Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Wed, 16 Jun 2021 13:26:30 +0200
From:   Peter Zijlstra <peterz@infradead.org>
To:     Frederic Weisbecker <frederic@kernel.org>
Cc:     Thomas Gleixner <tglx@linutronix.de>,
        LKML <linux-kernel@vger.kernel.org>,
        "Eric W . Biederman" <ebiederm@xmission.com>,
        Oleg Nesterov <oleg@redhat.com>, Ingo Molnar <mingo@kernel.org>
Subject: Re: [PATCH 2/6] posix-cpu-timers: Don't start process wide cputime
 counter if timer is disabled
Message-ID: <YMnf5vW3MUyuKUa5@hirez.programming.kicks-ass.net>
References: <20210604113159.26177-1-frederic@kernel.org>
 <20210604113159.26177-3-frederic@kernel.org>
 <YMm7iafJ1mberGIg@hirez.programming.kicks-ass.net>
 <20210616105116.GA801071@lothringen>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20210616105116.GA801071@lothringen>
Precedence: bulk

On Wed, Jun 16, 2021 at 12:51:16PM +0200, Frederic Weisbecker wrote:
> On Wed, Jun 16, 2021 at 10:51:21AM +0200, Peter Zijlstra wrote:

> > The cpu_timer_enqueue() is in arm_timer() and the condition for calling
> > that is:
> > 
> >   'new_expires != 0 && val < new_expires'
> > 
> > Which is not the same as the one you add.
> 
> There are two different things here:
> 
> 1) the threadgroup cputime counter, activated by cpu_clock_sample_group(clkid,
> p, true)
> 
> 2) the expiration set (+ the callback enqueued) in arm_timer()
> 
> The issue here is that we go through 1) but not through 2)

Correct, but then I would think the cleanup would need the same
conditions as 2, and not something slightly different, which is what
confused me.

> > I'm thinking the fundamental problem here is the disconnect between
> > cpu_timer_enqueue() and pct->timers_active ?
> 
> You're right it's the core issue. But what prevents the whole to be
> fundamentally connected is a circular dependency: we need to know the
> threadgroup cputime before arming the timer, but we would need to know
> if we arm the timer before starting the threadgroup cputime counter
> 
> To sum up, the current sequence is:
> 
> * fetch the threadgroup cputime AND start the whole threadgroup counter
> 
> * arm the timer if it isn't zero and it hasn't yet expired
> 
> While the ideal sequence should be:
> 
> * fetch the threadgroup cputime (without starting the whole threadgroup counter
>   yet)
> 
> * arm the timer if it isn't zero and it hasn't yet expired
> 
> * iff we armed the timer, start the whole theadgroup counter
> 
> But that means re-iterating the whole threadgroup and update atomically
> the group counter with each task's time.

Right, so by the time patch #5 comes around, you seem to be at the point
where you can do:

 * fetch cputime and start threadgroup counter

 * possibly arm timer

 * if expired:
   - fire now
   - if armed, disarm (which leads to stop)

Which is the other 'obvious' solution to not starting it.