On Fri, Sep 06, 2019 at 10:18:49AM -0400, Michael S. Tsirkin wrote:
> On Fri, Sep 06, 2019 at 10:17:05AM -0400, Vivek Goyal wrote:
> > On Fri, Sep 06, 2019 at 11:52:10AM +0100, Stefan Hajnoczi wrote:
> > > On Thu, Sep 05, 2019 at 03:48:49PM -0400, Vivek Goyal wrote:
> > > > +static void virtio_fs_drain_queue(struct virtio_fs_vq *fsvq)
> > > > +{
> > > > +	WARN_ON(fsvq->in_flight < 0);
> > > > +
> > > > +	/* Wait for in flight requests to finish.*/
> > > > +	while (1) {
> > > > +		spin_lock(&fsvq->lock);
> > > > +		if (!fsvq->in_flight) {
> > > > +			spin_unlock(&fsvq->lock);
> > > > +			break;
> > > > +		}
> > > > +		spin_unlock(&fsvq->lock);
> > > > +		usleep_range(1000, 2000);
> > > > +	}
> > > 
> > > I think all contexts that call this allow sleeping so we could avoid
> > > usleep here.
> > 
> > usleep_range() is supposed to be used from non-atomic context.
> > 
> > https://github.com/torvalds/linux/blob/master/Documentation/timers/timers-howto.rst
> > 
> > What construct you are thinking of?
> > 
> > Vivek
> 
> completion + signal on vq callback?

Yes.  Time-based sleep() is sub-optimal because we could wake up exactly
when in_flight is decremented from the vq callback.  This avoids
unnecessary sleep wakeups and the extra time spent sleeping after
in_flight has been decremented.

Stefan