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(croak confess carp);
42 my $nextq; # queue for next_tick
43 my $WaitPids; # list of [ pid, callback, callback_arg ]
44 my $later_queue; # callbacks
45 my $EXPMAP; # fd -> [ idle_time, $self ]
46 our $EXPTIME = 180; # 3 minutes
47 my ($later_timer, $reap_timer, $exp_timer);
49 %DescriptorMap, # fd (num) -> PublicInbox::DS object
50 $Epoll, # Global epoll fd (or DSKQXS ref)
51 $_io, # IO::Handle for Epoll
52 @ToClose, # sockets to close when event loop is done
54 $PostLoopCallback, # subref to call at the end of each loop, if defined (global)
56 $LoopTimeout, # timeout of event loop in milliseconds
57 $DoneInit, # if we've done the one-time module init yet
64 #####################################################################
65 ### C L A S S M E T H O D S
66 #####################################################################
68 =head2 C<< CLASS->Reset() >>
79 $reap_timer = $later_timer = $exp_timer = undef;
81 $LoopTimeout = -1; # no timeout by default
84 $PostLoopCallback = undef;
87 $_io = undef; # closes real $Epoll FD
88 $Epoll = undef; # may call DSKQXS::DESTROY
90 *EventLoop = *FirstTimeEventLoop;
93 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
95 Set the loop timeout for the event loop to some value in milliseconds.
97 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
102 return $LoopTimeout = $_[1] + 0;
105 =head2 C<< CLASS->AddTimer( $seconds, $coderef ) >>
107 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
108 are not guaranteed to fire at the exact time you ask for.
110 Returns a timer object which you can call C<< $timer->cancel >> on if you need to.
114 my ($class, $secs, $coderef) = @_;
116 my $fire_time = now() + $secs;
118 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
120 if (!@Timers || $fire_time >= $Timers[-1][0]) {
121 push @Timers, $timer;
125 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
126 # but it was compared against calendar queues, heaps, naive push/sort,
127 # and a bunch of other versions, and found to be fastest with a large
128 # variety of datasets.)
129 for (my $i = 0; $i < @Timers; $i++) {
130 if ($Timers[$i][0] > $fire_time) {
131 splice(@Timers, $i, 0, $timer);
136 die "Shouldn't get here.";
139 # keeping this around in case we support other FD types for now,
140 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
141 sub set_cloexec ($) {
144 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
145 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
146 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
154 if (PublicInbox::Syscall::epoll_defined()) {
155 $Epoll = epoll_create();
156 set_cloexec($Epoll) if (defined($Epoll) && $Epoll >= 0);
159 for (qw(DSKQXS DSPoll)) {
160 $cls = "PublicInbox::$_";
161 last if eval "require $cls";
163 $cls->import(qw(epoll_ctl epoll_wait));
166 *EventLoop = *EpollEventLoop;
