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; IO::KQueue->import; 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
60 #####################################################################
61 ### C L A S S M E T H O D S
62 #####################################################################
64 =head2 C<< CLASS->Reset() >>
72 $LoopTimeout = -1; # no timeout by default
75 $PostLoopCallback = undef;
78 # NOTE kqueue is close-on-fork, and we don't account for it, yet
79 # OTOH, we (public-inbox) don't need this sub outside of tests...
80 POSIX::close($$KQueue) if !$_io && $KQueue && $$KQueue >= 0;
83 $_io = undef; # close $Epoll
86 *EventLoop = *FirstTimeEventLoop;
89 =head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
91 Set the loop timeout for the event loop to some value in milliseconds.
93 A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return
98 return $LoopTimeout = $_[1] + 0;
101 =head2 C<< CLASS->AddTimer( $seconds, $coderef ) >>
103 Add a timer to occur $seconds from now. $seconds may be fractional, but timers
104 are not guaranteed to fire at the exact time you ask for.
106 Returns a timer object which you can call C<< $timer->cancel >> on if you need to.
110 my ($class, $secs, $coderef) = @_;
113 my $timer = bless([0, $coderef], 'PublicInbox::DS::Timer');
114 unshift(@Timers, $timer);
118 my $fire_time = now() + $secs;
120 my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer";
122 if (!@Timers || $fire_time >= $Timers[-1][0]) {
123 push @Timers, $timer;
127 # Now, where do we insert? (NOTE: this appears slow, algorithm-wise,
128 # but it was compared against calendar queues, heaps, naive push/sort,
129 # and a bunch of other versions, and found to be fastest with a large
130 # variety of datasets.)
131 for (my $i = 0; $i < @Timers; $i++) {
132 if ($Timers[$i][0] > $fire_time) {
133 splice(@Timers, $i, 0, $timer);
138 die "Shouldn't get here.";
141 # keeping this around in case we support other FD types for now,
142 # epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
143 sub set_cloexec ($) {
146 $_io = IO::Handle->new_from_fd($fd, 'r+') or return;
147 defined(my $fl = fcntl($_io, F_GETFD, 0)) or return;
148 fcntl($_io, F_SETFD, $fl | FD_CLOEXEC);
157 $KQueue = IO::KQueue->new();
158 $HaveKQueue = defined $KQueue;
160 *EventLoop = *KQueueEventLoop;
163 elsif (PublicInbox::Syscall::epoll_defined()) {
164 $Epoll = eval { epoll_create(1024); };
165 $HaveEpoll = defined $Epoll && $Epoll >= 0;
168 *EventLoop = *EpollEventLoop;
173 =head2 C<< CLASS->EventLoop() >>
175 Start processing IO events. In most daemon programs this never exits. See
176 C<PostLoopCallback> below for how to exit the loop.
179 sub FirstTimeEventLoop {
185 EpollEventLoop($class);
186 } elsif ($HaveKQueue) {
187 KQueueEventLoop($class);
191 sub now () { clock_gettime(CLOCK_MONOTONIC) }
193 # runs timers and returns milliseconds for next one, or next event loop
195 return $LoopTimeout unless @Timers;
200 while (@Timers && $Timers[0][0] <= $now) {
201 my $to_run = shift(@Timers);
202 $to_run->[1]->($now) if $to_run->[1];
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 ### The epoll-based event loop. Gets installed as EventLoop if IO::Epoll loads
230 my $timeout = RunTimers();
232 # get up to 1000 events
233 my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
234 for ($i=0; $i<$evcount; $i++) {
235 # it's possible epoll_wait returned many events, including some at the end
236 # that ones in the front triggered unregister-interest actions. if we
237 # can't find the %sock entry, it's because we're no longer interested
239 $DescriptorMap{$events[$i]->[0]}->event_step;
241 return unless PostEventLoop();
246 ### The kqueue-based event loop. Gets installed as EventLoop if IO::KQueue works
248 sub KQueueEventLoop {
252 my $timeout = RunTimers();
253 my @ret = eval { $KQueue->kevent($timeout) };
255 # workaround https://rt.cpan.org/Ticket/Display.html?id=116615
256 if ($err =~ /Interrupted system call/) {
263 foreach my $kev (@ret) {
264 $DescriptorMap{$kev->[0]}->event_step;
266 return unless PostEventLoop();
272 =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >>
274 Sets post loop callback function. Pass a subref and it will be
275 called every time the event loop finishes.
277 Return 1 (or any true value) from the sub to make the loop continue, 0 or false
280 The callback function will be passed two parameters: \%DescriptorMap
283 sub SetPostLoopCallback {
284 my ($class, $ref) = @_;
287 $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef;
290 # Internal function: run the post-event callback, send read events
291 # for pushed-back data, and close pending connections. returns 1
292 # if event loop should continue, or 0 to shut it all down.
294 # now we can close sockets that wanted to close during our event processing.
295 # (we didn't want to close them during the loop, as we didn't want fd numbers
296 # being reused and confused during the event loop)
297 while (my $sock = shift @ToClose) {
298 my $fd = fileno($sock);
300 # close the socket. (not a PublicInbox::DS close)
303 # and now we can finally remove the fd from the map. see
304 # comment above in ->close.
305 delete $DescriptorMap{$fd};
309 # by default we keep running, unless a postloop callback (either per-object
310 # or global) cancels it
311 my $keep_running = 1;
313 # now we're at the very end, call callback if defined
314 if (defined $PostLoopCallback) {
315 $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
318 return $keep_running;
321 #####################################################################
322 ### PublicInbox::DS-the-object code
323 #####################################################################
325 =head2 OBJECT METHODS
327 =head2 C<< CLASS->new( $socket ) >>
329 Create a new PublicInbox::DS subclass object for the given I<socket> which will
330 react to events on it during the C<EventLoop>.
332 This is normally (always?) called from your subclass via:
334 $class->SUPER::new($socket);
338 my ($self, $sock, $ev) = @_;
339 $self = fields::new($self) unless ref $self;
341 $self->{sock} = $sock;
342 my $fd = fileno($sock);
344 Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || ""))
347 $self->{event_watch} = $ev;
353 if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
354 if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) {
355 $self->{event_watch} = ($ev &= ~EPOLLEXCLUSIVE);
358 die "couldn't add epoll watch for $fd: $!\n";
361 elsif ($HaveKQueue) {
362 my $f = $ev & EPOLLIN ? EV_ENABLE() : EV_DISABLE();
363 $KQueue->EV_SET($fd, EVFILT_READ(), EV_ADD() | $f);
364 $f = $ev & EPOLLOUT ? EV_ENABLE() : EV_DISABLE();
365 $KQueue->EV_SET($fd, EVFILT_WRITE(), EV_ADD() | $f);
368 Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})")
369 if $DescriptorMap{$fd};
371 $DescriptorMap{$fd} = $self;
376 #####################################################################
377 ### I N S T A N C E M E T H O D S
378 #####################################################################
380 =head2 C<< $obj->close >>
387 my $sock = delete $self->{sock} or return;
389 # we need to flush our write buffer, as there may
390 # be self-referential closures (sub { $client->close })
391 # preventing the object from being destroyed
392 delete $self->{wbuf};
394 # if we're using epoll, we have to remove this from our epoll fd so we stop getting
395 # notifications about it
397 my $fd = fileno($sock);
398 epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and
399 confess("EPOLL_CTL_DEL: $!");
402 # we explicitly don't delete from DescriptorMap here until we
403 # actually close the socket, as we might be in the middle of
404 # processing an epoll_wait/etc that returned hundreds of fds, one
405 # of which is not yet processed and is what we're closing. if we
406 # keep it in DescriptorMap, then the event harnesses can just
407 # looked at $pob->{sock} == undef and ignore it. but if it's an
408 # un-accounted for fd, then it (understandably) freak out a bit
409 # and emit warnings, thinking their state got off.
411 # defer closing the actual socket until the event loop is done
412 # processing this round of events. (otherwise we might reuse fds)
413 push @ToClose, $sock;
418 # portable, non-thread-safe sendfile emulation (no pread, yet)
419 sub psendfile ($$$) {
420 my ($sock, $fh, $off) = @_;
422 sysseek($fh, $$off, SEEK_SET) or return;
423 defined(my $to_write = sysread($fh, my $buf, 16384)) or return;
425 while ($to_write > 0) {
426 if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) {
430 return if $written == 0;
438 # returns 1 if done, 0 if incomplete
439 sub flush_write ($) {
441 my $wbuf = $self->{wbuf} or return 1;
442 my $sock = $self->{sock} or return 1;
445 while (my $bref = $wbuf->[0]) {
446 if (ref($bref) ne 'CODE') {
447 my $off = delete($self->{wbuf_off}) // 0;
449 my $w = psendfile($sock, $bref, \$off);
455 } elsif ($! == EAGAIN) {
456 $self->{wbuf_off} = $off;
457 watch_write($self, 1);
463 } else { #($ref eq 'CODE') {
469 delete $self->{wbuf};
470 $self->watch_write(0);
474 sub write_in_full ($$$$) {
475 my ($fh, $bref, $len, $off) = @_;
478 my $w = syswrite($fh, $$bref, $len, $off);
479 return ($rv ? $rv : $w) unless $w; # undef or 0
488 my ($bref, $off) = @_;
489 # open(my $fh, '+>>', undef) doesn't set O_APPEND
490 my ($fh, $path) = tempfile('wbuf-XXXXXXX', TMPDIR => 1);
491 open $fh, '+>>', $path or die "open: $!";
493 my $to_write = bytes::length($$bref) - $off;
494 my $w = write_in_full($fh, $bref, $to_write, $off);
495 die "write_in_full ($to_write): $!" unless defined $w;
496 $w == $to_write ? $fh : die("short write $w < $to_write");
499 =head2 C<< $obj->write( $data ) >>
501 Write the specified data to the underlying handle. I<data> may be scalar,
502 scalar ref, code ref (to run when there).
503 Returns 1 if writes all went through, or 0 if there are writes in queue. If
504 it returns 1, caller should stop waiting for 'writable' events)
508 my ($self, $data) = @_;
510 # nobody should be writing to closed sockets, but caller code can
511 # do two writes within an event, have the first fail and
512 # disconnect the other side (whose destructor then closes the
513 # calling object, but it's still in a method), and then the
514 # now-dead object does its second write. that is this case. we
515 # just lie and say it worked. it'll be dead soon and won't be
517 my $sock = $self->{sock} or return 1;
519 my $bref = $ref ? $data : \$data;
520 if (my $wbuf = $self->{wbuf}) { # already buffering, can't write more...
521 if ($ref eq 'CODE') {
524 my $last = $wbuf->[-1];
525 if (ref($last) eq 'GLOB') { # append to tmp file buffer
526 write_in_full($last, $bref, bytes::length($$bref), 0);
528 push @$wbuf, tmpbuf($bref, 0);
532 } elsif ($ref eq 'CODE') {
536 my $to_write = bytes::length($$bref);
537 my $written = syswrite($sock, $$bref, $to_write);
539 if (defined $written) {
540 return 1 if $written == $to_write;
541 } elsif ($! == EAGAIN) {
546 $self->{wbuf} = [ tmpbuf($bref, $written) ];
547 watch_write($self, 1);
552 use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0;
556 my $sock = $self->{sock} or return 1;
558 if (MSG_MORE && !$self->{wbuf}) {
559 my $n = send($sock, $_[1], MSG_MORE);
561 my $nlen = bytes::length($_[1]) - $n;
562 return 1 if $nlen == 0; # all done!
564 # queue up the unwritten substring:
565 $self->{wbuf} = [ tmpbuf(\($_[1]), $n) ];
566 watch_write($self, 1);
570 $self->write(\($_[1]));
573 sub watch_chg ($$$) {
574 my ($self, $bits, $set) = @_;
575 my $sock = $self->{sock} or return;
576 my $cur = $self->{event_watch};
583 return if $changes == $cur;
584 my $fd = fileno($sock);
586 epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $changes) and
587 confess("EPOLL_CTL_MOD $!");
588 } elsif ($HaveKQueue) {
589 my $flag = $set ? EV_ENABLE() : EV_DISABLE();
590 $KQueue->EV_SET($fd, EVFILT_READ(), $flag) if $bits & EPOLLIN;
591 $KQueue->EV_SET($fd, EVFILT_WRITE(), $flag) if $bits & EPOLLOUT;
593 $self->{event_watch} = $changes;
596 =head2 C<< $obj->watch_read( $boolean ) >>
598 Turn 'readable' event notification on or off.
601 sub watch_read ($$) { watch_chg($_[0], EPOLLIN, $_[1]) };
603 =head2 C<< $obj->watch_write( $boolean ) >>
605 Turn 'writable' event notification on or off.
608 sub watch_write ($$) { watch_chg($_[0], EPOLLOUT, $_[1]) };
610 package PublicInbox::DS::Timer;
611 # [$abs_float_firetime, $coderef];
618 =head1 AUTHORS (Danga::Socket)
620 Brad Fitzpatrick <brad@danga.com> - author
622 Michael Granger <ged@danga.com> - docs, testing
624 Mark Smith <junior@danga.com> - contributor, heavy user, testing
626 Matt Sergeant <matt@sergeant.org> - kqueue support, docs, timers, other bits