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;
Message-ID: <f5d20f6c-5363-231b-b208-b577a59b4ae9@kernel.dk>
Date:   Wed, 20 Jul 2022 11:24:14 -0600
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux aarch64; rv:91.0) Gecko/20100101
 Thunderbird/91.10.0
Subject: Re: [LKP] Re: [io_uring] 584b0180f0:
 phoronix-test-suite.fio.SequentialWrite.IO_uring.Yes.Yes.1MB.DefaultTestDirectory.mb_s
 -10.2% regression
Content-Language: en-US
To:     Yin Fengwei <fengwei.yin@intel.com>,
        kernel test robot <oliver.sang@intel.com>
Cc:     LKML <linux-kernel@vger.kernel.org>, io-uring@vger.kernel.org,
        lkp@lists.01.org, lkp@intel.com
References: <20220527092432.GE11731@xsang-OptiPlex-9020>
 <2085bfef-a91c-8adb-402b-242e8c5d5c55@kernel.dk>
 <0d60aa42-a519-12ad-3c69-72ed12398865@intel.com>
 <26d913ea-7aa0-467d-4caf-a93f8ca5b3ff@kernel.dk>
 <9df150bb-f4fd-7857-aea8-b2c7a06a8791@intel.com>
 <7146c853-0ff8-3c92-c872-ce6615baab40@kernel.dk>
 <81af5cdf-1a13-db2c-7b7b-cfd86f1271e6@intel.com>
 <74d1f308-de03-fd5e-b7f0-0e17980f988e@kernel.dk>
 <2ec953da-78fd-df01-44cf-6f33a5e40864@intel.com>
From:   Jens Axboe <axboe@kernel.dk>
In-Reply-To: <2ec953da-78fd-df01-44cf-6f33a5e40864@intel.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 7/19/22 2:58 AM, Yin Fengwei wrote:
> Hi Jens,
> 
> On 7/19/2022 10:29 AM, Jens Axboe wrote:
>> I'll poke at this tomorrow.
> 
> Just FYI. Another finding (test is based on commit 584b0180f0):
> If the code block is put to different function, the fio performance result is
> different:

I think this turned out to be a little bit of a goose chase. What's
happening here is that later kernels defer the file assignment, which
means it isn't set if a request is queued with IOSQE_ASYNC. That in
turn, for writes, means that we don't hash it on io-wq insertion, and
then it doesn't get serialized with other writes to that file.

I'll come up with a patch for this that you can test.

-- 
Jens Axboe