Re: [PATCH 1/1] [SCSI] Fix a bug in deriving the FLUSH_TIMEOUT from the basic I/O timeout

[Jens Axboe <axboe@kernel.dk>]

On 2014-06-06 11:52, James Bottomley wrote:
> On Fri, 2014-06-06 at 12:18 -0500, Mike Christie wrote:
>> On 6/5/14, 9:53 PM, KY Srinivasan wrote:
>>>
>>>
>>>> -----Original Message-----
>>>> From: Mike Christie [mailto:michaelc@cs.wisc.edu]
>>>> Sent: Thursday, June 5, 2014 6:33 PM
>>>> To: KY Srinivasan
>>>> Cc: James Bottomley; linux-kernel@vger.kernel.org; apw@canonical.com;
>>>> devel@linuxdriverproject.org; hch@infradead.org; linux-
>>>> scsi@vger.kernel.org; ohering@suse.com; gregkh@linuxfoundation.org;
>>>> jasowang@redhat.com
>>>> Subject: Re: [PATCH 1/1] [SCSI] Fix a bug in deriving the FLUSH_TIMEOUT
>>>> from the basic I/O timeout
>>>>
>>>> On 06/04/2014 12:15 PM, KY Srinivasan wrote:
>>>>>
>>>>>
>>>>>> -----Original Message-----
>>>>>> From: James Bottomley [mailto:jbottomley@parallels.com]
>>>>>> Sent: Wednesday, June 4, 2014 10:02 AM
>>>>>> To: KY Srinivasan
>>>>>> Cc: linux-kernel@vger.kernel.org; apw@canonical.com;
>>>>>> devel@linuxdriverproject.org; hch@infradead.org; linux-
>>>>>> scsi@vger.kernel.org; ohering@suse.com; gregkh@linuxfoundation.org;
>>>>>> jasowang@redhat.com
>>>>>> Subject: Re: [PATCH 1/1] [SCSI] Fix a bug in deriving the
>>>>>> FLUSH_TIMEOUT from the basic I/O timeout
>>>>>>
>>>>>> On Wed, 2014-06-04 at 09:33 -0700, K. Y. Srinivasan wrote:
>>>>>>> Commit ID: 7e660100d85af860e7ad763202fff717adcdaacd added code to
>>>>>>> derive the FLUSH_TIMEOUT from the basic I/O timeout. However, this
>>>>>>> patch did not use the basic I/O timeout of the device. Fix this bug.
>>>>>>>
>>>>>>> Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
>>>>>>> ---
>>>>>>>    drivers/scsi/sd.c |    4 +++-
>>>>>>>    1 files changed, 3 insertions(+), 1 deletions(-)
>>>>>>>
>>>>>>> diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c index
>>>>>>> e9689d5..54150b1 100644
>>>>>>> --- a/drivers/scsi/sd.c
>>>>>>> +++ b/drivers/scsi/sd.c
>>>>>>> @@ -832,7 +832,9 @@ static int sd_setup_write_same_cmnd(struct
>>>>>>> scsi_device *sdp, struct request *rq)
>>>>>>>
>>>>>>>    static int scsi_setup_flush_cmnd(struct scsi_device *sdp, struct
>>>>>>> request *rq)  {
>>>>>>> -	rq->timeout *= SD_FLUSH_TIMEOUT_MULTIPLIER;
>>>>>>> +	int timeout = sdp->request_queue->rq_timeout;
>>>>>>> +
>>>>>>> +	rq->timeout = (timeout * SD_FLUSH_TIMEOUT_MULTIPLIER);
>>>>>>
>>>>>> Could you share where you found this to be a problem?  It looks like
>>>>>> a bug in block because all inbound requests being prepared should
>>>>>> have a timeout set, so block would be the place to fix it.
>>>>>
>>>>> Perhaps; what I found was that the value in rq->timeout was 0 coming
>>>>> into this function and thus multiplying obviously has no effect.
>>>>>
>>>>
>>>> I think you are right. We hit this problem because we are doing:
>>>>
>>>> scsi_request_fn -> blk_peek_request -> sd_prep_fn ->
>>>> scsi_setup_flush_cmnd.
>>>>
>>>> At this time request->timeout is zero so the multiplication does nothing. See
>>>> how sd_setup_write_same_cmnd will set the request->timeout at this time.
>>>>
>>>> Then in scsi_request_fn we do:
>>>>
>>>> scsi_request_fn -> blk_start_request -> blk_add_timer.
>>>>
>>>> At this time it will set the request->timeout if something like req block pc
>>>> users (like scsi_execute() or block/scsi_ioctl.c) or the write same code
>>>> mentioned above have not set the timeout.
>>>
>>> I don't think this is a recent change. Prior to this commit, we were setting the timeout
>>> value in this function; it just happened to be a different constant unrelated to the I/O
>>> timeout.
>>>
>>
>> Yeah, it looks like when 7e660100d85af860e7ad763202fff717adcdaacd was
>> merged we were supposed to initialize it like in your patch in this thread.
>>
>> I guess we could do your patch in this thread, or if we want the block
>> layer to initialize the timeout before the prep_fn callout is called
>> then we would need to have the blk-flush.c code to that when it sets up
>> the request. If we do the latter, do we want the discard and write same
>> code to initialize the request's timeout before the prep_fn callout is
>> called too?
>
> I looked through the call chain; it seems to be intentional behaviour on
> the part of block.  Just from an mq point of view, it would make better
> code if we unconditionally initialised rq->timeout early and allowed
> prep to modify it and then dumped the if(!req->timeout) in
> blk_add_timer(), but it's a marginal if condition that would compile to
> a conditional store on sensible architectures, so losing the conditional
> probably isn't worth worrying about.
>
> Cc'd Jens for his opinion with the block patch

I just committed this one earlier today:

http://git.kernel.dk/?p=linux-block.git;a=commit;h=f6be4fb4bcb396fc3b1c134b7863351972de081f

since I ran into the same thing on nvme. Either approach is fine with 
me, as they both allow override of the timeout before insertion. But 
we've always done the rq->timeout = 0 init, so I think we should just 
reinstate that behavior.

-- 
Jens Axboe