#
# 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
-
+# advantage of newer kernels.
+#
+# 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:
+# https://rt.cpan.org/Public/Dist/Display.html?Name=Danga-Socket
package PublicInbox::DS;
use strict;
use bytes;
use POSIX ();
use Time::HiRes ();
-
-use vars qw{$VERSION};
-$VERSION = "1.61";
+use IO::Handle qw();
+use Fcntl qw(FD_CLOEXEC F_SETFD F_GETFD);
use warnings;
-no warnings qw(deprecated);
use PublicInbox::Syscall qw(:epoll);
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
+ 'wbuf', # arrayref of scalars, scalarrefs, or coderefs to write
+ 'wbuf_off', # offset into first element of wbuf to start writing at
'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
);
-use Errno qw(EINPROGRESS EWOULDBLOCK EISCONN ENOTSOCK
- EPIPE EAGAIN EBADF ECONNRESET ENOPROTOOPT);
-use Socket qw(IPPROTO_TCP);
+use Errno qw(EAGAIN EINVAL);
use Carp qw(croak confess);
use constant DebugLevel => 0;
$HaveKQueue,
%DescriptorMap, # fd (num) -> PublicInbox::DS object
$Epoll, # Global epoll fd (for epoll mode only)
- $KQueue, # Global kqueue fd (for kqueue 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
- %OtherFds, # A hash of "other" (non-PublicInbox::DS) file
- # descriptors for the event loop to track.
$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
sub Reset {
%DescriptorMap = ();
@ToClose = ();
- %OtherFds = ();
$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;
-
- *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() >>
+ # 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;
-Return the list of sockets that are awaiting close() at the end of the
-current event loop.
+ $_io = undef; # close $Epoll
+ $Epoll = undef;
-=cut
-sub ToClose { return @ToClose; }
-
-=head2 C<< CLASS->OtherFds( [%fdmap] ) >>
-
-Get/set the hash of file descriptors that need processing in parallel with
-the registered PublicInbox::DS objects.
-
-=cut
-sub OtherFds {
- my $class = shift;
- if ( @_ ) { %OtherFds = @_ }
- return wantarray ? %OtherFds : \%OtherFds;
-}
-
-=head2 C<< CLASS->AddOtherFds( [%fdmap] ) >>
-
-Add fds to the OtherFds hash for processing.
-
-=cut
-sub AddOtherFds {
- my $class = shift;
- %OtherFds = ( %OtherFds, @_ ); # FIXME investigate what happens on dupe fds
- return wantarray ? %OtherFds : \%OtherFds;
+ *EventLoop = *FirstTimeEventLoop;
}
=head2 C<< CLASS->SetLoopTimeout( $timeout ) >>
die "Shouldn't get here.";
}
+# keeping this around in case we support other FD types for now,
+# epoll_create1(EPOLL_CLOEXEC) requires Linux 2.6.27+...
+sub set_cloexec ($) {
+ 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 $DoneInit;
if ($HAVE_KQUEUE) {
$KQueue = IO::KQueue->new();
- $HaveKQueue = $KQueue >= 0;
+ $HaveKQueue = defined $KQueue;
if ($HaveKQueue) {
*EventLoop = *KQueueEventLoop;
}
$Epoll = eval { epoll_create(1024); };
$HaveEpoll = defined $Epoll && $Epoll >= 0;
if ($HaveEpoll) {
+ set_cloexec($Epoll);
*EventLoop = *EpollEventLoop;
}
}
return $timeout;
}
+sub event_step ($) {
+ my ($self) = @_;
+ return if $self->{closed};
+
+ my $wbuf = $self->{wbuf};
+ if (@$wbuf) {
+ $self->event_write;
+ return if $self->{closed} || scalar(@$wbuf);
+ }
+
+ # only read more requests if we've drained the write buffer,
+ # otherwise we can be buffering infinitely w/o backpressure
+ $self->event_read;
+}
+
### The epoll-based event loop. Gets installed as EventLoop if IO::Epoll loads
### okay.
sub EpollEventLoop {
my $class = shift;
- foreach my $fd ( keys %OtherFds ) {
- if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, EPOLLIN) == -1) {
- warn "epoll_ctl(): failure adding fd=$fd; $! (", $!+0, ")\n";
- }
- }
-
while (1) {
my @events;
my $i;
# get up to 1000 events
my $evcount = epoll_wait($Epoll, 1000, $timeout, \@events);
- EVENT:
for ($i=0; $i<$evcount; $i++) {
my $ev = $events[$i];
# 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) {
- if (my $code = $OtherFds{$ev->[0]}) {
- $code->($state);
- } else {
- 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 && ! $pob->{closed};
- }
+ event_step($DescriptorMap{$ev->[0]});
}
return unless PostEventLoop();
}
# modifies the array in place with the even elements being
# replaced with the event masks that occured.
my @poll;
- foreach my $fd ( keys %OtherFds ) {
- push @poll, $fd, POLLIN;
- }
while ( my ($fd, $sock) = each %DescriptorMap ) {
push @poll, $fd, $sock->{event_watch};
}
my ($fd, $state) = splice(@poll, 0, 2);
next unless $state;
- $pob = $DescriptorMap{$fd};
-
- if (!$pob) {
- if (my $code = $OtherFds{$fd}) {
- $code->($state);
- }
- 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 && ! $pob->{closed};
+ event_step($DescriptorMap{$fd});
}
return unless PostEventLoop();
sub KQueueEventLoop {
my $class = shift;
- foreach my $fd (keys %OtherFds) {
- $KQueue->EV_SET($fd, IO::KQueue::EVFILT_READ(), IO::KQueue::EV_ADD());
- }
-
while (1) {
my $timeout = RunTimers();
my @ret = eval { $KQueue->kevent($timeout) };
foreach my $kev (@ret) {
my ($fd, $filter, $flags, $fflags) = @$kev;
- my PublicInbox::DS $pob = $DescriptorMap{$fd};
- if (!$pob) {
- if (my $code = $OtherFds{$fd}) {
- $code->($filter);
- } else {
- 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 {
- $pob->event_hup;
- }
- }
+ event_step($DescriptorMap{$fd});
}
return unless PostEventLoop();
}
Return 1 (or any true value) from the sub to make the loop continue, 0 or false
and it will exit.
-The callback function will be passed two parameters: \%DescriptorMap, \%OtherFds.
+The callback function will be passed two parameters: \%DescriptorMap
=cut
sub SetPostLoopCallback {
my ($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
# or global) cancels it
my $keep_running = 1;
- # per-object post-loop-callbacks
- for my $plc (values %PLCMap) {
- $keep_running &&= $plc->(\%DescriptorMap, \%OtherFds);
- }
-
# now we're at the very end, call callback if defined
if (defined $PostLoopCallback) {
- $keep_running &&= $PostLoopCallback->(\%DescriptorMap, \%OtherFds);
+ $keep_running &&= $PostLoopCallback->(\%DescriptorMap);
}
return $keep_running;
my ($self, $sock, $exclusive) = @_;
$self = fields::new($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->{wbuf} = [];
+ $self->{wbuf_off} = 0;
$self->{closed} = 0;
my $ev = $self->{event_watch} = POLLERR|POLLHUP|POLLNVAL;
}
retry:
if (epoll_ctl($Epoll, EPOLL_CTL_ADD, $fd, $ev)) {
- if ($!{EINVAL} && ($ev & $EPOLLEXCLUSIVE)) {
+ if ($! == EINVAL && ($ev & $EPOLLEXCLUSIVE)) {
$EPOLLEXCLUSIVE = 0; # old kernel
$ev = $self->{event_watch} = EPOLLIN|EPOLLERR|EPOLLHUP;
goto retry;
### I N S T A N C E M E T H O D S
#####################################################################
-=head2 C<< $obj->steal_socket() >>
-
-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.
-
-=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();
+=head2 C<< $obj->close >>
- # 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<reason> 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;
+ if (my $sock = delete $self->{sock}) {
+ push @ToClose, $sock;
}
return 0;
# we need to flush our write buffer, as there may
# be self-referential closures (sub { $client->close })
# preventing the object from being destroyed
- $self->{write_buf} = [];
+ @{$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) . ")");
- }
+ if ($HaveEpoll && $self->{sock}) {
+ my $fd = fileno($self->{sock});
+ epoll_ctl($Epoll, EPOLL_CTL_DEL, $fd, $self->{event_watch}) and
+ confess("EPOLL_CTL_DEL: $!");
}
- # 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}};
-
# 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
# looked at $pob->{closed} 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;
}
=head2 C<< $obj->sock() >>
return $self->{sock};
}
-=head2 C<< $obj->set_writer_func( CODEREF ) >>
-
-Sets a function to use instead of C<syswrite()> when writing data to the socket.
-
-=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;
-}
-
=head2 C<< $obj->write( $data ) >>
Write the specified data to the underlying handle. I<data> may be scalar,
# just queue data if there's already a wait
my $need_queue;
+ my $wbuf = $self->{wbuf};
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($$bref) : 1;
+ if (scalar @$wbuf) {
+ push @$wbuf, $bref;
return 0;
}
WRITE:
while (1) {
- return 1 unless $bref ||= $self->{write_buf}[0];
+ return 1 unless $bref ||= $wbuf->[0];
my $len;
eval {
if ($@) {
if (UNIVERSAL::isa($bref, "CODE")) {
unless ($need_queue) {
- $self->{write_buf_size}--; # code refs are worth 1
- shift @{$self->{write_buf}};
+ shift @$wbuf;
}
$bref->();
die "Write error: $@ <$bref>";
}
- my $to_write = $len - $self->{write_buf_offset};
- my $written;
- if (my $wtr = $self->{writer_func}) {
- $written = $wtr->($bref, $to_write, $self->{write_buf_offset});
- } else {
- $written = syswrite($self->{sock}, $$bref, $to_write, $self->{write_buf_offset});
- }
+ my $to_write = $len - $self->{wbuf_off};
+ my $written = syswrite($self->{sock}, $$bref, $to_write,
+ $self->{wbuf_off});
if (! defined $written) {
- if ($! == EPIPE) {
- return $self->close("EPIPE");
- } elsif ($! == EAGAIN) {
+ if ($! == EAGAIN) {
# since connection has stuff to write, it should now be
# interested in pending writes:
if ($need_queue) {
- push @{$self->{write_buf}}, $bref;
- $self->{write_buf_size} += $len;
+ push @$wbuf, $bref;
}
- $self->{write_set_watch} = 1 unless $self->{event_watch} & POLLOUT;
$self->watch_write(1);
return 0;
- } elsif ($! == ECONNRESET) {
- return $self->close("ECONNRESET");
}
- DebugLevel >= 1 && $self->debugmsg("Closing connection ($self) due to write error: $!\n");
-
- return $self->close("write_error");
+ return $self->close;
} elsif ($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} += $len;
+ push @$wbuf, $bref;
}
# 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->{wbuf_off} += $written;
$self->on_incomplete_write;
return 0;
} elsif ($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;
- }
+ $self->{wbuf_off} = 0;
+ $self->watch_write(0);
# this was our only write, so we can return immediately
# since we avoided incrementing the buffer size or
# can't be anything else to write.
return 1 if $need_queue;
- $self->{write_buf_size} -= $written;
- shift @{$self->{write_buf}};
+ shift @$wbuf;
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);
}
-=head2 C<< $obj->read( $bytecount ) >>
-
-Read at most I<bytecount> 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;
-}
-
=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.
+provide an implementation of this. The default implementation is a noop
+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"; }
+sub event_read {} # noop
=head2 C<< $obj->event_write() >>
$event &= ~POLLIN if ! $val;
$event |= POLLIN if $val;
+ my $fd = fileno($self->{sock});
# If it changed, set it
if ($event != $self->{event_watch}) {
if ($HaveKQueue) {
- $KQueue->EV_SET($self->{fd}, IO::KQueue::EVFILT_READ(),
+ $KQueue->EV_SET($fd, IO::KQueue::EVFILT_READ(),
$val ? IO::KQueue::EV_ENABLE() : IO::KQueue::EV_DISABLE());
}
elsif ($HaveEpoll) {
- epoll_ctl($Epoll, EPOLL_CTL_MOD, $self->{fd}, $event)
- and $self->dump_error("couldn't modify epoll settings for $self->{fd} " .
- "from $self->{event_watch} -> $event: $! (" . ($!+0) . ")");
+ epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $event) and
+ confess("EPOLL_CTL_MOD: $!");
}
$self->{event_watch} = $event;
}
$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;
- }
+ my $fd = fileno($self->{sock});
# If it changed, set it
if ($event != $self->{event_watch}) {
if ($HaveKQueue) {
- $KQueue->EV_SET($self->{fd}, IO::KQueue::EVFILT_WRITE(),
+ $KQueue->EV_SET($fd, IO::KQueue::EVFILT_WRITE(),
$val ? IO::KQueue::EV_ENABLE() : IO::KQueue::EV_DISABLE());
}
elsif ($HaveEpoll) {
- epoll_ctl($Epoll, EPOLL_CTL_MOD, $self->{fd}, $event)
- and $self->dump_error("couldn't modify epoll settings for $self->{fd} " .
- "from $self->{event_watch} -> $event: $! (" . ($!+0) . ")");
+ epoll_ctl($Epoll, EPOLL_CTL_MOD, $fd, $event) and
+ confess "EPOLL_CTL_MOD: $!";
}
$self->{event_watch} = $event;
}