Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753153AbdDLGzz (ORCPT ); Wed, 12 Apr 2017 02:55:55 -0400 Received: from mail.santannapisa.it ([193.205.80.98]:55054 "EHLO mail.santannapisa.it" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751079AbdDLGzw (ORCPT ); Wed, 12 Apr 2017 02:55:52 -0400 Date: Wed, 12 Apr 2017 08:55:44 +0200 From: Luca Abeni To: Xunlei Pang Cc: Daniel Bristot de Oliveira , Xunlei Pang , linux-kernel@vger.kernel.org, Peter Zijlstra , Juri Lelli , Ingo Molnar , Steven Rostedt , Tommaso Cucinotta , =?UTF-8?B?UsO0bXVsbw==?= Silva de Oliveira , Mathieu Poirier Subject: Re: [PATCH] sched/deadline: Throttle a constrained task activated if overflow Message-ID: <20170412085544.6a54658f@luca> In-Reply-To: <58EDBAC4.7020902@redhat.com> References: <1491816131-20268-1-git-send-email-xlpang@redhat.com> <4d24b94c-1e57-b82c-ff0f-099157ba5526@redhat.com> <58EDBAC4.7020902@redhat.com> X-Mailer: Claws Mail 3.13.2 (GTK+ 2.24.30; x86_64-pc-linux-gnu) MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 3905 Lines: 103 Hi, On Wed, 12 Apr 2017 13:27:32 +0800 Xunlei Pang wrote: [...] > The more I read the code, the more I am confused why > dl_entity_overflow() is needed, if the task is before its deadline, > just let it run. Sorry for jumping in this thread; I did not read all of the previous emails, but I think I can answer this question :) dl_entity_overflow() is needed to check if a deadline task that is waking up can use its current runtime and scheduling deadline without breaking the guarantees for other deadline tasks. If the relative deadline of the tasks is equal to their period, this check is mathematically correct (and its correctness has been formally proved in some papers - see for example http://disi.unitn.it/~abeni/tr-98-01.pdf). If the relative deadline is different from the period, then the check is an approximation (and this is the big issue here). I am still not sure about what is the best thing to do in this case. > E.g. For (runtime 2ms, deadline 4ms, period 8ms), > for some reason was preempted after running a very short time 0.1ms, > after 0.9ms it was scheduled back and immediately sleep 1ms, when it > is awakened, remaining runtime is 1.9ms, remaining > deadline(deadline-now) within this period is 2ms, but > dl_entity_overflow() is true. However, clearly it can be correctly > run 1.9ms before deadline comes wthin this period. Sure, in this case the task can run for 1.9ms before the deadline... But doing so, it might break the guarantees for other deadline tasks. This is what the check is trying to avoid. > > We can add a condition in dl_runtime_exceeded(), if its deadline is > passed, then zero its runtime if positive, and a new period begins. > > I did some tests with the following patch, seems it works well, > please correct me if I am wrong. --- > kernel/sched/deadline.c | 16 ++++++++++++---- > 1 file changed, 12 insertions(+), 4 deletions(-) > > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c > index a2ce590..600c530 100644 > --- a/kernel/sched/deadline.c > +++ b/kernel/sched/deadline.c > @@ -498,8 +498,7 @@ static void update_dl_entity(struct > sched_dl_entity *dl_se, struct dl_rq *dl_rq = dl_rq_of_se(dl_se); > struct rq *rq = rq_of_dl_rq(dl_rq); > > - if (dl_time_before(dl_se->deadline, rq_clock(rq)) || > - dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { > + if (!dl_time_before(rq_clock(rq), dl_se->deadline)) { > dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; > dl_se->runtime = pi_se->dl_runtime; > } I think this will break the guarantees for tasks with relative deadline equal to period (think about a task with runtime 5ms, period 10ms and relative deadline 10ms... What happens if the task executes for 4.9ms, then blocks and immediately wakes up?) Luca > @@ -722,13 +721,22 @@ static inline void > dl_check_constrained_dl(struct sched_dl_entity *dl_se) > dl_time_before(rq_clock(rq), dl_next_period(dl_se))) { if > (unlikely(dl_se->dl_boosted || !start_dl_timer(p))) return; > + > + if (dl_se->runtime > 0) > + dl_se->runtime = 0; > + > dl_se->dl_throttled = 1; > } > } > > static > -int dl_runtime_exceeded(struct sched_dl_entity *dl_se) > +int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se) > { > + if (!dl_time_before(rq_clock(rq), dl_se->deadline)) { > + if (dl_se->runtime > 0) > + dl_se->runtime = 0; > + } > + > return (dl_se->runtime <= 0); > } > > @@ -779,7 +787,7 @@ static void update_curr_dl(struct rq *rq) > dl_se->runtime -= delta_exec; > > throttle: > - if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) { > + if (dl_runtime_exceeded(rq, dl_se) || dl_se->dl_yielded) { > dl_se->dl_throttled = 1; > __dequeue_task_dl(rq, curr, 0); > if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))