Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Subject: Re: [PATCH -next 2/3] md/raid10: convert resync_lock to use seqlock
To:     Yu Kuai <yukuai1@huaweicloud.com>, song@kernel.org
Cc:     linux-raid@vger.kernel.org, linux-kernel@vger.kernel.org,
        yi.zhang@huawei.com, "yukuai (C)" <yukuai3@huawei.com>
References: <20220829131502.165356-1-yukuai1@huaweicloud.com>
 <20220829131502.165356-3-yukuai1@huaweicloud.com>
 <3d8859bc-80d6-08b0-fd40-8874df4d3419@linux.dev>
 <1891ec2c-0ccc-681e-31de-fdd28eebce82@huaweicloud.com>
From:   Guoqing Jiang <guoqing.jiang@linux.dev>
Message-ID: <82f11462-454c-4a5e-d3a2-e71479960eaf@linux.dev>
Date:   Fri, 2 Sep 2022 18:16:44 +0800
MIME-Version: 1.0
In-Reply-To: <1891ec2c-0ccc-681e-31de-fdd28eebce82@huaweicloud.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 8bit
Content-Language: en-US
Precedence: bulk


On 9/2/22 6:02 PM, Yu Kuai wrote:
> Hi,
>
> 在 2022/09/02 17:42, Guoqing Jiang 写道:
>> Hi,
>>
>> On 8/29/22 9:15 PM, Yu Kuai wrote:
>>> +static bool wait_barrier_nolock(struct r10conf *conf)
>>> +{
>>> +    unsigned int seq = raw_read_seqcount(&conf->resync_lock.seqcount);
>>> +
>>> +    if (seq & 1)
>>> +        return false;
>>> +
>>> +    if (READ_ONCE(conf->barrier))
>>> +        return false;
>>> +
>>> +    atomic_inc(&conf->nr_pending);
>>> +    if (!read_seqcount_retry(&conf->resync_lock.seqcount, seq))
>>
>> I think 'seq' is usually get from read_seqcount_begin.
>
> read_seqcount_begin will loop untill "req & 1" failed, I'm afraid this
> will cause high cpu usage in come cases.
>
> What I try to do here is just try once, and fall back to hold lock and
> wait if failed.

Thanks for the explanation.

I'd suggest to try with read_seqcount_begin/read_seqcount_retry pattern
because it is a common usage in kernel I think, then check whether the
performance drops or not.  Maybe it is related to lockdep issue, but I am
not sure.

Thanks,
Guoqing