Re: [Qemu-devel] [PATCH 2/2] virtio-balloon: resume collecting stats on vmload

[Roman Kagan]

On Tue, Sep 06, 2016 at 04:45:13AM +0300, Michael S. Tsirkin wrote:
> On Mon, Sep 05, 2016 at 11:02:36AM +0300, Roman Kagan wrote:
> > On Sat, Sep 03, 2016 at 01:53:53AM +0300, Michael S. Tsirkin wrote:
> > > On Fri, Sep 02, 2016 at 10:21:58AM +0300, Roman Kagan wrote:
> > > > On Thu, Sep 01, 2016 at 10:26:54PM +0300, Michael S. Tsirkin wrote:
> > > > > I'm sorry - I don't like this patch. This means that
> > > > > virtio_balloon_receive_stats will be called and will poke
> > > > > at the ring even if the ring was never kicked.
> > > > 
> > > > I'm not sure I understand what the problem is with that:
> > > > virtio_balloon_receive_stats just returns early if virtio_queue_empty(),
> > > > which is no more poking at the ring than is already done in virtio_load.
> > > 
> > > Generally we should not look at ring until there was a kick.
> > 
> > How else would you recover ->inuse?
> > 
> > Roman.
> 
> We seem to do this:
> +            vdev->vq[i].inuse = vdev->vq[i].last_avail_idx -
> +                                vdev->vq[i].used_idx;

Right, but .used_idx is fetched from vring two lines above:

        vdev->vq[i].used_idx = vring_used_idx(&vdev->vq[i]);

> Is this wrong?

And no, I don't think it is wrong :)  I guess there's just no other way
to obtain it.

Roman.