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 (for now) unmaintained Danga::Socket 1.61.
7 # Unused features will be removed, and updates will be made to take
8 # advantage of newer kernels.
10 # API changes to diverge from Danga::Socket will happen to better
11 # accomodate new features and improve scalability. Do not expect
12 # this to be a stable API like Danga::Socket.
13 # Bugs encountered (and likely fixed) are reported to
14 # bug-Danga-Socket@rt.cpan.org 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<< CLASS->AddTimer( $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.
109 Returns a timer object which you can call C<< $timer->cancel >> on if you need to.
113 my ($class, $secs, $coderef) = @_;
115 my $fire_time = now() + $secs;
117 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
119 if (!@Timers || $fire_time >= $Timers[-1][0]) {
120 push @Timers, $timer;
124 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
125 # but it was compared against calendar queues, heaps, naive push/sort,
126 # and a bunch of other versions, and found to be fastest with a large
127 # variety of datasets.)
128 for (my $i = 0; $i < @Timers; $i++) {
129 if ($Timers[$i][0] > $fire_time) {
130 splice(@Timers, $i, 0, $timer);
135 die "Shouldn't get here.";
138 # keeping this around in case we support other FD types for now,
139 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
140 sub set_cloexec ($) {
143 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
144 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
145 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
153 if (PublicInbox::Syscall::epoll_defined()) {
154 $Epoll = epoll_create();
155 set_cloexec($Epoll) if (defined($Epoll) && $Epoll >= 0);
158 for (qw(DSKQXS DSPoll)) {
159 $cls = "PublicInbox::$_";
160 last if eval "require $cls";
162 $cls->import(qw(epoll_ctl epoll_wait));
165 *EventLoop = *EpollEventLoop;
168 =head2 C<< CLASS->EventLoop() >>
170 Start processing IO events. In most daemon programs this never exits. See
171 C<PostLoopCallback> below for how to exit the loop.
174 sub FirstTimeEventLoop {
182 sub now () { clock_gettime(CLOCK_MONOTONIC) }
188 # we avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak:
189 # https://rt.perl.org/Public/Bug/Display.html?id=114340
198 # runs timers and returns milliseconds for next one, or next event loop
202 return ((@$nextq || @ToClose) ? 0 : $LoopTimeout) unless @Timers;
207 while (@Timers && $Timers[0][0] <= $now) {
208 my $to_run = shift(@Timers);
209 $to_run->[1]->($now) if $to_run->[1];
212 # timers may enqueue into nextq:
213 return 0 if (@$nextq || @ToClose);
215 return $LoopTimeout unless @Timers;
217 # convert time to an even number of milliseconds, adding 1
218 # extra, otherwise floating point fun can occur and we'll
219 # call RunTimers like 20-30 times, each returning a timeout
220 # of 0.0000212 seconds
221 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
223 # -1 is an infinite timeout, so prefer a real timeout
224 return $timeout if $LoopTimeout == -1;
226 # otherwise pick the lower of our regular timeout and time until
228 return $LoopTimeout if $LoopTimeout < $timeout;
232 # We can't use waitpid(-1) safely here since it can hit ``, system(),
233 # and other things. So we scan the $WaitPids list, which is hopefully
239 foreach my $ary (@$tmp) {
240 my ($pid, $cb, $arg) = @$ary;
241 my $ret = waitpid($pid, WNOHANG);
243 push @$WaitPids, $ary;
245 eval { $cb->($arg, $pid) };
249 # we may not be donea, and we may miss our
250 $reap_timer = AddTimer(undef, 1, \&reap_pids);
254 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
255 sub enqueue_reap ($) { push @$nextq, \&reap_pids };
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);
340 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
345 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
346 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
347 $ev &= ~EPOLLEXCLUSIVE;
350 die "couldn't add epoll watch for $fd: $!\n";
352 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
353 if $DescriptorMap{$fd};
355 $DescriptorMap{$fd} = $self;
360 #####################################################################
361 ### I N S T A N C E M E T H O D S
362 #####################################################################
364 sub requeue ($) { push @$nextq, $_[0] }
366 =head2 C<< $obj->close >>
373 my $sock = delete $self->{sock} or return;
375 # we need to flush our write buffer, as there may
376 # be self-referential closures (sub { $client->close })
377 # preventing the object from being destroyed
378 delete $self->{wbuf};
380 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
381 # notifications about it
382 my $fd = fileno($sock);
383 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
384 confess("EPOLL_CTL_DEL: $!");
386 # we explicitly don't delete from DescriptorMap here until we
387 # actually close the socket, as we might be in the middle of
388 # processing an epoll_wait/etc that returned hundreds of fds, one
389 # of which is not yet processed and is what we're closing. if we
390 # keep it in DescriptorMap, then the event harnesses can just
391 # looked at $pob->{sock} == undef and ignore it. but if it's an
392 # un-accounted for fd, then it (understandably) freak out a bit
393 # and emit warnings, thinking their state got off.
395 # defer closing the actual socket until the event loop is done
396 # processing this round of events. (otherwise we might reuse fds)
397 push @ToClose, $sock;
402 # portable, non-thread-safe sendfile emulation (no pread, yet)
403 sub psendfile ($$$) {
404 my ($sock, $fh, $off) = @_;
406 seek($fh, $$off, SEEK_SET) or return;
407 defined(my $to_write = read($fh, my $buf, 16384)) or return;
409 while ($to_write > 0) {
410 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
414 return if $written == 0;
422 sub epbit ($$) { # (sock, default)
423 ref($_[0]) eq 'IO::Socket::SSL' ? PublicInbox::TLS::epollbit() : $_[1];
426 # returns 1 if done, 0 if incomplete
427 sub flush_write ($) {
429 my $wbuf = $self->{wbuf} or return 1;
430 my $sock = $self->{sock};
433 while (my $bref = $wbuf->[0]) {
434 if (ref($bref) ne 'CODE') {
435 my $off = delete($self->{wbuf_off}) // 0;
437 my $w = psendfile($sock, $bref, \$off);
443 } elsif ($! == EAGAIN) {
444 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
445 $self->{wbuf_off} = $off;
451 } else { #($ref eq 'CODE') {
453 my $before = scalar(@$wbuf);
456 # bref may be enqueueing more CODE to call (see accept_tls_step)
457 return 0 if (scalar(@$wbuf) > $before);
461 delete $self->{wbuf};
466 my ($self, $rbuf) = @_;
467 if ($$rbuf eq '') { # who knows how long till we can read again
468 delete $self->{rbuf};
470 $self->{rbuf} = $rbuf;
474 sub do_read ($$$;$) {
475 my ($self, $rbuf, $len, $off) = @_;
476 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
477 return ($r == 0 ? $self->close : $r) if defined $r;
478 # common for clients to break connections without warning,
479 # would be too noisy to log here:
481 epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT);
482 rbuf_idle($self, $rbuf);
489 # drop the socket if we hit unrecoverable errors on our system which
490 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
497 # n.b.: use ->write/->read for this buffer to allow compatibility with
498 # PerlIO::mmap or PerlIO::scalar if needed
500 my ($self, $bref, $off) = @_;
501 my $fh = tmpfile('wbuf', $self->{sock}, 1) or
502 return drop($self, "tmpfile $!");
504 my $len = bytes::length($$bref) - $off;
505 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
509 =head2 C<< $obj->write( $data ) >>
511 Write the specified data to the underlying handle. I<data> may be scalar,
512 scalar ref, code ref (to run when there).
513 Returns 1 if writes all went through, or 0 if there are writes in queue. If
514 it returns 1, caller should stop waiting for 'writable' events)
518 my ($self, $data) = @_;
520 # nobody should be writing to closed sockets, but caller code can
521 # do two writes within an event, have the first fail and
522 # disconnect the other side (whose destructor then closes the
523 # calling object, but it's still in a method), and then the
524 # now-dead object does its second write. that is this case. we
525 # just lie and say it worked. it'll be dead soon and won't be
527 my $sock = $self->{sock} or return 1;
529 my $bref = $ref ? $data : \$data;
530 my $wbuf = $self->{wbuf};
531 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
532 if ($ref eq 'CODE') {
535 my $last = $wbuf->[-1];
536 if (ref($last) eq 'GLOB') { # append to tmp file buffer
537 $last->print($$bref) or return drop($self, "print: $!");
539 my $tmpio = tmpio($self, $bref, 0) or return 0;
544 } elsif ($ref eq 'CODE') {
548 my $to_write = bytes::length($$bref);
549 my $written = syswrite($sock, $$bref, $to_write);
551 if (defined $written) {
552 return 1 if $written == $to_write;
553 requeue($self); # runs: event_step -> flush_write
554 } elsif ($! == EAGAIN) {
555 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
561 # deal with EAGAIN or partial write:
562 my $tmpio = tmpio($self, $bref, $written) or return 0;
564 # wbuf may be an empty array if we're being called inside
565 # ->flush_write via CODE bref:
566 push @{$self->{wbuf} ||= []}, $tmpio;
571 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
575 my $sock = $self->{sock} or return 1;
576 my $wbuf = $self->{wbuf};
578 if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) &&
579 ref($sock) ne 'IO::Socket::SSL') {
580 my $n = send($sock, $_[1], MSG_MORE);
582 my $nlen = bytes::length($_[1]) - $n;
583 return 1 if $nlen == 0; # all done!
584 # queue up the unwritten substring:
585 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
586 $self->{wbuf} //= $wbuf //= [];
588 epwait($sock, EPOLLOUT|EPOLLONESHOT);
593 # don't redispatch into NNTPdeflate::write
594 PublicInbox::DS::write($self, \($_[1]));
598 my ($sock, $ev) = @_;
599 epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and
600 confess("EPOLL_CTL_MOD $!");
603 # return true if complete, false if incomplete (or failure)
604 sub accept_tls_step ($) {
606 my $sock = $self->{sock} or return;
607 return 1 if $sock->accept_SSL;
608 return $self->close if $! != EAGAIN;
609 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
610 unshift @{$self->{wbuf} ||= []}, \&accept_tls_step;
614 # return true if complete, false if incomplete (or failure)
615 sub shutdn_tls_step ($) {
617 my $sock = $self->{sock} or return;
618 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
619 return $self->close if $! != EAGAIN;
620 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
621 unshift @{$self->{wbuf} ||= []}, \&shutdn_tls_step;
625 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
626 # or fork w/o exec, so no inadvertant socket sharing
629 my $sock = $self->{sock} or return;
630 if (ref($sock) eq 'IO::Socket::SSL') {
631 shutdn_tls_step($self);
637 # must be called with eval, PublicInbox::DS may not be loaded (see t/qspawn.t)
639 my ($pid, $cb, $arg) = @_;
641 push @$WaitPids, [ $pid, $cb, $arg ];
643 # We could've just missed our SIGCHLD, cover it, here:
644 requeue(\&reap_pids);
646 die "Not in EventLoop\n";
651 my $run = $later_queue;
652 $later_timer = undef;
659 push @$later_queue, $cb;
660 $later_timer //= AddTimer(undef, 60, \&_run_later);
666 my $old = $now - $exp;
668 while (my ($fd, $v) = each %$EXPMAP) {
669 my ($idle_time, $ds_obj) = @$v;
670 if ($idle_time < $old) {
671 if (!$ds_obj->shutdn) {
679 $exp_timer = scalar(keys %new) ? later(\&expire_old) : undef;
682 sub update_idle_time {
684 my $sock = $self->{sock} or return;
685 $EXPMAP->{fileno($sock)} = [ now(), $self ];
686 $exp_timer //= later(\&expire_old);
690 my ($self, $now) = @_;
691 my $sock = $self->{sock} or return;
692 my $ary = $EXPMAP->{fileno($sock)} or return;
693 my $exp_at = $ary->[0] + $EXPTIME;
697 package PublicInbox::DS::Timer;
698 # [$abs_float_firetime, $coderef];
705 =head1 AUTHORS (Danga::Socket)
707 Brad Fitzpatrick <brad@danga.com> - author
709 Michael Granger <ged@danga.com> - docs, testing
711 Mark Smith <junior@danga.com> - contributor, heavy user, testing
713 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits