From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751369AbbAMIKU (ORCPT ); Tue, 13 Jan 2015 03:10:20 -0500 Received: from service87.mimecast.com ([91.220.42.44]:34531 "EHLO service87.mimecast.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751086AbbAMIKT convert rfc822-to-8bit (ORCPT ); Tue, 13 Jan 2015 03:10:19 -0500 Message-ID: <54B4D2DF.9010308@arm.com> Date: Tue, 13 Jan 2015 08:10:07 +0000 From: Juri Lelli User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101 Thunderbird/31.3.0 MIME-Version: 1.0 To: tkhai@yandex.ru, Luca Abeni CC: Peter Zijlstra , Ingo Molnar , "linux-kernel@vger.kernel.org" , juri.lelli@gmail.com Subject: Re: Another SCHED_DEADLINE bug (with bisection and possible fix) References: <20141230002738.6c12db31@utopia> <2629881420411885@web9m.yandex.ru> <54AAABFB.3060109@unitn.it> <1420499241.3361.14.camel@yandex.ru> <20150107080155.1d42d123@luca-1225C> <1420633741.12772.10.camel@yandex.ru> In-Reply-To: <1420633741.12772.10.camel@yandex.ru> X-OriginalArrivalTime: 13 Jan 2015 08:10:14.0713 (UTC) FILETIME=[5C378290:01D02F08] X-MC-Unique: 115011308101700201 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8BIT Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Hi all, really sorry for the huge delay in replying to this! :( On 07/01/2015 12:29, Kirill Tkhai wrote: > On Ср, 2015-01-07 at 08:01 +0100, Luca Abeni wrote: >> Hi Kirill, >> >> On Tue, 06 Jan 2015 02:07:21 +0300 >> Kirill Tkhai wrote: >> >>> On Пн, 2015-01-05 at 16:21 +0100, Luca Abeni wrote: >> [...] >>>> For reference, I attach the patch I am using locally (based on what >>>> I suggested in my previous mail) and seems to work fine here. >>>> >>>> Based on your comments, I suspect my patch can be further >>>> simplified by moving the call to init_dl_task_timer() in >>>> __sched_fork(). >>> >>> It seems this way has problems. The first one is that task may become >>> throttled again, and we will start dl_timer again. >> Well, in my understanding if I change the parameters of a >> SCHED_DEADLINE task when it is throttled, it stays throttled... So, the >> task might not become throttled again before the dl timer fires. >> So, I hoped this problem does not exist. But I might be wrong. > > You keep zeroing of dl_se->dl_throttled, and further enqueue_task() > places it on the dl_rq. So, further update_curr_dl() may make it throttled > again, and it will try to start dl_timer (which is already set). > >>> The second is that >>> it's better to minimize number of combination of situations we have. >>> Let's keep only one combination: timer is set <-> task is throttled. >> Yes, this was my goal too... So, if I change the parameters of a task >> when it is throttled, I leave dl_throttled set to 1 and I leave the >> timer active. > > As I see, > > dl_se->dl_throttled = 0; > > is still in __setparam_dl() after your patch, so you do not leave > it set to 1. > >> >> [...] >>>>> @@ -3250,16 +3251,19 @@ static void >>>>> __setparam_dl(struct task_struct *p, const struct sched_attr >>>>> *attr) { >>>>> struct sched_dl_entity *dl_se = &p->dl; >>>>> + struct hrtimer *timer = &dl_se->dl_timer; >>>>> + >>>>> + if (!hrtimer_active(timer) || >>>>> hrtimer_try_to_cancel(timer) != -1) { >>>> Just for the sake of curiosity, why trying to cancel the timer >>>> ("|| hrtimer_try_to_cancel(timer)") here? If it is active, cannot >>>> we leave it active (without touching dl_throttled, dl_new and >>>> dl_yielded)? >>>> >>>> I mean: if I try to change the parameters of a task when it is >>>> throttled, I'd like it to stay throttled until the end of the >>>> reservation period... Or am I missing something? >>> >>> I think that when people change task's parameters, they want the >>> kernel reacts on this immediately. For example, you want to kill >>> throttled deadline task. You change parameters, but nothing happens. >>> I think all developers had this use case when they were debugging >>> deadline class. >> I see... Different people have different requirements :) >> My goal was to do something like adaptive scheduling (or scheduling >> tasks with mode changes), so I did not want that changing the >> scheduling parameters of a task affected the scheduling of the other >> tasks... But if a task exits the throttled state when I change its >> parameters, it might consume much more than the reserved CPU time. >> Also, I suspect this kind of approach can be exploited by malicious >> users: if I create a task with runtime 30ms and period 100ms, and I >> change its scheduling parameters (to runtime=29ms and back) frequently >> enough, I can consume much more than 30% of the CPU time... >> Well, I'm inclined to agree to Luca's viewpoint. We should not change parameters of a throttled task or we may affect other tasks. >> Anyway, I am fine with every patch that fixes the bug :) > > Deadline class requires root privileges. So, I do not see a problem > here. Please, see __sched_setscheduler(). > > If in the future we allow non-privileged users to increase deadline, > we will reflect that in __setparam_dl() too. > I'd say it is better to implement the right behavior even for root, so that we will find it right when we'll grant access to non root users too. Also, even if root can do everything, we always try not to break guarantees that come with admission control (root or non root that is). Thanks a lot, - Juri