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 POSIX qw(WNOHANG);
27 use Fcntl qw(SEEK_SET :DEFAULT O_APPEND);
28 use Time::HiRes qw(clock_gettime CLOCK_MONOTONIC);
29 use parent qw(Exporter);
30 our @EXPORT_OK = qw(now msg_more);
32 use Scalar::Util qw(blessed);
33 use PublicInbox::Syscall qw(:epoll);
34 use PublicInbox::Tmpfile;
35 use Errno qw(EAGAIN EINVAL);
36 use Carp qw(confess carp);
38 my $nextq; # queue for next_tick
39 my $wait_pids; # list of [ pid, callback, callback_arg ]
40 my $later_queue; # list of callbacks to run at some later interval
41 my $EXPMAP; # fd -> idle_time
42 our $EXPTIME = 180; # 3 minutes
43 my ($later_timer, $reap_armed, $exp_timer);
44 my $ToClose; # sockets to close when event loop is done
46 %DescriptorMap, # fd (num) -> PublicInbox::DS object
47 $Epoll, # Global epoll fd (or DSKQXS ref)
48 $_io, # IO::Handle for Epoll
50 $PostLoopCallback, # subref to call at the end of each loop, if defined (global)
52 $LoopTimeout, # timeout of event loop in milliseconds
59 #####################################################################
60 ### C L A S S M E T H O D S
61 #####################################################################
63 =head2 C<< CLASS->Reset() >>
70 $in_loop = $wait_pids = $later_queue = $reap_armed = undef;
72 $nextq = $ToClose = $later_timer = $exp_timer = undef;
73 $LoopTimeout = -1; # no timeout by default
76 $PostLoopCallback = undef;
78 $_io = undef; # closes real $Epoll FD
79 $Epoll = undef; # may call DSKQXS::DESTROY
82 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
84 Set the loop timeout for the event loop to some value in milliseconds.
86 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
90 sub SetLoopTimeout { $LoopTimeout = $_[1] + 0 }
92 =head2 C<< PublicInbox::DS::add_timer( $seconds, $coderef, $arg) >>
94 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
95 are not guaranteed to fire at the exact time you ask for.
98 sub add_timer ($$;$) {
99 my ($secs, $coderef, $arg) = @_;
101 my $fire_time = now() + $secs;
103 my $timer = [$fire_time, $coderef, $arg];
105 if (!@Timers || $fire_time >= $Timers[-1][0]) {
106 push @Timers, $timer;
110 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
111 # but it was compared against calendar queues, heaps, naive push/sort,
112 # and a bunch of other versions, and found to be fastest with a large
113 # variety of datasets.)
114 for (my $i = 0; $i < @Timers; $i++) {
115 if ($Timers[$i][0] > $fire_time) {
116 splice(@Timers, $i, 0, $timer);
121 die "Shouldn't get here.";
124 # keeping this around in case we support other FD types for now,
125 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
126 sub set_cloexec ($) {
129 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
130 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
131 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
134 # caller sets return value to $Epoll
137 if (PublicInbox::Syscall::epoll_defined()) {
138 my $fd = epoll_create();
139 set_cloexec($fd) if (defined($fd) && $fd >= 0);
143 for (qw(DSKQXS DSPoll)) {
144 $cls = "PublicInbox::$_";
145 last if eval "require $cls";
147 $cls->import(qw(epoll_ctl epoll_wait));
152 =head2 C<< CLASS->EventLoop() >>
154 Start processing IO events. In most daemon programs this never exits. See
155 C<PostLoopCallback> below for how to exit the loop.
159 sub now () { clock_gettime(CLOCK_MONOTONIC) }
162 my $q = $nextq or return;
165 # we avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak:
166 # https://rt.perl.org/Public/Bug/Display.html?id=114340
175 # runs timers and returns milliseconds for next one, or next event loop
179 return (($nextq || $ToClose) ? 0 : $LoopTimeout) unless @Timers;
184 while (@Timers && $Timers[0][0] <= $now) {
185 my $to_run = shift(@Timers);
186 $to_run->[1]->($to_run->[2]);
189 # timers may enqueue into nextq:
190 return 0 if ($nextq || $ToClose);
192 return $LoopTimeout unless @Timers;
194 # convert time to an even number of milliseconds, adding 1
195 # extra, otherwise floating point fun can occur and we'll
196 # call RunTimers like 20-30 times, each returning a timeout
197 # of 0.0000212 seconds
198 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
200 # -1 is an infinite timeout, so prefer a real timeout
201 ($LoopTimeout < 0 || $LoopTimeout >= $timeout) ? $timeout : $LoopTimeout;
204 # We can't use waitpid(-1) safely here since it can hit ``, system(),
205 # and other things. So we scan the $wait_pids list, which is hopefully
206 # not too big. We keep $wait_pids small by not calling dwaitpid()
207 # until we've hit EOF when reading the stdout of the child.
211 my $tmp = $wait_pids or return;
213 foreach my $ary (@$tmp) {
214 my ($pid, $cb, $arg) = @$ary;
215 my $ret = waitpid($pid, WNOHANG);
217 push @$wait_pids, $ary; # autovivifies @$wait_pids
219 eval { $cb->($arg, $pid) };
222 # we may not be done, yet, and could've missed/masked a SIGCHLD:
223 $reap_armed //= requeue(\&reap_pids) if $wait_pids;
226 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
227 sub enqueue_reap () { $reap_armed //= requeue(\&reap_pids) }
229 sub in_loop () { $in_loop }
231 # Internal function: run the post-event callback, send read events
232 # for pushed-back data, and close pending connections. returns 1
233 # if event loop should continue, or 0 to shut it all down.
234 sub PostEventLoop () {
235 # now we can close sockets that wanted to close during our event
236 # processing. (we didn't want to close them during the loop, as we
237 # didn't want fd numbers being reused and confused during the event
239 if (my $close_now = $ToClose) {
240 $ToClose = undef; # will be autovivified on push
241 @$close_now = map { fileno($_) } @$close_now;
243 # order matters, destroy expiry times, first:
244 delete @$EXPMAP{@$close_now};
246 # ->DESTROY methods may populate ToClose
247 delete @DescriptorMap{@$close_now};
250 # by default we keep running, unless a postloop callback cancels it
251 $PostLoopCallback ? $PostLoopCallback->(\%DescriptorMap) : 1;
255 $Epoll //= _InitPoller();
259 my $timeout = RunTimers();
261 # get up to 1000 events
262 epoll_wait($Epoll, 1000, $timeout, \@events);
263 for my $fd (@events) {
264 # it's possible epoll_wait returned many events, including some at the end
265 # that ones in the front triggered unregister-interest actions. if we
266 # can't find the %sock entry, it's because we're no longer interested
268 $DescriptorMap{$fd}->event_step;
270 } while (PostEventLoop());
274 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
276 Sets post loop callback function. Pass a subref and it will be
277 called every time the event loop finishes.
279 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
282 The callback function will be passed two parameters: \%DescriptorMap
285 sub SetPostLoopCallback {
286 my ($class, $ref) = @_;
289 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
292 #####################################################################
293 ### PublicInbox::DS-the-object code
294 #####################################################################
296 =head2 OBJECT METHODS
298 =head2 C<< CLASS->new( $socket ) >>
300 Create a new PublicInbox::DS subclass object for the given I<socket> which will
301 react to events on it during the C<EventLoop>.
303 This is normally (always?) called from your subclass via:
305 $class->SUPER::new($socket);
309 my ($self, $sock, $ev) = @_;
310 $self->{sock} = $sock;
311 my $fd = fileno($sock);
313 $Epoll //= _InitPoller();
315 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
316 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
317 $ev &= ~EPOLLEXCLUSIVE;
320 die "couldn't add epoll watch for $fd: $!\n";
322 confess("DescriptorMap{$fd} defined ($DescriptorMap{$fd})")
323 if defined($DescriptorMap{$fd});
325 $DescriptorMap{$fd} = $self;
329 #####################################################################
330 ### I N S T A N C E M E T H O D S
331 #####################################################################
333 sub requeue ($) { push @$nextq, $_[0] } # autovivifies
335 =head2 C<< $obj->close >>
342 my $sock = delete $self->{sock} or return;
344 # we need to flush our write buffer, as there may
345 # be self-referential closures (sub { $client->close })
346 # preventing the object from being destroyed
347 delete $self->{wbuf};
349 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
350 # notifications about it
351 my $fd = fileno($sock);
352 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
353 confess("EPOLL_CTL_DEL: $!");
355 # we explicitly don't delete from DescriptorMap here until we
356 # actually close the socket, as we might be in the middle of
357 # processing an epoll_wait/etc that returned hundreds of fds, one
358 # of which is not yet processed and is what we're closing. if we
359 # keep it in DescriptorMap, then the event harnesses can just
360 # looked at $pob->{sock} == undef and ignore it. but if it's an
361 # un-accounted for fd, then it (understandably) freak out a bit
362 # and emit warnings, thinking their state got off.
364 # defer closing the actual socket until the event loop is done
365 # processing this round of events. (otherwise we might reuse fds)
366 push @$ToClose, $sock; # autovivifies $ToClose
371 # portable, non-thread-safe sendfile emulation (no pread, yet)
372 sub send_tmpio ($$) {
373 my ($sock, $tmpio) = @_;
375 sysseek($tmpio->[0], $tmpio->[1], SEEK_SET) or return;
376 my $n = $tmpio->[2] // 65536;
377 $n = 65536 if $n > 65536;
378 defined(my $to_write = sysread($tmpio->[0], my $buf, $n)) or return;
380 while ($to_write > 0) {
381 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
385 return if $written == 0;
389 $tmpio->[1] += $written; # offset
390 $tmpio->[2] -= $written if defined($tmpio->[2]); # length
394 sub epbit ($$) { # (sock, default)
395 $_[0]->can('stop_SSL') ? PublicInbox::TLS::epollbit() : $_[1];
398 # returns 1 if done, 0 if incomplete
399 sub flush_write ($) {
401 my $sock = $self->{sock} or return;
402 my $wbuf = $self->{wbuf} or return 1;
405 while (my $bref = $wbuf->[0]) {
406 if (ref($bref) ne 'CODE') {
408 my $w = send_tmpio($sock, $bref); # bref is tmpio
414 } elsif ($! == EAGAIN) {
415 my $ev = epbit($sock, EPOLLOUT) or return $self->close;
416 epwait($sock, $ev | EPOLLONESHOT);
422 } else { #(ref($bref) eq 'CODE') {
424 my $before = scalar(@$wbuf);
427 # bref may be enqueueing more CODE to call (see accept_tls_step)
428 return 0 if (scalar(@$wbuf) > $before);
432 delete $self->{wbuf};
437 my ($self, $rbuf) = @_;
438 if ($$rbuf eq '') { # who knows how long till we can read again
439 delete $self->{rbuf};
441 $self->{rbuf} = $rbuf;
445 sub do_read ($$$;$) {
446 my ($self, $rbuf, $len, $off) = @_;
447 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
448 return ($r == 0 ? $self->close : $r) if defined $r;
449 # common for clients to break connections without warning,
450 # would be too noisy to log here:
452 my $ev = epbit($sock, EPOLLIN) or return $self->close;
453 epwait($sock, $ev | EPOLLONESHOT);
454 rbuf_idle($self, $rbuf);
461 # drop the socket if we hit unrecoverable errors on our system which
462 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
469 # n.b.: use ->write/->read for this buffer to allow compatibility with
470 # PerlIO::mmap or PerlIO::scalar if needed
472 my ($self, $bref, $off) = @_;
473 my $fh = tmpfile('wbuf', $self->{sock}, O_APPEND) or
474 return drop($self, "tmpfile $!");
476 my $len = bytes::length($$bref) - $off;
477 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
478 [ $fh, 0 ] # [1] = offset, [2] = length, not set by us
481 =head2 C<< $obj->write( $data ) >>
483 Write the specified data to the underlying handle. I<data> may be scalar,
484 scalar ref, code ref (to run when there).
