From mboxrd@z Thu Jan 1 00:00:00 1970 Content-Type: multipart/mixed; boundary="===============2542398828987898150==" MIME-Version: 1.0 From: Jiang Liu To: lkp@lists.01.org Subject: Re: [x86/MSI] 52f518a3a7c: -30.5% netperf.Throughput_tps Date: Tue, 16 Jun 2015 10:09:29 +0800 Message-ID: <557F8559.4040105@linux.intel.com> In-Reply-To: List-Id: --===============2542398828987898150== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On 2015/6/16 1:52, Thomas Gleixner wrote: > On Mon, 15 Jun 2015, Huang Ying wrote: > = >> FYI, we noticed the below changes on >> >> git://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git master >> commit 52f518a3a7c2f80551a38d38be28bc9f335e713c ("x86/MSI: Use hierarchi= cal irqdomains to manage MSI interrupts") >> > = > I really appreciate this testing effort, but the information provided > is not really helpful. > = > I asked this before. Can you pretty please, upload ALL relevant > information (.config, full dmesg, below stats, /proc/interrupts ...) > to some place where everyone interested can download them? > = > Then themail contains a useful link instead of 200k waste of network band= width. Hi Ying and Thomas, I guess this report discloses a regression in hierarchy irqdomain, and which should have been fixed by the patch posted at: lkml.org/lkml/2015/6/1/80 The root cause is that, with hierarchy irqdomain enabled, there are multiple irq_data associated with one irq. And function irq_move_irq() on x86 uses a wrong copy of irq_data to check whether there's pending irq migration operation. So all irq migration /set_affinity operations will get pending for ever. This may affect network performance due to interrupt load balance issue. And the patch set posted at www.gossamer-threads.com/lists/linux/kernel/2185533 should have solved all such regressions. Thanks! Gerry > = > Thanks, > = > tglx = >=20 --===============2542398828987898150==-- From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751799AbbFPCJl (ORCPT ); Mon, 15 Jun 2015 22:09:41 -0400 Received: from mga11.intel.com ([192.55.52.93]:56365 "EHLO mga11.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751062AbbFPCJd (ORCPT ); Mon, 15 Jun 2015 22:09:33 -0400 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.13,622,1427785200"; d="scan'208";a="747314541" Message-ID: <557F8559.4040105@linux.intel.com> Date: Tue, 16 Jun 2015 10:09:29 +0800 From: Jiang Liu Organization: Intel User-Agent: Mozilla/5.0 (Windows NT 6.2; WOW64; rv:31.0) Gecko/20100101 Thunderbird/31.7.0 MIME-Version: 1.0 To: Thomas Gleixner , Huang Ying CC: LKML , LKP ML Subject: Re: [lkp] [x86/MSI] 52f518a3a7c: -30.5% netperf.Throughput_tps References: <557E786D.4080906@intel.com> In-Reply-To: Content-Type: text/plain; charset=windows-1252 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 2015/6/16 1:52, Thomas Gleixner wrote: > On Mon, 15 Jun 2015, Huang Ying wrote: > >> FYI, we noticed the below changes on >> >> git://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git master >> commit 52f518a3a7c2f80551a38d38be28bc9f335e713c ("x86/MSI: Use hierarchical irqdomains to manage MSI interrupts") >> > > I really appreciate this testing effort, but the information provided > is not really helpful. > > I asked this before. Can you pretty please, upload ALL relevant > information (.config, full dmesg, below stats, /proc/interrupts ...) > to some place where everyone interested can download them? > > Then themail contains a useful link instead of 200k waste of network bandwidth. Hi Ying and Thomas, I guess this report discloses a regression in hierarchy irqdomain, and which should have been fixed by the patch posted at: lkml.org/lkml/2015/6/1/80 The root cause is that, with hierarchy irqdomain enabled, there are multiple irq_data associated with one irq. And function irq_move_irq() on x86 uses a wrong copy of irq_data to check whether there's pending irq migration operation. So all irq migration /set_affinity operations will get pending for ever. This may affect network performance due to interrupt load balance issue. And the patch set posted at www.gossamer-threads.com/lists/linux/kernel/2185533 should have solved all such regressions. Thanks! Gerry > > Thanks, > > tglx >