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 ]
46 %DescriptorMap, # fd (num) -> PublicInbox::DS object
47 $Epoll, # Global epoll fd (or DSKQXS ref)
48 $_io, # IO::Handle for Epoll
49 @ToClose, # sockets to close when event loop is done
51 $PostLoopCallback, # subref to call at the end of each loop, if defined (global)
53 $LoopTimeout, # timeout of event loop in milliseconds
54 $DoneInit, # if we've done the one-time module init yet
61 #####################################################################
62 ### C L A S S M E T H O D S
63 #####################################################################
65 =head2 C<< CLASS->Reset() >>
75 $LoopTimeout = -1; # no timeout by default
78 $PostLoopCallback = undef;
81 $_io = undef; # closes real $Epoll FD
82 $Epoll = undef; # may call DSKQXS::DESTROY
84 *EventLoop = *FirstTimeEventLoop;
87 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
89 Set the loop timeout for the event loop to some value in milliseconds.
91 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
96 return $LoopTimeout = $_[1] + 0;
99 =head2 C<< CLASS->AddTimer( $seconds, $coderef ) >>
101 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
102 are not guaranteed to fire at the exact time you ask for.
104 Returns a timer object which you can call C<< $timer->cancel >> on if you need to.
108 my ($class, $secs, $coderef) = @_;
110 my $fire_time = now() + $secs;
112 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
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 $cls->import(qw(epoll_ctl epoll_wait));
160 *EventLoop = *EpollEventLoop;
163 =head2 C<< CLASS->EventLoop() >>
165 Start processing IO events. In most daemon programs this never exits. See
166 C<PostLoopCallback> below for how to exit the loop.
169 sub FirstTimeEventLoop {
177 sub now () { clock_gettime(CLOCK_MONOTONIC) }
183 # we avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak:
184 # https://rt.perl.org/Public/Bug/Display.html?id=114340
193 # runs timers and returns milliseconds for next one, or next event loop
197 return ((@$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 $WaitPids list, which is hopefully
234 foreach my $ary (@$tmp) {
235 my ($pid, $cb, $arg) = @$ary;
236 my $ret = waitpid($pid, WNOHANG);
238 push @$WaitPids, $ary;
240 eval { $cb->($arg, $pid) };
244 # we may not be donea, and we may miss our
245 $reap_timer = AddTimer(undef, 1, \&reap_pids);
249 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
250 sub enqueue_reap ($) { push @$nextq, \&reap_pids };
257 my $timeout = RunTimers();
259 # get up to 1000 events
260 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
261 for ($i=0; $i<$evcount; $i++) {
262 # it's possible epoll_wait returned many events, including some at the end
263 # that ones in the front triggered unregister-interest actions. if we
264 # can't find the %sock entry, it's because we're no longer interested
266 $DescriptorMap{$events[$i]->[0]}->event_step;
268 return unless PostEventLoop();
272 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
274 Sets post loop callback function. Pass a subref and it will be
275 called every time the event loop finishes.
277 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
280 The callback function will be passed two parameters: \%DescriptorMap
283 sub SetPostLoopCallback {
284 my ($class, $ref) = @_;
287 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
290 # Internal function: run the post-event callback, send read events
291 # for pushed-back data, and close pending connections. returns 1
292 # if event loop should continue, or 0 to shut it all down.
294 # now we can close sockets that wanted to close during our event processing.
295 # (we didn't want to close them during the loop, as we didn't want fd numbers
296 # being reused and confused during the event loop)
297 delete($DescriptorMap{fileno($_)}) for @ToClose;
298 @ToClose = (); # let refcounting drop everything all at once
300 # by default we keep running, unless a postloop callback (either per-object
301 # or global) cancels it
302 my $keep_running = 1;
304 # now we're at the very end, call callback if defined
305 if (defined $PostLoopCallback) {
306 $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
309 return $keep_running;
312 #####################################################################
313 ### PublicInbox::DS-the-object code
314 #####################################################################
316 =head2 OBJECT METHODS
318 =head2 C<< CLASS->new( $socket ) >>
320 Create a new PublicInbox::DS subclass object for the given I<socket> which will
321 react to events on it during the C<EventLoop>.
323 This is normally (always?) called from your subclass via:
325 $class->SUPER::new($socket);
329 my ($self, $sock, $ev) = @_;
330 $self = fields::new($self) unless ref $self;
332 $self->{sock} = $sock;
333 my $fd = fileno($sock);
335 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
340 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
341 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
342 $ev &= ~EPOLLEXCLUSIVE;
345 die "couldn't add epoll watch for $fd: $!\n";
347 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
348 if $DescriptorMap{$fd};
350 $DescriptorMap{$fd} = $self;
355 #####################################################################
356 ### I N S T A N C E M E T H O D S
357 #####################################################################
359 sub requeue ($) { push @$nextq, $_[0] }
361 =head2 C<< $obj->close >>
368 my $sock = delete $self->{sock} or return;
370 # we need to flush our write buffer, as there may
371 # be self-referential closures (sub { $client->close })
372 # preventing the object from being destroyed
373 delete $self->{wbuf};
375 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
376 # notifications about it
377 my $fd = fileno($sock);
378 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
379 confess("EPOLL_CTL_DEL: $!");
381 # we explicitly don't delete from DescriptorMap here until we
382 # actually close the socket, as we might be in the middle of
383 # processing an epoll_wait/etc that returned hundreds of fds, one
384 # of which is not yet processed and is what we're closing. if we
385 # keep it in DescriptorMap, then the event harnesses can just
386 # looked at $pob->{sock} == undef and ignore it. but if it's an
387 # un-accounted for fd, then it (understandably) freak out a bit
388 # and emit warnings, thinking their state got off.
