Re: [PATCH RFC v3 09/16] virtio: set FEATURES_OK

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

On Thu, Oct 23, 2014 at 03:30:06PM +0200, Cornelia Huck wrote:
> On Thu, 23 Oct 2014 15:51:39 +0300
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> 
> > On Thu, Oct 23, 2014 at 02:28:08PM +0200, Cornelia Huck wrote:
> > > On Wed, 22 Oct 2014 21:44:44 +0300
> > > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > > 
> > > > set FEATURES_OK as per virtio 1.0 spec
> > > > 
> > > > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > > > ---
> > > >  include/uapi/linux/virtio_config.h |  2 ++
> > > >  drivers/virtio/virtio.c            | 29 ++++++++++++++++++++++-------
> > > >  2 files changed, 24 insertions(+), 7 deletions(-)
> > > > 
> > > 
> > > >  	dev->config->finalize_features(dev);
> > > > 
> > > > +	if (virtio_has_feature(dev, VIRTIO_F_VERSION_1)) {
> > > > +		add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK);
> > > > +		status = dev->config->get_status(dev);
> > > > +		if (!(status & VIRTIO_CONFIG_S_FEATURES_OK)) {
> > > > +			printk(KERN_ERR "virtio: device refuses features: %x\n",
> > > > +			       status);
> > > > +			err = -ENODEV;
> > > > +			goto err;
> > > > +		}
> > > > +	}
> > > > +
> > > 
> > > Ugh, I just realize that virtio-ccw has a problem with that mechanism :(
> > > 
> > > Up to now, the driver only propagated status to the device: For
> > > virtio-ccw, this was easily implemented via a ccw that transmitted
> > > "status" to the device. However, the "read back status" part now
> > > actually requires that the driver can get "status" from the device, or
> > > has a comparable way to find out that the device won't accept the
> > > status it tried to write.
> > 
> > Ugh, it actually caches the status in the transport :(
> 
> Well, it worked as long as this was unidirectional...
> 
> > 
> > 
> > > I can think of two solutions:
> > > 
> > > (1) Introduce a new ccw that actually reads the device status.
> > > (2) Make the WRITE_STATUS ccw fail (with a unit check) if the driver
> > >     sets FEATURES_OK after it tried to set features the device won't
> > >     accept.
> > > 
> > > (1) is probably more generic, while (2) is more straightforward to
> > > implement.
> > > 
> > > Good thing we actually try to finally implement this,
> > 
> > > I did not notice
> > > this problem during the review :(
> > 
> > Well, it's a nuisance, but the spec is out.
> > It seems to me a new command would be a substantive change so we can't
> > do this in errata.
> > 
> > Option (2) would require two statements for drivers and devices,
> > but since it's clearly the case for correct drivers/devices
> > that command does not fail, it follows that this
> > is not a substantive change so it can be fixed
> > in an errata.
> 
> It only adds a new failure case, so it's not really magic, agreed.
> 
> > 
> > So the new command would have to be optional, 
> 
> I think a new command should be tied to a new virtio-ccw revision.

Or a feature bit.

> > please open
> > two issues in the TC: one documenting that driver must check
> > WRITE_STATUS and device can fail WRITE_STATUS, and another
> > for adding READ_STATUS (which will have to wait until
> > the next CS).
> 
> I think I need to contemplate that a bit more. The problem with a new
> READ_STATUS is that it is, by nature, an asynchronous command (as all
> ccw commands are - the Linux guest driver uses a simple wrapper that
> makes it appear synchronous). This means, for example, that every
> status update now involves two channel programs instead of one, and
> that reading e.g. the status sysfs attribute in Linux now triggers
> channel I/O, which means several exits (at least ssch, interrupt
> injection, tsch; probably more depending on what the guest does)...

Well it seems more robust than just caching it within guest.
Anyway, let's start with issue tracker, it does not force any specific
solution.

-- 
MST