Received-SPF: pass (google.com: domain of linux-kernel+bounces-55340-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Sender: Tejun Heo <htejun@gmail.com>
Date: Tue, 6 Feb 2024 06:52:41 -1000
From: Tejun Heo <tj@kernel.org>
To: Yunlong Xing <yunlong.xing@unisoc.com>
Cc: jiangshanlai@gmail.com, yunlong.xing23@gmail.com,
	linux-kernel@vger.kernel.org, hongyu.jin@unisoc.com,
	zhiguo.niu@unisoc.com, zhengxu.zhang@unisoc.com
Subject: Re: [PATCH] workqueue: Fix pool->nr_running type back to atomic
Message-ID: <ZcJj2dUUyAY_7XFS@slm.duckdns.org>
References: <20240206080024.2373490-1-yunlong.xing@unisoc.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240206080024.2373490-1-yunlong.xing@unisoc.com>

Hello,

On Tue, Feb 06, 2024 at 04:00:24PM +0800, Yunlong Xing wrote:
> In CPU-hotplug test, when plug the core, set_cpus_allowed_ptr() restoring
> the cpus_mask of the per-cpu worker may fail, the cpus_mask of the worker
> remain wq_unbound_cpumask until the core hotpluged next time. so, workers
> in the same per-cpu pool can run concurrently and change nr_running at the
>  same time, atomic problem occur.

How would set_cpus_allowed_ptr() fail? That should trigger WARN_ON, right?
If set_cpus_allowed_ptr() fails, nr_running getting desynchronized is only a
part of the problem. We will end up running per-cpu work items which must
execute on the same CPU on foreign CPUs.

Thanks.

-- 
tejun