From mboxrd@z Thu Jan 1 00:00:00 1970 From: Patrick McHardy Subject: Re: NET_SCHED cbq dropping too many packets on a bonding interface Date: Thu, 15 May 2008 20:14:39 +0200 Message-ID: <482C7D8F.3040401@trash.net> References: <20080515091216.GA6550@ff.dom.local> <8ECDBB4EB5394859BFFACAAEE3A6EDB0@uglypunk> <482C6040.9030808@trash.net> <20080515180914.GA2936@ami.dom.local> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-15; format=flowed Content-Transfer-Encoding: 7bit Cc: Kingsley Foreman , Eric Dumazet , Andrew Morton , linux-kernel@vger.kernel.org, netdev@vger.kernel.org To: Jarek Poplawski Return-path: Received: from stinky.trash.net ([213.144.137.162]:49509 "EHLO stinky.trash.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751479AbYEOSOm (ORCPT ); Thu, 15 May 2008 14:14:42 -0400 In-Reply-To: <20080515180914.GA2936@ami.dom.local> Sender: netdev-owner@vger.kernel.org List-ID: Jarek Poplawski wrote: > On Thu, May 15, 2008 at 06:09:36PM +0200, Patrick McHardy wrote: >> Kingsley Foreman wrote: >>> i just rolled back the kernel to 2.6.24 and im seeing the same thing, >>> >>> I was using 2.6.22 before and didn't see the problem, txqueuelen on the >>> bond0 interface is 0 (the default) >> That might explain things, although it shouldn't have worked before >> either. >> >> CBQ creates default pfifo qdiscs for its leaves, these use a limit >> of txqueuelen or 1 if it is zero. So even small bursts will cause >> drops. Do things improve if you set txqueuelen to a larger value >> *before* configuring the qdiscs? > > Kingsley wrote to me that even after changing txqueuelen to 1000 the > "dropped" number didn't change much. A debugging patch with printks > around all "sch->qstats.dropps++" showed only the end of cbq_enqueue(). Thats where packets dropped by default pfifo would be accounted. Did you change txqueuelen before or after setting up the qdiscs? > I've asked to check tomorrow "pfifo limit 1000" for these drops too. That will clear it up. >> Another thing is that CBQ on bond will probably not work properly >> at all, it needs a real device since it measures the timing between >> dequeue events for idle time estimation. On software devices this >> doesn't work. > > Right, but these drops without any sign of overactions or overlimits > seem to show it's not about shaping (or it's not counted/documented > enough). Yes, these drops are probably unrelated, just thought I mention it.