From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: e1000e 2.6.30.1 massive packet loss Date: Thu, 16 Jul 2009 10:04:07 +0200 Message-ID: <4A5EDEF7.7090303@cosmosbay.com> References: <81bfc67a0907160049q3d43093cy52276f0fe1157d89@mail.gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: Linux Kernel Mailing List , Linux Netdev List To: Caleb Cushing Return-path: Received: from gw1.cosmosbay.com ([212.99.114.194]:59733 "EHLO gw1.cosmosbay.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752754AbZGPIEO (ORCPT ); Thu, 16 Jul 2009 04:04:14 -0400 In-Reply-To: <81bfc67a0907160049q3d43093cy52276f0fe1157d89@mail.gmail.com> Sender: netdev-owner@vger.kernel.org List-ID: Caleb Cushing a =C3=A9crit : > So from 2.6.30 to .1 I noticed that via an icmp test (mtr) 20-50% > packet loss web browsing became impractical (downgrading fixed the > issue). I'm wondering if the change to the driver (w/ msi) is the > cause? also wondering if this is known and will be fixed in .2 I can > provide more information if necessary. >=20 Added netdev to get more attention to this. What do you call 20-50% packet loss ? (Losses to icmp flood itself, or tcp losses ? Is the machine answering to icmp flood, or is it the origin ? Are you saying packet loss was not occurring with 2.6.30 ? Could you describe your methodology ? You could give us results of "cat /proc/interrupts" from both versions, because it sounds as an interrupt affinity problem. # cat /proc/interrupts # tc -s -d qdisc # ifconfig -a