[PATCH] inboxidle: avoid needless syscalls on refresh

[PATCH] inboxidle: avoid needless syscalls on refresh

[Eric Wong]

We don't have to replace a bunch of existing watches
with identical new ones.  On Linux with Linux::Inotify2
installed, this avoids a storm of inotify_add_watch(2)
and inotify_rm_watch(2) syscalls on SIGHUP with -imapd
and "-extindex --watch"
---
 lib/PublicInbox/InboxIdle.pm | 15 ++++++++++++---
 t/imapd.t                    | 26 +++++---------------------
 2 files changed, 17 insertions(+), 24 deletions(-)

diff --git a/lib/PublicInbox/InboxIdle.pm b/lib/PublicInbox/InboxIdle.pm
index 69da5f4f..35aed696 100644
--- a/lib/PublicInbox/InboxIdle.pm
+++ b/lib/PublicInbox/InboxIdle.pm
@@ -24,17 +24,26 @@ sub in2_arm ($$) { # PublicInbox::Config::each_inbox callback
 	my $dir = $ibx->{inboxdir};
 	my $inot = $self->{inot};
 	my $cur = $self->{pathmap}->{$dir} //= [];
+	my $lock = "$dir/".($ibx->version >= 2 ? 'inbox.lock' : 'ssoma.lock');
 
 	# transfer old subscriptions to the current inbox, cancel the old watch
-	if (my $old_ibx = $cur->[0]) {
+	my $old_ibx = $cur->[0];
+	$cur->[0] = $ibx;
+	if ($old_ibx) {
 		$ibx->{unlock_subs} and
 			die "BUG: $dir->{unlock_subs} should not exist";
 		$ibx->{unlock_subs} = $old_ibx->{unlock_subs};
+
+		# Linux::Inotify2::Watch::name matches if watches are the
+		# same, no point in replacing a watch of the same name
+		if ($cur->[1]->name eq $lock) {
+			$self->{on_unlock}->{$lock} = $ibx;
+			return;
+		}
+		# rare, name changed (v1 inbox converted to v2)
 		$cur->[1]->cancel; # Linux::Inotify2::Watch::cancel
 	}
-	$cur->[0] = $ibx;
 
-	my $lock = "$dir/".($ibx->version >= 2 ? 'inbox.lock' : 'ssoma.lock');
 	if (my $w = $cur->[1] = $inot->watch($lock, $IN_MODIFY)) {
 		$self->{on_unlock}->{$w->name} = $ibx;
 	} else {
diff --git a/t/imapd.t b/t/imapd.t
index 43ec200c..63a86e71 100644
--- a/t/imapd.t
+++ b/t/imapd.t
@@ -296,27 +296,11 @@ $pi_cfg->each_inbox(sub {
 
 	# ensure IDLE persists across HUP, w/o extra watches or FDs
 	$td->kill('HUP') or BAIL_OUT "failed to kill -imapd: $!";
-	SKIP: {
-		skip 'no inotify fdinfo (or support)', 2 if !@ino_info;
-		my (@tmp, %prev);
-		local $/ = "\n";
-		my $end = time + 5;
-		until (time > $end) {
-			select undef, undef, undef, 0.01;
-			open my $fh, '<', $ino_fdinfo or
-						BAIL_OUT "$ino_fdinfo: $!";
-			%prev = map { $_ => 1 } @ino_info;
-			@tmp = grep(/^inotify wd:/, <$fh>);
-			if (scalar(@tmp) == scalar(@ino_info)) {
-				delete @prev{@tmp};
-				last if scalar(keys(%prev)) == @ino_info;
-			}
-		}
-		is(scalar @tmp, scalar @ino_info,
-			'old inotify watches replaced');
-		is(scalar keys %prev, scalar @ino_info,
-			'no previous watches overlap');
-	};
+	for my $n (1..2) { # kick the event loop so we know HUP is done
+		my $m = $imap_client->new(%mic_opt);
+		ok($m->login && $m->IsAuthenticated && $m->logout,
+			"connection $n works after HUP");
+	}
 
 	open($fh, '<', 't/data/0001.patch') or BAIL_OUT("open: $!");
 	run_script(['-mda', '--no-precheck'], $env, { 0 => $fh }) or

Re: [PATCH] inboxidle: avoid needless syscalls on refresh

[Eric Wong]

Eric Wong <e@80x24.org> wrote:
> We don't have to replace a bunch of existing watches
> with identical new ones.  On Linux with Linux::Inotify2
> installed, this avoids a storm of inotify_add_watch(2)
> and inotify_rm_watch(2) syscalls on SIGHUP with -imapd
> and "-extindex --watch"

Addendum: this seems to fix missed wakeups after SIGHUP with
many (50K) inboxes, I'm not sure if there's some other threshold
that's missing or if there's a bug somewhere else...