Re: [Qemu-devel] [PATCH 1/2] virtio: add vhost-user-fs base device

[Cornelia Huck <cohuck@redhat.com>]

On Thu, 22 Aug 2019 09:52:37 +0100
Stefan Hajnoczi <stefanha@redhat.com> wrote:

> On Wed, Aug 21, 2019 at 08:11:18PM +0100, Dr. David Alan Gilbert wrote:
> > * Michael S. Tsirkin (mst@redhat.com) wrote:  
> > > On Fri, Aug 16, 2019 at 03:33:20PM +0100, Dr. David Alan Gilbert (git) wrote:  

> > > > +static void vuf_device_realize(DeviceState *dev, Error **errp)
> > > > +{
> > > > +    VirtIODevice *vdev = VIRTIO_DEVICE(dev);
> > > > +    VHostUserFS *fs = VHOST_USER_FS(dev);
> > > > +    unsigned int i;
> > > > +    size_t len;
> > > > +    int ret;
> > > > +
> > > > +    if (!fs->conf.chardev.chr) {
> > > > +        error_setg(errp, "missing chardev");
> > > > +        return;
> > > > +    }
> > > > +
> > > > +    if (!fs->conf.tag) {
> > > > +        error_setg(errp, "missing tag property");
> > > > +        return;
> > > > +    }
> > > > +    len = strlen(fs->conf.tag);
> > > > +    if (len == 0) {
> > > > +        error_setg(errp, "tag property cannot be empty");
> > > > +        return;
> > > > +    }
> > > > +    if (len > sizeof_field(struct virtio_fs_config, tag)) {
> > > > +        error_setg(errp, "tag property must be %zu bytes or less",
> > > > +                   sizeof_field(struct virtio_fs_config, tag));
> > > > +        return;
> > > > +    }
> > > > +
> > > > +    if (fs->conf.num_queues == 0) {
> > > > +        error_setg(errp, "num-queues property must be larger than 0");
> > > > +        return;
> > > > +    }  
> > > 
> > > The strange thing is that actual # of queues is this number + 2.
> > > And this affects an optimal number of vectors (see patch 2).
> > > Not sure what a good solution is - include the
> > > mandatory queues in the number?
> > > Needs to be documented in some way.  
> > 
> > Should we be doing nvectors the same way virtio-scsi-pci does it;
> > with a magic 'unspecified' default where it sets the nvectors based on
> > the number of queues?
> > 
> > I think my preference is not to show the users the mandatory queues.  
> 
> I agree.  Users want to control multiqueue, not on the absolute number
> of virtqueues including mandatory queues.

I agree as well, but let me advocate again for renaming this to
'num_request_queues' or similar to make it more obvious what this number
actually means.

> 
> > > > +
> > > > +    if (!is_power_of_2(fs->conf.queue_size)) {
> > > > +        error_setg(errp, "queue-size property must be a power of 2");
> > > > +        return;
> > > > +    }  
> > > 
> > > Hmm packed ring allows non power of 2 ...
> > > We need to look into a generic helper to support VQ
> > > size checks.  
> > 
> > Which would also have to include the negotiation of where it's doing
> > packaged ring?  
> 
> It's impossible to perform this check at .realize() time since the
> packed virtqueue layout is negotiated via a VIRTIO feature bit.  This
> puts us in the awkward position of either failing when the guest has
> already booted or rounding up the queue size for split ring layouts
> (with a warning message?).

I fear that is always going to be awkward if you allow to specify the
queue size via a property. Basically, you can do two things: fail to
accept FEATURES_OK if the queue size is not a power of 2 and the guest
did not negotiate packed ring, or disallow to set a non power of 2
value here, which is what the other devices with such a property
currently do (see also my other mail.) Would probably be good if all
devices used the same approach (when we introduced packed ring support.)