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;
Subject: Re: Question on io-wq
To:     Hillf Danton <hdanton@sina.com>
Cc:     Zhang Qiang <qiang.zhang@windriver.com>, viro@zeniv.linux.org.uk,
        io-uring@vger.kernel.org, Pavel Begunkov <asml.silence@gmail.com>,
        linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org
References: <98c41fba-87fe-b08d-2c8c-da404f91ef31@windriver.com>
 <20201023020514.2230-1-hdanton@sina.com>
From:   Jens Axboe <axboe@kernel.dk>
Message-ID: <e61d3c07-7db0-28d8-03a0-cae13698a634@kernel.dk>
Date:   Thu, 22 Oct 2020 20:24:16 -0600
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Thunderbird/68.10.0
MIME-Version: 1.0
In-Reply-To: <20201023020514.2230-1-hdanton@sina.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 10/22/20 8:05 PM, Hillf Danton wrote:
> On Thu, 22 Oct 2020 08:08:09 -0600 Jens Axboe wrote:
>> On 10/22/20 3:02 AM, Zhang,Qiang wrote:
>>>
>>> Hi Jens Axboe
>>>
>>> There are some problem in 'io_wqe_worker' thread, when the 
>>> 'io_wqe_worker' be create and  Setting the affinity of CPUs in NUMA 
>>> nodes, due to CPU hotplug, When the last CPU going down, the 
>>> 'io_wqe_worker' thread will run anywhere. when the CPU in the node goes 
>>> online again, we should restore their cpu bindings?
>>
>> Something like the below should help in ensuring affinities are
>> always correct - trigger an affinity set for an online CPU event. We
>> should not need to do it for offlining. Can you test it?
> 
> CPU affinity is intact because of nothing to do on offline, and scheduler
> will move the stray workers on to the correct NUMA node if any CPU goes
> online, so it's a bit hard to see what is going to be tested.

Test it yourself:

- Boot with > 1 NUMA node
- Start an io_uring, you now get 2 workers, each affinitized to a node
- Now offline all CPUs in one node
- Online one or more of the CPU in that same node

The end result is that the worker on the node that was offlined now
has a mask of the other node, plus the newly added CPU.

So your last statement isn't correct, which is what the original
reporter stated.

-- 
Jens Axboe