Re: [PATCH] blk-mq: use blk_mq_timeout_work to limit the max timeout

[Bart Van Assche]

On Wed, 2018-06-20 at 09:28 +0800, jianchao.wang wrote:
> Hi Bart
> 
> Thanks for your kindly response.
> 
> On 06/19/2018 11:18 PM, Bart Van Assche wrote:
> > On Tue, 2018-06-19 at 15:00 +0800, Jianchao Wang wrote:
> > > blk_rq_timeout is needed to limit the max timeout value, otherwise,
> > > a idle hctx cannot be deactivated timely in shared-tag case.
> > > 
> > > Fixes: 12f5b931 (blk-mq: Remove generation seqeunce)
> > > Signed-off-by: Jianchao Wang <jianchao.w.wang@xxxxxxxxxx>
> > > ---
> > >  block/blk-mq.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > > index 70c65bb..ccebe7b 100644
> > > --- a/block/blk-mq.c
> > > +++ b/block/blk-mq.c
> > > @@ -868,7 +868,7 @@ static void blk_mq_timeout_work(struct work_struct *work)
> > >  	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
> > >  
> > >  	if (next != 0) {
> > > -		mod_timer(&q->timeout, next);
> > > +		mod_timer(&q->timeout, blk_rq_timeout(round_jiffies_up(next)));
> > >  	} else {
> > >  		/*
> > >  		 * Request timeouts are handled as a forward rolling timer. If
> > 
> > Hello Jianchao,
> > 
> > What makes you think that it would be necessary to call blk_rq_timeout() from
> > blk_mq_timeout_work()? Have you noticed that blk_add_timer() already calls that
> > function? I think it is not necessary to call blk_rq_timeout() from
> > blk_mq_timeout_work() because it is guaranteed in that function that the next
> > timeout is less than BLK_MAX_TIMEOUT jiffies in the future.
> > 
> 
> blk_add_timer will not re-arm the timer if the timer's expire value is before the new rq's expire value.
> 
> Let's look at the following scenario.
> 
> 0                    +30s
> > __________________|___|
> 
> T0                T1   T2
> 
> T1 = T2 - 1 jiffies 
> 
> T0:  rq_a is issued and q->timer is armed and will expire at T2
>      then rq_a is completed.
> T1:  rq_b is issued and q->timer is not re-armed, because its next expire time is T2 < (T1 + 30s)
> 
> T2:  if rq_b have not been completed when timer expires at T2, timer would be re-armed based on the rq_b 
>      If we don't have blk_rq_timeout here, the next expire time is about T2 + 30s.

Hello Jianchao,

I disagree with the last sentence above. I think for your example blk_mq_req_expired()
will set next to T1 + 30s instead of T2 + 30s.

Bart.