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;
21 use Fcntl qw(FD_CLOEXEC F_SETFD F_GETFD SEEK_SET);
22 use Time::HiRes qw(clock_gettime CLOCK_MONOTONIC);
23 use parent qw(Exporter);
24 our @EXPORT_OK = qw(now msg_more write_in_full);
27 use PublicInbox::Syscall qw(:epoll);
29 use fields ('sock', # underlying socket
30 'wbuf', # arrayref of coderefs or GLOB refs
31 'wbuf_off', # offset into first element of wbuf to start writing at
32 'event_watch', # bitmask of events the client is interested in
36 use Errno qw(EAGAIN EINVAL);
37 use Carp qw(croak confess);
38 use File::Temp qw(tempfile);
40 our $HAVE_KQUEUE = eval { require IO::KQueue; 1 };
43 $HaveEpoll, # Flag -- is epoll available? initially undefined.
45 %DescriptorMap, # fd (num) -> PublicInbox::DS object
46 $Epoll, # Global epoll fd (for epoll mode only)
47 $KQueue, # Global kqueue fd ref (for kqueue mode only)
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
58 # this may be set to zero with old kernels
59 our $EPOLLEXCLUSIVE = EPOLLEXCLUSIVE;
62 #####################################################################
63 ### C L A S S M E T H O D S
64 #####################################################################
66 =head2 C<< CLASS->Reset() >>
74 $LoopTimeout = -1; # no timeout by default
77 $PostLoopCallback = undef;
80 # NOTE kqueue is close-on-fork, and we don't account for it, yet
81 # OTOH, we (public-inbox) don't need this sub outside of tests...
82 POSIX::close($$KQueue) if !$_io && $KQueue && $$KQueue >= 0;
85 $_io = undef; # close $Epoll
88 *EventLoop = *FirstTimeEventLoop;
91 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
93 Set the loop timeout for the event loop to some value in milliseconds.
95 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
100 return $LoopTimeout = $_[1] + 0;
103 =head2 C<< CLASS->AddTimer( $seconds, $coderef ) >>
105 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
106 are not guaranteed to fire at the exact time you ask for.
108 Returns a timer object which you can call C<< $timer->cancel >> on if you need to.
112 my ($class, $secs, $coderef) = @_;
115 my $timer = bless([0, $coderef], 'PublicInbox::DS::Timer');
116 unshift(@Timers, $timer);
120 my $fire_time = now() + $secs;
122 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
124 if (!@Timers || $fire_time >= $Timers[-1][0]) {
125 push @Timers, $timer;
129 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
130 # but it was compared against calendar queues, heaps, naive push/sort,
131 # and a bunch of other versions, and found to be fastest with a large
132 # variety of datasets.)
133 for (my $i = 0; $i < @Timers; $i++) {
134 if ($Timers[$i][0] > $fire_time) {
135 splice(@Timers, $i, 0, $timer);
140 die "Shouldn't get here.";
143 # keeping this around in case we support other FD types for now,
144 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
145 sub set_cloexec ($) {
148 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
149 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
150 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
159 $KQueue = IO::KQueue->new();
160 $HaveKQueue = defined $KQueue;
162 *EventLoop = *KQueueEventLoop;
165 elsif (PublicInbox::Syscall::epoll_defined()) {
166 $Epoll = eval { epoll_create(1024); };
167 $HaveEpoll = defined $Epoll && $Epoll >= 0;
170 *EventLoop = *EpollEventLoop;
174 if (!$HaveEpoll && !$HaveKQueue) {
176 *EventLoop = *PollEventLoop;
180 =head2 C<< CLASS->EventLoop() >>
182 Start processing IO events. In most daemon programs this never exits. See
183 C<PostLoopCallback> below for how to exit the loop.
186 sub FirstTimeEventLoop {
192 EpollEventLoop($class);
193 } elsif ($HaveKQueue) {
194 KQueueEventLoop($class);
196 PollEventLoop($class);
200 sub now () { clock_gettime(CLOCK_MONOTONIC) }
202 # runs timers and returns milliseconds for next one, or next event loop
204 return $LoopTimeout unless @Timers;
209 while (@Timers && $Timers[0][0] <= $now) {
210 my $to_run = shift(@Timers);
211 $to_run->[1]->($now) if $to_run->[1];
214 return $LoopTimeout unless @Timers;
216 # convert time to an even number of milliseconds, adding 1
217 # extra, otherwise floating point fun can occur and we'll
218 # call RunTimers like 20-30 times, each returning a timeout
219 # of 0.0000212 seconds
220 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
222 # -1 is an infinite timeout, so prefer a real timeout
223 return $timeout if $LoopTimeout == -1;
225 # otherwise pick the lower of our regular timeout and time until
227 return $LoopTimeout if $LoopTimeout < $timeout;
231 ### The epoll-based event loop. Gets installed as EventLoop if IO::Epoll loads
239 my $timeout = RunTimers();
241 # get up to 1000 events
242 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
243 for ($i=0; $i<$evcount; $i++) {
244 # it's possible epoll_wait returned many events, including some at the end
245 # that ones in the front triggered unregister-interest actions. if we
246 # can't find the %sock entry, it's because we're no longer interested
248 $DescriptorMap{$events[$i]->[0]}->event_step;
250 return unless PostEventLoop();
255 ### The fallback IO::Poll-based event loop. Gets installed as EventLoop if
256 ### IO::Epoll fails to load.
260 my PublicInbox::DS $pob;
263 my $timeout = RunTimers();
265 # the following sets up @poll as a series of ($poll,$event_mask)
266 # items, then uses IO::Poll::_poll, implemented in XS, which
267 # modifies the array in place with the even elements being
268 # replaced with the event masks that occured.
270 while ( my ($fd, $sock) = each %DescriptorMap ) {
271 push @poll, $fd, $sock->{event_watch};
274 # if nothing to poll, either end immediately (if no timeout)
275 # or just keep calling the callback
277 select undef, undef, undef, ($timeout / 1000);
278 return unless PostEventLoop();
282 my $count = IO::Poll::_poll($timeout, @poll);
283 unless ($count >= 0) {
284 return unless PostEventLoop();
288 # Fetch handles with read events
290 my ($fd, $state) = splice(@poll, 0, 2);
291 $DescriptorMap{$fd}->event_step if $state;
294 return unless PostEventLoop();
300 ### The kqueue-based event loop. Gets installed as EventLoop if IO::KQueue works
302 sub KQueueEventLoop {
306 my $timeout = RunTimers();
307 my @ret = eval { $KQueue->kevent($timeout) };
309 # workaround https://rt.cpan.org/Ticket/Display.html?id=116615
310 if ($err =~ /Interrupted system call/) {
317 foreach my $kev (@ret) {
318 $DescriptorMap{$kev->[0]}->event_step;
320 return unless PostEventLoop();
326 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
328 Sets post loop callback function. Pass a subref and it will be
329 called every time the event loop finishes.
331 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
334 The callback function will be passed two parameters: \%DescriptorMap
337 sub SetPostLoopCallback {
338 my ($class, $ref) = @_;
341 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
344 # Internal function: run the post-event callback, send read events
345 # for pushed-back data, and close pending connections. returns 1
346 # if event loop should continue, or 0 to shut it all down.
348 # now we can close sockets that wanted to close during our event processing.
349 # (we didn't want to close them during the loop, as we didn't want fd numbers
350 # being reused and confused during the event loop)
351 while (my $sock = shift @ToClose) {
352 my $fd = fileno($sock);
354 # close the socket. (not a PublicInbox::DS close)
357 # and now we can finally remove the fd from the map. see
358 # comment above in ->close.
359 delete $DescriptorMap{$fd};
363 # by default we keep running, unless a postloop callback (either per-object
364 # or global) cancels it
365 my $keep_running = 1;
367 # now we're at the very end, call callback if defined
368 if (defined $PostLoopCallback) {
369 $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
372 return $keep_running;
375 #####################################################################
376 ### PublicInbox::DS-the-object code
377 #####################################################################
379 =head2 OBJECT METHODS
381 =head2 C<< CLASS->new( $socket ) >>
383 Create a new PublicInbox::DS subclass object for the given I<socket> which will
384 react to events on it during the C<EventLoop>.
386 This is normally (always?) called from your subclass via:
388 $class->SUPER::new($socket);
392 my ($self, $sock, $exclusive) = @_;
393 $self = fields::new($self) unless ref $self;
395 $self->{sock} = $sock;
396 my $fd = fileno($sock);
398 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
401 my $ev = $self->{event_watch} = 0;
407 $ev = $self->{event_watch} = EPOLLIN|$EPOLLEXCLUSIVE;
410 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
411 if ($! == EINVAL && ($ev & $EPOLLEXCLUSIVE)) {
412 $EPOLLEXCLUSIVE = 0; # old kernel
413 $ev = $self->{event_watch} = EPOLLIN;
416 die "couldn't add epoll watch for $fd: $!\n";
419 elsif ($HaveKQueue) {
420 # Add them to the queue but disabled for now
421 $KQueue->EV_SET($fd, IO::KQueue::EVFILT_READ(),
422 IO::KQueue::EV_ADD() | IO::KQueue::EV_DISABLE());
423 $KQueue->EV_SET($fd, IO::KQueue::EVFILT_WRITE(),
424 IO::KQueue::EV_ADD() | IO::KQueue::EV_DISABLE());
427 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
428 if $DescriptorMap{$fd};
430 $DescriptorMap{$fd} = $self;
435 #####################################################################
436 ### I N S T A N C E M E T H O D S
437 #####################################################################
439 =head2 C<< $obj->close >>
446 my $sock = delete $self->{sock} or return;
448 # we need to flush our write buffer, as there may
449 # be self-referential closures (sub { $client->close })
450 # preventing the object from being destroyed
451 delete $self->{wbuf};
453 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
454 # notifications about it
456 my $fd = fileno($sock);
457 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
458 confess("EPOLL_CTL_DEL: $!");
461 # we explicitly don't delete from DescriptorMap here until we
462 # actually close the socket, as we might be in the middle of
463 # processing an epoll_wait/etc that returned hundreds of fds, one
464 # of which is not yet processed and is what we're closing. if we
465 # keep it in DescriptorMap, then the event harnesses can just
466 # looked at $pob->{sock} == undef and ignore it. but if it's an
467 # un-accounted for fd, then it (understandably) freak out a bit
468 # and emit warnings, thinking their state got off.
470 # defer closing the actual socket until the event loop is done
471 # processing this round of events. (otherwise we might reuse fds)
472 push @ToClose, $sock;
477 # portable, non-thread-safe sendfile emulation (no pread, yet)
478 sub psendfile ($$$) {
479 my ($sock, $fh, $off) = @_;
481 sysseek($fh, $$off, SEEK_SET) or return;
482 defined(my $to_write = sysread($fh, my $buf, 16384)) or return;
484 while ($to_write > 0) {
485 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
489 return if $written == 0;
497 # returns 1 if done, 0 if incomplete
498 sub flush_write ($) {
500 my $wbuf = $self->{wbuf} or return 1;
501 my $sock = $self->{sock} or return 1;
504 while (my $bref = $wbuf->[0]) {
505 if (ref($bref) ne 'CODE') {
506 my $off = delete($self->{wbuf_off}) // 0;
508 my $w = psendfile($sock, $bref, \$off);
514 } elsif ($! == EAGAIN) {
515 $self->{wbuf_off} = $off;
516 watch_write($self, 1);
522 } else { #($ref eq 'CODE') {
528 delete $self->{wbuf};
529 $self->watch_write(0);
533 sub write_in_full ($$$$) {
534 my ($fh, $bref, $len, $off) = @_;
537 my $w = syswrite($fh, $$bref, $len, $off);
538 return ($rv ? $rv : $w) unless $w; # undef or 0
547 my ($bref, $off) = @_;
548 # open(my $fh, '+>>', undef) doesn't set O_APPEND
549 my ($fh, $path) = tempfile('wbuf-XXXXXXX', TMPDIR => 1);
550 open $fh, '+>>', $path or die "open: $!";
552 my $to_write = bytes::length($$bref) - $off;
553 my $w = write_in_full($fh, $bref, $to_write, $off);
554 die "write_in_full ($to_write): $!" unless defined $w;
555 $w == $to_write ? $fh : die("short write $w < $to_write");
558 =head2 C<< $obj->write( $data ) >>
560 Write the specified data to the underlying handle. I<data> may be scalar,
561 scalar ref, code ref (to run when there).
562 Returns 1 if writes all went through, or 0 if there are writes in queue. If
563 it returns 1, caller should stop waiting for 'writable' events)
567 my ($self, $data) = @_;
569 # nobody should be writing to closed sockets, but caller code can
570 # do two writes within an event, have the first fail and
571 # disconnect the other side (whose destructor then closes the
572 # calling object, but it's still in a method), and then the
573 # now-dead object does its second write. that is this case. we
574 # just lie and say it worked. it'll be dead soon and won't be
576 my $sock = $self->{sock} or return 1;
578 my $bref = $ref ? $data : \$data;
579 if (my $wbuf = $self->{wbuf}) { # already buffering, can't write more...
580 if ($ref eq 'CODE') {
583 my $last = $wbuf->[-1];
584 if (ref($last) eq 'GLOB') { # append to tmp file buffer
585 write_in_full($last, $bref, bytes::length($$bref), 0);
587 push @$wbuf, tmpbuf($bref, 0);
591 } elsif ($ref eq 'CODE') {
595 my $to_write = bytes::length($$bref);
596 my $written = syswrite($sock, $$bref, $to_write);
598 if (defined $written) {
599 return 1 if $written == $to_write;
600 } elsif ($! == EAGAIN) {
605 $self->{wbuf} = [ tmpbuf($bref, $written) ];
606 watch_write($self, 1);
611 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
615 my $sock = $self->{sock} or return 1;
617 if (MSG_MORE && !$self->{wbuf}) {
618 my $n = send($sock, $_[1], MSG_MORE);
620 my $nlen = bytes::length($_[1]) - $n;
621 return 1 if $nlen == 0; # all done!
623 # queue up the unwritten substring:
624 $self->{wbuf} = [ tmpbuf(\($_[1]), $n) ];
625 watch_write($self, 1);
629 $self->write(\($_[1]));
632 =head2 C<< $obj->watch_read( $boolean ) >>
634 Turn 'readable' event notification on or off.
638 my PublicInbox::DS $self = shift;
639 my $sock = $self->{sock} or return;
642 my $event = $self->{event_watch};
644 $event &= ~EPOLLIN if ! $val;
645 $event |= EPOLLIN if $val;
647 my $fd = fileno($sock);
648 # If it changed, set it
649 if ($event != $self->{event_watch}) {
651 $KQueue->EV_SET($fd, IO::KQueue::EVFILT_READ(),
652 $val ? IO::KQueue::EV_ENABLE() : IO::KQueue::EV_DISABLE());
655 epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $event) and
656 confess("EPOLL_CTL_MOD: $!");
658 $self->{event_watch} = $event;
662 =head2 C<< $obj->watch_write( $boolean ) >>
664 Turn 'writable' event notification on or off.
668 my PublicInbox::DS $self = shift;
669 my $sock = $self->{sock} or return;
672 my $event = $self->{event_watch};
674 $event &= ~EPOLLOUT if ! $val;
675 $event |= EPOLLOUT if $val;
676 my $fd = fileno($sock);
678 # If it changed, set it
679 if ($event != $self->{event_watch}) {
681 $KQueue->EV_SET($fd, IO::KQueue::EVFILT_WRITE(),
682 $val ? IO::KQueue::EV_ENABLE() : IO::KQueue::EV_DISABLE());
685 epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $event) and
686 confess "EPOLL_CTL_MOD: $!";
688 $self->{event_watch} = $event;
692 package PublicInbox::DS::Timer;
693 # [$abs_float_firetime, $coderef];
700 =head1 AUTHORS (Danga::Socket)
702 Brad Fitzpatrick <brad@danga.com> - author
704 Michael Granger <ged@danga.com> - docs, testing
706 Mark Smith <junior@danga.com> - contributor, heavy user, testing
708 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits