From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jason Low Subject: Re: [PATCH 1/7] locking/rwsem: don't resched at the end of optimistic spinning Date: Mon, 04 Aug 2014 14:12:34 -0700 Message-ID: <1407186754.11985.7.camel@j-VirtualBox> References: <1407119782-41119-1-git-send-email-Waiman.Long@hp.com> <1407119782-41119-2-git-send-email-Waiman.Long@hp.com> <20140804075528.GI9918@twins.programming.kicks-ass.net> <53DFD2B3.3090101@hp.com> <20140804204824.GT3935@laptop> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20140804204824.GT3935@laptop> Sender: linux-kernel-owner@vger.kernel.org To: Peter Zijlstra Cc: Waiman Long , Ingo Molnar , linux-kernel@vger.kernel.org, linux-api@vger.kernel.org, linux-doc@vger.kernel.org, Davidlohr Bueso , Scott J Norton List-Id: linux-api@vger.kernel.org On Mon, 2014-08-04 at 22:48 +0200, Peter Zijlstra wrote: > On Mon, Aug 04, 2014 at 02:36:35PM -0400, Waiman Long wrote: > > On 08/04/2014 03:55 AM, Peter Zijlstra wrote: > > >On Sun, Aug 03, 2014 at 10:36:16PM -0400, Waiman Long wrote: > > >>For a fully preemptive kernel, a call to preempt_enable() could > > >>potentially trigger a task rescheduling event. In the case of rwsem > > >>optimistic spinning, the task has either gotten the lock or is going > > >>to sleep soon. So there is no point to do rescheduling here. > > >Uh what? Why shouldn't we preempt if we've gotten the lock? What if a > > >FIFO task just woke up? > > > > I didn't mean that we shouldn't preempt if there is a higher priority task. > > I am sure that there will be other preemption points along the way that a > > higher priority task can take over the CPU. I just want to say that doing it > > here may not be the best place especially if the task is going to sleep > > soon. > > > > If you think this patch does not make sense, I can remove it as other > > patches in the set has no dependency on this one. > > Yeah, its actively harmful, you delay preemption by an unspecified > amount of time in case of the spin-acquire. We've had such bugs in -rt > and they're not fun. > > Basically the only time you should use no_resched is if the very next > statement is schedule(). Right, we actually added an extra resched check at essentially the same point in the mutex spinning code :)