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);
28 use PublicInbox::Syscall qw(:epoll);
30 use fields ('sock', # underlying socket
31 'rbuf', # scalarref, usually undef
32 'wbuf', # arrayref of coderefs or GLOB refs
33 'wbuf_off', # offset into first element of wbuf to start writing at
36 use Errno qw(EAGAIN EINVAL EEXIST);
37 use Carp qw(croak confess carp);
40 my $nextq = []; # queue for next_tick
41 my $WaitPids = []; # list of [ pid, callback, callback_arg ]
44 %DescriptorMap, # fd (num) -> PublicInbox::DS object
45 $Epoll, # Global epoll fd (or DSKQXS ref)
46 $_io, # IO::Handle for Epoll
47 @ToClose, # sockets to close when event loop is done
49 $PostLoopCallback, # subref to call at the end of each loop, if defined (global)
51 $LoopTimeout, # timeout of event loop in milliseconds
52 $DoneInit, # if we've done the one-time module init yet
58 #####################################################################
59 ### C L A S S M E T H O D S
60 #####################################################################
62 =head2 C<< CLASS->Reset() >>
72 $LoopTimeout = -1; # no timeout by default
75 $PostLoopCallback = undef;
78 $_io = undef; # closes real $Epoll FD
79 $Epoll = undef; # may call DSKQXS::DESTROY
81 *EventLoop = *FirstTimeEventLoop;
84 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
86 Set the loop timeout for the event loop to some value in milliseconds.
88 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
93 return $LoopTimeout = $_[1] + 0;
96 =head2 C<< CLASS->AddTimer( $seconds, $coderef ) >>
98 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
99 are not guaranteed to fire at the exact time you ask for.
101 Returns a timer object which you can call C<< $timer->cancel >> on if you need to.
105 my ($class, $secs, $coderef) = @_;
107 my $fire_time = now() + $secs;
109 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
111 if (!@Timers || $fire_time >= $Timers[-1][0]) {
112 push @Timers, $timer;
116 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
117 # but it was compared against calendar queues, heaps, naive push/sort,
118 # and a bunch of other versions, and found to be fastest with a large
119 # variety of datasets.)
120 for (my $i = 0; $i < @Timers; $i++) {
121 if ($Timers[$i][0] > $fire_time) {
122 splice(@Timers, $i, 0, $timer);
127 die "Shouldn't get here.";
130 # keeping this around in case we support other FD types for now,
131 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
132 sub set_cloexec ($) {
135 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
136 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
137 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
145 if (PublicInbox::Syscall::epoll_defined()) {
146 $Epoll = epoll_create();
147 set_cloexec($Epoll) if (defined($Epoll) && $Epoll >= 0);
150 for (qw(DSKQXS DSPoll)) {
151 $cls = "PublicInbox::$_";
152 last if eval "require $cls";
154 $cls->import(qw(epoll_ctl epoll_wait));
157 *EventLoop = *EpollEventLoop;
160 =head2 C<< CLASS->EventLoop() >>
162 Start processing IO events. In most daemon programs this never exits. See
163 C<PostLoopCallback> below for how to exit the loop.
166 sub FirstTimeEventLoop {
174 sub now () { clock_gettime(CLOCK_MONOTONIC) }
180 if (ref($_) eq 'CODE') {
188 # runs timers and returns milliseconds for next one, or next event loop
192 return ((@$nextq || @ToClose) ? 0 : $LoopTimeout) unless @Timers;
197 while (@Timers && $Timers[0][0] <= $now) {
198 my $to_run = shift(@Timers);
199 $to_run->[1]->($now) if $to_run->[1];
202 # timers may enqueue into nextq:
203 return 0 if (@$nextq || @ToClose);
205 return $LoopTimeout unless @Timers;
207 # convert time to an even number of milliseconds, adding 1
208 # extra, otherwise floating point fun can occur and we'll
209 # call RunTimers like 20-30 times, each returning a timeout
210 # of 0.0000212 seconds
211 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
213 # -1 is an infinite timeout, so prefer a real timeout
214 return $timeout if $LoopTimeout == -1;
216 # otherwise pick the lower of our regular timeout and time until
218 return $LoopTimeout if $LoopTimeout < $timeout;
222 # We can't use waitpid(-1) safely here since it can hit ``, system(),
223 # and other things. So we scan the $WaitPids list, which is hopefully
229 foreach my $ary (@$tmp) {
230 my ($pid, $cb, $arg) = @$ary;
231 my $ret = waitpid($pid, WNOHANG);
233 push @$WaitPids, $ary;
235 eval { $cb->($arg, $pid) };
239 # we may not be donea, and we may miss our
240 $reap_timer = AddTimer(undef, 1, \&reap_pids);
244 # reentrant SIGCHLD handler (since reap_pids is not reentrant)
245 sub enqueue_reap ($) { push @$nextq, \&reap_pids };
248 local $SIG{CHLD} = \&enqueue_reap;
252 my $timeout = RunTimers();
254 # get up to 1000 events
255 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
256 for ($i=0; $i<$evcount; $i++) {
257 # it's possible epoll_wait returned many events, including some at the end
258 # that ones in the front triggered unregister-interest actions. if we
259 # can't find the %sock entry, it's because we're no longer interested
261 $DescriptorMap{$events[$i]->[0]}->event_step;
263 return unless PostEventLoop();
267 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
269 Sets post loop callback function. Pass a subref and it will be
270 called every time the event loop finishes.
272 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
275 The callback function will be passed two parameters: \%DescriptorMap
278 sub SetPostLoopCallback {
279 my ($class, $ref) = @_;
282 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
285 # Internal function: run the post-event callback, send read events
286 # for pushed-back data, and close pending connections. returns 1
287 # if event loop should continue, or 0 to shut it all down.
289 # now we can close sockets that wanted to close during our event processing.
290 # (we didn't want to close them during the loop, as we didn't want fd numbers
291 # being reused and confused during the event loop)
292 delete($DescriptorMap{fileno($_)}) for @ToClose;
293 @ToClose = (); # let refcounting drop everything all at once
295 # by default we keep running, unless a postloop callback (either per-object
296 # or global) cancels it
297 my $keep_running = 1;
299 # now we're at the very end, call callback if defined
300 if (defined $PostLoopCallback) {
301 $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
304 return $keep_running;
307 #####################################################################
308 ### PublicInbox::DS-the-object code
309 #####################################################################
311 =head2 OBJECT METHODS
313 =head2 C<< CLASS->new( $socket ) >>
315 Create a new PublicInbox::DS subclass object for the given I<socket> which will
316 react to events on it during the C<EventLoop>.
318 This is normally (always?) called from your subclass via:
320 $class->SUPER::new($socket);
324 my ($self, $sock, $ev) = @_;
325 $self = fields::new($self) unless ref $self;
327 $self->{sock} = $sock;
328 my $fd = fileno($sock);
330 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
335 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
336 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
337 $ev &= ~EPOLLEXCLUSIVE;
340 die "couldn't add epoll watch for $fd: $!\n";
342 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
343 if $DescriptorMap{$fd};
345 $DescriptorMap{$fd} = $self;
350 #####################################################################
351 ### I N S T A N C E M E T H O D S
352 #####################################################################
354 sub requeue ($) { push @$nextq, $_[0] }
356 =head2 C<< $obj->close >>
363 my $sock = delete $self->{sock} or return;
365 # we need to flush our write buffer, as there may
366 # be self-referential closures (sub { $client->close })
367 # preventing the object from being destroyed
368 delete $self->{wbuf};
370 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
371 # notifications about it
372 my $fd = fileno($sock);
373 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
374 confess("EPOLL_CTL_DEL: $!");
376 # we explicitly don't delete from DescriptorMap here until we
377 # actually close the socket, as we might be in the middle of
378 # processing an epoll_wait/etc that returned hundreds of fds, one
379 # of which is not yet processed and is what we're closing. if we
380 # keep it in DescriptorMap, then the event harnesses can just
381 # looked at $pob->{sock} == undef and ignore it. but if it's an
382 # un-accounted for fd, then it (understandably) freak out a bit
383 # and emit warnings, thinking their state got off.
385 # defer closing the actual socket until the event loop is done
386 # processing this round of events. (otherwise we might reuse fds)
387 push @ToClose, $sock;
392 # portable, non-thread-safe sendfile emulation (no pread, yet)
393 sub psendfile ($$$) {
394 my ($sock, $fh, $off) = @_;
396 seek($fh, $$off, SEEK_SET) or return;
397 defined(my $to_write = read($fh, my $buf, 16384)) or return;
399 while ($to_write > 0) {
400 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
404 return if $written == 0;
412 sub epbit ($$) { # (sock, default)
413 ref($_[0]) eq 'IO::Socket::SSL' ? PublicInbox::TLS::epollbit() : $_[1];
416 # returns 1 if done, 0 if incomplete
417 sub flush_write ($) {
419 my $wbuf = $self->{wbuf} or return 1;
420 my $sock = $self->{sock};
423 while (my $bref = $wbuf->[0]) {
424 if (ref($bref) ne 'CODE') {
425 my $off = delete($self->{wbuf_off}) // 0;
427 my $w = psendfile($sock, $bref, \$off);
433 } elsif ($! == EAGAIN) {
434 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
435 $self->{wbuf_off} = $off;
441 } else { #($ref eq 'CODE') {
443 my $before = scalar(@$wbuf);
446 # bref may be enqueueing more CODE to call (see accept_tls_step)
447 return 0 if (scalar(@$wbuf) > $before);
451 delete $self->{wbuf};
456 my ($self, $rbuf) = @_;
457 if ($$rbuf eq '') { # who knows how long till we can read again
458 delete $self->{rbuf};
460 $self->{rbuf} = $rbuf;
464 sub do_read ($$$;$) {
465 my ($self, $rbuf, $len, $off) = @_;
466 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
467 return ($r == 0 ? $self->close : $r) if defined $r;
468 # common for clients to break connections without warning,
469 # would be too noisy to log here:
471 epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT);
472 rbuf_idle($self, $rbuf);
479 # drop the socket if we hit unrecoverable errors on our system which
480 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
487 # n.b.: use ->write/->read for this buffer to allow compatibility with
488 # PerlIO::mmap or PerlIO::scalar if needed
490 my ($self, $bref, $off) = @_;
491 my $fh; # open(my $fh, '+>>', undef) doesn't set O_APPEND
493 my $fn = File::Spec->tmpdir . '/wbuf-' . rand;
494 if (sysopen($fh, $fn, O_RDWR|O_CREAT|O_EXCL|O_APPEND, 0600)) { # likely
495 unlink($fn) or return drop($self, "unlink($fn) $!");
496 } elsif ($! != EEXIST) { # EMFILE/ENFILE/ENOSPC/ENOMEM
497 return drop($self, "open: $!");
499 } until (defined $fh);
501 my $len = bytes::length($$bref) - $off;
502 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
506 =head2 C<< $obj->write( $data ) >>
508 Write the specified data to the underlying handle. I<data> may be scalar,
509 scalar ref, code ref (to run when there).
510 Returns 1 if writes all went through, or 0 if there are writes in queue. If
511 it returns 1, caller should stop waiting for 'writable' events)
515 my ($self, $data) = @_;
517 # nobody should be writing to closed sockets, but caller code can
518 # do two writes within an event, have the first fail and
519 # disconnect the other side (whose destructor then closes the
520 # calling object, but it's still in a method), and then the
521 # now-dead object does its second write. that is this case. we
522 # just lie and say it worked. it'll be dead soon and won't be
524 my $sock = $self->{sock} or return 1;
526 my $bref = $ref ? $data : \$data;
527 my $wbuf = $self->{wbuf};
528 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
529 if ($ref eq 'CODE') {
532 my $last = $wbuf->[-1];
533 if (ref($last) eq 'GLOB') { # append to tmp file buffer
534 $last->print($$bref) or return drop($self, "print: $!");
536 my $tmpio = tmpio($self, $bref, 0) or return 0;
541 } elsif ($ref eq 'CODE') {
545 my $to_write = bytes::length($$bref);
546 my $written = syswrite($sock, $$bref, $to_write);
548 if (defined $written) {
549 return 1 if $written == $to_write;
550 requeue($self); # runs: event_step -> flush_write
551 } elsif ($! == EAGAIN) {
552 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
558 # deal with EAGAIN or partial write:
559 my $tmpio = tmpio($self, $bref, $written) or return 0;
561 # wbuf may be an empty array if we're being called inside
562 # ->flush_write via CODE bref:
563 push @{$self->{wbuf} ||= []}, $tmpio;
568 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
572 my $sock = $self->{sock} or return 1;
574 if (MSG_MORE && !$self->{wbuf} && ref($sock) ne 'IO::Socket::SSL') {
575 my $n = send($sock, $_[1], MSG_MORE);
577 my $nlen = bytes::length($_[1]) - $n;
578 return 1 if $nlen == 0; # all done!
579 # queue up the unwritten substring:
580 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
581 $self->{wbuf} = [ $tmpio ];
582 epwait($sock, EPOLLOUT|EPOLLONESHOT);
586 $self->write(\($_[1]));
590 my ($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) = @_;
632 my $chld = $SIG{CHLD};
633 if (defined($chld) && $chld eq \&enqueue_reap) {
634 push @$WaitPids, [ $pid, $cb, $arg ];
636 # We could've just missed our SIGCHLD, cover it, here:
637 requeue(\&reap_pids);
639 die "Not in EventLoop\n";
643 package PublicInbox::DS::Timer;
644 # [$abs_float_firetime, $coderef];
651 =head1 AUTHORS (Danga::Socket)
653 Brad Fitzpatrick <brad@danga.com> - author
655 Michael Granger <ged@danga.com> - docs, testing
657 Mark Smith <junior@danga.com> - contributor, heavy user, testing
659 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits