Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
Date:   Wed, 16 Mar 2022 21:27:34 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Steven Rostedt <rostedt@goodmis.org>
Cc:     LKML <linux-kernel@vger.kernel.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Subject: Re: sched_core_balance() releasing interrupts with pi_lock held
Message-ID: <20220316202734.GJ8939@worktop.programming.kicks-ass.net>
References: <20220308161455.036e9933@gandalf.local.home>
 <20220315174606.02959816@gandalf.local.home>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220315174606.02959816@gandalf.local.home>
Precedence: bulk

On Tue, Mar 15, 2022 at 05:46:06PM -0400, Steven Rostedt wrote:
> On Tue, 8 Mar 2022 16:14:55 -0500
> Steven Rostedt <rostedt@goodmis.org> wrote:
> 
> > Hi Peter,
> 
> Have you had time to look into this?

Not since I talk to you on IRC about it last week.

Like I wrote, the balance_callback should be ran under whichever
rq->lock instance it gets queued under. As per:

  565790d28b1e ("sched: Fix balance_callback()")

Now, we only do queue_core_balance() from set_next_task_idle(), which
*should* only happen from pick_next_task(), and as such the callback
should only ever get called from finish_lock_switch() or the 'prev ==
next' case in __schedule().

Neither of these two sites holds pi_lock.


This is about as far as I got explaining things, and it being late, it's
about as far as I got looking at things.

Now that also makes conceptual sense, we only want to pull a core-cookie
task when we're scheduling an idle task.

Now, clearly this gets triggered from the PI path, but that's not making
immediate sense to me, it would mean we're boosting the idle task, which
is wrong too.

So it would be useful for someone that can reproduce this to provide a
trace of where queue_core_balance() gets called, because that *should*
only be in __schedule().