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);
28 use PublicInbox::Syscall qw(:epoll);
30 use fields ('sock', # underlying socket
31 'wbuf', # arrayref of coderefs or GLOB refs
32 'wbuf_off', # offset into first element of wbuf to start writing at
35 use Errno qw(EAGAIN EINVAL);
36 use Carp qw(croak confess carp);
37 use File::Temp qw(tempfile);
39 our $HAVE_KQUEUE = eval { require IO::KQueue; IO::KQueue->import; 1 };
42 $HaveEpoll, # Flag -- is epoll available? initially undefined.
44 %DescriptorMap, # fd (num) -> PublicInbox::DS object
45 $Epoll, # Global epoll fd (for epoll mode only)
46 $KQueue, # Global kqueue fd ref (for kqueue mode only)
47 $_io, # IO::Handle for Epoll
48 @ToClose, # sockets to close when event loop is done
50 $PostLoopCallback, # subref to call at the end of each loop, if defined (global)
52 $LoopTimeout, # timeout of event loop in milliseconds
53 $DoneInit, # if we've done the one-time module init yet
59 #####################################################################
60 ### C L A S S M E T H O D S
61 #####################################################################
63 =head2 C<< CLASS->Reset() >>
71 $LoopTimeout = -1; # no timeout by default
74 $PostLoopCallback = undef;
77 # NOTE kqueue is close-on-fork, and we don't account for it, yet
78 # OTOH, we (public-inbox) don't need this sub outside of tests...
79 POSIX::close($$KQueue) if !$_io && $KQueue && $$KQueue >= 0;
82 $_io = undef; # close $Epoll
85 *EventLoop = *FirstTimeEventLoop;
88 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
90 Set the loop timeout for the event loop to some value in milliseconds.
92 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
97 return $LoopTimeout = $_[1] + 0;
100 =head2 C<< CLASS->AddTimer( $seconds, $coderef ) >>
102 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
103 are not guaranteed to fire at the exact time you ask for.
105 Returns a timer object which you can call C<< $timer->cancel >> on if you need to.
109 my ($class, $secs, $coderef) = @_;
112 my $timer = bless([0, $coderef], 'PublicInbox::DS::Timer');
113 unshift(@Timers, $timer);
117 my $fire_time = now() + $secs;
119 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
121 if (!@Timers || $fire_time >= $Timers[-1][0]) {
122 push @Timers, $timer;
126 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
127 # but it was compared against calendar queues, heaps, naive push/sort,
128 # and a bunch of other versions, and found to be fastest with a large
129 # variety of datasets.)
130 for (my $i = 0; $i < @Timers; $i++) {
131 if ($Timers[$i][0] > $fire_time) {
132 splice(@Timers, $i, 0, $timer);
137 die "Shouldn't get here.";
140 # keeping this around in case we support other FD types for now,
141 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
142 sub set_cloexec ($) {
145 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
146 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
147 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
156 $KQueue = IO::KQueue->new();
157 $HaveKQueue = defined $KQueue;
159 *EventLoop = *KQueueEventLoop;
162 elsif (PublicInbox::Syscall::epoll_defined()) {
163 $Epoll = eval { epoll_create(1024); };
164 $HaveEpoll = defined $Epoll && $Epoll >= 0;
167 *EventLoop = *EpollEventLoop;