169 =head2 C<< CLASS->EventLoop() >>
171 Start processing IO events. In most daemon programs this never exits. See
172 C<PostLoopCallback> below for how to exit the loop.
175 sub FirstTimeEventLoop {
183 sub now () { clock_gettime(CLOCK_MONOTONIC) }
189 # we avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak:
190 # https://rt.perl.org/Public/Bug/Display.html?id=114340
199 # runs timers and returns milliseconds for next one, or next event loop
203 return ((@$nextq || @ToClose) ? 0 : $LoopTimeout) unless @Timers;
208 while (@Timers && $Timers[0][0] <= $now) {
209 my $to_run = shift(@Timers);
210 $to_run->[1]->($now) if $to_run->[1];
213 # timers may enqueue into nextq:
214 return 0 if (@$nextq || @ToClose);
216 return $LoopTimeout unless @Timers;
218 # convert time to an even number of milliseconds, adding 1
219 # extra, otherwise floating point fun can occur and we'll
220 # call RunTimers like 20-30 times, each returning a timeout
221 # of 0.0000212 seconds
222 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
224 # -1 is an infinite timeout, so prefer a real timeout
225 return $timeout if $LoopTimeout == -1;
227 # otherwise pick the lower of our regular timeout and time until
229 return $LoopTimeout if $LoopTimeout < $timeout;
233 # We can't use waitpid(-1) safely here since it can hit ``, system(),
234 # and other things. So we scan the $WaitPids list, which is hopefully
240 foreach my $ary (@$tmp) {
241 my ($pid, $cb, $arg) = @$ary;
242 my $ret = waitpid($pid, WNOHANG);
244 push @$WaitPids, $ary;
246 eval { $cb->($arg, $pid) };
250 # we may not be donea, and we may miss our
251 $reap_timer = AddTimer(undef, 1, \&reap_pids);
255 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
256 sub enqueue_reap ($) { push @$nextq, \&reap_pids };
263 my $timeout = RunTimers();
265 # get up to 1000 events
266 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
267 for ($i=0; $i<$evcount; $i++) {
268 # it's possible epoll_wait returned many events, including some at the end
269 # that ones in the front triggered unregister-interest actions. if we
270 # can't find the %sock entry, it's because we're no longer interested
272 $DescriptorMap{$events[$i]->[0]}->event_step;
274 } while (PostEventLoop());
278 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
280 Sets post loop callback function. Pass a subref and it will be
281 called every time the event loop finishes.
283 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
286 The callback function will be passed two parameters: \%DescriptorMap
289 sub SetPostLoopCallback {
290 my ($class, $ref) = @_;
293 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
296 # Internal function: run the post-event callback, send read events
297 # for pushed-back data, and close pending connections. returns 1
298 # if event loop should continue, or 0 to shut it all down.
300 # now we can close sockets that wanted to close during our event processing.
301 # (we didn't want to close them during the loop, as we didn't want fd numbers
302 # being reused and confused during the event loop)
303 delete($DescriptorMap{fileno($_)}) for @ToClose;
304 @ToClose = (); # let refcounting drop everything all at once
306 # by default we keep running, unless a postloop callback (either per-object
307 # or global) cancels it
308 my $keep_running = 1;
310 # now we're at the very end, call callback if defined
311 if (defined $PostLoopCallback) {
312 $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
315 return $keep_running;
318 #####################################################################
319 ### PublicInbox::DS-the-object code
320 #####################################################################
322 =head2 OBJECT METHODS
324 =head2 C<< CLASS->new( $socket ) >>
326 Create a new PublicInbox::DS subclass object for the given I<socket> which will
327 react to events on it during the C<EventLoop>.
329 This is normally (always?) called from your subclass via:
331 $class->SUPER::new($socket);
335 my ($self, $sock, $ev) = @_;
336 $self = fields::new($self) unless ref $self;
338 $self->{sock} = $sock;
339 my $fd = fileno($sock);
341 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
346 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
347 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
348 $ev &= ~EPOLLEXCLUSIVE;
351 die "couldn't add epoll watch for $fd: $!\n";
353 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
354 if $DescriptorMap{$fd};
356 $DescriptorMap{$fd} = $self;
361 #####################################################################
362 ### I N S T A N C E M E T H O D S
363 #####################################################################
365 sub requeue ($) { push @$nextq, $_[0] }
367 =head2 C<< $obj->close >>
374 my $sock = delete $self->{sock} or return;
376 # we need to flush our write buffer, as there may
377 # be self-referential closures (sub { $client->close })
378 # preventing the object from being destroyed
379 delete $self->{wbuf};
381 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
382 # notifications about it
383 my $fd = fileno($sock);
384 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
385 confess("EPOLL_CTL_DEL: $!");
387 # we explicitly don't delete from DescriptorMap here until we
388 # actually close the socket, as we might be in the middle of
389 # processing an epoll_wait/etc that returned hundreds of fds, one
390 # of which is not yet processed and is what we're closing. if we
391 # keep it in DescriptorMap, then the event harnesses can just
392 # looked at $pob->{sock} == undef and ignore it. but if it's an
393 # un-accounted for fd, then it (understandably) freak out a bit
394 # and emit warnings, thinking their state got off.
396 # defer closing the actual socket until the event loop is done
397 # processing this round of events. (otherwise we might reuse fds)
398 push @ToClose, $sock;
403 # portable, non-thread-safe sendfile emulation (no pread, yet)
404 sub psendfile ($$$) {
405 my ($sock, $fh, $off) = @_;
407 seek($fh, $$off, SEEK_SET) or return;
408 defined(my $to_write = read($fh, my $buf, 16384)) or return;
410 while ($to_write > 0) {
411 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
415 return if $written == 0;
423 sub epbit ($$) { # (sock, default)
424 ref($_[0]) eq 'IO::Socket::SSL' ? PublicInbox::TLS::epollbit() : $_[1];
427 # returns 1 if done, 0 if incomplete
428 sub flush_write ($) {
430 my $wbuf = $self->{wbuf} or return 1;
431 my $sock = $self->{sock};
434 while (my $bref = $wbuf->[0]) {
435 if (ref($bref) ne 'CODE') {
436 my $off = delete($self->{wbuf_off}) // 0;
438 my $w = psendfile($sock, $bref, \$off);
444 } elsif ($! == EAGAIN) {
445 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
446 $self->{wbuf_off} = $off;
452 } else { #($ref eq 'CODE') {
454 my $before = scalar(@$wbuf);
457 # bref may be enqueueing more CODE to call (see accept_tls_step)
458 return 0 if (scalar(@$wbuf) > $before);
462 delete $self->{wbuf};
467 my ($self, $rbuf) = @_;
468 if ($$rbuf eq '') { # who knows how long till we can read again
469 delete $self->{rbuf};
471 $self->{rbuf} = $rbuf;
475 sub do_read ($$$;$) {
476 my ($self, $rbuf, $len, $off) = @_;
477 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
478 return ($r == 0 ? $self->close : $r) if defined $r;
479 # common for clients to break connections without warning,
480 # would be too noisy to log here:
482 epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT);
483 rbuf_idle($self, $rbuf);
490 # drop the socket if we hit unrecoverable errors on our system which
491 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
498 # n.b.: use ->write/->read for this buffer to allow compatibility with
499 # PerlIO::mmap or PerlIO::scalar if needed
501 my ($self, $bref, $off) = @_;
502 my $fh = tmpfile('wbuf', $self->{sock}, 1) or
503 return drop($self, "tmpfile $!");
505 my $len = bytes::length($$bref) - $off;
506 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
510 =head2 C<< $obj->write( $data ) >>
512 Write the specified data to the underlying handle. I<data> may be scalar,
513 scalar ref, code ref (to run when there).
514 Returns 1 if writes all went through, or 0 if there are writes in queue. If
515 it returns 1, caller should stop waiting for 'writable' events)
519 my ($self, $data) = @_;
521 # nobody should be writing to closed sockets, but caller code can
522 # do two writes within an event, have the first fail and
523 # disconnect the other side (whose destructor then closes the
524 # calling object, but it's still in a method), and then the
525 # now-dead object does its second write. that is this case. we
526 # just lie and say it worked. it'll be dead soon and won't be
528 my $sock = $self->{sock} or return 1;
530 my $bref = $ref ? $data : \$data;
531 my $wbuf = $self->{wbuf};
532 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
533 if ($ref eq 'CODE') {
536 my $last = $wbuf->[-1];
537 if (ref($last) eq 'GLOB') { # append to tmp file buffer
538 $last->print($$bref) or return drop($self, "print: $!");
540 my $tmpio = tmpio($self, $bref, 0) or return 0;
545 } elsif ($ref eq 'CODE') {
549 my $to_write = bytes::length($$bref);
550 my $written = syswrite($sock, $$bref, $to_write);
552 if (defined $written) {
553 return 1 if $written == $to_write;
554 requeue($self); # runs: event_step -> flush_write
555 } elsif ($! == EAGAIN) {
556 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
562 # deal with EAGAIN or partial write:
563 my $tmpio = tmpio($self, $bref, $written) or return 0;
565 # wbuf may be an empty array if we're being called inside
566 # ->flush_write via CODE bref:
567 push @{$self->{wbuf} ||= []}, $tmpio;
572 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
576 my $sock = $self->{sock} or return 1;
577 my $wbuf = $self->{wbuf};
579 if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) &&
580 ref($sock) ne 'IO::Socket::SSL') {
581 my $n = send($sock, $_[1], MSG_MORE);
583 my $nlen = bytes::length($_[1]) - $n;
584 return 1 if $nlen == 0; # all done!
585 # queue up the unwritten substring:
586 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
587 $self->{wbuf} //= $wbuf //= [];
589 epwait($sock, EPOLLOUT|EPOLLONESHOT);
594 # don't redispatch into NNTPdeflate::write
595 PublicInbox::DS::write($self, \($_[1]));
599 my ($sock, $ev) = @_;
600 epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and
601 confess("EPOLL_CTL_MOD $!");
604 # return true if complete, false if incomplete (or failure)
605 sub accept_tls_step ($) {
607 my $sock = $self->{sock} or return;
608 return 1 if $sock->accept_SSL;
609 return $self->close if $! != EAGAIN;
610 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
611 unshift @{$self->{wbuf} ||= []}, \&accept_tls_step;
615 # return true if complete, false if incomplete (or failure)
616 sub shutdn_tls_step ($) {
618 my $sock = $self->{sock} or return;
619 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
620 return $self->close if $! != EAGAIN;
621 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
622 unshift @{$self->{wbuf} ||= []}, \&shutdn_tls_step;
626 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
627 # or fork w/o exec, so no inadvertant socket sharing
630 my $sock = $self->{sock} or return;
631 if (ref($sock) eq 'IO::Socket::SSL') {
632 shutdn_tls_step($self);
638 # must be called with eval, PublicInbox::DS may not be loaded (see t/qspawn.t)
640 my ($pid, $cb, $arg) = @_;
642 push @$WaitPids, [ $pid, $cb, $arg ];
644 # We could've just missed our SIGCHLD, cover it, here:
645 requeue(\&reap_pids);
647 die "Not in EventLoop\n";
652 my $run = $later_queue;
653 $later_timer = undef;
660 push @$later_queue, $cb;
661 $later_timer //= AddTimer(undef, 60, \&_run_later);
667 my $old = $now - $exp;
669 while (my ($fd, $v) = each %$EXPMAP) {
670 my ($idle_time, $ds_obj) = @$v;
671 if ($idle_time < $old) {
672 if (!$ds_obj->shutdn) {
680 $exp_timer = scalar(keys %new) ? later(\&expire_old) : undef;
683 sub update_idle_time {
685 my $sock = $self->{sock} or return;
686 $EXPMAP->{fileno($sock)} = [ now(), $self ];
687 $exp_timer //= later(\&expire_old);
691 my ($self, $now) = @_;
692 my $sock = $self->{sock} or return;
693 my $ary = $EXPMAP->{fileno($sock)} or return;
694 my $exp_at = $ary->[0] + $EXPTIME;
698 package PublicInbox::DS::Timer;
699 # [$abs_float_firetime, $coderef];
706 =head1 AUTHORS (Danga::Socket)
708 Brad Fitzpatrick <brad@danga.com> - author
710 Michael Granger <ged@danga.com> - docs, testing
712 Mark Smith <junior@danga.com> - contributor, heavy user, testing
714 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits