user/dev discussion of public-inbox itself
 help / color / Atom feed
From: Eric Wong <e@80x24.org>
To: meta@public-inbox.org
Subject: [PATCH 10/11] httpd/async: switch to buffering-as-fast-as-possible
Date: Sat, 29 Jun 2019 19:59:50 +0000
Message-ID: <20190629195951.32160-11-e@80x24.org> (raw)
In-Reply-To: <20190629195951.32160-1-e@80x24.org>

With DS buffering to a temporary file nowadays, applying
backpressure to git-http-backend(1) hurts overall memory
usage of the system.  Instead, try to get git-http-backend(1)
to finish as quickly as possible and use edge-triggered
notifications to reduce wakeups on our end.
---
 lib/PublicInbox/DS.pm          |  6 ------
 lib/PublicInbox/HTTP.pm        |  7 -------
 lib/PublicInbox/HTTPD/Async.pm | 40 +++++++++++++++++-----------------------
 3 files changed, 17 insertions(+), 36 deletions(-)

diff --git a/lib/PublicInbox/DS.pm b/lib/PublicInbox/DS.pm
index b2f59983..a8236023 100644
--- a/lib/PublicInbox/DS.pm
+++ b/lib/PublicInbox/DS.pm
@@ -571,12 +571,6 @@ sub epwait ($$) {
     0;
 }
 
-sub watch ($$) {
-    my ($self, $ev) = @_;
-    my $sock = $self->{sock} or return;
-    epwait($sock, $ev);
-}
-
 # return true if complete, false if incomplete (or failure)
 sub accept_tls_step ($) {
     my ($self) = @_;
diff --git a/lib/PublicInbox/HTTP.pm b/lib/PublicInbox/HTTP.pm
index 680be72b..5546ac46 100644
--- a/lib/PublicInbox/HTTP.pm
+++ b/lib/PublicInbox/HTTP.pm
@@ -467,11 +467,4 @@ sub busy () {
 	($self->{rbuf} || $self->{env} || $self->{wbuf});
 }
 
-# fires after pending writes are complete:
-sub restart_pass ($) {
-	$_[0]->{forward}->restart_read; # see PublicInbox::HTTPD::Async
-}
-
-sub enqueue_restart_pass ($) { $_[0]->write(\&restart_pass) }
-
 1;
diff --git a/lib/PublicInbox/HTTPD/Async.pm b/lib/PublicInbox/HTTPD/Async.pm
index 35d17150..a468ed91 100644
--- a/lib/PublicInbox/HTTPD/Async.pm
+++ b/lib/PublicInbox/HTTPD/Async.pm
@@ -4,14 +4,15 @@
 # XXX This is a totally unstable API for public-inbox internal use only
 # This is exposed via the 'pi-httpd.async' key in the PSGI env hash.
 # The name of this key is not even stable!
-# Currently is is intended for use with read-only pipes.
+# Currently intended for use with read-only pipes with expensive
+# processes such as git-http-backend(1), cgit(1)
 package PublicInbox::HTTPD::Async;
 use strict;
 use warnings;
 use base qw(PublicInbox::DS);
 use fields qw(cb cleanup);
-require PublicInbox::EvCleanup;
 use Errno qw(EAGAIN);
+use PublicInbox::Syscall qw(EPOLLIN EPOLLET);
 
 sub new {
 	my ($class, $io, $cb, $cleanup) = @_;
@@ -26,14 +27,12 @@ sub new {
 
 	my $self = fields::new($class);
 	IO::Handle::blocking($io, 0);
-	$self->SUPER::new($io, PublicInbox::DS::EPOLLIN());
+	$self->SUPER::new($io, EPOLLIN | EPOLLET);
 	$self->{cb} = $cb;
 	$self->{cleanup} = $cleanup;
 	$self;
 }
 
-sub restart_read ($) { $_[0]->watch(PublicInbox::DS::EPOLLIN()) }
-
 sub main_cb ($$$) {
 	my ($http, $fh, $bref) = @_;
 	sub {
@@ -41,25 +40,15 @@ sub main_cb ($$$) {
 		my $r = sysread($self->{sock}, $$bref, 8192);
 		if ($r) {
 			$fh->write($$bref); # may call $http->close
-
 			if ($http->{sock}) { # !closed
-				if ($http->{wbuf}) {
-					# HTTP client could not keep up, so
-					# stop reading and buffering.
-					$self->watch(0);
-
-					# Tell the HTTP socket to restart us
-					# when HTTP client is done draining
-					# $http->{wbuf}:
-					$http->enqueue_restart_pass;
-				}
-				# stay in EPOLLIN, but let other clients
-				# get some work done, too.
+				$self->requeue;
+				# let other clients get some work done, too
 				return;
 			}
-			# fall through to close below...
-		} elsif (!defined $r) {
-			return restart_read($self) if $! == EAGAIN;
+
+			# else: fall through to close below...
+		} elsif (!defined $r && $! == EAGAIN) {
+			return; # EPOLLET means we'll be notified
 		}
 
 		# Done! Error handling will happen in $fh->close
@@ -75,10 +64,15 @@ sub async_pass {
 	# will automatically close this ($self) object.
 	$http->{forward} = $self;
 	$fh->write($$bref); # PublicInbox:HTTP::{chunked,identity}_wcb
-	$self->{cb} = main_cb($http, $fh, $bref);
+	my $cb = $self->{cb} = main_cb($http, $fh, $bref);
+	$cb->($self); # either hit EAGAIN or ->requeue to keep EPOLLET happy
 }
 
-sub event_step { $_[0]->{cb}->(@_) }
+sub event_step {
+	# {cb} may be undef after ->requeue due to $http->close happening
+	my $cb = $_[0]->{cb} or return;
+	$cb->(@_);
+}
 
 sub close {
 	my $self = $_[0];
-- 
EW


  parent reply index

Thread overview: 12+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-06-29 19:59 [PATCH 00/11] ds: more updates Eric Wong
2019-06-29 19:59 ` [PATCH 01/11] ds: share lazy rbuf handling between HTTP and NNTP Eric Wong
2019-06-29 19:59 ` [PATCH 02/11] ds: move requeue logic over from NNTP Eric Wong
2019-06-29 19:59 ` [PATCH 03/11] http: use requeue instead of watch_in1 Eric Wong
2019-06-29 19:59 ` [PATCH 04/11] listener: use edge-triggered notifications Eric Wong
2019-06-29 19:59 ` [PATCH 05/11] ds: handle deferred DS->close after timers Eric Wong
2019-06-29 19:59 ` [PATCH 06/11] ds: consolidate IO::Socket::SSL checks Eric Wong
2019-06-29 19:59 ` [PATCH 07/11] http: support HTTPS (kinda) Eric Wong
2019-06-29 19:59 ` [PATCH 08/11] parentpipe: document and use one-shot wakeups Eric Wong
2019-06-29 19:59 ` [PATCH 09/11] parentpipe: make the ->close call more obvious Eric Wong
2019-06-29 19:59 ` Eric Wong [this message]
2019-06-29 19:59 ` [PATCH 11/11] http: use bigger, but shorter-lived buffers for pipes Eric Wong

Reply instructions:

You may reply publically to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

  List information: https://public-inbox.org/README

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20190629195951.32160-11-e@80x24.org \
    --to=e@80x24.org \
    --cc=meta@public-inbox.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

user/dev discussion of public-inbox itself

Archives are clonable:
	git clone --mirror https://public-inbox.org/meta
	git clone --mirror http://czquwvybam4bgbro.onion/meta
	git clone --mirror http://hjrcffqmbrq6wope.onion/meta
	git clone --mirror http://ou63pmih66umazou.onion/meta

Newsgroups are available over NNTP:
	nntp://news.public-inbox.org/inbox.comp.mail.public-inbox.meta
	nntp://ou63pmih66umazou.onion/inbox.comp.mail.public-inbox.meta
	nntp://czquwvybam4bgbro.onion/inbox.comp.mail.public-inbox.meta
	nntp://hjrcffqmbrq6wope.onion/inbox.comp.mail.public-inbox.meta
	nntp://news.gmane.org/gmane.mail.public-inbox.general

 note: .onion URLs require Tor: https://www.torproject.org/

AGPL code for this site: git clone https://public-inbox.org/ public-inbox