[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20181120140925.GV2131@hirez.programming.kicks-ass.net>
Date: Tue, 20 Nov 2018 15:09:25 +0100
From: Peter Zijlstra <peterz@...radead.org>
To: Frederic Weisbecker <frederic@...nel.org>
Cc: LKML <linux-kernel@...r.kernel.org>,
Wanpeng Li <wanpengli@...cent.com>,
Thomas Gleixner <tglx@...utronix.de>,
Yauheni Kaliuta <yauheni.kaliuta@...hat.com>,
Ingo Molnar <mingo@...nel.org>, Rik van Riel <riel@...hat.com>
Subject: Re: [PATCH 18/25] vtime: Track nice-ness on top of context switch
On Wed, Nov 14, 2018 at 03:46:02AM +0100, Frederic Weisbecker wrote:
> We need to read the nice value of the task running on any CPU, possibly
> remotely, in order to correctly support kcpustat on nohz_full.
> Unfortunately we can't just read task_nice(tsk) when tsk runs on another
> CPU because its nice value may be concurrently changed. There could be a
> risk that a recently modified nice value is thought to apply for a longer
> while than is supposed to.
>
> For example if a task runs at T0 with nice = -10, then its nice value
> is changed at T0 + 1 second with nice = 10, a reader at T0 + 1 second
> could think that the task had this "nice == 10" value since the beginning
> (T0) and spuriously account 1 second nice time on kcpustat instead of 1
> second user time.
>
> So we need to track the nice value changes under vtime seqcount. Start
> with context switches and account the vtime nice-ness on top of it.
Huh, what!? That doesn't make any sense..
Powered by blists - more mailing lists