From mboxrd@z Thu Jan  1 00:00:00 1970
From: Eric Dumazet <eric.dumazet@gmail.com>
Subject: Re: [PATCH v2 net-next] tcp: avoid expensive pskb_expand_head()
 calls
Date: Thu, 19 Apr 2012 16:10:52 +0200
Message-ID: <1334844652.2395.187.camel@edumazet-glaptop>
References: <1334653608.6226.11.camel@edumazet-laptop>
	 <1334654187.2696.2.camel@jtkirshe-mobl> <4F8D93E1.9090000@intel.com>
	 <1334681204.2472.41.camel@edumazet-glaptop>
	 <1334698722.2472.71.camel@edumazet-glaptop>
	 <1334764184.2472.299.camel@edumazet-glaptop>
	 <CADVnQy=BkhSBHyN2hyBy=_H64oM8sJvyZZfEjK1x7PYzoLv=5w@mail.gmail.com>
	 <1334776707.2472.316.camel@edumazet-glaptop>
	 <1334778707.2472.333.camel@edumazet-glaptop>
	 <alpine.DEB.2.00.1204191401240.735@wel-95.cs.helsinki.fi>
	 <1334835018.2395.66.camel@edumazet-glaptop>
	 <1334841481.2395.175.camel@edumazet-glaptop>
	 <1334843527.2395.182.camel@edumazet-glaptop>
Mime-Version: 1.0
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 7bit
Cc: Neal Cardwell <ncardwell@google.com>,
	David Miller <davem@davemloft.net>,
	netdev <netdev@vger.kernel.org>,
	Tom Herbert <therbert@google.com>,
	Maciej =?UTF-8?Q?=C5=BBenczykowski?= <maze@google.com>,
	Yuchung Cheng <ycheng@google.com>
To: Ilpo =?ISO-8859-1?Q?J=E4rvinen?= <ilpo.jarvinen@helsinki.fi>
Return-path: <netdev-owner@vger.kernel.org>
Received: from mail-bk0-f46.google.com ([209.85.214.46]:50258 "EHLO
	mail-bk0-f46.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
	with ESMTP id S1755447Ab2DSOK5 (ORCPT
	<rfc822;netdev@vger.kernel.org>); Thu, 19 Apr 2012 10:10:57 -0400
Received: by bkcik5 with SMTP id ik5so6400762bkc.19
        for <netdev@vger.kernel.org>; Thu, 19 Apr 2012 07:10:56 -0700 (PDT)
In-Reply-To: <1334843527.2395.182.camel@edumazet-glaptop>
Sender: netdev-owner@vger.kernel.org
List-ID: <netdev.vger.kernel.org>

On Thu, 2012-04-19 at 15:52 +0200, Eric Dumazet wrote:

> And disabling GRO on receiver definitely demonstrates the problem, even
> with a single flow. (and performance drops from 9410 Mbit to 6050 Mbit)

That insane. 

Performance drops so much because we _drop_ incoming ACKS :

<     TCPSackShifted: 39117
<     TCPSackMerged: 16500
<     TCPSackShiftFallback: 5092
<     TCPBacklogDrop: 27965
---
>     TCPSackShifted: 35122
>     TCPSackMerged: 16368
>     TCPSackShiftFallback: 4889
>     TCPBacklogDrop: 23247

Hmm, maybe we should reduce skb->truesize for small packets before
queueing them in socket backlog...