Received-SPF: pass (google.com: domain of linux-kernel+bounces-106465-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
From: Valentin Schneider <vschneid@redhat.com>
To: Daniel Vacek <neelx@redhat.com>, Ingo Molnar <mingo@redhat.com>, Peter
 Zijlstra <peterz@infradead.org>, Juri Lelli <juri.lelli@redhat.com>,
 Vincent Guittot <vincent.guittot@linaro.org>, Dietmar Eggemann
 <dietmar.eggemann@arm.com>, Steven Rostedt <rostedt@goodmis.org>, Ben
 Segall <bsegall@google.com>, Mel Gorman <mgorman@suse.de>, Daniel Bristot
 de Oliveira <bristot@redhat.com>
Cc: Daniel Vacek <neelx@redhat.com>, stable@vger.kernel.org, Bill Peters
 <wpeters@atpco.net>, Ingo Molnar <mingo@kernel.org>,
 linux-kernel@vger.kernel.org
Subject: Re: [PATCH] sched/core: fix affine_move_task failure case
In-Reply-To: <20240318111750.3097906-1-neelx@redhat.com>
References: <20240318111750.3097906-1-neelx@redhat.com>
Date: Mon, 18 Mar 2024 18:34:37 +0100
Message-ID: <xhsmhfrwncuky.mognet@vschneid-thinkpadt14sgen2i.remote.csb>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain

On 18/03/24 12:17, Daniel Vacek wrote:
> Bill Peters reported CPU hangs while offlining/onlining CPUs on s390.
>
> Analyzing the vmcore data shows `stop_one_cpu_nowait()` in `affine_move_task()`
> can fail when racing with off-/on-lining resulting in a deadlock waiting for
> the pending migration stop work completion which is never done.
>
> Fix this by correctly handling such a condition.
>

IIUC the problem is that the dest_cpu and its stopper thread can be taken
down by take_cpu_down(), and affine_move_task() currently isn't aware of
that. I thought we had tested this vs hotplug, but oh well...

> Fixes: 9e81889c7648 ("sched: Fix affine_move_task() self-concurrency")
> Cc: stable@vger.kernel.org
> Reported-by: Bill Peters <wpeters@atpco.net>
> Tested-by: Bill Peters <wpeters@atpco.net>
> Signed-off-by: Daniel Vacek <neelx@redhat.com>
> ---
>  kernel/sched/core.c | 13 +++++++++++--
>  1 file changed, 11 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 9116bcc903467..d0ff5c611a1c8 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -3069,8 +3069,17 @@ static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flag
>               preempt_disable();
>               task_rq_unlock(rq, p, rf);
>               if (!stop_pending) {
> -			stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
> -					    &pending->arg, &pending->stop_work);
> +			stop_pending =
> +				stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
> +						    &pending->arg, &pending->stop_work);
> +
> +			if (!stop_pending) {
> +				rq = task_rq_lock(p, rf);
> +				pending->stop_pending = false;
> +				p->migration_pending = NULL;
> +				task_rq_unlock(rq, p, rf);
> +				complete_all(&pending->done);
> +			}

This can leave the task @p on a now-illegal CPU; consider a task affined to
CPUs 0-1, migrate_disable(); then affined to CPUs 2-3, then in
migrate_enable() the dest_cpu is chosen as 3 but that's racing with it
being brought down. The stop_one_cpu_nowait() fails, and we leave the task
on CPUs 0-1.

Issuing a redo of affine_move_task() with a different dest_cpu doesn't
sound great, and while very unlikely that doesn't have forward progress
guarantees.

Unfortunately we can't hold the hotplug lock during the affinity change of
migrate_enable(), as migrate_enable() isn't allowed to block.

Now, the CPU selection in __set_cpus_allowed_ptr_locked() that is passed
down to affine_move_task() relies on the active mask, which itself is
cleared in sched_cpu_deactivate() and is followed by a
synchronize_rcu().

What if we made the affinity change of migrate_enable() an RCU read-side
section? Then if a CPU X is observed as active in
  migrate_enable()->__set_cpus_allowed_ptr_locked()
, then its' hotplug state cannot go lower than CPUHP_AP_ACTIVE until the task is
migrated away.

Something like the below. Thoughts?
---
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 929fce69f555e..c6d128711d1a9 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2450,8 +2450,11 @@ void migrate_enable(void)
 	 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
 	 */
 	guard(preempt)();
-	if (p->cpus_ptr != &p->cpus_mask)
+	if (p->cpus_ptr != &p->cpus_mask) {
+		guard(rcu)();
 		__set_cpus_allowed_ptr(p, &ac);
+	}
+
 	/*
 	 * Mustn't clear migration_disabled() until cpus_ptr points back at the
 	 * regular cpus_mask, otherwise things that race (eg.