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(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
42 %DescriptorMap, # fd (num) -> PublicInbox::DS object
43 $Epoll, # Global epoll fd (or DSKQXS ref)
44 $_io, # IO::Handle for Epoll
45 @ToClose, # sockets to close when event loop is done
47 $PostLoopCallback, # subref to call at the end of each loop, if defined (global)
49 $LoopTimeout, # timeout of event loop in milliseconds
50 $DoneInit, # if we've done the one-time module init yet
56 #####################################################################
57 ### C L A S S M E T H O D S
58 #####################################################################
60 =head2 C<< CLASS->Reset() >>
68 $LoopTimeout = -1; # no timeout by default
71 $PostLoopCallback = undef;
74 $_io = undef; # closes real $Epoll FD
75 $Epoll = undef; # may call DSKQXS::DESTROY
77 *EventLoop = *FirstTimeEventLoop;
80 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
82 Set the loop timeout for the event loop to some value in milliseconds.
84 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
89 return $LoopTimeout = $_[1] + 0;
92 =head2 C<< CLASS->AddTimer( $seconds, $coderef ) >>
94 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
95 are not guaranteed to fire at the exact time you ask for.
97 Returns a timer object which you can call C<< $timer->cancel >> on if you need to.
101 my ($class, $secs, $coderef) = @_;
103 my $fire_time = now() + $secs;
105 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
107 if (!@Timers || $fire_time >= $Timers[-1][0]) {
108 push @Timers, $timer;
112 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
113 # but it was compared against calendar queues, heaps, naive push/sort,
114 # and a bunch of other versions, and found to be fastest with a large
115 # variety of datasets.)
116 for (my $i = 0; $i < @Timers; $i++) {
117 if ($Timers[$i][0] > $fire_time) {
118 splice(@Timers, $i, 0, $timer);
123 die "Shouldn't get here.";
126 # keeping this around in case we support other FD types for now,
127 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
128 sub set_cloexec ($) {
131 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
132 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
133 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
141 if (PublicInbox::Syscall::epoll_defined()) {
142 $Epoll = epoll_create();
143 set_cloexec($Epoll) if (defined($Epoll) && $Epoll >= 0);
146 for (qw(DSKQXS DSPoll)) {
147 $cls = "PublicInbox::$_";
148 last if eval "require $cls";
150 $cls->import(qw(epoll_ctl epoll_wait));
153 *EventLoop = *EpollEventLoop;
156 =head2 C<< CLASS->EventLoop() >>
158 Start processing IO events. In most daemon programs this never exits. See
159 C<PostLoopCallback> below for how to exit the loop.
162 sub FirstTimeEventLoop {
170 sub now () { clock_gettime(CLOCK_MONOTONIC) }
176 if (ref($_) eq 'CODE') {
184 # runs timers and returns milliseconds for next one, or next event loop
188 return ((@$nextq || @ToClose) ? 0 : $LoopTimeout) unless @Timers;
193 while (@Timers && $Timers[0][0] <= $now) {
194 my $to_run = shift(@Timers);
195 $to_run->[1]->($now) if $to_run->[1];
198 # timers may enqueue into nextq:
199 return 0 if (@$nextq || @ToClose);
201 return $LoopTimeout unless @Timers;
203 # convert time to an even number of milliseconds, adding 1
204 # extra, otherwise floating point fun can occur and we'll
205 # call RunTimers like 20-30 times, each returning a timeout
206 # of 0.0000212 seconds
207 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
209 # -1 is an infinite timeout, so prefer a real timeout
210 return $timeout if $LoopTimeout == -1;
212 # otherwise pick the lower of our regular timeout and time until
214 return $LoopTimeout if $LoopTimeout < $timeout;
222 my $timeout = RunTimers();
224 # get up to 1000 events
225 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
226 for ($i=0; $i<$evcount; $i++) {
227 # it's possible epoll_wait returned many events, including some at the end
228 # that ones in the front triggered unregister-interest actions. if we
229 # can't find the %sock entry, it's because we're no longer interested
231 $DescriptorMap{$events[$i]->[0]}->event_step;
233 return unless PostEventLoop();
237 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
239 Sets post loop callback function. Pass a subref and it will be
240 called every time the event loop finishes.
242 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
245 The callback function will be passed two parameters: \%DescriptorMap
248 sub SetPostLoopCallback {
249 my ($class, $ref) = @_;
252 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
255 # Internal function: run the post-event callback, send read events
256 # for pushed-back data, and close pending connections. returns 1
257 # if event loop should continue, or 0 to shut it all down.
259 # now we can close sockets that wanted to close during our event processing.
260 # (we didn't want to close them during the loop, as we didn't want fd numbers
261 # being reused and confused during the event loop)
262 delete($DescriptorMap{fileno($_)}) for @ToClose;
263 @ToClose = (); # let refcounting drop everything all at once
265 # by default we keep running, unless a postloop callback (either per-object
266 # or global) cancels it
267 my $keep_running = 1;
269 # now we're at the very end, call callback if defined
270 if (defined $PostLoopCallback) {
271 $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
274 return $keep_running;
277 #####################################################################
278 ### PublicInbox::DS-the-object code
279 #####################################################################
281 =head2 OBJECT METHODS
283 =head2 C<< CLASS->new( $socket ) >>
285 Create a new PublicInbox::DS subclass object for the given I<socket> which will
286 react to events on it during the C<EventLoop>.
288 This is normally (always?) called from your subclass via:
290 $class->SUPER::new($socket);
294 my ($self, $sock, $ev) = @_;
295 $self = fields::new($self) unless ref $self;
297 $self->{sock} = $sock;
298 my $fd = fileno($sock);
300 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
305 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
306 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
307 $ev &= ~EPOLLEXCLUSIVE;
310 die "couldn't add epoll watch for $fd: $!\n";
312 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
313 if $DescriptorMap{$fd};
315 $DescriptorMap{$fd} = $self;
320 #####################################################################
321 ### I N S T A N C E M E T H O D S
322 #####################################################################
324 sub requeue ($) { push @$nextq, $_[0] }
326 =head2 C<< $obj->close >>
333 my $sock = delete $self->{sock} or return;
335 # we need to flush our write buffer, as there may
336 # be self-referential closures (sub { $client->close })
337 # preventing the object from being destroyed
338 delete $self->{wbuf};
340 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
341 # notifications about it
342 my $fd = fileno($sock);
343 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
344 confess("EPOLL_CTL_DEL: $!");
346 # we explicitly don't delete from DescriptorMap here until we
347 # actually close the socket, as we might be in the middle of
348 # processing an epoll_wait/etc that returned hundreds of fds, one
349 # of which is not yet processed and is what we're closing. if we
350 # keep it in DescriptorMap, then the event harnesses can just
351 # looked at $pob->{sock} == undef and ignore it. but if it's an
352 # un-accounted for fd, then it (understandably) freak out a bit
353 # and emit warnings, thinking their state got off.
355 # defer closing the actual socket until the event loop is done
356 # processing this round of events. (otherwise we might reuse fds)
357 push @ToClose, $sock;
362 # portable, non-thread-safe sendfile emulation (no pread, yet)
363 sub psendfile ($$$) {
364 my ($sock, $fh, $off) = @_;
366 seek($fh, $$off, SEEK_SET) or return;
367 defined(my $to_write = read($fh, my $buf, 16384)) or return;
369 while ($to_write > 0) {
370 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
374 return if $written == 0;
382 sub epbit ($$) { # (sock, default)
383 ref($_[0]) eq 'IO::Socket::SSL' ? PublicInbox::TLS::epollbit() : $_[1];
386 # returns 1 if done, 0 if incomplete
387 sub flush_write ($) {
389 my $wbuf = $self->{wbuf} or return 1;
390 my $sock = $self->{sock};
393 while (my $bref = $wbuf->[0]) {
394 if (ref($bref) ne 'CODE') {
395 my $off = delete($self->{wbuf_off}) // 0;
397 my $w = psendfile($sock, $bref, \$off);
403 } elsif ($! == EAGAIN) {
404 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
405 $self->{wbuf_off} = $off;
411 } else { #($ref eq 'CODE') {
413 my $before = scalar(@$wbuf);
416 # bref may be enqueueing more CODE to call (see accept_tls_step)
417 return 0 if (scalar(@$wbuf) > $before);
421 delete $self->{wbuf};
426 my ($self, $rbuf) = @_;
427 if ($$rbuf eq '') { # who knows how long till we can read again
428 delete $self->{rbuf};
430 $self->{rbuf} = $rbuf;
434 sub do_read ($$$;$) {
435 my ($self, $rbuf, $len, $off) = @_;
436 my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0);
437 return ($r == 0 ? $self->close : $r) if defined $r;
438 # common for clients to break connections without warning,
439 # would be too noisy to log here:
441 epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT);
442 rbuf_idle($self, $rbuf);
449 # drop the socket if we hit unrecoverable errors on our system which
450 # require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE...
457 # n.b.: use ->write/->read for this buffer to allow compatibility with
458 # PerlIO::mmap or PerlIO::scalar if needed
460 my ($self, $bref, $off) = @_;
461 my $fh; # open(my $fh, '+>>', undef) doesn't set O_APPEND
463 my $fn = File::Spec->tmpdir . '/wbuf-' . rand;
464 if (sysopen($fh, $fn, O_RDWR|O_CREAT|O_EXCL|O_APPEND, 0600)) { # likely
465 unlink($fn) or return drop($self, "unlink($fn) $!");
466 } elsif ($! != EEXIST) { # EMFILE/ENFILE/ENOSPC/ENOMEM
467 return drop($self, "open: $!");
469 } until (defined $fh);
471 my $len = bytes::length($$bref) - $off;
472 $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!");
476 =head2 C<< $obj->write( $data ) >>
478 Write the specified data to the underlying handle. I<data> may be scalar,
479 scalar ref, code ref (to run when there).
480 Returns 1 if writes all went through, or 0 if there are writes in queue. If
481 it returns 1, caller should stop waiting for 'writable' events)
485 my ($self, $data) = @_;
487 # nobody should be writing to closed sockets, but caller code can
488 # do two writes within an event, have the first fail and
489 # disconnect the other side (whose destructor then closes the
490 # calling object, but it's still in a method), and then the
491 # now-dead object does its second write. that is this case. we
492 # just lie and say it worked. it'll be dead soon and won't be
494 my $sock = $self->{sock} or return 1;
496 my $bref = $ref ? $data : \$data;
497 my $wbuf = $self->{wbuf};
498 if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more...
499 if ($ref eq 'CODE') {
502 my $last = $wbuf->[-1];
503 if (ref($last) eq 'GLOB') { # append to tmp file buffer
504 $last->print($$bref) or return drop($self, "print: $!");
506 my $tmpio = tmpio($self, $bref, 0) or return 0;
511 } elsif ($ref eq 'CODE') {
515 my $to_write = bytes::length($$bref);
516 my $written = syswrite($sock, $$bref, $to_write);
518 if (defined $written) {
519 return 1 if $written == $to_write;
520 requeue($self); # runs: event_step -> flush_write
521 } elsif ($! == EAGAIN) {
522 epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT);
528 # deal with EAGAIN or partial write:
529 my $tmpio = tmpio($self, $bref, $written) or return 0;
531 # wbuf may be an empty array if we're being called inside
532 # ->flush_write via CODE bref:
533 push @{$self->{wbuf} ||= []}, $tmpio;
538 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
542 my $sock = $self->{sock} or return 1;
544 if (MSG_MORE && !$self->{wbuf} && ref($sock) ne 'IO::Socket::SSL') {
545 my $n = send($sock, $_[1], MSG_MORE);
547 my $nlen = bytes::length($_[1]) - $n;
548 return 1 if $nlen == 0; # all done!
549 # queue up the unwritten substring:
550 my $tmpio = tmpio($self, \($_[1]), $n) or return 0;
551 $self->{wbuf} = [ $tmpio ];
552 epwait($sock, EPOLLOUT|EPOLLONESHOT);
556 $self->write(\($_[1]));
560 my ($sock, $ev) = @_;
561 epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and
562 confess("EPOLL_CTL_MOD $!");
565 # return true if complete, false if incomplete (or failure)
566 sub accept_tls_step ($) {
568 my $sock = $self->{sock} or return;
569 return 1 if $sock->accept_SSL;
570 return $self->close if $! != EAGAIN;
571 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
572 unshift @{$self->{wbuf} ||= []}, \&accept_tls_step;
576 # return true if complete, false if incomplete (or failure)
577 sub shutdn_tls_step ($) {
579 my $sock = $self->{sock} or return;
580 return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1);
581 return $self->close if $! != EAGAIN;
582 epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT);
583 unshift @{$self->{wbuf} ||= []}, \&shutdn_tls_step;
587 # don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC
588 # or fork w/o exec, so no inadvertant socket sharing
591 my $sock = $self->{sock} or return;
592 if (ref($sock) eq 'IO::Socket::SSL') {
593 shutdn_tls_step($self);
598 package PublicInbox::DS::Timer;
599 # [$abs_float_firetime, $coderef];
606 =head1 AUTHORS (Danga::Socket)
608 Brad Fitzpatrick <brad@danga.com> - author
610 Michael Granger <ged@danga.com> - docs, testing
612 Mark Smith <junior@danga.com> - contributor, heavy user, testing
614 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits