll fd (or DSKQXS ref)
51 $_io, # IO::Handle for Epoll
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() >>
77 $nextq = $ToClose = $reap_timer = $later_timer = $exp_timer = undef;
78 $LoopTimeout = -1; # no timeout by default
81 $PostLoopCallback = undef;
84 $_io = undef; # closes real $Epoll FD
85 $Epoll = undef; # may call DSKQXS::DESTROY
87 *EventLoop = *FirstTimeEventLoop;
90 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
92 Set the loop timeout for the event loop to some value in milliseconds.
94 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
99 return $LoopTimeout = $_[1] + 0;
102 =head2 C<< PublicInbox::DS::add_timer( $seconds, $coderef ) >>
104 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
105 are not guaranteed to fire at the exact time you ask for.
109 my ($secs, $coderef) = @_;
111 my $fire_time = now() + $secs;
113 my $timer = [$fire_time, $coderef];
115 if (!@Timers || $fire_time >= $Timers[-1][0]) {
116 push @Timers, $timer;
120 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
121 # but it was compared against calendar queues, heaps, naive push/sort,
122 # and a bunch of other versions, and found to be fastest with a large
123 # variety of datasets.)
124 for (my $i = 0; $i < @Timers; $i++) {
125 if ($Timers[$i][0] > $fire_time) {
126 splice(@Timers, $i, 0, $timer);
131 die "Shouldn't get here.";
134 # keeping this around in case we support other FD types for now,
135 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
136 sub set_cloexec ($) {
139 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
140 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
141 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
149 if (PublicInbox::Syscall::epoll_defined()) {
150 $Epoll = epoll_create();
151 set_cloexec($Epoll) if (defined($Epoll) && $Epoll >= 0);
154 for (qw(DSKQXS DSPoll)) {
155 $cls = "PublicInbox::$_";
156 last if eval "require $cls";
158 $cls->import(qw(epoll_ctl epoll_wait));
161 *EventLoop = *EpollEventLoop;
164 =head2 C<< CLASS->EventLoop() >>
166 Start processing IO events. In most daemon programs this never exits. See
167 C<PostLoopCallback> below for how to exit the loop.
170 sub FirstTimeEventLoop {
178 sub now () { clock_gettime(CLOCK_MONOTONIC) }
181 my $q = $nextq or return;
184 # we avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak:
185 # https://rt.perl.org/Public/Bug/Display.html?id=114340
194 # runs timers and returns milliseconds for next one, or next event loop
198 return (($nextq || $ToClose) ? 0 : $LoopTimeout) unless @Timers;
203 while (@Timers && $Timers[0][0] <= $now) {
204 my $to_run = shift(@Timers);
205 $to_run->[1]->($now) if $to_run->[1];
208 # timers may enqueue into nextq:
209 return 0 if ($nextq || $ToClose);
211 return $LoopTimeout unless @Timers;
213 # convert time to an even number of milliseconds, adding 1
214 # extra, otherwise floating point fun can occur and we'll
215 # call RunTimers like 20-30 times, each returning a timeout
216 # of 0.0000212 seconds
217 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
219 # -1 is an infinite timeout, so prefer a real timeout
220 return $timeout if $LoopTimeout == -1;
222 # otherwise pick the lower of our regular timeout and time until
224 return $LoopTimeout if $LoopTimeout < $timeout;
228 # We can't use waitpid(-1) safely here since it can hit ``, system(),
229 # and other things. So we scan the $wait_pids list, which is hopefully
230 # not too big. We keep $wait_pids small by not calling dwaitpid()
231 # until we've hit EOF when reading the stdout of the child.
233 my $tmp = $wait_pids or return;
234 $wait_pids = $reap_timer = undef;
235 foreach my $ary (@$tmp) {
236 my ($pid, $cb, $arg) = @$ary;
237 my $ret = waitpid($pid, WNOHANG);
239 push @$wait_pids, $ary; # autovivifies @$wait_pids
241 eval { $cb->($arg, $pid) };
244 # we may not be done, yet, and could've missed/masked a SIGCHLD:
245 $reap_timer = add_timer(1, \&reap_pids) if $wait_pids;
248 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
249 sub enqueue_reap ($) { push @$nextq, \&reap_pids }; # autovivifies
251 sub in_loop () { $in_loop }
253 # Internal function: run the post-event callback, send read events
254 # for pushed-back data, and close pending connections. returns 1
255 # if event loop should continue, or 0 to shut it all down.
256 sub PostEventLoop () {
257 # now we can close sockets that wanted to close during our event
258 # processing. (we didn't want to close them during the loop, as we
259 # didn't want fd numbers being reused and confused during the event
261 if (my $close_now = $ToClose) {
262 $ToClose = undef; # will be autovivified on push
263 # ->DESTROY methods may populate ToClose
264 delete($DescriptorMap{fileno($_)}) for @$close_now;
265 # let refcounting drop everything in $close_now at once
268 # by default we keep running, unless a postloop callback cancels it
269 $PostLoopCallback ? $PostLoopCallback->(\%DescriptorMap) : 1;
277 my $timeout = RunTimers();
279 # get up to 1000 events
280 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
281 for ($i=0; $i<$evcount; $i++) {
282 # it's possible epoll_wait returned many events, including some at the end
283 # that ones in the front triggered unregister-interest actions. if we
284 # can't find the %sock entry, it's because we're no longer interested
286 $DescriptorMap{$events[$i]->[0]}->event_step;
288 } while (PostEventLoop());
292 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
294 Sets post loop callback function. Pass a subref and it will be
295 called every time the event loop finishes.
297 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
300 The callback function will be passed two parameters: \%DescriptorMap
303 sub SetPostLoopCallback {
304 my ($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<EventLoop>.
321 This is normally (always?) called from your subclass via:
323 $class->SUPER::new($socket);
327 my ($self, $sock, $ev) = @_;
328 $self = fields::new($self) unless ref $self;
330 $self->{sock} = $sock;
331 my $fd = fileno($sock);
335 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
336 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
337 $ev &= ~EPOLLEXCLUSIVE;
340 die "couldn't add epoll watch for $fd: $!\n";
342 confess("DescriptorMap{$fd} defined ($DescriptorMap{$fd})")
343 if defined($DescriptorMap{$fd});
345 $DescriptorMap{$fd} = $self;
349 #####################################################################
350 ### I N S T A N C E M E T H O D S
351 #####################################################################
353 sub requeue ($) { push @$nextq, $_[0] } # autovivifies
355 =head2 C<< $obj->close >>
362 my $sock = delete $self->{sock} or return;
364 # we need to flush our write buffer, as there may
365 # be self-referential closures (sub { $client->close })
366 # preventing the object from being destroyed
367 delete $self->{wbuf};
369 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
370 # notifications about it
371 my $fd = fileno($sock);
372 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
373 confess("EPOLL_CTL_DEL: $!");
375 # we explicitly don't delete from DescriptorMap here until we
376 # actually close the socket, as we might be in the middle of
377 # processing an epoll_wait/etc that returned hundreds of fds, one
378 # of which is not yet processed and is what we're closing. if we
379 # keep it in DescriptorMap, then the event harnesses can just
380 # looked at $pob->{sock} == undef and ignore it. but if it's an
381 # un-accounted for fd, then it (understandably) freak out a bit
382 # and emit warnings, thinking their state got off.
384 # defer closing the actual socket until the event loop is done
385 # processing this round of events. (otherwise we might reuse fds)
386 push @$ToClose, $sock; # autovivifies $ToClose
391 # portable, non-thread-safe sendfile emulation (no pread, yet)
392 sub psendfile ($$$) {
393 my ($sock, $fh, $off) = @_;
395 seek($fh, $$off, SEEK_SET) or return;
396 defined(my $to_write = read($fh, my $buf, 16384)) or return;
398 while ($to_write > 0) {
399 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
403 return if $written == 0;
411 sub epbit ($$) { # (sock, default)
412 ref($_[0]) eq 'IO::Socket::SSL' ? PublicInbox::TLS::epollbit() : $_[1];
415 # returns 1 if done, 0 if incomplete
416 sub flush_write ($) {
418 my $wbuf = $self->{wbuf} or return 1;
419 my $sock = $self->{sock};
422 while (my $bref = $wbuf->[0]) {
423 if (ref($bref) ne 'CODE') {
424 my $off = delete($self->{wbuf_off}) // 0;
426 my $w = psendfile($sock, $bref, \$off);
432 } elsif ($! == EAGAIN) {
433 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
434 $self->{wbuf_off} = $off;
440 } else { #($ref eq 'CODE') {
442 my $before = scalar(@$wbuf);
445 # bref may be enqueueing more CODE to call (see accept_tls_step)
446 return 0 if (scalar(@$wbuf) > $before);
450 delete $self->{wbuf};
455 my ($self, $rbuf) = @_;
456 if ($$rbuf eq '') { # who knows how long till we can read again
457 delete $self->{rbuf};
459 $self->{rbuf} = $rbuf;
463 sub do_read ($$$;$) {
464 my ($self, $rbuf, $len, $off) = @_;
465 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
466 return ($r == 0 ? $self->close : $r) if defined $r;
467 # common for clients to break connections without warning,
468 # would be too noisy to log here:
470 epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT);
471 rbuf_idle($self, $rbuf);
478 # drop the socket if we hit unrecoverable errors on our system which
479 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
486 # n.b.: use ->write/->read for this buffer to allow compatibility with
487 # PerlIO::mmap or PerlIO::scalar if needed
489 my ($self, $bref, $off) = @_;
490 my $fh = tmpfile('wbuf', $self->{sock}, 1) or
491 return drop($self, "tmpfile $!");
493 my $len = bytes::length($$bref) - $off;
494 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
498 =head2 C<< $obj->write( $data ) >>
500 Write the specified data to the underlying handle. I<data> may be scalar,
501 scalar ref, code ref (to run when there).
502 Returns 1 if writes all went through, or 0 if there are writes in queue. If
503 it returns 1, caller should stop waiting for 'writable' events)
507 my ($self, $data) = @_;
509 # nobody should be writing to closed sockets, but caller code can
510 # do two writes within an event, have the first fail and
511 # disconnect the other side (whose destructor then closes the
512 # calling object, but it's still in a method), and then the
513 # now-dead object does its second write. that is this case. we
514 # just lie and say it worked. it'll be dead soon and won't be
516 my $sock = $self->{sock} or return 1;
518 my $bref = $ref ? $data : \$data;
519 my $wbuf = $self->{wbuf};
520 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
521 if ($ref eq 'CODE') {
524 my $last = $wbuf->[-1];
525 if (ref($last) eq 'GLOB') { # append to tmp file buffer
526 $last->print($$bref) or return drop($self, "print: $!");
528 my $tmpio = tmpio($self, $bref, 0) or return 0;
533 } elsif ($ref eq 'CODE') {
537 my $to_write = bytes::length($$bref);
538 my $written = syswrite($sock, $$bref, $to_write);
540 if (defined $written) {
541 return 1 if $written == $to_write;
542 requeue($self); # runs: event_step -> flush_write
543 } elsif ($! == EAGAIN) {
544 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
550 # deal with EAGAIN or partial write:
551 my $tmpio = tmpio($self, $bref, $written) or return 0;
553 # wbuf may be an empty array if we're being called inside
554 # ->flush_write via CODE bref:
555 push @{$self->{wbuf}}, $tmpio; # autovivifies
560 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
564 my $sock = $self->{sock} or return 1;
565 my $wbuf = $self->{wbuf};
567 if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) &&
568 ref($sock) ne 'IO::Socket::SSL') {
569 my $n = send($sock, $_[1], MSG_MORE);
571 my $nlen = bytes::length($_[1]) - $n;
572 return 1 if $nlen == 0; # all done!
573 # queue up the unwritten substring:
574 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
575 push @{$self->{wbuf}}, $tmpio; # autovivifies
576 epwait($sock, EPOLLOUT|EPOLLONESHOT);
581 # don't redispatch into NNTPdeflate::write
582 PublicInbox::DS::write($self, \($_[1]));
586 my ($sock, $ev) = @_;
587 epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and
588 confess("EPOLL_CTL_MOD $!");
591 # return true if complete, false if incomplete (or failure)
592 sub accept_tls_step ($) {
594 my $sock = $self->{sock} or return;
595 return 1 if $sock->accept_SSL;
596 return $self->close if $! != EAGAIN;
597 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
598 unshift(@{$self->{wbuf}}, \&accept_tls_step); # autovivifies
602 # return true if complete, false if incomplete (or failure)
603 sub shutdn_tls_step ($) {
605 my $sock = $self->{sock} or return;
606 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
607 return $self->close if $! != EAGAIN;
608 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
609 unshift(@{$self->{wbuf}}, \&shutdn_tls_step); # autovivifies
613 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
614 # or fork w/o exec, so no inadvertant socket sharing
617 my $sock = $self->{sock} or return;
618 if (ref($sock) eq 'IO::Socket::SSL') {
619 shutdn_tls_step($self);
625 # must be called with eval, PublicInbox::DS may not be loaded (see t/qspawn.t)
627 die "Not in EventLoop\n" unless $in_loop;
628 push @$wait_pids, [ @_ ]; # [ $pid, $cb, $arg ]
630 # We could've just missed our SIGCHLD, cover it, here:
631 requeue(\&reap_pids);
635 my $run = $later_queue;
636 $later_timer = undef;
643 push @$later_queue, $cb;
644 $later_timer //= add_timer(60, \&_run_later);
650 my $old = $now - $exp;
652 while (my ($fd, $v) = each %$EXPMAP) {
653 my ($idle_time, $ds_obj) = @$v;
654 if ($idle_time < $old) {
655 if (!$ds_obj->shutdn) {
663 $exp_timer = scalar(keys %new) ? later(\&expire_old) : undef;
666 sub update_idle_time {
668 my $sock = $self->{sock} or return;
669 $EXPMAP->{fileno($sock)} = [ now(), $self ];
670 $exp_timer //= later(\&expire_old);
674 my ($self, $now) = @_;
675 my $sock = $self->{sock} or return;
676 my $ary = $EXPMAP->{fileno($sock)} or return;
677 my $exp_at = $ary->[0] + $EXPTIME;
683 =head1 AUTHORS (Danga::Socket)
685 Brad Fitzpatrick <brad@danga.com> - author
687 Michael Granger <ged@danga.com> - docs, testing
689 Mark Smith <junior@danga.com> - contributor, heavy user, testing
691 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits