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: <405ef35e-5998-789f-8577-897b4cdf7e71@linux.dev>
Date:   Thu, 3 Aug 2023 23:35:34 +0800
MIME-Version: 1.0
Subject: Re: [PATCH v2 1/4] blk-flush: flush_rq should inherit first_rq's
 cmd_flags
Content-Language: en-US
To:     Christoph Hellwig <hch@lst.de>, Hannes Reinecke <hare@suse.de>
Cc:     axboe@kernel.dk, ming.lei@redhat.com, linux-block@vger.kernel.org,
        linux-kernel@vger.kernel.org, zhouchengming@bytedance.com
References: <20230725130102.3030032-1-chengming.zhou@linux.dev>
 <20230725130102.3030032-2-chengming.zhou@linux.dev>
 <20230731060957.GA30409@lst.de>
 <de6ba52c-0f14-670f-7262-93f2aced926a@suse.de> <20230801110432.GA2781@lst.de>
 <20230801110607.GB2781@lst.de>
From:   Chengming Zhou <chengming.zhou@linux.dev>
In-Reply-To: <20230801110607.GB2781@lst.de>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 2023/8/1 19:06, Christoph Hellwig wrote:
> On Tue, Aug 01, 2023 at 01:04:32PM +0200, Christoph Hellwig wrote:
>> On Mon, Jul 31, 2023 at 06:28:01PM +0200, Hannes Reinecke wrote:
>>> The flush machinery is sending flushes before and/or after the original 
>>> request (preflush/postflush). For blocked transports (ie during FC RSCN 
>>> handling) the transport will error out commands depending on the FAILFAST 
>>> setting. If FAILFAST is set the SCSI layer gets an STS_TRANSPORT error 
>>> (causing the I/O to be retried), but STS_ERROR if not set (causing I/O to 
>>> failed).
>>>
>>> So if the FAILFAST setting is _not_ aligned between flush_rq and the 
>>> original we'll get an error on the flush rq and a retry on the original rq, 
>>> causing the entire command to fail.
>>>
>>> I guess we need to align them.
>>
>> But you can't, because multiple pre/postflushes are coalesced into a
>> single outstanding flush request.  They can and will not match quite
>> commonly.
> 
> And if you mean the REQ_FAILFAST_TRANSPORT added by dm - this will
> never even see the flush state machine, as that is run in dm-mpath
> which then inserts the fully built flush request into the lower request
> queue.  At least for request based multipath, bio could hit it.

Yes, multiple pre/postflushes are coalesced into a single flush request.
So we can't figure out which request to use.

From the above explanation, can we just drop this inherit logic? It seems
strange or wrong here.

Thanks.