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
34 use Errno qw(EAGAIN EINVAL);
35 use Carp qw(croak confess);
36 use File::Temp qw(tempfile);
38 our $HAVE_KQUEUE = eval { require IO::KQueue; IO::KQueue->import; 1 };
41 $HaveEpoll, # Flag -- is epoll available? initially undefined.
43 %DescriptorMap, # fd (num) -> PublicInbox::DS object
44 $Epoll, # Global epoll fd (for epoll mode only)
45 $KQueue, # Global kqueue fd ref (for kqueue mode only)
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() >>
70 $LoopTimeout = -1; # no timeout by default
73 $PostLoopCallback = undef;
76 # NOTE kqueue is close-on-fork, and we don't account for it, yet
77 # OTOH, we (public-inbox) don't need this sub outside of tests...
78 POSIX::close($$KQueue) if !$_io && $KQueue && $$KQueue >= 0;
81 $_io = undef; # close $Epoll
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) = @_;
111 my $timer = bless([0, $coderef], 'PublicInbox::DS::Timer');
112 unshift(@Timers, $timer);
116 my $fire_time = now() + $secs;
118 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
120 if (!@Timers || $fire_time >= $Timers[-1][0]) {
121 push @Timers, $timer;
125 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
126 # but it was compared against calendar queues, heaps, naive push/sort,
127 # and a bunch of other versions, and found to be fastest with a large
128 # variety of datasets.)
129 for (my $i = 0; $i < @Timers; $i++) {
130 if ($Timers[$i][0] > $fire_time) {
131 splice(@Timers, $i, 0, $timer);
136 die "Shouldn't get here.";
139 # keeping this around in case we support other FD types for now,
140 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
141 sub set_cloexec ($) {
144 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
145 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
146 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
155 $KQueue = IO::KQueue->new();
156 $HaveKQueue = defined $KQueue;
158 *EventLoop = *KQueueEventLoop;
161 elsif (PublicInbox::Syscall::epoll_defined()) {
162 $Epoll = eval { epoll_create(1024); };
163 $HaveEpoll = defined $Epoll && $Epoll >= 0;
166 *EventLoop = *EpollEventLoop;
171 =head2 C<< CLASS->EventLoop() >>
173 Start processing IO events. In most daemon programs this never exits. See
174 C<PostLoopCallback> below for how to exit the loop.
177 sub FirstTimeEventLoop {
183 EpollEventLoop($class);
184 } elsif ($HaveKQueue) {
185 KQueueEventLoop($class);
189 sub now () { clock_gettime(CLOCK_MONOTONIC) }
191 # runs timers and returns milliseconds for next one, or next event loop
193 return $LoopTimeout unless @Timers;
198 while (@Timers && $Timers[0][0] <= $now) {
199 my $to_run = shift(@Timers);
200 $to_run->[1]->($now) if $to_run->[1];
203 return $LoopTimeout unless @Timers;
205 # convert time to an even number of milliseconds, adding 1
206 # extra, otherwise floating point fun can occur and we'll
207 # call RunTimers like 20-30 times, each returning a timeout
208 # of 0.0000212 seconds
209 my $timeout = int(($Timers[0][0] - $now) * 1000) + 1;
211 # -1 is an infinite timeout, so prefer a real timeout
212 return $timeout if $LoopTimeout == -1;
214 # otherwise pick the lower of our regular timeout and time until
216 return $LoopTimeout if $LoopTimeout < $timeout;
220 ### The epoll-based event loop. Gets installed as EventLoop if IO::Epoll loads
228 my $timeout = RunTimers();
230 # get up to 1000 events
231 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
232 for ($i=0; $i<$evcount; $i++) {
233 # it's possible epoll_wait returned many events, including some at the end
234 # that ones in the front triggered unregister-interest actions. if we
235 # can't find the %sock entry, it's because we're no longer interested
237 $DescriptorMap{$events[$i]->[0]}->event_step;
239 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();
270 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
272 Sets post loop callback function. Pass a subref and it will be
273 called every time the event loop finishes.
275 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
278 The callback function will be passed two parameters: \%DescriptorMap
281 sub SetPostLoopCallback {
282 my ($class, $ref) = @_;
285 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
288 # Internal function: run the post-event callback, send read events
289 # for pushed-back data, and close pending connections. returns 1
290 # if event loop should continue, or 0 to shut it all down.
292 # now we can close sockets that wanted to close during our event processing.
293 # (we didn't want to close them during the loop, as we didn't want fd numbers
294 # being reused and confused during the event loop)
295 while (my $sock = shift @ToClose) {
296 my $fd = fileno($sock);
298 # close the socket. (not a PublicInbox::DS close)
301 # and now we can finally remove the fd from the map. see
302 # comment above in ->close.
303 delete $DescriptorMap{$fd};
307 # by default we keep running, unless a postloop callback (either per-object
308 # or global) cancels it
309 my $keep_running = 1;
311 # now we're at the very end, call callback if defined
312 if (defined $PostLoopCallback) {
313 $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
316 return $keep_running;
319 # map EPOLL* bits to kqueue EV_* flags for EV_SET
323 my $fl = EV_ADD() | EV_ENABLE();
324 ($ev & EPOLLONESHOT) ? ($fl|EV_ONESHOT()) : $fl;
330 #####################################################################
331 ### PublicInbox::DS-the-object code
332 #####################################################################
334 =head2 OBJECT METHODS
336 =head2 C<< CLASS->new( $socket ) >>
338 Create a new PublicInbox::DS subclass object for the given I<socket> which will
339 react to events on it during the C<EventLoop>.
341 This is normally (always?) called from your subclass via:
343 $class->SUPER::new($socket);
347 my ($self, $sock, $ev) = @_;
348 $self = fields::new($self) unless ref $self;
350 $self->{sock} = $sock;
351 my $fd = fileno($sock);
353 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
360 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
361 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
362 $ev &= ~EPOLLEXCLUSIVE;
365 die "couldn't add epoll watch for $fd: $!\n";
368 elsif ($HaveKQueue) {
369 $KQueue->EV_SET($fd, EVFILT_READ(), EV_ADD() | kq_flag(EPOLLIN, $ev));
370 $KQueue->EV_SET($fd, EVFILT_WRITE(), EV_ADD() | kq_flag(EPOLLOUT, $ev));
373 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
374 if $DescriptorMap{$fd};
376 $DescriptorMap{$fd} = $self;
381 #####################################################################
382 ### I N S T A N C E M E T H O D S
383 #####################################################################
385 =head2 C<< $obj->close >>
392 my $sock = delete $self->{sock} or return;
394 # we need to flush our write buffer, as there may
395 # be self-referential closures (sub { $client->close })
396 # preventing the object from being destroyed
397 delete $self->{wbuf};
399 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
400 # notifications about it
402 my $fd = fileno($sock);
403 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
404 confess("EPOLL_CTL_DEL: $!");
407 # we explicitly don't delete from DescriptorMap here until we
408 # actually close the socket, as we might be in the middle of
409 # processing an epoll_wait/etc that returned hundreds of fds, one
410 # of which is not yet processed and is what we're closing. if we
411 # keep it in DescriptorMap, then the event harnesses can just
412 # looked at $pob->{sock} == undef and ignore it. but if it's an
413 # un-accounted for fd, then it (understandably) freak out a bit
414 # and emit warnings, thinking their state got off.
416 # defer closing the actual socket until the event loop is done
417 # processing this round of events. (otherwise we might reuse fds)
418 push @ToClose, $sock;
423 # portable, non-thread-safe sendfile emulation (no pread, yet)
424 sub psendfile ($$$) {
425 my ($sock, $fh, $off) = @_;
427 sysseek($fh, $$off, SEEK_SET) or return;
428 defined(my $to_write = sysread($fh, my $buf, 16384)) or return;
430 while ($to_write > 0) {
431 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
435 return if $written == 0;
443 # returns 1 if done, 0 if incomplete
444 sub flush_write ($) {
446 my $wbuf = $self->{wbuf} or return 1;
447 my $sock = $self->{sock} or return 1;
450 while (my $bref = $wbuf->[0]) {
451 if (ref($bref) ne 'CODE') {
452 my $off = delete($self->{wbuf_off}) // 0;
454 my $w = psendfile($sock, $bref, \$off);
460 } elsif ($! == EAGAIN) {
461 $self->{wbuf_off} = $off;
462 watch($self, EPOLLOUT|EPOLLONESHOT);
468 } else { #($ref eq 'CODE') {
474 delete $self->{wbuf};
478 sub write_in_full ($$$$) {
479 my ($fh, $bref, $len, $off) = @_;
482 my $w = syswrite($fh, $$bref, $len, $off);
483 return ($rv ? $rv : $w) unless $w; # undef or 0
492 my ($bref, $off) = @_;
493 # open(my $fh, '+>>', undef) doesn't set O_APPEND
494 my ($fh, $path) = tempfile('wbuf-XXXXXXX', TMPDIR => 1);
495 open $fh, '+>>', $path or die "open: $!";
497 my $to_write = bytes::length($$bref) - $off;
498 my $w = write_in_full($fh, $bref, $to_write, $off);
499 die "write_in_full ($to_write): $!" unless defined $w;
500 $w == $to_write ? $fh : die("short write $w < $to_write");
503 =head2 C<< $obj->write( $data ) >>
505 Write the specified data to the underlying handle. I<data> may be scalar,
506 scalar ref, code ref (to run when there).
507 Returns 1 if writes all went through, or 0 if there are writes in queue. If
508 it returns 1, caller should stop waiting for 'writable' events)
512 my ($self, $data) = @_;
514 # nobody should be writing to closed sockets, but caller code can
515 # do two writes within an event, have the first fail and
516 # disconnect the other side (whose destructor then closes the
517 # calling object, but it's still in a method), and then the
518 # now-dead object does its second write. that is this case. we
519 # just lie and say it worked. it'll be dead soon and won't be
521 my $sock = $self->{sock} or return 1;
523 my $bref = $ref ? $data : \$data;
524 if (my $wbuf = $self->{wbuf}) { # already buffering, can't write more...
525 if ($ref eq 'CODE') {
528 my $last = $wbuf->[-1];
529 if (ref($last) eq 'GLOB') { # append to tmp file buffer
530 write_in_full($last, $bref, bytes::length($$bref), 0);
532 push @$wbuf, tmpbuf($bref, 0);
536 } elsif ($ref eq 'CODE') {
540 my $to_write = bytes::length($$bref);
541 my $written = syswrite($sock, $$bref, $to_write);
543 if (defined $written) {
544 return 1 if $written == $to_write;
545 } elsif ($! == EAGAIN) {
550 $self->{wbuf} = [ tmpbuf($bref, $written) ];
551 watch($self, EPOLLOUT|EPOLLONESHOT);
556 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
560 my $sock = $self->{sock} or return 1;
562 if (MSG_MORE && !$self->{wbuf}) {
563 my $n = send($sock, $_[1], MSG_MORE);
565 my $nlen = bytes::length($_[1]) - $n;
566 return 1 if $nlen == 0; # all done!
568 # queue up the unwritten substring:
569 $self->{wbuf} = [ tmpbuf(\($_[1]), $n) ];
570 watch($self, EPOLLOUT|EPOLLONESHOT);
574 $self->write(\($_[1]));
578 my ($self, $ev) = @_;
579 my $sock = $self->{sock} or return;
580 my $fd = fileno($sock);
582 epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $ev) and
583 confess("EPOLL_CTL_MOD $!");
584 } elsif ($HaveKQueue) {
585 $KQueue->EV_SET($fd, EVFILT_READ(), kq_flag(EPOLLIN, $ev));
586 $KQueue->EV_SET($fd, EVFILT_WRITE(), kq_flag(EPOLLOUT, $ev));
590 sub watch_in1 ($) { watch($_[0], EPOLLIN | EPOLLONESHOT) }
592 package PublicInbox::DS::Timer;
593 # [$abs_float_firetime, $coderef];
600 =head1 AUTHORS (Danga::Socket)
602 Brad Fitzpatrick <brad@danga.com> - author
604 Michael Granger <ged@danga.com> - docs, testing
606 Mark Smith <junior@danga.com> - contributor, heavy user, testing
608 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits