On Tue, Dec 17, 2013 at 02:05:13PM +0100, Willy Tarreau wrote: > On Tue, Dec 17, 2013 at 10:44:12AM +0100, Guillaume Castagnino wrote: > > Le mardi 17 décembre 2013 10:32:30 Sander Klein a écrit : > > > Hi, > > > > > > I've enabled http-keep-alive in my config and now haproxy continuously > > > peaks at 100% CPU usage where without http-keep-alive it only uses > > > 10-13% CPU. > > > > > > Is this normal/expected behavior? > > > > Hi, > > > > Indeed, I can confirm this behaviour when enabling server-side > > keepalive. > > So it looks like the simple idle connection manager I did yesterday > is still not perfect :-/ > I tried to trigger this case but could not manage to make it fail, > so I considered that was OK.
OK I could reproduce it. I only focused on closed connections yesterday, and not the maitained ones :-( Here's the fix. Thanks guys! Willy
>From ea90063cbcc03506510522809d564112aa1a9ce9 Mon Sep 17 00:00:00 2001 From: Willy Tarreau <[email protected]> Date: Tue, 17 Dec 2013 14:21:48 +0100 Subject: BUG/MEDIUM: stream-int: fix the keep-alive idle connection handler Commit 2737562 (MEDIUM: stream-int: implement a very simplistic idle connection manager) implemented an idle connection handler. In the case where all data is drained from the server, it fails to disable polling, resulting in a busy spinning loop. Thanks to Sander Klein and Guillaume Castagnino for reporting this bug. No backport is needed. --- src/stream_interface.c | 2 ++ 1 files changed, 2 insertions(+), 0 deletions(-) diff --git a/src/stream_interface.c b/src/stream_interface.c index 5c4633b..bf0a400 100644 --- a/src/stream_interface.c +++ b/src/stream_interface.c @@ -508,6 +508,8 @@ static void si_idle_conn_null_cb(struct connection *conn) /* disable draining if we were called and have no drain function */ if (!conn->ctrl->drain) __conn_data_stop_recv(conn); + else if (!(conn->flags & CO_FL_SOCK_RD_SH)) + __conn_data_poll_recv(conn); } /* Callback to be used by connection I/O handlers when some activity is detected -- 1.7.1
