Message-Id: <20081117120029M.fujita.tomonori@lab.ntt.co.jp>
Date:	Mon, 17 Nov 2008 12:48:46 +0900
From:	FUJITA Tomonori <fujita.tomonori@....ntt.co.jp>
To:	jens.axboe@...cle.com
Cc:	fujita.tomonori@....ntt.co.jp, stern@...land.harvard.edu,
	James.Bottomley@...senPartnership.com, linux-scsi@...r.kernel.org,
	linux-kernel@...r.kernel.org
Subject: Re: Problems with the block-layer timeouts

On Thu, 13 Nov 2008 11:34:56 +0100
Jens Axboe <jens.axboe@...cle.com> wrote:

> On Wed, Nov 12 2008, FUJITA Tomonori wrote:
> > On Tue, 11 Nov 2008 20:19:36 +0100
> > Jens Axboe <jens.axboe@...cle.com> wrote:
> > 
> > > On Tue, Nov 11 2008, Alan Stern wrote:
> > > > On Tue, 11 Nov 2008, FUJITA Tomonori wrote:
> > > > 
> > > > > I don't worry about anything. I just think that these round_jiffies_up
> > > > > are pointless because they were added for the block-layer users that
> > > > > care about exact timeouts, however the block-layer doesn't export
> > > > > blk_add_timer() so the block-layer users can't control the exact time
> > > > > when the timer starts. So doing round_jiffies_up calculation per every
> > > > > request doesn't make sense for me.
> > > > 
> > > > In fact the round_jiffies_up() routines were added for other users as
> > > > well as the block layer.  However none of the others could be changed
> > > > until the routines were merged.  Now that the routines are in the 
> > > > mainline, you should see them start to be called in multiple places.
> > > > 
> > > > Also, the users of the block layer _don't_ care about exact timeouts.  
> > > > That's an important aspect of round_jiffies() and round_jiffies_up() --
> > > > you don't use them if you want an exact timeout.
> > > > 
> > > > The reason for using round_jiffies() is to insure that the timeout
> > > > will occur at a 1-second boundary.  If several timeouts are set for
> > > > about the same time and they all use round_jiffies() or
> > > > round_jiffies_up(), then they will all occur at the same tick instead
> > > > of spread out among several different ticks during the course of that
> > > > 1-second interval.  As a result, the system will need to wake up only
> > > > once to service all those timeouts, instead of waking up several
> > > > different times.  It is a power-saving scheme.
> > 
> > Hmm, but for 99.9% of the cases, the timeout of the block layer
> > doesn't expire, the timeout rarely happens. The power-saving scheme
> > can be applied to only 0.1%, but at the cost of the round_jiffies
> > overhead per every request.
> > 
> > If I understand correctly, round_jiffies() is designed for timers that
> > will expire, such as periodic checking. The power-saving scheme nicely
> > works for such usages.
> 
> Your understanding is correct. The overhead of round_jiffies() is not
> large, though.
> 
> I want to get rid of this in blk_delete_timer():
> 
>         if (list_empty(&q->timeout_list))
>                 del_timer(&q->timeout);
> 
> though and simply let the timer run even if the list is empty, since for
> sync sequential IO we'll be fiddling a much with the timer as we did
> before unifying it. And then the timer will expire every x seconds
> always and it becomes more important with the grouping.

I see. It depends on workloads but the above 'periodic expiration'
scheme might be better than the current one, I guess. It doesn't gives
large impact though.

Thanks for the clarification.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists