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;
MIME-Version: 1.0
References: <20220815164324.645550-1-haowenchao@huawei.com> <3ed022dd-3026-542a-c3b0-e4f4005208aa@linux.ibm.com>
In-Reply-To: <3ed022dd-3026-542a-c3b0-e4f4005208aa@linux.ibm.com>
From:   wenchao hao <haowenchao22@gmail.com>
Date:   Tue, 16 Aug 2022 12:36:38 +0800
Message-ID: <CAOptpSMqfiApXn2jgmaLu-CdqyGUr7yQ1v3AwQUDev4mFjaqJQ@mail.gmail.com>
Subject: Re: [PATCH] scsi: increase scsi device's iodone_cnt in scsi_timeout()
To:     Steffen Maier <maier@linux.ibm.com>
Cc:     "James E . J . Bottomley" <jejb@linux.ibm.com>,
        "Martin K . Petersen" <martin.petersen@oracle.com>,
        linux-scsi@vger.kernel.org, linux-kernel@vger.kernel.org,
        Wenchao Hao <haowenchao@huawei.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Aug 16, 2022 at 12:53 AM Steffen Maier <maier@linux.ibm.com> wrote:
>
> On 8/15/22 18:43, Wenchao Hao wrote:
> > The iodone_cnt might be less than iorequest_cnt because
> > we did not increase the iodone_cnt when a command is done
> > from timeout.
> >
> > Signed-off-by: Wenchao Hao <haowenchao@huawei.com>
> > ---
> >   drivers/scsi/scsi_error.c | 1 +
> >   1 file changed, 1 insertion(+)
> >
> > diff --git a/drivers/scsi/scsi_error.c b/drivers/scsi/scsi_error.c
> > index 448748e3fba5..d21ae0090166 100644
> > --- a/drivers/scsi/scsi_error.c
> > +++ b/drivers/scsi/scsi_error.c
> > @@ -355,6 +355,7 @@ enum blk_eh_timer_return scsi_timeout(struct request *req)
> >                */
> >               if (test_and_set_bit(SCMD_STATE_COMPLETE, &scmd->state))
> >                       return BLK_EH_RESET_TIMER;
> > +             atomic_inc(&scmd->device->iodone_cnt);
> >               if (scsi_abort_command(scmd) != SUCCESS) {
> >                       set_host_byte(scmd, DID_TIME_OUT);
> >                       scsi_eh_scmd_add(scmd);
>
> Not sure, but can't we still get a (late) regular completion even after the
> timeout happened (double accounting) and before we successfully aborted the
> command?
>
>
The flag SCMD_STATE_COMPLETE would take care of it, this would make
sure a scsi command
is done from timeout or regular completion exclusively.

When scsi_abort_command() begins to handle a command, the command
would be finished by
scsi_finish_command() or retried by requeue function.

If a command finished by scsi_finish_command(), it would not add the iodone_cnt.
If a command is requeued, it would call scsi_dispatch_cmd() again and
another iorequest_cnt would be
added.

According to above analysis, I think it's suitable to increase iodone_cnt here.