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 O_APPEND);
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 tmpio (autovivified))
35 # (tmpio = [ GLOB, offset, [ length ] ])
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; # list of callbacks to run at some later interval
44 my $EXPMAP; # fd -> idle_time
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() >>
74 $wait_pids = $later_queue = undef;
76 $nextq = $ToClose = $reap_timer = $later_timer = $exp_timer = undef;
77 $LoopTimeout = -1; # no timeout by default
80 $PostLoopCallback = undef;
83 $_io = undef; # closes real $Epoll FD
84 $Epoll = undef; # may call DSKQXS::DESTROY
86 *EventLoop = *FirstTimeEventLoop;
89 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
91 Set the loop timeout for the event loop to some value in milliseconds.
93 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
98 return $LoopTimeout = $_[1] + 0;
101 =head2 C<< PublicInbox::DS::add_timer( $seconds, $coderef ) >>
103 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
104 are not guaranteed to fire at the exact time you ask for.
108 my ($secs, $coderef) = @_;
110 my $fire_time = now() + $secs;
112 my $timer = [$fire_time, $coderef];
114 if (!@Timers || $fire_time >= $Timers[-1][0]) {
115 push @Timers, $timer;
119 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
120 # but it was compared against calendar queues, heaps, naive push/sort,
121 # and a bunch of other versions, and found to be fastest with a large
122 # variety of datasets.)
123 for (my $i = 0; $i < @Timers; $i++) {
124 if ($Timers[$i][0] > $fire_time) {
125 splice(@Timers, $i, 0, $timer);
130 die "Shouldn't get here.";
133 # keeping this around in case we support other FD types for now,
134 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
135 sub set_cloexec ($) {
138 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
139 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
140 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
148 if (PublicInbox::Syscall::epoll_defined()) {
149 $Epoll = epoll_create();
150 set_cloexec($Epoll) if (defined($Epoll) && $Epoll >= 0);
153 for (qw(DSKQXS DSPoll)) {
154 $cls = "PublicInbox::$_";
155 last if eval "require $cls";
157 $cls->import(qw(epoll_ctl epoll_wait));
160 *EventLoop = *EpollEventLoop;
163 =head2 C<< CLASS->EventLoop() >>
165 Start processing IO events. In most daemon programs this never exits. See
166 C<PostLoopCallback> below for how to exit the loop.
169 sub FirstTimeEventLoop {
177 sub now () { clock_gettime(CLOCK_MONOTONIC) }
180 my $q = $nextq or return;
183 # we avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak:
184 # https://rt.perl.org/Public/Bug/Display.html?id=114340
193 # runs timers and returns milliseconds for next one, or next event loop
197 return (($nextq || $ToClose) ? 0 : $LoopTimeout) unless @Timers;
202 while (@Timers && $Timers[0][0] <= $now) {
203 my $to_run = shift(@Timers);
204 $to_run->[1]->($now) if $to_run->[1];
207 # timers may enqueue into nextq:
208 return 0 if ($nextq || $ToClose);
210 return $LoopTimeout unless @Timers;
212 # convert time to an even number of milliseconds, adding 1
213 # extra, otherwise floating point fun can occur and we'll
214 # call RunTimers like 20-30 times, each returning a timeout
215 # of 0.0000212 seconds
216 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
218 # -1 is an infinite timeout, so prefer a real timeout
219 return $timeout if $LoopTimeout == -1;
221 # otherwise pick the lower of our regular timeout and time until
223 return $LoopTimeout if $LoopTimeout < $timeout;
227 # We can't use waitpid(-1) safely here since it can hit ``, system(),
228 # and other things. So we scan the $wait_pids list, which is hopefully
229 # not too big. We keep $wait_pids small by not calling dwaitpid()
230 # until we've hit EOF when reading the stdout of the child.
232 my $tmp = $wait_pids or return;
233 $wait_pids = $reap_timer = undef;
234 foreach my $ary (@$tmp) {
235 my ($pid, $cb, $arg) = @$ary;
236 my $ret = waitpid($pid, WNOHANG);
238 push @$wait_pids, $ary; # autovivifies @$wait_pids
240 eval { $cb->($arg, $pid) };
243 # we may not be done, yet, and could've missed/masked a SIGCHLD:
244 $reap_timer = add_timer(1, \&reap_pids) if $wait_pids;
247 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
248 sub enqueue_reap ($) { push @$nextq, \&reap_pids }; # autovivifies
250 sub in_loop () { $in_loop }
252 # Internal function: run the post-event callback, send read events
253 # for pushed-back data, and close pending connections. returns 1
254 # if event loop should continue, or 0 to shut it all down.
255 sub PostEventLoop () {
256 # now we can close sockets that wanted to close during our event
257 # processing. (we didn't want to close them during the loop, as we
258 # didn't want fd numbers being reused and confused during the event
260 if (my $close_now = $ToClose) {
261 $ToClose = undef; # will be autovivified on push
262 @$close_now = map { fileno($_) } @$close_now;
264 # order matters, destroy expiry times, first:
265 delete @$EXPMAP{@$close_now};
267 # ->DESTROY methods may populate ToClose
268 delete @DescriptorMap{@$close_now};
271 # by default we keep running, unless a postloop callback cancels it
272 $PostLoopCallback ? $PostLoopCallback->(\%DescriptorMap) : 1;
280 my $timeout = RunTimers();
282 # get up to 1000 events
283 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
284 for ($i=0; $i<$evcount; $i++) {
285 # it's possible epoll_wait returned many events, including some at the end
286 # that ones in the front triggered unregister-interest actions. if we
287 # can't find the %sock entry, it's because we're no longer interested
289 $DescriptorMap{$events[$i]->[0]}->event_step;
291 } while (PostEventLoop());
295 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
297 Sets post loop callback function. Pass a subref and it will be
298 called every time the event loop finishes.
300 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
303 The callback function will be passed two parameters: \%DescriptorMap
306 sub SetPostLoopCallback {
307 my ($class, $ref) = @_;
310 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
313 #####################################################################
314 ### PublicInbox::DS-the-object code
315 #####################################################################
317 =head2 OBJECT METHODS
319 =head2 C<< CLASS->new( $socket ) >>
321 Create a new PublicInbox::DS subclass object for the given I<socket> which will
322 react to events on it during the C<EventLoop>.
324 This is normally (always?) called from your subclass via:
326 $class->SUPER::new($socket);
330 my ($self, $sock, $ev) = @_;
331 $self = fields::new($self) unless ref $self;
333 $self->{sock} = $sock;
334 my $fd = fileno($sock);
338 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
339 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
340 $ev &= ~EPOLLEXCLUSIVE;
343 die "couldn't add epoll watch for $fd: $!\n";
345 confess("DescriptorMap{$fd} defined ($DescriptorMap{$fd})")
346 if defined($DescriptorMap{$fd});
348 $DescriptorMap{$fd} = $self;
352 #####################################################################
353 ### I N S T A N C E M E T H O D S
354 #####################################################################
356 sub requeue ($) { push @$nextq, $_[0] } # autovivifies
358 =head2 C<< $obj->close >>
365 my $sock = delete $self->{sock} or return;
367 # we need to flush our write buffer, as there may
368 # be self-referential closures (sub { $client->close })
369 # preventing the object from being destroyed
370 delete $self->{wbuf};
372 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
373 # notifications about it
374 my $fd = fileno($sock);
375 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
376 confess("EPOLL_CTL_DEL: $!");
378 # we explicitly don't delete from DescriptorMap here until we
379 # actually close the socket, as we might be in the middle of
380 # processing an epoll_wait/etc that returned hundreds of fds, one
381 # of which is not yet processed and is what we're closing. if we
382 # keep it in DescriptorMap, then the event harnesses can just
383 # looked at $pob->{sock} == undef and ignore it. but if it's an
384 # un-accounted for fd, then it (understandably) freak out a bit
385 # and emit warnings, thinking their state got off.
387 # defer closing the actual socket until the event loop is done
388 # processing this round of events. (otherwise we might reuse fds)
389 push @$ToClose, $sock; # autovivifies $ToClose
394 # portable, non-thread-safe sendfile emulation (no pread, yet)
395 sub send_tmpio ($$) {
396 my ($sock, $tmpio) = @_;
398 sysseek($tmpio->[0], $tmpio->[1], SEEK_SET) or return;
399 my $n = $tmpio->[2] // 65536;
400 $n = 65536 if $n > 65536;
401 defined(my $to_write = sysread($tmpio->[0], my $buf, $n)) or return;
403 while ($to_write > 0) {
404 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
408 return if $written == 0;
412 $tmpio->[1] += $written; # offset
413 $tmpio->[2] -= $written if defined($tmpio->[2]); # length
417 sub epbit ($$) { # (sock, default)
418 $_[0]->can('stop_SSL') ? PublicInbox::TLS::epollbit() : $_[1];
421 # returns 1 if done, 0 if incomplete
422 sub flush_write ($) {
424 my $wbuf = $self->{wbuf} or return 1;
425 my $sock = $self->{sock};
428 while (my $bref = $wbuf->[0]) {
429 if (ref($bref) ne 'CODE') {
431 my $w = send_tmpio($sock, $bref); # bref is tmpio
437 } elsif ($! == EAGAIN) {
438 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
444 } else { #(ref($bref) eq 'CODE') {
446 my $before = scalar(@$wbuf);
449 # bref may be enqueueing more CODE to call (see accept_tls_step)
450 return 0 if (scalar(@$wbuf) > $before);
454 delete $self->{wbuf};
459 my ($self, $rbuf) = @_;
460 if ($$rbuf eq '') { # who knows how long till we can read again
461 delete $self->{rbuf};
463 $self->{rbuf} = $rbuf;
467 sub do_read ($$$;$) {
468 my ($self, $rbuf, $len, $off) = @_;
469 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
470 return ($r == 0 ? $self->close : $r) if defined $r;
471 # common for clients to break connections without warning,
472 # would be too noisy to log here:
474 epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT);
475 rbuf_idle($self, $rbuf);
482 # drop the socket if we hit unrecoverable errors on our system which
483 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
490 # n.b.: use ->write/->read for this buffer to allow compatibility with
491 # PerlIO::mmap or PerlIO::scalar if needed
493 my ($self, $bref, $off) = @_;
494 my $fh = tmpfile('wbuf', $self->{sock}, O_APPEND) or
495 return drop($self, "tmpfile $!");
497 my $len = bytes::length($$bref) - $off;
498 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
499 [ $fh, 0 ] # [1] = offset, [2] = length, not set by us
502 =head2 C<< $obj->write( $data ) >>
504 Write the specified data to the underlying handle. I<data> may be scalar,
505 scalar ref, code ref (to run when there).
506 Returns 1 if writes all went through, or 0 if there are writes in queue. If
507 it returns 1, caller should stop waiting for 'writable' events)
511 my ($self, $data) = @_;
513 # nobody should be writing to closed sockets, but caller code can
514 # do two writes within an event, have the first fail and
515 # disconnect the other side (whose destructor then closes the
516 # calling object, but it's still in a method), and then the
517 # now-dead object does its second write. that is this case. we
518 # just lie and say it worked. it'll be dead soon and won't be
520 my $sock = $self->{sock} or return 1;
522 my $bref = $ref ? $data : \$data;
523 my $wbuf = $self->{wbuf};
524 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
525 if ($ref eq 'CODE') {
528 my $tmpio = $wbuf->[-1];
529 if ($tmpio && !defined($tmpio->[2])) { # append to tmp file buffer
530 $tmpio->[0]->print($$bref) or return drop($self, "print: $!");
532 my $tmpio = tmpio($self, $bref, 0) or return 0;
537 } elsif ($ref eq 'CODE') {
541 my $to_write = bytes::length($$bref);
542 my $written = syswrite($sock, $$bref, $to_write);
544 if (defined $written) {
545 return 1 if $written == $to_write;
546 requeue($self); # runs: event_step -> flush_write
547 } elsif ($! == EAGAIN) {
548 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
554 # deal with EAGAIN or partial write:
555 my $tmpio = tmpio($self, $bref, $written) or return 0;
557 # wbuf may be an empty array if we're being called inside
558 # ->flush_write via CODE bref:
559 push @{$self->{wbuf}}, $tmpio; # autovivifies
564 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
568 my $sock = $self->{sock} or return 1;
569 my $wbuf = $self->{wbuf};
571 if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) &&
572 !$sock->can('stop_SSL')) {
573 my $n = send($sock, $_[1], MSG_MORE);
575 my $nlen = bytes::length($_[1]) - $n;
576 return 1 if $nlen == 0; # all done!
577 # queue up the unwritten substring:
578 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
579 push @{$self->{wbuf}}, $tmpio; # autovivifies
580 epwait($sock, EPOLLOUT|EPOLLONESHOT);
585 # don't redispatch into NNTPdeflate::write
586 PublicInbox::DS::write($self, \($_[1]));
590 my ($sock, $ev) = @_;
591 epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and
592 confess("EPOLL_CTL_MOD $!");
595 # return true if complete, false if incomplete (or failure)
596 sub accept_tls_step ($) {
598 my $sock = $self->{sock} or return;
599 return 1 if $sock->accept_SSL;
600 return $self->close if $! != EAGAIN;
601 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
602 unshift(@{$self->{wbuf}}, \&accept_tls_step); # autovivifies
606 # return true if complete, false if incomplete (or failure)
607 sub shutdn_tls_step ($) {
609 my $sock = $self->{sock} or return;
610 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
611 return $self->close if $! != EAGAIN;
612 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
613 unshift(@{$self->{wbuf}}, \&shutdn_tls_step); # autovivifies
617 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
618 # or fork w/o exec, so no inadvertent socket sharing
621 my $sock = $self->{sock} or return;
622 if ($sock->can('stop_SSL')) {
623 shutdn_tls_step($self);
629 # must be called with eval, PublicInbox::DS may not be loaded (see t/qspawn.t)
631 die "Not in EventLoop\n" unless $in_loop;
632 push @$wait_pids, [ @_ ]; # [ $pid, $cb, $arg ]
634 # We could've just missed our SIGCHLD, cover it, here:
635 requeue(\&reap_pids);
639 my $run = $later_queue or return;
640 $later_timer = $later_queue = undef;
645 push @$later_queue, $_[0]; # autovivifies @$later_queue
646 $later_timer //= add_timer(60, \&_run_later);
652 my $old = $now - $exp;
654 while (my ($fd, $idle_at) = each %$EXPMAP) {
655 if ($idle_at < $old) {
656 my $ds_obj = $DescriptorMap{$fd};
657 $new{$fd} = $idle_at if !$ds_obj->shutdn;
659 $new{$fd} = $idle_at;
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();
670 $exp_timer //= later(\&expire_old);
674 my ($self, $now) = @_;
675 my $sock = $self->{sock} or return;
676 my $idle_at = $EXPMAP->{fileno($sock)} or return;
677 ($idle_at + $EXPTIME) > $now;
682 =head1 AUTHORS (Danga::Socket)
684 Brad Fitzpatrick <brad@danga.com> - author
686 Michael Granger <ged@danga.com> - docs, testing
688 Mark Smith <junior@danga.com> - contributor, heavy user, testing
690 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits