Re: [PATCH] cfq-iosched: replace sync_flight by rq_in_driver[BLK_RW_SYNC]

[Jens Axboe <jens.axboe@oracle.com>]

On Wed, Dec 30 2009, Corrado Zoccolo wrote:
> On Wed, Dec 30, 2009 at 7:47 PM, Jens Axboe <jens.axboe@oracle.com> wrote:
> > On Wed, Dec 30 2009, Corrado Zoccolo wrote:
> >> According to my intuition (and brief testing), sync_flight is always
> >> equal to rq_in_driver[BLK_RW_SYNC] at the point of usage, so it can
> >> be removed and replaced by the other.
> >
> > They are not fully identical. ->sync_flight is incremented on insertion
> > on the dispatch list, ->rq_in_driver not until the request is activated
> > (eg the driver has retrieved it and wants to dispatch to the hardware).
> Usually (only exceptions are forced dispatch, or when a conflict in the
> rb tree is found), a request is activated as soon as cfq returns from
> cfq_dispatch_requests.

Yes, but then it may be deactivated immediately for requeue.

> > They will usually be identical, but that may not be true for requeues
> > for instance.
> 
> For our purpose, it is sufficient that in cfq_may_dispatch, they are either
> both 0 or both non-0.
> Since we have sync_flight >= rq_in_driver[1], the only question is:
> can the number of requests in the driver drop to 0 with requests still
> in flight?

It's mostly a theoretical issue, but yes it could happen. I'm assuming
you mean ->rq_in_driver[1] == 0 while ->sync_flight != 0. But then we
are into the area of some starvation problem, in hardware or in the
kernel. So it's not likely, but still.

> I'm asking because to drain async requests, we are using the rq_in_driver
> counter instead. Maybe they need the same treatment.

In theory, yes the same applies there. Normal operations would not have
that distinction between activated and on dispatch list.

-- 
Jens Axboe