X-Git-Url: http://www.git.stargrave.org/?a=blobdiff_plain;f=lib%2FPublicInbox%2FDS.pm;h=cea25d90e37516c354e94410148f8299fa64cb15;hb=eb9c415ba4421cecb5157967c843dc7f8720e916;hp=737f4c7a8ca366fd44535014a9f9ba1092d3a906;hpb=e36db4b8acdad26ef70a45d8912e8f3c5f2d9910;p=public-inbox.git diff --git a/lib/PublicInbox/DS.pm b/lib/PublicInbox/DS.pm index 737f4c7a..cea25d90 100644 --- a/lib/PublicInbox/DS.pm +++ b/lib/PublicInbox/DS.pm @@ -3,63 +3,61 @@ # # This license differs from the rest of public-inbox # -# This is a fork of the (for now) unmaintained Danga::Socket 1.61. -# Unused features will be removed, and updates will be made to take -# advantage of newer kernels - +# This is a fork of the unmaintained Danga::Socket (1.61) with +# significant changes. See Documentation/technical/ds.txt in our +# source for details. +# +# Do not expect this to be a stable API like Danga::Socket, +# but it will evolve to suite our needs and to take advantage of +# newer Linux and *BSD features. +# Bugs encountered were reported to bug-Danga-Socket@rt.cpan.org, +# fixed in Danga::Socket 1.62 and visible at: +# https://rt.cpan.org/Public/Dist/Display.html?Name=Danga-Socket package PublicInbox::DS; use strict; use bytes; -use POSIX (); -use Time::HiRes (); - +use POSIX qw(WNOHANG); +use IO::Handle qw(); +use Fcntl qw(SEEK_SET :DEFAULT); +use Time::HiRes qw(clock_gettime CLOCK_MONOTONIC); +use parent qw(Exporter); +our @EXPORT_OK = qw(now msg_more); use warnings; +use 5.010_001; +use Scalar::Util qw(blessed); use PublicInbox::Syscall qw(:epoll); +use PublicInbox::Tmpfile; use fields ('sock', # underlying socket - 'fd', # numeric file descriptor - 'write_buf', # arrayref of scalars, scalarrefs, or coderefs to write - 'write_buf_offset', # offset into first array of write_buf to start writing at - 'write_buf_size', # total length of data in all write_buf items - 'write_set_watch', # bool: true if we internally set watch_write rather than by a subclass - 'closed', # bool: socket is closed - 'event_watch', # bitmask of events the client is interested in (POLLIN,OUT,etc.) - 'writer_func', # subref which does writing. must return bytes written (or undef) and set $! on errors + 'rbuf', # scalarref, usually undef + 'wbuf', # arrayref of coderefs or GLOB refs + 'wbuf_off', # offset into first element of wbuf to start writing at ); -use Errno qw(EINPROGRESS EWOULDBLOCK EISCONN ENOTSOCK - EPIPE EAGAIN EBADF ECONNRESET ENOPROTOOPT); -use Carp qw(croak confess); - -use constant DebugLevel => 0; - -use constant POLLIN => 1; -use constant POLLOUT => 4; -use constant POLLERR => 8; -use constant POLLHUP => 16; -use constant POLLNVAL => 32; - -our $HAVE_KQUEUE = eval { require IO::KQueue; 1 }; +use Errno qw(EAGAIN EINVAL); +use Carp qw(confess carp); +my $nextq; # queue for next_tick +my $WaitPids; # list of [ pid, callback, callback_arg ] +my $later_queue; # callbacks +my $EXPMAP; # fd -> [ idle_time, $self ] +our $EXPTIME = 180; # 3 minutes +my ($later_timer, $reap_timer, $exp_timer); our ( - $HaveEpoll, # Flag -- is epoll available? initially undefined. - $HaveKQueue, %DescriptorMap, # fd (num) -> PublicInbox::DS object - $Epoll, # Global epoll fd (for epoll mode only) - $KQueue, # Global kqueue fd (for kqueue mode only) + $Epoll, # Global epoll fd (or DSKQXS ref) + $_io, # IO::Handle for Epoll @ToClose, # sockets to close when event loop is done $PostLoopCallback, # subref to call at the end of each loop, if defined (global) - %PLCMap, # fd (num) -> PostLoopCallback (per-object) $LoopTimeout, # timeout of event loop in milliseconds $DoneInit, # if we've done the one-time module init yet @Timers, # timers + $in_loop, ); -# this may be set to zero with old kernels -our $EPOLLEXCLUSIVE = EPOLLEXCLUSIVE; Reset(); ##################################################################### @@ -73,38 +71,24 @@ Reset all state =cut sub Reset { %DescriptorMap = (); + $nextq = []; + $WaitPids = []; + $later_queue = []; + $EXPMAP = {}; + $reap_timer = $later_timer = $exp_timer = undef; @ToClose = (); $LoopTimeout = -1; # no timeout by default @Timers = (); $PostLoopCallback = undef; - %PLCMap = (); $DoneInit = 0; - POSIX::close($Epoll) if defined $Epoll && $Epoll >= 0; - POSIX::close($KQueue) if defined $KQueue && $KQueue >= 0; + $_io = undef; # closes real $Epoll FD + $Epoll = undef; # may call DSKQXS::DESTROY *EventLoop = *FirstTimeEventLoop; } -=head2 C<< CLASS->HaveEpoll() >> - -Returns a true value if this class will use IO::Epoll for async IO. - -=cut -sub HaveEpoll { - _InitPoller(); - return $HaveEpoll; -} - -=head2 C<< CLASS->ToClose() >> - -Return the list of sockets that are awaiting close() at the end of the -current event loop. - -=cut -sub ToClose { return @ToClose; } - =head2 C<< CLASS->SetLoopTimeout( $timeout ) >> Set the loop timeout for the event loop to some value in milliseconds. @@ -117,33 +101,18 @@ sub SetLoopTimeout { return $LoopTimeout = $_[1] + 0; } -=head2 C<< CLASS->DebugMsg( $format, @args ) >> - -Print the debugging message specified by the C-style I and -I - -=cut -sub DebugMsg { - my ( $class, $fmt, @args ) = @_; - chomp $fmt; - printf STDERR ">>> $fmt\n", @args; -} - -=head2 C<< CLASS->AddTimer( $seconds, $coderef ) >> +=head2 C<< PublicInbox::DS::add_timer( $seconds, $coderef ) >> Add a timer to occur $seconds from now. $seconds may be fractional, but timers are not guaranteed to fire at the exact time you ask for. -Returns a timer object which you can call C<< $timer->cancel >> on if you need to. - =cut -sub AddTimer { - my $class = shift; +sub add_timer ($$) { my ($secs, $coderef) = @_; - my $fire_time = Time::HiRes::time() + $secs; + my $fire_time = now() + $secs; - my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer"; + my $timer = [$fire_time, $coderef]; if (!@Timers || $fire_time >= $Timers[-1][0]) { push @Timers, $timer; @@ -164,30 +133,34 @@ sub AddTimer { die "Shouldn't get here."; } +# keeping this around in case we support other FD types for now, +# epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+... +sub set_cloexec ($) { + my ($fd) = @_; + + $_io = IO::Handle->new_from_fd($fd, 'r+') or return; + defined(my $fl = fcntl($_io, F_GETFD, 0)) or return; + fcntl($_io, F_SETFD, $fl | FD_CLOEXEC); +} + sub _InitPoller { return if $DoneInit; $DoneInit = 1; - if ($HAVE_KQUEUE) { - $KQueue = IO::KQueue->new(); - $HaveKQueue = $KQueue >= 0; - if ($HaveKQueue) { - *EventLoop = *KQueueEventLoop; - } - } - elsif (PublicInbox::Syscall::epoll_defined()) { - $Epoll = eval { epoll_create(1024); }; - $HaveEpoll = defined $Epoll && $Epoll >= 0; - if ($HaveEpoll) { - *EventLoop = *EpollEventLoop; + if (PublicInbox::Syscall::epoll_defined()) { + $Epoll = epoll_create(); + set_cloexec($Epoll) if (defined($Epoll) && $Epoll >= 0); + } else { + my $cls; + for (qw(DSKQXS DSPoll)) { + $cls = "PublicInbox::$_"; + last if eval "require $cls"; } + $cls->import(qw(epoll_ctl epoll_wait)); + $Epoll = $cls->new; } - - if (!$HaveEpoll && !$HaveKQueue) { - require IO::Poll; - *EventLoop = *PollEventLoop; - } + *EventLoop = *EpollEventLoop; } =head2 C<< CLASS->EventLoop() >> @@ -201,20 +174,32 @@ sub FirstTimeEventLoop { _InitPoller(); - if ($HaveEpoll) { - EpollEventLoop($class); - } elsif ($HaveKQueue) { - KQueueEventLoop($class); - } else { - PollEventLoop($class); + EventLoop($class); +} + +sub now () { clock_gettime(CLOCK_MONOTONIC) } + +sub next_tick () { + my $q = $nextq; + $nextq = []; + for (@$q) { + # we avoid "ref" on blessed refs to workaround a Perl 5.16.3 leak: + # https://rt.perl.org/Public/Bug/Display.html?id=114340 + if (blessed($_)) { + $_->event_step; + } else { + $_->(); + } } } # runs timers and returns milliseconds for next one, or next event loop sub RunTimers { - return $LoopTimeout unless @Timers; + next_tick(); + + return ((@$nextq || @ToClose) ? 0 : $LoopTimeout) unless @Timers; - my $now = Time::HiRes::time(); + my $now = now(); # Run expired timers while (@Timers && $Timers[0][0] <= $now) { @@ -222,6 +207,9 @@ sub RunTimers { $to_run->[1]->($now) if $to_run->[1]; } + # timers may enqueue into nextq: + return 0 if (@$nextq || @ToClose); + return $LoopTimeout unless @Timers; # convert time to an even number of milliseconds, adding 1 @@ -239,155 +227,51 @@ sub RunTimers { return $timeout; } -### The epoll-based event loop. Gets installed as EventLoop if IO::Epoll loads -### okay. -sub EpollEventLoop { - my $class = shift; +# We can't use waitpid(-1) safely here since it can hit ``, system(), +# and other things. So we scan the $WaitPids list, which is hopefully +# not too big. +sub reap_pids { + my $tmp = $WaitPids; + $WaitPids = []; + $reap_timer = undef; + foreach my $ary (@$tmp) { + my ($pid, $cb, $arg) = @$ary; + my $ret = waitpid($pid, WNOHANG); + if ($ret == 0) { + push @$WaitPids, $ary; + } elsif ($cb) { + eval { $cb->($arg, $pid) }; + } + } + if (@$WaitPids) { + # we may not be donea, and we may miss our + $reap_timer = add_timer(1, \&reap_pids); + } +} - while (1) { +# reentrant SIGCHLD handler (since reap_pids is not reentrant) +sub enqueue_reap ($) { push @$nextq, \&reap_pids }; + +sub in_loop () { $in_loop } + +sub EpollEventLoop { + local $in_loop = 1; + do { my @events; my $i; my $timeout = RunTimers(); # get up to 1000 events my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events); - EVENT: for ($i=0; $i<$evcount; $i++) { - my $ev = $events[$i]; - # it's possible epoll_wait returned many events, including some at the end # that ones in the front triggered unregister-interest actions. if we # can't find the %sock entry, it's because we're no longer interested # in that event. - my PublicInbox::DS $pob = $DescriptorMap{$ev->[0]}; - my $code; - my $state = $ev->[1]; - - # if we didn't find a Perlbal::Socket subclass for that fd, try other - # pseudo-registered (above) fds. - if (! $pob) { - my $fd = $ev->[0]; - warn "epoll() returned fd $fd w/ state $state for which we have no mapping. removing.\n"; - epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0); - POSIX::close($fd); - next; - } - - DebugLevel >= 1 && $class->DebugMsg("Event: fd=%d (%s), state=%d \@ %s\n", - $ev->[0], ref($pob), $ev->[1], time); - - # standard non-profiling codepat - $pob->event_read if $state & EPOLLIN && ! $pob->{closed}; - $pob->event_write if $state & EPOLLOUT && ! $pob->{closed}; - if ($state & (EPOLLERR|EPOLLHUP)) { - $pob->event_err if $state & EPOLLERR && ! $pob->{closed}; - $pob->event_hup if $state & EPOLLHUP && ! $pob->{closed}; - } - } - return unless PostEventLoop(); - } - exit 0; -} - -### The fallback IO::Poll-based event loop. Gets installed as EventLoop if -### IO::Epoll fails to load. -sub PollEventLoop { - my $class = shift; - - my PublicInbox::DS $pob; - - while (1) { - my $timeout = RunTimers(); - - # the following sets up @poll as a series of ($poll,$event_mask) - # items, then uses IO::Poll::_poll, implemented in XS, which - # modifies the array in place with the even elements being - # replaced with the event masks that occured. - my @poll; - while ( my ($fd, $sock) = each %DescriptorMap ) { - push @poll, $fd, $sock->{event_watch}; - } - - # if nothing to poll, either end immediately (if no timeout) - # or just keep calling the callback - unless (@poll) { - select undef, undef, undef, ($timeout / 1000); - return unless PostEventLoop(); - next; - } - - my $count = IO::Poll::_poll($timeout, @poll); - unless ($count >= 0) { - return unless PostEventLoop(); - next; - } - - # Fetch handles with read events - while (@poll) { - my ($fd, $state) = splice(@poll, 0, 2); - next unless $state; - - $pob = $DescriptorMap{$fd}; - - if (!$pob) { - next; - } - - $pob->event_read if $state & POLLIN && ! $pob->{closed}; - $pob->event_write if $state & POLLOUT && ! $pob->{closed}; - $pob->event_err if $state & POLLERR && ! $pob->{closed}; - $pob->event_hup if $state & POLLHUP && ! $pob->{closed}; + $DescriptorMap{$events[$i]->[0]}->event_step; } - - return unless PostEventLoop(); - } - - exit 0; -} - -### The kqueue-based event loop. Gets installed as EventLoop if IO::KQueue works -### okay. -sub KQueueEventLoop { - my $class = shift; - - while (1) { - my $timeout = RunTimers(); - my @ret = eval { $KQueue->kevent($timeout) }; - if (my $err = $@) { - # workaround https://rt.cpan.org/Ticket/Display.html?id=116615 - if ($err =~ /Interrupted system call/) { - @ret = (); - } else { - die $err; - } - } - - foreach my $kev (@ret) { - my ($fd, $filter, $flags, $fflags) = @$kev; - my PublicInbox::DS $pob = $DescriptorMap{$fd}; - if (!$pob) { - warn "kevent() returned fd $fd for which we have no mapping. removing.\n"; - POSIX::close($fd); # close deletes the kevent entry - next; - } - - DebugLevel >= 1 && $class->DebugMsg("Event: fd=%d (%s), flags=%d \@ %s\n", - $fd, ref($pob), $flags, time); - - $pob->event_read if $filter == IO::KQueue::EVFILT_READ() && !$pob->{closed}; - $pob->event_write if $filter == IO::KQueue::EVFILT_WRITE() && !$pob->{closed}; - if ($flags == IO::KQueue::EV_EOF() && !$pob->{closed}) { - if ($fflags) { - $pob->event_err; - } else { - $pob->event_hup; - } - } - } - return unless PostEventLoop(); - } - - exit(0); + } while (PostEventLoop()); + _run_later(); } =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >> @@ -404,18 +288,8 @@ The callback function will be passed two parameters: \%DescriptorMap sub SetPostLoopCallback { my ($class, $ref) = @_; - if (ref $class) { - # per-object callback - my PublicInbox::DS $self = $class; - if (defined $ref && ref $ref eq 'CODE') { - $PLCMap{$self->{fd}} = $ref; - } else { - delete $PLCMap{$self->{fd}}; - } - } else { - # global callback - $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef; - } + # global callback + $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef; } # Internal function: run the post-event callback, send read events @@ -425,27 +299,13 @@ sub PostEventLoop { # now we can close sockets that wanted to close during our event processing. # (we didn't want to close them during the loop, as we didn't want fd numbers # being reused and confused during the event loop) - while (my $sock = shift @ToClose) { - my $fd = fileno($sock); - - # close the socket. (not a PublicInbox::DS close) - $sock->close; - - # and now we can finally remove the fd from the map. see - # comment above in _cleanup. - delete $DescriptorMap{$fd}; - } - + delete($DescriptorMap{fileno($_)}) for @ToClose; + @ToClose = (); # let refcounting drop everything all at once # by default we keep running, unless a postloop callback (either per-object # or global) cancels it my $keep_running = 1; - # per-object post-loop-callbacks - for my $plc (values %PLCMap) { - $keep_running &&= $plc->(\%DescriptorMap); - } - # now we're at the very end, call callback if defined if (defined $PostLoopCallback) { $keep_running &&= $PostLoopCallback->(\%DescriptorMap); @@ -471,52 +331,25 @@ This is normally (always?) called from your subclass via: =cut sub new { - my ($self, $sock, $exclusive) = @_; + my ($self, $sock, $ev) = @_; $self = fields::new($self) unless ref $self; - $self->{sock} = $sock; + $self->{sock} = $sock; my $fd = fileno($sock); - Carp::cluck("undef sock and/or fd in PublicInbox::DS->new. sock=" . ($sock || "") . ", fd=" . ($fd || "")) - unless $sock && $fd; - - $self->{fd} = $fd; - $self->{write_buf} = []; - $self->{write_buf_offset} = 0; - $self->{write_buf_size} = 0; - $self->{closed} = 0; - - my $ev = $self->{event_watch} = POLLERR|POLLHUP|POLLNVAL; - _InitPoller(); - if ($HaveEpoll) { - if ($exclusive) { - $ev = $self->{event_watch} = EPOLLIN|EPOLLERR|EPOLLHUP|$EPOLLEXCLUSIVE; - } -retry: - if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) { - if ($!{EINVAL} && ($ev & $EPOLLEXCLUSIVE)) { - $EPOLLEXCLUSIVE = 0; # old kernel - $ev = $self->{event_watch} = EPOLLIN|EPOLLERR|EPOLLHUP; - goto retry; - } - die "couldn't add epoll watch for $fd: $!\n"; + if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) { + if ($! == EINVAL && ($ev & EPOLLEXCLUSIVE)) { + $ev &= ~EPOLLEXCLUSIVE; + goto retry; } + die "couldn't add epoll watch for $fd: $!\n"; } - elsif ($HaveKQueue) { - # Add them to the queue but disabled for now - $KQueue->EV_SET($fd, IO::KQueue::EVFILT_READ(), - IO::KQueue::EV_ADD() | IO::KQueue::EV_DISABLE()); - $KQueue->EV_SET($fd, IO::KQueue::EVFILT_WRITE(), - IO::KQueue::EV_ADD() | IO::KQueue::EV_DISABLE()); - } - - Carp::cluck("PublicInbox::DS::new blowing away existing descriptor map for fd=$fd ($DescriptorMap{$fd})") - if $DescriptorMap{$fd}; + confess("DescriptorMap{$fd} defined ($DescriptorMap{$fd})") + if defined($DescriptorMap{$fd}); $DescriptorMap{$fd} = $self; - return $self; } @@ -524,128 +357,161 @@ retry: ### I N S T A N C E M E T H O D S ##################################################################### -=head2 C<< $obj->steal_socket() >> +sub requeue ($) { push @$nextq, $_[0] } -Basically returns our socket and makes it so that we don't try to close it, -but we do remove it from epoll handlers. THIS CLOSES $self. It is the same -thing as calling close, except it gives you the socket to use. +=head2 C<< $obj->close >> -=cut -sub steal_socket { - my PublicInbox::DS $self = $_[0]; - return if $self->{closed}; - - # cleanup does most of the work of closing this socket - $self->_cleanup(); - - # now undef our internal sock and fd structures so we don't use them - my $sock = $self->{sock}; - $self->{sock} = undef; - return $sock; -} - -=head2 C<< $obj->close( [$reason] ) >> - -Close the socket. The I argument will be used in debugging messages. +Close the socket. =cut sub close { - my PublicInbox::DS $self = $_[0]; - return if $self->{closed}; - - # print out debugging info for this close - if (DebugLevel) { - my ($pkg, $filename, $line) = caller; - my $reason = $_[1] || ""; - warn "Closing \#$self->{fd} due to $pkg/$filename/$line ($reason)\n"; - } - - # this does most of the work of closing us - $self->_cleanup(); - - # defer closing the actual socket until the event loop is done - # processing this round of events. (otherwise we might reuse fds) - if ($self->{sock}) { - push @ToClose, $self->{sock}; - $self->{sock} = undef; - } - - return 0; -} - -### METHOD: _cleanup() -### Called by our closers so we can clean internal data structures. -sub _cleanup { - my PublicInbox::DS $self = $_[0]; - - # we're effectively closed; we have no fd and sock when we leave here - $self->{closed} = 1; + my ($self) = @_; + my $sock = delete $self->{sock} or return; # we need to flush our write buffer, as there may # be self-referential closures (sub { $client->close }) # preventing the object from being destroyed - $self->{write_buf} = []; + delete $self->{wbuf}; # if we're using epoll, we have to remove this from our epoll fd so we stop getting # notifications about it - if ($HaveEpoll && $self->{fd}) { - if (epoll_ctl($Epoll, EPOLL_CTL_DEL, $self->{fd}, $self->{event_watch}) != 0) { - # dump_error prints a backtrace so we can try to figure out why this happened - $self->dump_error("epoll_ctl(): failure deleting fd=$self->{fd} during _cleanup(); $! (" . ($!+0) . ")"); - } - } - - # now delete from mappings. this fd no longer belongs to us, so we don't want - # to get alerts for it if it becomes writable/readable/etc. - delete $PLCMap{$self->{fd}}; + my $fd = fileno($sock); + epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and + confess("EPOLL_CTL_DEL: $!"); # we explicitly don't delete from DescriptorMap here until we # actually close the socket, as we might be in the middle of # processing an epoll_wait/etc that returned hundreds of fds, one # of which is not yet processed and is what we're closing. if we # keep it in DescriptorMap, then the event harnesses can just - # looked at $pob->{closed} and ignore it. but if it's an + # looked at $pob->{sock} == undef and ignore it. but if it's an # un-accounted for fd, then it (understandably) freak out a bit # and emit warnings, thinking their state got off. - # and finally get rid of our fd so we can't use it anywhere else - $self->{fd} = undef; + # defer closing the actual socket until the event loop is done + # processing this round of events. (otherwise we might reuse fds) + push @ToClose, $sock; + + return 0; } -=head2 C<< $obj->sock() >> +# portable, non-thread-safe sendfile emulation (no pread, yet) +sub psendfile ($$$) { + my ($sock, $fh, $off) = @_; + + seek($fh, $$off, SEEK_SET) or return; + defined(my $to_write = read($fh, my $buf, 16384)) or return; + my $written = 0; + while ($to_write > 0) { + if (defined(my $w = syswrite($sock, $buf, $to_write, $written))) { + $written += $w; + $to_write -= $w; + } else { + return if $written == 0; + last; + } + } + $$off += $written; + $written; +} + +sub epbit ($$) { # (sock, default) + ref($_[0]) eq 'IO::Socket::SSL' ? PublicInbox::TLS::epollbit() : $_[1]; +} -Returns the underlying IO::Handle for the object. +# returns 1 if done, 0 if incomplete +sub flush_write ($) { + my ($self) = @_; + my $wbuf = $self->{wbuf} or return 1; + my $sock = $self->{sock}; -=cut -sub sock { - my PublicInbox::DS $self = shift; - return $self->{sock}; +next_buf: + while (my $bref = $wbuf->[0]) { + if (ref($bref) ne 'CODE') { + my $off = delete($self->{wbuf_off}) // 0; + while ($sock) { + my $w = psendfile($sock, $bref, \$off); + if (defined $w) { + if ($w == 0) { + shift @$wbuf; + goto next_buf; + } + } elsif ($! == EAGAIN) { + epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT); + $self->{wbuf_off} = $off; + return 0; + } else { + return $self->close; + } + } + } else { #($ref eq 'CODE') { + shift @$wbuf; + my $before = scalar(@$wbuf); + $bref->($self); + + # bref may be enqueueing more CODE to call (see accept_tls_step) + return 0 if (scalar(@$wbuf) > $before); + } + } # while @$wbuf + + delete $self->{wbuf}; + 1; # all done } -=head2 C<< $obj->set_writer_func( CODEREF ) >> +sub rbuf_idle ($$) { + my ($self, $rbuf) = @_; + if ($$rbuf eq '') { # who knows how long till we can read again + delete $self->{rbuf}; + } else { + $self->{rbuf} = $rbuf; + } +} + +sub do_read ($$$;$) { + my ($self, $rbuf, $len, $off) = @_; + my $r = sysread(my $sock = $self->{sock}, $$rbuf, $len, $off // 0); + return ($r == 0 ? $self->close : $r) if defined $r; + # common for clients to break connections without warning, + # would be too noisy to log here: + if ($! == EAGAIN) { + epwait($sock, epbit($sock, EPOLLIN) | EPOLLONESHOT); + rbuf_idle($self, $rbuf); + 0; + } else { + $self->close; + } +} -Sets a function to use instead of C when writing data to the socket. +# drop the socket if we hit unrecoverable errors on our system which +# require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE... +sub drop { + my $self = shift; + carp(@_); + $self->close; +} -=cut -sub set_writer_func { - my PublicInbox::DS $self = shift; - my $wtr = shift; - Carp::croak("Not a subref") unless !defined $wtr || UNIVERSAL::isa($wtr, "CODE"); - $self->{writer_func} = $wtr; +# n.b.: use ->write/->read for this buffer to allow compatibility with +# PerlIO::mmap or PerlIO::scalar if needed +sub tmpio ($$$) { + my ($self, $bref, $off) = @_; + my $fh = tmpfile('wbuf', $self->{sock}, 1) or + return drop($self, "tmpfile $!"); + $fh->autoflush(1); + my $len = bytes::length($$bref) - $off; + $fh->write($$bref, $len, $off) or return drop($self, "write ($len): $!"); + $fh } =head2 C<< $obj->write( $data ) >> Write the specified data to the underlying handle. I may be scalar, -scalar ref, code ref (to run when there), or undef just to kick-start. +scalar ref, code ref (to run when there). Returns 1 if writes all went through, or 0 if there are writes in queue. If it returns 1, caller should stop waiting for 'writable' events) =cut sub write { - my PublicInbox::DS $self; - my $data; - ($self, $data) = @_; + my ($self, $data) = @_; # nobody should be writing to closed sockets, but caller code can # do two writes within an event, have the first fail and @@ -654,309 +520,174 @@ sub write { # now-dead object does its second write. that is this case. we # just lie and say it worked. it'll be dead soon and won't be # hurt by this lie. - return 1 if $self->{closed}; - - my $bref; - - # just queue data if there's already a wait - my $need_queue; - - if (defined $data) { - $bref = ref $data ? $data : \$data; - if ($self->{write_buf_size}) { - push @{$self->{write_buf}}, $bref; - $self->{write_buf_size} += ref $bref eq "SCALAR" ? length($$bref) : 1; - return 0; - } - - # this flag says we're bypassing the queue system, knowing we're the - # only outstanding write, and hoping we don't ever need to use it. - # if so later, though, we'll need to queue - $need_queue = 1; - } - - WRITE: - while (1) { - return 1 unless $bref ||= $self->{write_buf}[0]; - - my $len; - eval { - $len = length($$bref); # this will die if $bref is a code ref, caught below - }; - if ($@) { - if (UNIVERSAL::isa($bref, "CODE")) { - unless ($need_queue) { - $self->{write_buf_size}--; # code refs are worth 1 - shift @{$self->{write_buf}}; - } - $bref->(); - - # code refs are just run and never get reenqueued - # (they're one-shot), so turn off the flag indicating the - # outstanding data needs queueing. - $need_queue = 0; - - undef $bref; - next WRITE; + my $sock = $self->{sock} or return 1; + my $ref = ref $data; + my $bref = $ref ? $data : \$data; + my $wbuf = $self->{wbuf}; + if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more... + if ($ref eq 'CODE') { + push @$wbuf, $bref; + } else { + my $last = $wbuf->[-1]; + if (ref($last) eq 'GLOB') { # append to tmp file buffer + $last->print($$bref) or return drop($self, "print: $!"); + } else { + my $tmpio = tmpio($self, $bref, 0) or return 0; + push @$wbuf, $tmpio; } - die "Write error: $@ <$bref>"; } - - my $to_write = $len - $self->{write_buf_offset}; - my $written; - if (my $wtr = $self->{writer_func}) { - $written = $wtr->($bref, $to_write, $self->{write_buf_offset}); + return 0; + } elsif ($ref eq 'CODE') { + $bref->($self); + return 1; + } else { + my $to_write = bytes::length($$bref); + my $written = syswrite($sock, $$bref, $to_write); + + if (defined $written) { + return 1 if $written == $to_write; + requeue($self); # runs: event_step -> flush_write + } elsif ($! == EAGAIN) { + epwait($sock, epbit($sock, EPOLLOUT) | EPOLLONESHOT); + $written = 0; } else { - $written = syswrite($self->{sock}, $$bref, $to_write, $self->{write_buf_offset}); + return $self->close; } - if (! defined $written) { - if ($! == EPIPE) { - return $self->close("EPIPE"); - } elsif ($! == EAGAIN) { - # since connection has stuff to write, it should now be - # interested in pending writes: - if ($need_queue) { - push @{$self->{write_buf}}, $bref; - $self->{write_buf_size} += $len; - } - $self->{write_set_watch} = 1 unless $self->{event_watch} & POLLOUT; - $self->watch_write(1); - return 0; - } elsif ($! == ECONNRESET) { - return $self->close("ECONNRESET"); - } + # deal with EAGAIN or partial write: + my $tmpio = tmpio($self, $bref, $written) or return 0; - DebugLevel >= 1 && $self->debugmsg("Closing connection ($self) due to write error: $!\n"); + # wbuf may be an empty array if we're being called inside + # ->flush_write via CODE bref: + push @{$self->{wbuf} ||= []}, $tmpio; + return 0; + } +} - return $self->close("write_error"); - } elsif ($written != $to_write) { - DebugLevel >= 2 && $self->debugmsg("Wrote PARTIAL %d bytes to %d", - $written, $self->{fd}); - if ($need_queue) { - push @{$self->{write_buf}}, $bref; - $self->{write_buf_size} += $len; - } - # since connection has stuff to write, it should now be - # interested in pending writes: - $self->{write_buf_offset} += $written; - $self->{write_buf_size} -= $written; - $self->on_incomplete_write; +use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0; + +sub msg_more ($$) { + my $self = $_[0]; + my $sock = $self->{sock} or return 1; + my $wbuf = $self->{wbuf}; + + if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) && + ref($sock) ne 'IO::Socket::SSL') { + my $n = send($sock, $_[1], MSG_MORE); + if (defined $n) { + my $nlen = bytes::length($_[1]) - $n; + return 1 if $nlen == 0; # all done! + # queue up the unwritten substring: + my $tmpio = tmpio($self, \($_[1]), $n) or return 0; + $self->{wbuf} //= $wbuf //= []; + push @$wbuf, $tmpio; + epwait($sock, EPOLLOUT|EPOLLONESHOT); return 0; - } elsif ($written == $to_write) { - DebugLevel >= 2 && $self->debugmsg("Wrote ALL %d bytes to %d (nq=%d)", - $written, $self->{fd}, $need_queue); - $self->{write_buf_offset} = 0; - - if ($self->{write_set_watch}) { - $self->watch_write(0); - $self->{write_set_watch} = 0; - } - - # this was our only write, so we can return immediately - # since we avoided incrementing the buffer size or - # putting it in the buffer. we also know there - # can't be anything else to write. - return 1 if $need_queue; - - $self->{write_buf_size} -= $written; - shift @{$self->{write_buf}}; - undef $bref; - next WRITE; } } -} -sub on_incomplete_write { - my PublicInbox::DS $self = shift; - $self->{write_set_watch} = 1 unless $self->{event_watch} & POLLOUT; - $self->watch_write(1); + # don't redispatch into NNTPdeflate::write + PublicInbox::DS::write($self, \($_[1])); } -=head2 C<< $obj->read( $bytecount ) >> - -Read at most I bytes from the underlying handle; returns scalar -ref on read, or undef on connection closed. - -=cut -sub read { - my PublicInbox::DS $self = shift; - return if $self->{closed}; - my $bytes = shift; - my $buf; - my $sock = $self->{sock}; - - # if this is too high, perl quits(!!). reports on mailing lists - # don't seem to point to a universal answer. 5MB worked for some, - # crashed for others. 1MB works for more people. let's go with 1MB - # for now. :/ - my $req_bytes = $bytes > 1048576 ? 1048576 : $bytes; - - my $res = sysread($sock, $buf, $req_bytes, 0); - DebugLevel >= 2 && $self->debugmsg("sysread = %d; \$! = %d", $res, $!); - - if (! $res && $! != EWOULDBLOCK) { - # catches 0=conn closed or undef=error - DebugLevel >= 2 && $self->debugmsg("Fd \#%d read hit the end of the road.", $self->{fd}); - return undef; - } - - return \$buf; +sub epwait ($$) { + my ($sock, $ev) = @_; + epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and + confess("EPOLL_CTL_MOD $!"); } -=head2 (VIRTUAL) C<< $obj->event_read() >> - -Readable event handler. Concrete deriviatives of PublicInbox::DS should -provide an implementation of this. The default implementation will die if -called. - -=cut -sub event_read { die "Base class event_read called for $_[0]\n"; } - -=head2 (VIRTUAL) C<< $obj->event_err() >> - -Error event handler. Concrete deriviatives of PublicInbox::DS should -provide an implementation of this. The default implementation will die if -called. - -=cut -sub event_err { die "Base class event_err called for $_[0]\n"; } - -=head2 (VIRTUAL) C<< $obj->event_hup() >> - -'Hangup' event handler. Concrete deriviatives of PublicInbox::DS should -provide an implementation of this. The default implementation will die if -called. - -=cut -sub event_hup { die "Base class event_hup called for $_[0]\n"; } - -=head2 C<< $obj->event_write() >> - -Writable event handler. Concrete deriviatives of PublicInbox::DS may wish to -provide an implementation of this. The default implementation calls -C with an C. - -=cut -sub event_write { - my $self = shift; - $self->write(undef); +# return true if complete, false if incomplete (or failure) +sub accept_tls_step ($) { + my ($self) = @_; + my $sock = $self->{sock} or return; + return 1 if $sock->accept_SSL; + return $self->close if $! != EAGAIN; + epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT); + unshift @{$self->{wbuf} ||= []}, \&accept_tls_step; + 0; } -=head2 C<< $obj->watch_read( $boolean ) >> - -Turn 'readable' event notification on or off. - -=cut -sub watch_read { - my PublicInbox::DS $self = shift; - return if $self->{closed} || !$self->{sock}; - - my $val = shift; - my $event = $self->{event_watch}; - - $event &= ~POLLIN if ! $val; - $event |= POLLIN if $val; +# return true if complete, false if incomplete (or failure) +sub shutdn_tls_step ($) { + my ($self) = @_; + my $sock = $self->{sock} or return; + return $self->close if $sock->stop_SSL(SSL_fast_shutdown => 1); + return $self->close if $! != EAGAIN; + epwait($sock, PublicInbox::TLS::epollbit() | EPOLLONESHOT); + unshift @{$self->{wbuf} ||= []}, \&shutdn_tls_step; + 0; +} - # If it changed, set it - if ($event != $self->{event_watch}) { - if ($HaveKQueue) { - $KQueue->EV_SET($self->{fd}, IO::KQueue::EVFILT_READ(), - $val ? IO::KQueue::EV_ENABLE() : IO::KQueue::EV_DISABLE()); - } - elsif ($HaveEpoll) { - epoll_ctl($Epoll, EPOLL_CTL_MOD, $self->{fd}, $event) - and $self->dump_error("couldn't modify epoll settings for $self->{fd} " . - "from $self->{event_watch} -> $event: $! (" . ($!+0) . ")"); - } - $self->{event_watch} = $event; +# don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC +# or fork w/o exec, so no inadvertant socket sharing +sub shutdn ($) { + my ($self) = @_; + my $sock = $self->{sock} or return; + if (ref($sock) eq 'IO::Socket::SSL') { + shutdn_tls_step($self); + } else { + $self->close; } } -=head2 C<< $obj->watch_write( $boolean ) >> +# must be called with eval, PublicInbox::DS may not be loaded (see t/qspawn.t) +sub dwaitpid ($$$) { + my ($pid, $cb, $arg) = @_; + if ($in_loop) { + push @$WaitPids, [ $pid, $cb, $arg ]; -Turn 'writable' event notification on or off. - -=cut -sub watch_write { - my PublicInbox::DS $self = shift; - return if $self->{closed} || !$self->{sock}; - - my $val = shift; - my $event = $self->{event_watch}; - - $event &= ~POLLOUT if ! $val; - $event |= POLLOUT if $val; - - if ($val && caller ne __PACKAGE__) { - # A subclass registered interest, it's now responsible for this. - $self->{write_set_watch} = 0; - } - - # If it changed, set it - if ($event != $self->{event_watch}) { - if ($HaveKQueue) { - $KQueue->EV_SET($self->{fd}, IO::KQueue::EVFILT_WRITE(), - $val ? IO::KQueue::EV_ENABLE() : IO::KQueue::EV_DISABLE()); - } - elsif ($HaveEpoll) { - epoll_ctl($Epoll, EPOLL_CTL_MOD, $self->{fd}, $event) - and $self->dump_error("couldn't modify epoll settings for $self->{fd} " . - "from $self->{event_watch} -> $event: $! (" . ($!+0) . ")"); - } - $self->{event_watch} = $event; + # We could've just missed our SIGCHLD, cover it, here: + requeue(\&reap_pids); + } else { + die "Not in EventLoop\n"; } } -=head2 C<< $obj->dump_error( $message ) >> +sub _run_later () { + my $run = $later_queue; + $later_timer = undef; + $later_queue = []; + $_->() for @$run; +} -Prints to STDERR a backtrace with information about this socket and what lead -up to the dump_error call. +sub later ($) { + my ($cb) = @_; + push @$later_queue, $cb; + $later_timer //= add_timer(60, \&_run_later); +} -=cut -sub dump_error { - my $i = 0; - my @list; - while (my ($file, $line, $sub) = (caller($i++))[1..3]) { - push @list, "\t$file:$line called $sub\n"; +sub expire_old () { + my $now = now(); + my $exp = $EXPTIME; + my $old = $now - $exp; + my %new; + while (my ($fd, $v) = each %$EXPMAP) { + my ($idle_time, $ds_obj) = @$v; + if ($idle_time < $old) { + if (!$ds_obj->shutdn) { + $new{$fd} = $v; + } + } else { + $new{$fd} = $v; + } } - - warn "ERROR: $_[1]\n" . - "\t$_[0] = " . $_[0]->as_string . "\n" . - join('', @list); + $EXPMAP = \%new; + $exp_timer = scalar(keys %new) ? later(\&expire_old) : undef; } -=head2 C<< $obj->debugmsg( $format, @args ) >> - -Print the debugging message specified by the C-style I and -I. - -=cut -sub debugmsg { - my ( $self, $fmt, @args ) = @_; - confess "Not an object" unless ref $self; - - chomp $fmt; - printf STDERR ">>> $fmt\n", @args; +sub update_idle_time { + my ($self) = @_; + my $sock = $self->{sock} or return; + $EXPMAP->{fileno($sock)} = [ now(), $self ]; + $exp_timer //= later(\&expire_old); } -=head2 C<< $obj->as_string() >> - -Returns a string describing this socket. - -=cut -sub as_string { - my PublicInbox::DS $self = shift; - my $rw = "(" . ($self->{event_watch} & POLLIN ? 'R' : '') . - ($self->{event_watch} & POLLOUT ? 'W' : '') . ")"; - my $ret = ref($self) . "$rw: " . ($self->{closed} ? "closed" : "open"); - return $ret; -} - -package PublicInbox::DS::Timer; -# [$abs_float_firetime, $coderef]; -sub cancel { - $_[0][1] = undef; +sub not_idle_long { + my ($self, $now) = @_; + my $sock = $self->{sock} or return; + my $ary = $EXPMAP->{fileno($sock)} or return; + my $exp_at = $ary->[0] + $EXPTIME; + $exp_at > $now; } 1;