X-Git-Url: http://www.git.stargrave.org/?a=blobdiff_plain;f=lib%2FPublicInbox%2FDS.pm;h=eb468f5720f698d2f3f786a6c23e7e226a42d378;hb=d07ba9c30800225052d17ccca458afbfa05a8ff0;hp=9e24ed78d10f2ab9891340b8df5fb78d4827c95a;hpb=8c619bae62a3d468505716d58d3559278883eee7;p=public-inbox.git diff --git a/lib/PublicInbox/DS.pm b/lib/PublicInbox/DS.pm index 9e24ed78..fee31e3d 100644 --- a/lib/PublicInbox/DS.pm +++ b/lib/PublicInbox/DS.pm @@ -3,63 +3,55 @@ # # 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. # -# API changes to diverge from Danga::Socket will happen to better -# accomodate new features and improve scalability. Do not expect -# this to be a stable API like Danga::Socket. -# Bugs encountered (and likely fixed) are reported to -# bug-Danga-Socket@rt.cpan.org and visible at: +# 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 +# +# fields: +# sock: underlying socket +# rbuf: scalarref, usually undef +# wbuf: arrayref of coderefs or tmpio (autovivified)) +# (tmpio = [ GLOB, offset, [ length ] ]) package PublicInbox::DS; use strict; -use bytes; -use POSIX (); -use Time::HiRes (); -use IO::Handle qw(); -use Fcntl qw(FD_CLOEXEC F_SETFD F_GETFD); - -use warnings; - +use v5.10.1; +use parent qw(Exporter); +use bytes qw(length substr); # FIXME(?): needed for PublicInbox::NNTP +use POSIX qw(WNOHANG sigprocmask SIG_SETMASK); +use Fcntl qw(SEEK_SET :DEFAULT O_APPEND); +use Time::HiRes qw(clock_gettime CLOCK_MONOTONIC); +use Scalar::Util qw(blessed); use PublicInbox::Syscall qw(:epoll); - -use fields ('sock', # underlying socket - 'wbuf', # arrayref of scalars, scalarrefs, or coderefs to write - 'wbuf_off', # offset into first element of wbuf to start writing at - 'event_watch', # bitmask of events the client is interested in (POLLIN,OUT,etc.) - ); - -use Errno qw(EAGAIN EINVAL); -use Carp qw(croak confess); - -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 PublicInbox::Tmpfile; +use Errno qw(EAGAIN EINVAL); +use Carp qw(carp croak); +our @EXPORT_OK = qw(now msg_more dwaitpid add_timer add_uniq_timer); + +my %Stack; +my $nextq; # queue for next_tick +my $wait_pids; # list of [ pid, callback, callback_arg ] +my $reap_armed; +my $ToClose; # sockets to close when event loop is done 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 ref (for kqueue mode only) - $_io, # IO::Handle for Epoll - @ToClose, # sockets to close when event loop is done + $Epoll, # Global epoll fd (or DSKQXS ref) + $ep_io, # IO::Handle for Epoll $PostLoopCallback, # subref to call at the end of each loop, if defined (global) $LoopTimeout, # timeout of event loop in milliseconds - $DoneInit, # if we've done the one-time module init yet @Timers, # timers + %UniqTimer, + $in_loop, ); -# this may be set to zero with old kernels -our $EPOLLEXCLUSIVE = EPOLLEXCLUSIVE; Reset(); ##################################################################### @@ -72,23 +64,25 @@ Reset all state =cut sub Reset { - %DescriptorMap = (); - @ToClose = (); - $LoopTimeout = -1; # no timeout by default - @Timers = (); - - $PostLoopCallback = undef; - $DoneInit = 0; - - # NOTE kqueue is close-on-fork, and we don't account for it, yet - # OTOH, we (public-inbox) don't need this sub outside of tests... - POSIX::close($$KQueue) if !$_io && $KQueue && $$KQueue >= 0; - $KQueue = undef; - - $_io = undef; # close $Epoll - $Epoll = undef; - - *EventLoop = *FirstTimeEventLoop; + do { + $in_loop = undef; # first in case DESTROY callbacks use this + %DescriptorMap = (); + @Timers = (); + %UniqTimer = (); + $PostLoopCallback = undef; + + # we may be iterating inside one of these on our stack + my @q = delete @Stack{keys %Stack}; + for my $q (@q) { @$q = () } + $wait_pids = $nextq = $ToClose = undef; + $ep_io = undef; # closes real $Epoll FD + $Epoll = undef; # may call DSKQXS::DESTROY + } while (@Timers || keys(%Stack) || $nextq || $wait_pids || + $ToClose || keys(%DescriptorMap) || + $PostLoopCallback || keys(%UniqTimer)); + + $reap_armed = undef; + $LoopTimeout = -1; # no timeout by default } =head2 C<< CLASS->SetLoopTimeout( $timeout ) >> @@ -99,224 +93,204 @@ A timeout of 0 (zero) means poll forever. A timeout of -1 means poll and return immediately. =cut -sub SetLoopTimeout { - return $LoopTimeout = $_[1] + 0; -} - -=head2 C<< CLASS->AddTimer( $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 $class = shift; - my ($secs, $coderef) = @_; - - my $fire_time = Time::HiRes::time() + $secs; - - my $timer = bless [$fire_time, $coderef], "PublicInbox::DS::Timer"; - - if (!@Timers || $fire_time >= $Timers[-1][0]) { - push @Timers, $timer; - return $timer; - } - - # Now, where do we insert? (NOTE: this appears slow, algorithm-wise, - # but it was compared against calendar queues, heaps, naive push/sort, - # and a bunch of other versions, and found to be fastest with a large - # variety of datasets.) - for (my $i = 0; $i < @Timers; $i++) { - if ($Timers[$i][0] > $fire_time) { - splice(@Timers, $i, 0, $timer); - return $timer; - } - } - - die "Shouldn't get here."; +sub SetLoopTimeout { $LoopTimeout = $_[1] + 0 } + +sub _add_named_timer { + my ($name, $secs, $coderef, @args) = @_; + my $fire_time = now() + $secs; + my $timer = [$fire_time, $name, $coderef, @args]; + + if (!@Timers || $fire_time >= $Timers[-1][0]) { + push @Timers, $timer; + return $timer; + } + + # Now, where do we insert? (NOTE: this appears slow, algorithm-wise, + # but it was compared against calendar queues, heaps, naive push/sort, + # and a bunch of other versions, and found to be fastest with a large + # variety of datasets.) + for (my $i = 0; $i < @Timers; $i++) { + if ($Timers[$i][0] > $fire_time) { + splice(@Timers, $i, 0, $timer); + return $timer; + } + } + die "Shouldn't get here."; } -# keeping this around in case we support other FD types for now, -# epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+... -sub set_cloexec ($) { - my ($fd) = @_; +sub add_timer { _add_named_timer(undef, @_) } - $_io = IO::Handle->new_from_fd($fd, 'r+') or return; - defined(my $fl = fcntl($_io, F_GETFD, 0)) or return; - fcntl($_io, F_SETFD, $fl | FD_CLOEXEC); +sub add_uniq_timer { # ($name, $secs, $coderef, @args) = @_; + $UniqTimer{$_[0]} //= _add_named_timer(@_); } -sub _InitPoller -{ - return if $DoneInit; - $DoneInit = 1; - - if ($HAVE_KQUEUE) { - $KQueue = IO::KQueue->new(); - $HaveKQueue = defined $KQueue; - if ($HaveKQueue) { - *EventLoop = *KQueueEventLoop; - } - } - elsif (PublicInbox::Syscall::epoll_defined()) { - $Epoll = eval { epoll_create(1024); }; - $HaveEpoll = defined $Epoll && $Epoll >= 0; - if ($HaveEpoll) { - set_cloexec($Epoll); - *EventLoop = *EpollEventLoop; - } - } - - if (!$HaveEpoll && !$HaveKQueue) { - require IO::Poll; - *EventLoop = *PollEventLoop; - } +# caller sets return value to $Epoll +sub _InitPoller () { + if (PublicInbox::Syscall::epoll_defined()) { + my $fd = epoll_create(); + die "epoll_create: $!" if $fd < 0; + open($ep_io, '+<&=', $fd) or return; + my $fl = fcntl($ep_io, F_GETFD, 0); + fcntl($ep_io, F_SETFD, $fl | FD_CLOEXEC); + $fd; + } else { + my $cls; + for (qw(DSKQXS DSPoll)) { + $cls = "PublicInbox::$_"; + last if eval "require $cls"; + } + $cls->import(qw(epoll_ctl epoll_wait)); + $cls->new; + } } -=head2 C<< CLASS->EventLoop() >> - -Start processing IO events. In most daemon programs this never exits. See -C below for how to exit the loop. - -=cut -sub FirstTimeEventLoop { - my $class = shift; - - _InitPoller(); - - if ($HaveEpoll) { - EpollEventLoop($class); - } elsif ($HaveKQueue) { - KQueueEventLoop($class); - } else { - PollEventLoop($class); - } +sub now () { clock_gettime(CLOCK_MONOTONIC) } + +sub next_tick () { + my $q = $nextq or return; + $nextq = undef; + $Stack{cur_runq} = $q; + for my $obj (@$q) { + # 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($obj)) { + $obj->event_step; + } else { + $obj->(); + } + } + delete $Stack{cur_runq}; } # runs timers and returns milliseconds for next one, or next event loop sub RunTimers { - return $LoopTimeout unless @Timers; + next_tick(); - my $now = Time::HiRes::time(); + return (($nextq || $ToClose) ? 0 : $LoopTimeout) unless @Timers; - # Run expired timers - while (@Timers && $Timers[0][0] <= $now) { - my $to_run = shift(@Timers); - $to_run->[1]->($now) if $to_run->[1]; - } + my $now = now(); - return $LoopTimeout unless @Timers; + # Run expired timers + while (@Timers && $Timers[0][0] <= $now) { + my $to_run = shift(@Timers); + delete $UniqTimer{$to_run->[1] // ''}; + $to_run->[2]->(@$to_run[3..$#$to_run]); + } - # convert time to an even number of milliseconds, adding 1 - # extra, otherwise floating point fun can occur and we'll - # call RunTimers like 20-30 times, each returning a timeout - # of 0.0000212 seconds - my $timeout = int(($Timers[0][0] - $now) * 1000) + 1; + # timers may enqueue into nextq: + return 0 if ($nextq || $ToClose); - # -1 is an infinite timeout, so prefer a real timeout - return $timeout if $LoopTimeout == -1; + return $LoopTimeout unless @Timers; - # otherwise pick the lower of our regular timeout and time until - # the next timer - return $LoopTimeout if $LoopTimeout < $timeout; - return $timeout; -} + # convert time to an even number of milliseconds, adding 1 + # extra, otherwise floating point fun can occur and we'll + # call RunTimers like 20-30 times, each returning a timeout + # of 0.0000212 seconds + my $timeout = int(($Timers[0][0] - $now) * 1000) + 1; -### The epoll-based event loop. Gets installed as EventLoop if IO::Epoll loads -### okay. -sub EpollEventLoop { - my $class = shift; - - while (1) { - my @events; - my $i; - my $timeout = RunTimers(); - - # get up to 1000 events - my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events); - for ($i=0; $i<$evcount; $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. - $DescriptorMap{$events[$i]->[0]}->event_step; - } - return unless PostEventLoop(); - } - exit 0; + # -1 is an infinite timeout, so prefer a real timeout + ($LoopTimeout < 0 || $LoopTimeout >= $timeout) ? $timeout : $LoopTimeout } -### 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; +sub sig_setmask { sigprocmask(SIG_SETMASK, @_) or die "sigprocmask: $!" } - 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}; - } +sub block_signals () { + my $oldset = POSIX::SigSet->new; + my $newset = POSIX::SigSet->new; + $newset->fillset or die "fillset: $!"; + sig_setmask($newset, $oldset); + $oldset; +} - # if nothing to poll, either end immediately (if no timeout) - # or just keep calling the callback - unless (@poll) { - select undef, undef, undef, ($timeout / 1000); - return unless PostEventLoop(); - next; - } +# We can't use waitpid(-1) safely here since it can hit ``, system(), +# and other things. So we scan the $wait_pids list, which is hopefully +# not too big. We keep $wait_pids small by not calling dwaitpid() +# until we've hit EOF when reading the stdout of the child. + +sub reap_pids { + $reap_armed = undef; + my $tmp = $wait_pids or return; + $wait_pids = undef; + $Stack{reap_runq} = $tmp; + my $oldset = block_signals(); + foreach my $ary (@$tmp) { + my ($pid, $cb, $arg) = @$ary; + my $ret = waitpid($pid, WNOHANG); + if ($ret == 0) { + push @$wait_pids, $ary; # autovivifies @$wait_pids + } elsif ($ret == $pid) { + if ($cb) { + eval { $cb->($arg, $pid) }; + warn "E: dwaitpid($pid) in_loop: $@" if $@; + } + } else { + warn "waitpid($pid, WNOHANG) = $ret, \$!=$!, \$?=$?"; + } + } + sig_setmask($oldset); + delete $Stack{reap_runq}; +} - my $count = IO::Poll::_poll($timeout, @poll); - unless ($count >= 0) { - return unless PostEventLoop(); - next; - } +# reentrant SIGCHLD handler (since reap_pids is not reentrant) +sub enqueue_reap () { $reap_armed //= requeue(\&reap_pids) } - # Fetch handles with read events - while (@poll) { - my ($fd, $state) = splice(@poll, 0, 2); - $DescriptorMap{$fd}->event_step if $state; - } +sub in_loop () { $in_loop } - return unless PostEventLoop(); - } - - exit 0; +# Internal function: run the post-event callback, send read events +# for pushed-back data, and close pending connections. returns 1 +# if event loop should continue, or 0 to shut it all down. +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) + if (my $close_now = $ToClose) { + $ToClose = undef; # will be autovivified on push + @$close_now = map { fileno($_) } @$close_now; + + # ->DESTROY methods may populate ToClose + delete @DescriptorMap{@$close_now}; + } + + # by default we keep running, unless a postloop callback cancels it + $PostLoopCallback ? $PostLoopCallback->(\%DescriptorMap) : 1; } -### 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) { - $DescriptorMap{$kev->[0]}->event_step; - } - return unless PostEventLoop(); - } - - exit(0); +# Start processing IO events. In most daemon programs this never exits. See +# C for how to exit the loop. +sub event_loop (;$$) { + my ($sig, $oldset) = @_; + $Epoll //= _InitPoller(); + require PublicInbox::Sigfd if $sig; + my $sigfd = PublicInbox::Sigfd->new($sig, 1) if $sig; + local @SIG{keys %$sig} = values(%$sig) if $sig && !$sigfd; + local $SIG{PIPE} = 'IGNORE'; + if (!$sigfd && $sig) { + # wake up every second to accept signals if we don't + # have signalfd or IO::KQueue: + sig_setmask($oldset); + PublicInbox::DS->SetLoopTimeout(1000); + } + $_[0] = $sigfd = $sig = undef; # $_[0] == sig + local $in_loop = 1; + my @events; + do { + my $timeout = RunTimers(); + + # get up to 1000 events + epoll_wait($Epoll, 1000, $timeout, \@events); + for my $fd (@events) { + # 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. + + # guard stack-not-refcounted w/ Carp + @DB::args + my $obj = $DescriptorMap{$fd}; + $obj->event_step; + } + } while (PostEventLoop()); } =head2 C<< CLASS->SetPostLoopCallback( CODEREF ) >> @@ -337,37 +311,6 @@ sub SetPostLoopCallback { $PostLoopCallback = (defined $ref && ref $ref eq 'CODE') ? $ref : undef; } -# Internal function: run the post-event callback, send read events -# for pushed-back data, and close pending connections. returns 1 -# if event loop should continue, or 0 to shut it all down. -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 ->close. - delete $DescriptorMap{$fd}; - } - - - # by default we keep running, unless a postloop callback (either per-object - # or global) cancels it - my $keep_running = 1; - - # now we're at the very end, call callback if defined - if (defined $PostLoopCallback) { - $keep_running &&= $PostLoopCallback->(\%DescriptorMap); - } - - return $keep_running; -} - ##################################################################### ### PublicInbox::DS-the-object code ##################################################################### @@ -377,7 +320,7 @@ sub PostEventLoop { =head2 C<< CLASS->new( $socket ) >> Create a new PublicInbox::DS subclass object for the given I which will -react to events on it during the C. +react to events on it during the C. This is normally (always?) called from your subclass via: @@ -385,56 +328,50 @@ This is normally (always?) called from your subclass via: =cut sub new { - my ($self, $sock, $exclusive) = @_; - $self = fields::new($self) unless ref $self; - + my ($self, $sock, $ev) = @_; $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->{wbuf} = []; - $self->{wbuf_off} = 0; - - my $ev = $self->{event_watch} = POLLERR|POLLHUP|POLLNVAL; - - _InitPoller(); - - if ($HaveEpoll) { - if ($exclusive) { - $ev = $self->{event_watch} = EPOLLIN|EPOLLERR|EPOLLHUP|$EPOLLEXCLUSIVE; - } + $Epoll //= _InitPoller(); 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 "EPOLL_CTL_ADD $self/$sock/$fd: $!"; } - 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}; + croak("FD:$fd in use by $DescriptorMap{$fd} (for $self/$sock)") + if defined($DescriptorMap{$fd}); $DescriptorMap{$fd} = $self; - return $self; } +# for IMAP, NNTP, and POP3 which greet clients upon connect +sub greet { + my ($self, $sock) = @_; + my $ev = EPOLLIN; + my $wbuf; + if ($sock->can('accept_SSL') && !$sock->accept_SSL) { + return CORE::close($sock) if $! != EAGAIN; + $ev = PublicInbox::TLS::epollbit() or return CORE::close($sock); + $wbuf = [ \&accept_tls_step, $self->can('do_greet')]; + } + new($self, $sock, $ev | EPOLLONESHOT); + if ($wbuf) { + $self->{wbuf} = $wbuf; + } else { + $self->do_greet; + } + $self; +} ##################################################################### ### I N S T A N C E M E T H O D S ##################################################################### +sub requeue ($) { push @$nextq, $_[0] } # autovivifies + =head2 C<< $obj->close >> Close the socket. @@ -447,15 +384,13 @@ sub close { # we need to flush our write buffer, as there may # be self-referential closures (sub { $client->close }) # preventing the object from being destroyed - @{$self->{wbuf}} = (); + 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) { - my $fd = fileno($sock); - epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, $self->{event_watch}) and - confess("EPOLL_CTL_DEL: $!"); - } + my $fd = fileno($sock); + epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, 0) and + croak("EPOLL_CTL_DEL($self/$sock): $!"); # we explicitly don't delete from DescriptorMap here until we # actually close the socket, as we might be in the middle of @@ -468,23 +403,131 @@ sub close { # defer closing the actual socket until the event loop is done # processing this round of events. (otherwise we might reuse fds) - push @ToClose, $sock; + push @$ToClose, $sock; # autovivifies $ToClose return 0; } +# portable, non-thread-safe sendfile emulation (no pread, yet) +sub send_tmpio ($$) { + my ($sock, $tmpio) = @_; + + sysseek($tmpio->[0], $tmpio->[1], SEEK_SET) or return; + my $n = $tmpio->[2] // 65536; + $n = 65536 if $n > 65536; + defined(my $to_write = sysread($tmpio->[0], my $buf, $n)) 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; + } + } + $tmpio->[1] += $written; # offset + $tmpio->[2] -= $written if defined($tmpio->[2]); # length + $written; +} + +sub epbit ($$) { # (sock, default) + $_[0]->can('stop_SSL') ? PublicInbox::TLS::epollbit() : $_[1]; +} + +# returns 1 if done, 0 if incomplete +sub flush_write ($) { + my ($self) = @_; + my $sock = $self->{sock} or return; + my $wbuf = $self->{wbuf} or return 1; + +next_buf: + while (my $bref = $wbuf->[0]) { + if (ref($bref) ne 'CODE') { + while ($sock) { + my $w = send_tmpio($sock, $bref); # bref is tmpio + if (defined $w) { + if ($w == 0) { + shift @$wbuf; + goto next_buf; + } + } elsif ($! == EAGAIN) { + my $ev = epbit($sock, EPOLLOUT) or return $self->close; + epwait($sock, $ev | EPOLLONESHOT); + return 0; + } else { + return $self->close; + } + } + } else { #(ref($bref) eq 'CODE') { + shift @$wbuf; + my $before = scalar(@$wbuf); + $bref->($self); + + # bref may be enqueueing more CODE to call (see accept_tls_step) + return 0 if (scalar(@$wbuf) > $before); + } + } # while @$wbuf + + delete $self->{wbuf}; + 1; # all done +} + +sub rbuf_idle ($$) { + 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}, $$rbuf, $len, $off // 0); + return ($r == 0 ? $self->close : $r) if defined $r; + # common for clients to break connections without warning, + # would be too noisy to log here: + if ($! == EAGAIN) { + my $ev = epbit($sock, EPOLLIN) or return $self->close; + epwait($sock, $ev | EPOLLONESHOT); + rbuf_idle($self, $rbuf); + 0; + } else { + $self->close; + } +} + +# drop the socket if we hit unrecoverable errors on our system which +# require BOFH attention: ENOSPC, EFBIG, EIO, EMFILE, ENFILE... +sub drop { + my $self = shift; + carp(@_); + $self->close; +} + +sub tmpio ($$$) { + my ($self, $bref, $off) = @_; + my $fh = tmpfile('wbuf', $self->{sock}, O_APPEND) or + return drop($self, "tmpfile $!"); + $fh->autoflush(1); + my $len = length($$bref) - $off; + my $n = syswrite($fh, $$bref, $len, $off) // + return drop($self, "write ($len): $!"); + $n == $len or return drop($self, "wrote $n < $len bytes"); + [ $fh, 0 ] # [1] = offset, [2] = length, not set by us +} + =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 @@ -493,164 +536,189 @@ 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 unless $self->{sock}; - - my $bref; - - # just queue data if there's already a wait - my $need_queue; + my $sock = $self->{sock} or return 1; + my $ref = ref $data; + my $bref = $ref ? $data : \$data; my $wbuf = $self->{wbuf}; - - if (defined $data) { - $bref = ref $data ? $data : \$data; - if (scalar @$wbuf) { + if ($wbuf && scalar(@$wbuf)) { # already buffering, can't write more... + if ($ref eq 'CODE') { push @$wbuf, $bref; - 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 ||= $wbuf->[0]; - - my $len; - eval { - $len = length($$bref); # this will die if $bref is a code ref, caught below - }; - if ($@) { - if (UNIVERSAL::isa($bref, "CODE")) { - unless ($need_queue) { - shift @$wbuf; - } - $bref->(); - - # code refs are just run and never get reenqueued - # (they're one-shot), so turn off the flag indicating the - # outstanding data needs queueing. - $need_queue = 0; - - undef $bref; - next WRITE; + } else { + my $tmpio = $wbuf->[-1]; + if ($tmpio && !defined($tmpio->[2])) { # append to tmp file buffer + $tmpio->[0]->print($$bref) or return drop($self, "print: $!"); + } else { + my $tmpio = tmpio($self, $bref, 0) or return 0; + push @$wbuf, $tmpio; } - die "Write error: $@ <$bref>"; + } + return 0; + } elsif ($ref eq 'CODE') { + $bref->($self); + return 1; + } else { + my $to_write = 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) { + my $ev = epbit($sock, EPOLLOUT) or return $self->close; + epwait($sock, $ev | EPOLLONESHOT); + $written = 0; + } else { + return $self->close; } - my $to_write = $len - $self->{wbuf_off}; - my $written = syswrite($self->{sock}, $$bref, $to_write, - $self->{wbuf_off}); + # deal with EAGAIN or partial write: + my $tmpio = tmpio($self, $bref, $written) or return 0; - if (! defined $written) { - if ($! == EAGAIN) { - # since connection has stuff to write, it should now be - # interested in pending writes: - if ($need_queue) { - push @$wbuf, $bref; - } - $self->watch_write(1); - return 0; - } + # wbuf may be an empty array if we're being called inside + # ->flush_write via CODE bref: + push @{$self->{wbuf}}, $tmpio; # autovivifies + return 0; + } +} - return $self->close; - } elsif ($written != $to_write) { - if ($need_queue) { - push @$wbuf, $bref; - } - # since connection has stuff to write, it should now be - # interested in pending writes: - $self->{wbuf_off} += $written; - $self->on_incomplete_write; - return 0; - } elsif ($written == $to_write) { - $self->{wbuf_off} = 0; - $self->watch_write(0); +use constant MSG_MORE => ($^O eq 'linux') ? 0x8000 : 0; - # this was our only write, so we can return immediately - # since we avoided incrementing the buffer size or - # putting it in the buffer. we also know there - # can't be anything else to write. - return 1 if $need_queue; +sub msg_more ($$) { + my $self = $_[0]; + my $sock = $self->{sock} or return 1; + my $wbuf = $self->{wbuf}; - shift @$wbuf; - undef $bref; - next WRITE; + if (MSG_MORE && (!defined($wbuf) || !scalar(@$wbuf)) && + !$sock->can('stop_SSL')) { + my $n = send($sock, $_[1], MSG_MORE); + if (defined $n) { + my $nlen = length($_[1]) - $n; + return 1 if $nlen == 0; # all done! + # queue up the unwritten substring: + my $tmpio = tmpio($self, \($_[1]), $n) or return 0; + push @{$self->{wbuf}}, $tmpio; # autovivifies + epwait($sock, EPOLLOUT|EPOLLONESHOT); + return 0; } } -} -sub on_incomplete_write { - my PublicInbox::DS $self = shift; - $self->watch_write(1); + # don't redispatch into NNTPdeflate::write + PublicInbox::DS::write($self, \($_[1])); } -=head2 C<< $obj->watch_read( $boolean ) >> - -Turn 'readable' event notification on or off. +sub epwait ($$) { + my ($sock, $ev) = @_; + epoll_ctl($Epoll, EPOLL_CTL_MOD, fileno($sock), $ev) and + croak("EPOLL_CTL_MOD($sock): $!"); +} -=cut -sub watch_read { - my PublicInbox::DS $self = shift; +# return true if complete, false if incomplete (or failure) +sub accept_tls_step ($) { + my ($self) = @_; my $sock = $self->{sock} or return; - - my $val = shift; - my $event = $self->{event_watch}; - - $event &= ~POLLIN if ! $val; - $event |= POLLIN if $val; - - my $fd = fileno($sock); - # If it changed, set it - if ($event != $self->{event_watch}) { - if ($HaveKQueue) { - $KQueue->EV_SET($fd, IO::KQueue::EVFILT_READ(), - $val ? IO::KQueue::EV_ENABLE() : IO::KQueue::EV_DISABLE()); - } - elsif ($HaveEpoll) { - epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $event) and - confess("EPOLL_CTL_MOD: $!"); - } - $self->{event_watch} = $event; - } + return 1 if $sock->accept_SSL; + return $self->close if $! != EAGAIN; + my $ev = PublicInbox::TLS::epollbit() or return $self->close; + epwait($sock, $ev | EPOLLONESHOT); + unshift(@{$self->{wbuf}}, \&accept_tls_step); # autovivifies + 0; } -=head2 C<< $obj->watch_write( $boolean ) >> - -Turn 'writable' event notification on or off. +# 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; + my $ev = PublicInbox::TLS::epollbit() or return $self->close; + epwait($sock, $ev | EPOLLONESHOT); + unshift(@{$self->{wbuf}}, \&shutdn_tls_step); # autovivifies + 0; +} -=cut -sub watch_write { - my PublicInbox::DS $self = shift; +# don't bother with shutdown($sock, 2), we don't fork+exec w/o CLOEXEC +# or fork w/o exec, so no inadvertent socket sharing +sub shutdn ($) { + my ($self) = @_; my $sock = $self->{sock} or return; + if ($sock->can('stop_SSL')) { + shutdn_tls_step($self); + } else { + $self->close; + } +} - my $val = shift; - my $event = $self->{event_watch}; +sub zflush {} # overridden by NNTPdeflate and IMAPdeflate + +sub long_response_done {} # overridden by Net::NNTP + +sub long_step { + my ($self) = @_; + # wbuf is unset or empty, here; {long} may add to it + my ($fd, $cb, $t0, @args) = @{$self->{long_cb}}; + my $more = eval { $cb->($self, @args) }; + if ($@ || !$self->{sock}) { # something bad happened... + delete $self->{long_cb}; + my $elapsed = now() - $t0; + $@ and $self->err("%s during long response[$fd] - %0.6f", + $@, $elapsed); + $self->out(" deferred[$fd] aborted - %0.6f", $elapsed); + $self->close; + } elsif ($more) { # $self->{wbuf}: + # control passed to ibx_async_cat if $more == \undef + requeue_once($self) if !ref($more); + } else { # all done! + delete $self->{long_cb}; + $self->long_response_done; + my $elapsed = now() - $t0; + my $fd = fileno($self->{sock}); + $self->out(" deferred[$fd] done - %0.6f", $elapsed); + my $wbuf = $self->{wbuf}; # do NOT autovivify + requeue($self) unless $wbuf && @$wbuf; + } +} - $event &= ~POLLOUT if ! $val; - $event |= POLLOUT if $val; - my $fd = fileno($sock); +sub requeue_once { + my ($self) = @_; + # COMPRESS users all share the same DEFLATE context. + # Flush it here to ensure clients don't see each other's data + $self->zflush; + + # no recursion, schedule another call ASAP, + # but only after all pending writes are done. + # autovivify wbuf. wbuf may be populated by $cb, + # no need to rearm if so: (push returns new size of array) + requeue($self) if push(@{$self->{wbuf}}, \&long_step) == 1; +} - # If it changed, set it - if ($event != $self->{event_watch}) { - if ($HaveKQueue) { - $KQueue->EV_SET($fd, IO::KQueue::EVFILT_WRITE(), - $val ? IO::KQueue::EV_ENABLE() : IO::KQueue::EV_DISABLE()); - } - elsif ($HaveEpoll) { - epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $event) and - confess "EPOLL_CTL_MOD: $!"; - } - $self->{event_watch} = $event; - } +sub long_response ($$;@) { + my ($self, $cb, @args) = @_; # cb returns true if more, false if done + my $sock = $self->{sock} or return; + # make sure we disable reading during a long response, + # clients should not be sending us stuff and making us do more + # work while we are stream a response to them + $self->{long_cb} = [ fileno($sock), $cb, now(), @args ]; + long_step($self); # kick off! + undef; } -package PublicInbox::DS::Timer; -# [$abs_float_firetime, $coderef]; -sub cancel { - $_[0][1] = undef; +sub dwaitpid ($;$$) { + my ($pid, $cb, $arg) = @_; + if ($in_loop) { + push @$wait_pids, [ $pid, $cb, $arg ]; + # We could've just missed our SIGCHLD, cover it, here: + enqueue_reap(); + } else { + my $ret = waitpid($pid, 0); + if ($ret == $pid) { + if ($cb) { + eval { $cb->($arg, $pid) }; + carp "E: dwaitpid($pid) !in_loop: $@" if $@; + } + } else { + carp "waitpid($pid, 0) = $ret, \$!=$!, \$?=$?"; + } + } } 1;