From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751997Ab0IHRMG (ORCPT ); Wed, 8 Sep 2010 13:12:06 -0400 Received: from one.firstfloor.org ([213.235.205.2]:57141 "EHLO one.firstfloor.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750845Ab0IHRL7 (ORCPT ); Wed, 8 Sep 2010 13:11:59 -0400 From: Andi Kleen To: Jan Blunck Cc: Peter Zijlstra , Linux-Kernel Mailinglist , linux-rt-users@vger.kernel.org, Sven-Thorsten Dietrich , Michael Galbraith Subject: Re: [RFC 3/4] Disable scheduler tick when we are running SCHED_FIFO tasks References: <1283948964-6418-1-git-send-email-jblunck@suse.de> <1283948964-6418-4-git-send-email-jblunck@suse.de> <1283949943.23762.16.camel@laptop> <20100908142826.GI7436@bolzano.suse.de> Date: Wed, 08 Sep 2010 19:11:53 +0200 In-Reply-To: <20100908142826.GI7436@bolzano.suse.de> (Jan Blunck's message of "Wed, 8 Sep 2010 16:28:26 +0200") Message-ID: <87iq2gw46e.fsf@basil.nowhere.org> User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/23.1 (gnu/linux) MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Jan Blunck writes: > On Wed, Sep 08, Peter Zijlstra wrote: > >> On Wed, 2010-09-08 at 14:29 +0200, Jan Blunck wrote: >> > This patch is disabling the scheduler tick to go off when there is a task >> > with SCHED_FIFO policy running. Since these tasks are not timesliced anyway >> > we only care about timers, softirqs and such stuff just like when we disable >> > the tick during idle periods. >> > >> >> Also, doesn't this break any and all jiffies users? >> > > Sure. It was enough to make FTQ run. > >> And you need to restart the tick on call_rcu() and everything else that >> requires the tick for processing. >> > > I realized that before as well but somehow forgot about it. Checking for a > disabled timer tick on syscall entry would help. Like we do on irq entry as > well. That doesn't help the gettimeofday/clock_gettime vDSOs which access jiffie like state in user space. I guess you could force the vdso to call into the kernel during such a region, but you cannot force it for an process that is already context switched in this path when the option is first enabled. I guess it would work if you force all the vDSOs to always do a system call at boot, but that's pretty costly. -Andi -- ak@linux.intel.com -- Speaking for myself only.