485 Returns 1 if writes all went through, or 0 if there are writes in queue. If
486 it returns 1, caller should stop waiting for 'writable' events)
490 my ($self, $data) = @_;
492 # nobody should be writing to closed sockets, but caller code can
493 # do two writes within an event, have the first fail and
494 # disconnect the other side (whose destructor then closes the
495 # calling object, but it's still in a method), and then the
496 # now-dead object does its second write. that is this case. we
497 # just lie and say it worked. it'll be dead soon and won't be
499 my $sock = $self->{sock} or return 1;
501 my $bref = $ref ? $data : \$data;
502 my $wbuf = $self->{wbuf};
503 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
504 if ($ref eq 'CODE') {
507 my $tmpio = $wbuf->[-1];
508 if ($tmpio && !defined($tmpio->[2])) { # append to tmp file buffer
509 $tmpio->[0]->print($$bref) or return drop($self, "print: $!");
511 my $tmpio = tmpio($self, $bref, 0) or return 0;
516 } elsif ($ref eq 'CODE') {
520 my $to_write = bytes::length($$bref);
521 my $written = syswrite($sock, $$bref, $to_write);
523 if (defined $written) {
524 return 1 if $written == $to_write;
525 requeue($self); # runs: event_step -> flush_write
526 } elsif ($! == EAGAIN) {
527 my $ev = epbit($sock, EPOLLOUT) or return $self->close;
528 epwait($sock, $ev | EPOLLONESHOT);
534 # deal with EAGAIN or partial write:
535 my $tmpio = tmpio($self, $bref, $written) or return 0;
537 # wbuf may be an empty array if we're being called inside
538 # ->flush_write via CODE bref:
539 push @{$self->{wbuf}}, $tmpio; # autovivifies
544 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
548 my $sock = $self->{sock} or return 1;
549 my $wbuf = $self->{wbuf};
551 if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) &&
552 !$sock->can('stop_SSL')) {
553 my $n = send($sock, $_[1], MSG_MORE);
555 my $nlen = bytes::length($_[1]) - $n;
556 return 1 if $nlen == 0; # all done!
557 # queue up the unwritten substring:
558 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
559 push @{$self->{wbuf}}, $tmpio; # autovivifies
560 epwait($sock, EPOLLOUT|EPOLLONESHOT);
565 # don't redispatch into NNTPdeflate::write
566 PublicInbox::DS::write($self, \($_[1]));
570 my ($sock, $ev) = @_;
571 epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and
572 confess("EPOLL_CTL_MOD $!");
575 # return true if complete, false if incomplete (or failure)
576 sub accept_tls_step ($) {
578 my $sock = $self->{sock} or return;
579 return 1 if $sock->accept_SSL;
580 return $self->close if $! != EAGAIN;
581 my $ev = PublicInbox::TLS::epollbit() or return $self->close;
582 epwait($sock, $ev | EPOLLONESHOT);
583 unshift(@{$self->{wbuf}}, \&accept_tls_step); # autovivifies
587 # return true if complete, false if incomplete (or failure)
588 sub shutdn_tls_step ($) {
590 my $sock = $self->{sock} or return;
591 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
592 return $self->close if $! != EAGAIN;
593 my $ev = PublicInbox::TLS::epollbit() or return $self->close;
594 epwait($sock, $ev | EPOLLONESHOT);
595 unshift(@{$self->{wbuf}}, \&shutdn_tls_step); # autovivifies
599 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
600 # or fork w/o exec, so no inadvertent socket sharing
603 my $sock = $self->{sock} or return;
604 if ($sock->can('stop_SSL')) {
605 shutdn_tls_step($self);
611 # must be called with eval, PublicInbox::DS may not be loaded (see t/qspawn.t)
613 die "Not in EventLoop\n" unless $in_loop;
614 push @$wait_pids, [ @_ ]; # [ $pid, $cb, $arg ]
616 # We could've just missed our SIGCHLD, cover it, here:
621 my $run = $later_queue or return;
622 $later_timer = $later_queue = undef;
627 push @$later_queue, $_[0]; # autovivifies @$later_queue
628 $later_timer //= add_timer(60, \&_run_later);
634 my $old = $now - $exp;
636 while (my ($fd, $idle_at) = each %$EXPMAP) {
637 if ($idle_at < $old) {
638 my $ds_obj = $DescriptorMap{$fd};
639 $new{$fd} = $idle_at if !$ds_obj->shutdn;
641 $new{$fd} = $idle_at;
645 $exp_timer = scalar(keys %new) ? later(\&expire_old) : undef;
648 sub update_idle_time {
650 my $sock = $self->{sock} or return;
651 $EXPMAP->{fileno($sock)} = now();
652 $exp_timer //= later(\&expire_old);
656 my ($self, $now) = @_;
657 my $sock = $self->{sock} or return;
658 my $idle_at = $EXPMAP->{fileno($sock)} or return;
659 ($idle_at + $EXPTIME) > $now;
664 =head1 AUTHORS (Danga::Socket)
666 Brad Fitzpatrick <brad@danga.com> - author
668 Michael Granger <ged@danga.com> - docs, testing
670 Mark Smith <junior@danga.com> - contributor, heavy user, testing
672 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits