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;
From:   Valentin Schneider <vschneid@redhat.com>
To:     paulmck@kernel.org
Cc:     linux-kernel@vger.kernel.org, mingo@redhat.com,
        peterz@infradead.org, juri.lelli@redhat.com,
        vincent.guittot@linaro.org, dietmar.eggemann@arm.com,
        rostedt@goodmis.org, bsegall@google.com, mgorman@suse.de,
        bristot@redhat.com
Subject: Re: "Dying CPU not properly vacated" splat
In-Reply-To: <20220426000328.GY4285@paulmck-ThinkPad-P17-Gen-1>
References: <20220421193821.GA173010@paulmck-ThinkPad-P17-Gen-1>
 <xhsmh4k2h9m26.mognet@vschneid.remote.csb>
 <20220425173320.GX4285@paulmck-ThinkPad-P17-Gen-1>
 <xhsmh1qxkakof.mognet@vschneid.remote.csb>
 <20220426000328.GY4285@paulmck-ThinkPad-P17-Gen-1>
Date:   Tue, 26 Apr 2022 15:48:06 +0100
Message-ID: <xhsmhy1zr99zt.mognet@vschneid.remote.csb>
MIME-Version: 1.0
Content-Type: text/plain
Precedence: bulk

On 25/04/22 17:03, Paul E. McKenney wrote:
> On Mon, Apr 25, 2022 at 10:59:44PM +0100, Valentin Schneider wrote:
>> On 25/04/22 10:33, Paul E. McKenney wrote:
>> >
>> > So what did rcu_torture_reader() do wrong here?  ;-)
>> >
>>
>> So on teardown, CPUHP_AP_SCHED_WAIT_EMPTY->sched_cpu_wait_empty() waits for
>> the rq to be empty. Tasks must *not* be enqueued onto that CPU after that
>> step has been run - if there are per-CPU tasks bound to that CPU, they must
>> be unbound in their respective hotplug callback.
>>
>> For instance for workqueue.c, we have workqueue_offline_cpu() as a hotplug
>> callback which invokes unbind_workers(cpu), the interesting bit being:
>>
>>                 for_each_pool_worker(worker, pool) {
>>                         kthread_set_per_cpu(worker->task, -1);
>>                         WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
>>                 }
>>
>> The rcu_torture_reader() kthreads aren't bound to any particular CPU are
>> they? I can't find any code that would indicate they are - and in that case
>> it means we have a problem with is_cpu_allowed() or related.
>
> I did not intend that the rcu_torture_reader() kthreads be bound, and
> I am not seeing anything that binds them.
>
> Thoughts?  (Other than that validating any alleged fix will be quite
> "interesting".)
>
>                                                       Thanx, Paul

IIUC the bogus scenario is is_cpu_allowed() lets one of those kthreads be
enqueued on the outgoing CPU *after* CPUHP_AP_SCHED_WAIT_EMPTY.teardown() has
been run, and hilarity ensues.

The cpu_dying() condition should prevent a regular kthread from getting
enqueued there, most of the details have been evinced from my brain but I
recall we got the ordering conditions right...

The only other "obvious" thing here is migrate_disable() which lets the
enqueue happen, but then balance_push()->select_fallback_rq() should punt
it away on context switch.

I need to rediscover those paths, I don't see any obvious clue right now.