Re: [RFC] add sysfs to dynamically control blk request tag maintenance

[Jens Axboe <axboe@suse.de>]

On Fri, Oct 07 2005, Arjan van de Ven wrote:
> On Fri, 2005-10-07 at 09:41 +0200, Jens Axboe wrote:
> > Ok that makes more sense! But it's a little worrying that
> > blk_queue_end_tag() would show up as hot in the profile, it is actually
> > quite lean.
> 
> it probably just is the first one to touch the IO structures after the
> completion, and thus gets the penalty for the cachemiss. Something has
> to have that after io completion (the io started usually > 10 msec ago
> after all, and usually on another cpu at that) and my experience is that
> it's one of those jello elephants; you can only move it around but not
> really avoid it.

That thought did occur to me, but I don't really see how that can be the
case. The ->queue_tags should be cache hot if you repeatedly call that
function, since that will never change. The request itself has been
touched by scsi_end_request() already, so unless the layout is really
bad we shouldn't need to fetch a lot of cache lines there. That leaves
__test_and_clear_bit(), I guess that must be it.

-- 
Jens Axboe