Re: [PATCH 1/2] vhost: Reduce vhost_work_flush() wakeup latency

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, Aug 15, 2013 at 08:13:14AM +0200, Bart Van Assche wrote:
> On 08/15/13 03:30, Asias He wrote:
> >On Wed, Aug 14, 2013 at 08:58:25PM +0300, Michael S. Tsirkin wrote:
> >>On Wed, Aug 14, 2013 at 05:22:36PM +0200, Bart Van Assche wrote:
> >>>On 08/14/13 13:37, Michael S. Tsirkin wrote:
> >>>>On Wed, Aug 14, 2013 at 09:02:32AM +0200, Bart Van Assche wrote:
> >>>>>If the TIF_NEED_RESCHED task flag is set, wake up any vhost_work_flush()
> >>>>>waiters before rescheduling instead of after rescheduling.
> >>>>>
> >>>>>Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> >>>>>Cc: Michael S. Tsirkin <mst@redhat.com>
> >>>>>Cc: Asias He <asias@redhat.com>
> >>>>
> >>>>Why exactly? It's not like flush needs to be extra fast ...
> >>>
> >>>I'm not worried about how fast a flush is processed either. But I
> >>>think this patch is a nice cleanup of the vhost_worker() function.
> >>>It eliminates the uninitialized_var() construct and moves the
> >>>assignment of "done_seq" below the read of "queue_seq".
> >>>
> >>>Bart.
> >>
> >>I'm not worried about uninitialized_var - it's just a compiler bug.
> >>done_seq below read is nice, but this is at the cost of sprinkling
> >>lock/unlock, lock/unlock all over the code.  Maybe we can come up with
> >>something that doesn't have this property.
> >
> >The extra locking introduced here does not look good to me neither.
> 
> Please note that although there are additional locking statements in
> the code, there are no additional locking calls at runtime.
> 
> Bart.

That's true but this is supposed to be a cleanup.

-- 
MST