390 # defer closing the actual socket until the event loop is done
391 # processing this round of events. (otherwise we might reuse fds)
392 push @ToClose, $sock;
397 # portable, non-thread-safe sendfile emulation (no pread, yet)
398 sub psendfile ($$$) {
399 my ($sock, $fh, $off) = @_;
401 seek($fh, $$off, SEEK_SET) or return;
402 defined(my $to_write = read($fh, my $buf, 16384)) or return;
404 while ($to_write > 0) {
405 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
409 return if $written == 0;
417 sub epbit ($$) { # (sock, default)
418 ref($_[0]) eq 'IO::Socket::SSL' ? PublicInbox::TLS::epollbit() : $_[1];
421 # returns 1 if done, 0 if incomplete
422 sub flush_write ($) {
424 my $wbuf = $self->{wbuf} or return 1;
425 my $sock = $self->{sock};
428 while (my $bref = $wbuf->[0]) {
429 if (ref($bref) ne 'CODE') {
430 my $off = delete($self->{wbuf_off}) // 0;
432 my $w = psendfile($sock, $bref, \$off);
438 } elsif ($! == EAGAIN) {
439 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
440 $self->{wbuf_off} = $off;
446 } else { #($ref eq 'CODE') {
448 my $before = scalar(@$wbuf);
451 # bref may be enqueueing more CODE to call (see accept_tls_step)
452 return 0 if (scalar(@$wbuf) > $before);
456 delete $self->{wbuf};
461 my ($self, $rbuf) = @_;
462 if ($$rbuf eq '') { # who knows how long till we can read again
463 delete $self->{rbuf};
465 $self->{rbuf} = $rbuf;
469 sub do_read ($$$;$) {
470 my ($self, $rbuf, $len, $off) = @_;
471 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
472 return ($r == 0 ? $self->close : $r) if defined $r;
473 # common for clients to break connections without warning,
474 # would be too noisy to log here:
476 epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT);
477 rbuf_idle($self, $rbuf);
484 # drop the socket if we hit unrecoverable errors on our system which
485 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
492 # n.b.: use ->write/->read for this buffer to allow compatibility with
493 # PerlIO::mmap or PerlIO::scalar if needed
495 my ($self, $bref, $off) = @_;
496 my $fh = tmpfile('wbuf', $self->{sock}, 1) or
497 return drop($self, "tmpfile $!");
499 my $len = bytes::length($$bref) - $off;
500 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
504 =head2 C<< $obj->write( $data ) >>
506 Write the specified data to the underlying handle. I<data> may be scalar,
507 scalar ref, code ref (to run when there).
508 Returns 1 if writes all went through, or 0 if there are writes in queue. If
509 it returns 1, caller should stop waiting for 'writable' events)
513 my ($self, $data) = @_;
515 # nobody should be writing to closed sockets, but caller code can
516 # do two writes within an event, have the first fail and
517 # disconnect the other side (whose destructor then closes the
518 # calling object, but it's still in a method), and then the
519 # now-dead object does its second write. that is this case. we
520 # just lie and say it worked. it'll be dead soon and won't be
522 my $sock = $self->{sock} or return 1;
524 my $bref = $ref ? $data : \$data;
525 my $wbuf = $self->{wbuf};
526 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
527 if ($ref eq 'CODE') {
530 my $last = $wbuf->[-1];
531 if (ref($last) eq 'GLOB') { # append to tmp file buffer
532 $last->print($$bref) or return drop($self, "print: $!");
534 my $tmpio = tmpio($self, $bref, 0) or return 0;
539 } elsif ($ref eq 'CODE') {
543 my $to_write = bytes::length($$bref);
544 my $written = syswrite($sock, $$bref, $to_write);
546 if (defined $written) {
547 return 1 if $written == $to_write;
548 requeue($self); # runs: event_step -> flush_write
549 } elsif ($! == EAGAIN) {
550 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
556 # deal with EAGAIN or partial write:
557 my $tmpio = tmpio($self, $bref, $written) or return 0;
559 # wbuf may be an empty array if we're being called inside
560 # ->flush_write via CODE bref:
561 push @{$self->{wbuf} ||= []}, $tmpio;
566 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
570 my $sock = $self->{sock} or return 1;
572 if (MSG_MORE && !$self->{wbuf} && ref($sock) ne 'IO::Socket::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 $self->{wbuf} = [ $tmpio ];
580 epwait($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 @{$self->{wbuf} ||= []}, \&accept_tls_step;
606 # return true if complete, false if incomplete (or failure)
607 sub shutdn_tls_step ($) {
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 @{$self->{wbuf} ||= []}, \&shutdn_tls_step;
617 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
618 # or fork w/o exec, so no inadvertant socket sharing
621 my $sock = $self->{sock} or return;
622 if (ref($sock) eq 'IO::Socket::SSL') {
623 shutdn_tls_step($self);
629 # must be called with eval, PublicInbox::DS may not be loaded (see t/qspawn.t)
631 my ($pid, $cb, $arg) = @_;
633 push @$WaitPids, [ $pid, $cb, $arg ];
635 # We could've just missed our SIGCHLD, cover it, here:
636 requeue(\&reap_pids);
638 die "Not in EventLoop\n";
642 package PublicInbox::DS::Timer;
643 # [$abs_float_firetime, $coderef];
650 =head1 AUTHORS (Danga::Socket)
652 Brad Fitzpatrick <brad@danga.com> - author
654 Michael Granger <ged@danga.com> - docs, testing
656 Mark Smith <junior@danga.com> - contributor, heavy user, testing
658 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits