user/dev discussion of public-inbox itself
 help / color / mirror / code / Atom feed
* [PATCH] listener: support publicinboxdaemon.multiaccept in config
@ 2019-05-13  2:56 Eric Wong
  2019-06-25  6:41 ` Eric Wong
  0 siblings, 1 reply; 3+ messages in thread
From: Eric Wong @ 2019-05-13  2:56 UTC (permalink / raw)
  To: meta

Similar to the nginx "multi_accept" parameter, this allows
single-worker deployments to accept() multiple clients whenever
the listen socket reports readiness via epoll_wait/poll/kevent.

Unlike nginx, we also parse an integer parameter to determine
how many times in a row we call accept().

Single-process deployments with no shared sockets can safely set
"true", here; meaning accept() will be retried indefinitely as
long as there are clients connecting.

Multi-process deployments with many workers should leave this at
the default ("false", or "0") or try a small positive integer.
Using large values with many workers can lead to imbalanced
connections between workers.

We also can't reliably detect whether a listen socket is shared
between multiple processes.  Sockets can be shared via socket
activation in systemd (or similar) and TTIN/TTOU signals can
adjust worker count of our daemons.
---
 lib/PublicInbox/Config.pm   | 29 +++++++++++++++++++++++++++--
 lib/PublicInbox/Listener.pm | 14 +++++++-------
 2 files changed, 34 insertions(+), 9 deletions(-)

diff --git a/lib/PublicInbox/Config.pm b/lib/PublicInbox/Config.pm
index 09f9179..938685a 100644
--- a/lib/PublicInbox/Config.pm
+++ b/lib/PublicInbox/Config.pm
@@ -13,6 +13,9 @@ use warnings;
 require PublicInbox::Inbox;
 use PublicInbox::Spawn qw(popen_rd);
 
+my $FALSE_RE = qr/\A(?:false|no|off|0)\z/;
+my $TRUE_RE = qr/\A(?:true|yes|on|1)\z/;
+
 sub _array ($) { ref($_[0]) eq 'ARRAY' ? $_[0] : [ $_[0] ] }
 
 # returns key-value pairs of config directives in a hash
@@ -54,6 +57,28 @@ sub new {
 		$self->{css} = _array($css);
 	}
 
+	if (defined(my $ma = $self->{'publicinboxdaemon.multiaccept'})) {
+		# multiaccept=0        accept once (default)
+		# multiaccept=(1..X)   accept an extra X times (or EAGAIN)
+		# multiaccept=true     accept until EAGAIN
+		if ($ma =~ $TRUE_RE) {
+			$ma = -1 if $ma ne '1';
+		} elsif ($ma =~ $FALSE_RE) {
+			$ma = 0;
+		} elsif ($ma =~ /\A(\d+)\z/) {
+			# positive integer value, leave as-is
+		} else {
+			warn
+"publicinboxdaemon.multiaccept=$ma not understood\n";
+			$ma = undef;
+		}
+
+		defined($ma) and eval {
+			no warnings 'once';
+			$PublicInbox::Listener::multi_accept = $ma;
+		};
+	}
+
 	$self;
 }
 
@@ -379,9 +404,9 @@ sub _fill {
 	foreach my $k (qw(obfuscate)) {
 		my $v = $self->{"$pfx.$k"};
 		defined $v or next;
-		if ($v =~ /\A(?:false|no|off|0)\z/) {
+		if ($v =~ $FALSE_RE) {
 			$ibx->{$k} = 0;
-		} elsif ($v =~ /\A(?:true|yes|on|1)\z/) {
+		} elsif ($v =~ $TRUE_RE) {
 			$ibx->{$k} = 1;
 		} else {
 			warn "Ignoring $pfx.$k=$v in config, not boolean\n";
diff --git a/lib/PublicInbox/Listener.pm b/lib/PublicInbox/Listener.pm
index a75a6fd..c4d1a55 100644
--- a/lib/PublicInbox/Listener.pm
+++ b/lib/PublicInbox/Listener.pm
@@ -10,6 +10,8 @@ use Socket qw(SOL_SOCKET SO_KEEPALIVE IPPROTO_TCP TCP_NODELAY);
 use fields qw(post_accept);
 require IO::Handle;
 
+our $multi_accept = 0;
+
 sub new ($$$) {
 	my ($class, $s, $cb) = @_;
 	setsockopt($s, SOL_SOCKET, SO_KEEPALIVE, 1);
@@ -26,16 +28,14 @@ sub new ($$$) {
 sub event_read {
 	my ($self) = @_;
 	my $sock = $self->{sock};
+	my $n = $multi_accept;
+	my ($addr, $c);
 
-	# no loop here, we want to fairly distribute clients
-	# between multiple processes sharing the same socket
-	# XXX our event loop needs better granularity for
-	# a single accept() here to be, umm..., acceptable
-	# on high-traffic sites.
-	if (my $addr = accept(my $c, $sock)) {
+	do {
+		$addr = accept($c, $sock) or return;
 		IO::Handle::blocking($c, 0); # no accept4 :<
 		$self->{post_accept}->($c, $addr, $sock);
-	}
+	} while ($n--);
 }
 
 1;
-- 
EW


^ permalink raw reply related	[flat|nested] 3+ messages in thread

* Re: [PATCH] listener: support publicinboxdaemon.multiaccept in config
  2019-05-13  2:56 [PATCH] listener: support publicinboxdaemon.multiaccept in config Eric Wong
@ 2019-06-25  6:41 ` Eric Wong
  2023-04-12 10:17   ` [PATCH v2] listener: support multi-accept like nginx Eric Wong
  0 siblings, 1 reply; 3+ messages in thread
From: Eric Wong @ 2019-06-25  6:41 UTC (permalink / raw)
  To: meta

Eric Wong <e@80x24.org> wrote:
> Similar to the nginx "multi_accept" parameter, this allows
> single-worker deployments to accept() multiple clients whenever
> the listen socket reports readiness via epoll_wait/poll/kevent.
> 
> Unlike nginx, we also parse an integer parameter to determine
> how many times in a row we call accept().

Given the ability to have per-listener cert/key options(*) for
OpenSSL, I think this parameter should be per-listener, too;
and thus specified in the command-line.

(*) which reminds me, I still need to write tests and document that...

^ permalink raw reply	[flat|nested] 3+ messages in thread

* [PATCH v2] listener: support multi-accept like nginx
  2019-06-25  6:41 ` Eric Wong
@ 2023-04-12 10:17   ` Eric Wong
  0 siblings, 0 replies; 3+ messages in thread
From: Eric Wong @ 2023-04-12 10:17 UTC (permalink / raw)
  To: meta

Eric Wong <e@80x24.org> wrote:
> Eric Wong <e@80x24.org> wrote:
> > Similar to the nginx "multi_accept" parameter, this allows
> > single-worker deployments to accept() multiple clients whenever
> > the listen socket reports readiness via epoll_wait/poll/kevent.
> 
> Given the ability to have per-listener cert/key options(*) for
> OpenSSL, I think this parameter should be per-listener, too;
> and thus specified in the command-line.

Done (after a few years :x).  Hopefully the explanation makes
sense, English brain is not good today :x

--------8<---------
Subject: [PATCH] listener: support multi-accept like nginx

While accepting a single connection at-a-time is likely best for
multi-worker and/or load-balanced deployments; accepting
multiple connections at once should be less bad on overloaded
single-worker systems.

We can't automatically pick the best value here since worker
counts are dynamic via SIGTTIN/SIGTTOU.  Process managers
(e.g. systemd) can also spawn multiple instances sharing a
single listener with no knowledge sharing between listeners.
---
 Documentation/public-inbox-daemon.pod | 17 ++++++++++
 lib/PublicInbox/Daemon.pm             |  8 +++--
 lib/PublicInbox/Listener.pm           | 45 +++++++++++++--------------
 3 files changed, 45 insertions(+), 25 deletions(-)

diff --git a/Documentation/public-inbox-daemon.pod b/Documentation/public-inbox-daemon.pod
index 81a79a10..71216833 100644
--- a/Documentation/public-inbox-daemon.pod
+++ b/Documentation/public-inbox-daemon.pod
@@ -115,6 +115,23 @@ per-listener C<cert=> option.  The private key may be
 concatenated into the path used by the cert, in which case this
 option is not needed.
 
+=item --multi-accept INTEGER
+
+By default, each worker accepts one connection at-a-time to maximize
+fairness and minimize contention across multiple processes on a
+shared listen socket.  Accepting multiple connections at once may be
+useful in constrained deployments with few, heavily-loaded workers.
+Negative values enables a worker to accept all available clients at
+once, possibly starving others in the process.  C<-1> behaves like
+C<multi_accept yes> in nginx; while C<0> (the default) is
+C<multi_accept no> in nginx.  Positive values allow
+fine-tuning without the runaway behavior of C<-1>.
+
+This may be specified on a per-listener basis via the C<multi-accept=>
+per-listener directive (e.g. C<-l http://127.0.0.1?multi-accept=1>).
+
+Default: 0
+
 =back
 
 =head1 SIGNALS
diff --git a/lib/PublicInbox/Daemon.pm b/lib/PublicInbox/Daemon.pm
index 57435421..30442227 100644
--- a/lib/PublicInbox/Daemon.pm
+++ b/lib/PublicInbox/Daemon.pm
@@ -136,6 +136,8 @@ sub load_mod ($;$$) {
 	}
 	my $err = $tlsd->{err};
 	$tlsd->{warn_cb} = sub { print $err @_ }; # for local $SIG{__WARN__}
+	$opt->{'multi-accept'} and
+		$xn{'multi-accept'} = $opt->{'multi-accept'}->[-1];
 	\%xn;
 }
 
@@ -167,6 +169,7 @@ EOF
 		'u|user=s' => \$user,
 		'g|group=s' => \$group,
 		'D|daemonize' => \$daemonize,
+		'multi-accept=i' => \$PublicInbox::Listener::MULTI_ACCEPT,
 		'cert=s' => \$default_cert,
 		'key=s' => \$default_key,
 		'help|h' => \(my $show_help),
@@ -251,7 +254,7 @@ EOF
 		$s->blocking(0);
 		my $sockname = sockname($s);
 		warn "# bound $scheme://$sockname\n";
-		$xnetd->{$sockname} //= load_mod($scheme);
+		$xnetd->{$sockname} //= load_mod($scheme, $opt);
 		$listener_names->{$sockname} = $s;
 		push @listeners, $s;
 	}
@@ -712,7 +715,8 @@ sub daemon_loop ($) {
 		defer_accept($_, $tls_cb ? 'dataready' : $xn->{af_default});
 
 		# this calls epoll_create:
-		PublicInbox::Listener->new($_, $tls_cb || $xn->{post_accept})
+		PublicInbox::Listener->new($_, $tls_cb || $xn->{post_accept},
+						$xn->{'multi-accept'})
 	} @listeners;
 	PublicInbox::DS::event_loop($sig, $oldset);
 }
diff --git a/lib/PublicInbox/Listener.pm b/lib/PublicInbox/Listener.pm
index 7cedc349..4669cf04 100644
--- a/lib/PublicInbox/Listener.pm
+++ b/lib/PublicInbox/Listener.pm
@@ -1,14 +1,15 @@
-# Copyright (C) 2015-2021 all contributors <meta@public-inbox.org>
+# Copyright (C) all contributors <meta@public-inbox.org>
 # License: AGPL-3.0+ <https://www.gnu.org/licenses/agpl-3.0.txt>
 #
 # Used by -nntpd for listen sockets
 package PublicInbox::Listener;
-use strict;
+use v5.12;
 use parent 'PublicInbox::DS';
 use Socket qw(SOL_SOCKET SO_KEEPALIVE IPPROTO_TCP TCP_NODELAY);
 use IO::Handle;
 use PublicInbox::Syscall qw(EPOLLIN EPOLLEXCLUSIVE);
 use Errno qw(EAGAIN ECONNABORTED);
+our $MULTI_ACCEPT = 0;
 
 # Warn on transient errors, mostly resource limitations.
 # EINTR would indicate the failure to set NonBlocking in systemd or similar
@@ -16,37 +17,35 @@ my %ERR_WARN = map {;
 	eval("Errno::$_()") => $_
 } qw(EMFILE ENFILE ENOBUFS ENOMEM EINTR);
 
-sub new ($$$) {
-	my ($class, $s, $cb) = @_;
+sub new {
+	my ($class, $s, $cb, $multi_accept) = @_;
 	setsockopt($s, SOL_SOCKET, SO_KEEPALIVE, 1);
 	setsockopt($s, IPPROTO_TCP, TCP_NODELAY, 1); # ignore errors on non-TCP
 	listen($s, 2**31 - 1); # kernel will clamp
 	my $self = bless { post_accept => $cb }, $class;
+	$self->{multi_accept} = $multi_accept //= $MULTI_ACCEPT;
 	$self->SUPER::new($s, EPOLLIN|EPOLLEXCLUSIVE);
 }
 
 sub event_step {
 	my ($self) = @_;
 	my $sock = $self->{sock} or return;
-
-	# no loop here, we want to fairly distribute clients
-	# between multiple processes sharing the same socket
-	# XXX our event loop needs better granularity for
-	# a single accept() here to be, umm..., acceptable
-	# on high-traffic sites.
-	if (my $addr = accept(my $c, $sock)) {
-		IO::Handle::blocking($c, 0); # no accept4 :<
-		eval { $self->{post_accept}->($c, $addr, $sock) };
-		warn "E: $@\n" if $@;
-	} elsif ($! == EAGAIN || $! == ECONNABORTED) {
-		# EAGAIN is common and likely
-		# ECONNABORTED is common with bad connections
-		return;
-	} elsif (my $sym = $ERR_WARN{int($!)}) {
-		warn "W: accept(): $! ($sym)\n";
-	} else {
-		warn "BUG?: accept(): $!\n";
-	}
+	my $n = $self->{multi_accept};
+	do {
+		if (my $addr = accept(my $c, $sock)) {
+			IO::Handle::blocking($c, 0); # no accept4 :<
+			eval { $self->{post_accept}->($c, $addr, $sock) };
+			warn "E: $@\n" if $@;
+		} elsif ($! == EAGAIN || $! == ECONNABORTED) {
+			# EAGAIN is common and likely
+			# ECONNABORTED is common with bad connections
+			return;
+		} elsif (my $sym = $ERR_WARN{int($!)}) {
+			warn "W: accept(): $! ($sym)\n";
+		} else {
+			warn "BUG?: accept(): $!\n";
+		}
+	} while ($n--);
 }
 
 1;

^ permalink raw reply related	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2023-04-12 10:17 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-05-13  2:56 [PATCH] listener: support publicinboxdaemon.multiaccept in config Eric Wong
2019-06-25  6:41 ` Eric Wong
2023-04-12 10:17   ` [PATCH v2] listener: support multi-accept like nginx Eric Wong

Code repositories for project(s) associated with this public inbox

	https://80x24.org/public-inbox.git

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).