172 =head2 C<< CLASS->EventLoop() >>
174 Start processing IO events. In most daemon programs this never exits. See
175 C<PostLoopCallback> below for how to exit the loop.
178 sub FirstTimeEventLoop {
184 EpollEventLoop($class);
185 } elsif ($HaveKQueue) {
186 KQueueEventLoop($class);
190 sub now () { clock_gettime(CLOCK_MONOTONIC) }
192 # runs timers and returns milliseconds for next one, or next event loop
194 return $LoopTimeout unless @Timers;
199 while (@Timers && $Timers[0][0] <= $now) {
200 my $to_run = shift(@Timers);
201 $to_run->[1]->($now) if $to_run->[1];
204 return $LoopTimeout unless @Timers;
206 # convert time to an even number of milliseconds, adding 1
207 # extra, otherwise floating point fun can occur and we'll
208 # call RunTimers like 20-30 times, each returning a timeout
209 # of 0.0000212 seconds
210 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
212 # -1 is an infinite timeout, so prefer a real timeout
213 return $timeout if $LoopTimeout == -1;
215 # otherwise pick the lower of our regular timeout and time until
217 return $LoopTimeout if $LoopTimeout < $timeout;
221 ### The epoll-based event loop. Gets installed as EventLoop if IO::Epoll loads
229 my $timeout = RunTimers();
231 # get up to 1000 events
232 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
233 for ($i=0; $i<$evcount; $i++) {
234 # it's possible epoll_wait returned many events, including some at the end
235 # that ones in the front triggered unregister-interest actions. if we
236 # can't find the %sock entry, it's because we're no longer interested
238 $DescriptorMap{$events[$i]->[0]}->event_step;
240 return unless PostEventLoop();
244 ### The kqueue-based event loop. Gets installed as EventLoop if IO::KQueue works
246 sub KQueueEventLoop {
250 my $timeout = RunTimers();
251 my @ret = eval { $KQueue->kevent($timeout) };
253 # workaround https://rt.cpan.org/Ticket/Display.html?id=116615
254 if ($err =~ /Interrupted system call/) {
261 foreach my $kev (@ret) {
262 $DescriptorMap{$kev->[0]}->event_step;
264 return unless PostEventLoop();
268 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
270 Sets post loop callback function. Pass a subref and it will be
271 called every time the event loop finishes.
273 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
276 The callback function will be passed two parameters: \%DescriptorMap
279 sub SetPostLoopCallback {
280 my ($class, $ref) = @_;
283 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
286 # Internal function: run the post-event callback, send read events
287 # for pushed-back data, and close pending connections. returns 1
288 # if event loop should continue, or 0 to shut it all down.
290 # now we can close sockets that wanted to close during our event processing.
291 # (we didn't want to close them during the loop, as we didn't want fd numbers
292 # being reused and confused during the event loop)
293 while (my $sock = shift @ToClose) {
294 my $fd = fileno($sock);
296 # close the socket. (not a PublicInbox::DS close)
299 # and now we can finally remove the fd from the map. see
300 # comment above in ->close.
301 delete $DescriptorMap{$fd};
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 # map EPOLL* bits to kqueue EV_* flags for EV_SET
321 my $fl = EV_ADD() | EV_ENABLE();
322 ($ev & EPOLLONESHOT) ? ($fl|EV_ONESHOT()) : $fl;
324 EV_ADD() | EV_DISABLE();
328 #####################################################################
329 ### PublicInbox::DS-the-object code
330 #####################################################################
332 =head2 OBJECT METHODS
334 =head2 C<< CLASS->new( $socket ) >>
336 Create a new PublicInbox::DS subclass object for the given I<socket> which will
337 react to events on it during the C<EventLoop>.
339 This is normally (always?) called from your subclass via:
341 $class->SUPER::new($socket);
345 my ($self, $sock, $ev) = @_;
346 $self = fields::new($self) unless ref $self;
348 $self->{sock} = $sock;
349 my $fd = fileno($sock);
351 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
358 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
359 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
360 $ev &= ~EPOLLEXCLUSIVE;
363 die "couldn't add epoll watch for $fd: $!\n";
366 elsif ($HaveKQueue) {
367 $KQueue->EV_SET($fd, EVFILT_READ(), kq_flag(EPOLLIN, $ev));
368 $KQueue->EV_SET($fd, EVFILT_WRITE(), kq_flag(EPOLLOUT, $ev));
371 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
372 if $DescriptorMap{$fd};
374 $DescriptorMap{$fd} = $self;
379 #####################################################################
380 ### I N S T A N C E M E T H O D S
381 #####################################################################
383 =head2 C<< $obj->close >>
390 my $sock = delete $self->{sock} or return;
392 # we need to flush our write buffer, as there may
393 # be self-referential closures (sub { $client->close })
394 # preventing the object from being destroyed
395 delete $self->{wbuf};
397 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
398 # notifications about it
400 my $fd = fileno($sock);
401 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
402 confess("EPOLL_CTL_DEL: $!");
405 # we explicitly don't delete from DescriptorMap here until we
406 # actually close the socket, as we might be in the middle of
407 # processing an epoll_wait/etc that returned hundreds of fds, one
408 # of which is not yet processed and is what we're closing. if we
409 # keep it in DescriptorMap, then the event harnesses can just
410 # looked at $pob->{sock} == undef and ignore it. but if it's an
411 # un-accounted for fd, then it (understandably) freak out a bit
412 # and emit warnings, thinking their state got off.
414 # defer closing the actual socket until the event loop is done
415 # processing this round of events. (otherwise we might reuse fds)
416 push @ToClose, $sock;
421 # portable, non-thread-safe sendfile emulation (no pread, yet)
422 sub psendfile ($$$) {
423 my ($sock, $fh, $off) = @_;
425 seek($fh, $$off, SEEK_SET) or return;
426 defined(my $to_write = read($fh, my $buf, 16384)) or return;
428 while ($to_write > 0) {
429 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
433 return if $written == 0;
441 # returns 1 if done, 0 if incomplete
442 sub flush_write ($) {
444 my $wbuf = $self->{wbuf} or return 1;
445 my $sock = $self->{sock};
448 while (my $bref = $wbuf->[0]) {
449 if (ref($bref) ne 'CODE') {
450 my $off = delete($self->{wbuf_off}) // 0;
452 my $w = psendfile($sock, $bref, \$off);
458 } elsif ($! == EAGAIN) {
459 $self->{wbuf_off} = $off;
460 watch($self, EPOLLOUT|EPOLLONESHOT);
466 } else { #($ref eq 'CODE') {
468 my $before = scalar(@$wbuf);
471 # bref may be enqueueing more CODE to call (see accept_tls_step)
472 return 0 if (scalar(@$wbuf) > $before);
476 delete $self->{wbuf};
481 my ($self, $rbuf, $len, $off) = @_;
482 my $r = sysread($self->{sock}, $$rbuf, $len, $off);
483 return ($r == 0 ? $self->close : $r) if defined $r;
484 # common for clients to break connections without warning,
485 # would be too noisy to log here:
486 if (ref($self) eq 'IO::Socket::SSL') {
487 my $ev = PublicInbox::TLS::epollbit() or return $self->close;
488 watch($self, $ev | EPOLLONESHOT);
489 } elsif ($! == EAGAIN) {
490 watch($self, EPOLLIN | EPOLLONESHOT);
496 # drop the socket if we hit unrecoverable errors on our system which
497 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
504 # n.b.: use ->write/->read for this buffer to allow compatibility with
505 # PerlIO::mmap or PerlIO::scalar if needed
507 my ($self, $bref, $off) = @_;
508 # open(my $fh, '+>>', undef) doesn't set O_APPEND
509 my ($fh, $path) = eval { tempfile('wbuf-XXXXXXX', TMPDIR => 1) };
510 $fh or return drop($self, "tempfile: $@");
511 open($fh, '+>>', $path) or return drop($self, "open: $!");
513 unlink($path) or return drop($self, "unlink: $!");
514 my $len = bytes::length($$bref) - $off;
515 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
519 =head2 C<< $obj->write( $data ) >>
521 Write the specified data to the underlying handle. I<data> may be scalar,
522 scalar ref, code ref (to run when there).
523 Returns 1 if writes all went through, or 0 if there are writes in queue. If
524 it returns 1, caller should stop waiting for 'writable' events)
528 my ($self, $data) = @_;
530 # nobody should be writing to closed sockets, but caller code can
531 # do two writes within an event, have the first fail and
532 # disconnect the other side (whose destructor then closes the
533 # calling object, but it's still in a method), and then the
534 # now-dead object does its second write. that is this case. we
535 # just lie and say it worked. it'll be dead soon and won't be
537 my $sock = $self->{sock} or return 1;
539 my $bref = $ref ? $data : \$data;
540 my $wbuf = $self->{wbuf};
541 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
542 if ($ref eq 'CODE') {
545 my $last = $wbuf->[-1];
546 if (ref($last) eq 'GLOB') { # append to tmp file buffer
547 $last->print($$bref) or return drop($self, "print: $!");
549 my $tmpio = tmpio($self, $bref, 0) or return 0;
554 } elsif ($ref eq 'CODE') {
558 my $to_write = bytes::length($$bref);
559 my $written = syswrite($sock, $$bref, $to_write);
561 if (defined $written) {
562 return 1 if $written == $to_write;
563 } elsif ($! == EAGAIN) {
568 my $tmpio = tmpio($self, $bref, $written) or return 0;
569 $self->{wbuf} = [ $tmpio ];
570 watch($self, EPOLLOUT|EPOLLONESHOT);
575 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
579 my $sock = $self->{sock} or return 1;
581 if (MSG_MORE && !$self->{wbuf} && ref($sock) ne 'IO::Socket::SSL') {
582 my $n = send($sock, $_[1], MSG_MORE);
584 my $nlen = bytes::length($_[1]) - $n;
585 return 1 if $nlen == 0; # all done!
586 # queue up the unwritten substring:
587 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
588 $self->{wbuf} = [ $tmpio ];
589 watch($self, EPOLLOUT|EPOLLONESHOT);
593 $self->write(\($_[1]));
597 my ($self, $ev) = @_;
598 my $sock = $self->{sock} or return;
599 my $fd = fileno($sock);
601 epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $ev) and
602 confess("EPOLL_CTL_MOD $!");
603 } elsif ($HaveKQueue) {
604 $KQueue->EV_SET($fd, EVFILT_READ(), kq_flag(EPOLLIN, $ev));
605 $KQueue->EV_SET($fd, EVFILT_WRITE(), kq_flag(EPOLLOUT, $ev));
610 sub watch_in1 ($) { watch($_[0], EPOLLIN | EPOLLONESHOT) }
612 # return true if complete, false if incomplete (or failure)
613 sub accept_tls_step ($) {
615 my $sock = $self->{sock} or return;
616 return 1 if $sock->accept_SSL;
617 return $self->close if $! != EAGAIN;
618 if (my $ev = PublicInbox::TLS::epollbit()) {
619 unshift @{$self->{wbuf} ||= []}, \&accept_tls_step;
620 return watch($self, $ev | EPOLLONESHOT);
622 drop($self, 'BUG? EAGAIN but '.PublicInbox::TLS::err());
625 sub shutdn_tls_step ($) {
627 my $sock = $self->{sock} or return;
628 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
629 return $self->close if $! != EAGAIN;
630 if (my $ev = PublicInbox::TLS::epollbit()) {
631 unshift @{$self->{wbuf} ||= []}, \&shutdn_tls_step;
632 return watch($self, $ev | EPOLLONESHOT);
634 drop($self, 'BUG? EAGAIN but '.PublicInbox::TLS::err());
637 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
638 # or fork w/o exec, so no inadvertant socket sharing
641 my $sock = $self->{sock} or return;
642 if (ref($sock) eq 'IO::Socket::SSL') {
643 shutdn_tls_step($self);
649 package PublicInbox::DS::Timer;
650 # [$abs_float_firetime, $coderef];
657 =head1 AUTHORS (Danga::Socket)
659 Brad Fitzpatrick <brad@danga.com> - author
661 Michael Granger <ged@danga.com> - docs, testing
663 Mark Smith <junior@danga.com> - contributor, heavy user, testing
665 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits