Re: [PATCH v5] vhost: introduce mdev based hardware backend

[Jason Wang]

On 2019/11/6 äå10:49, Tiwei Bie wrote:
> > > > > > +	default:
> > > > > > +		/*
> > > > > > +		 * VHOST_SET_MEM_TABLE, VHOST_SET_LOG_BASE, and
> > > > > > +		 * VHOST_SET_LOG_FD are not used yet.
> > > > > > +		 */
> > > > > If we don't even use them, there's probably no need to call
> > > > > vhost_dev_ioctl(). This may help to avoid confusion when we want to develop
> > > > > new API for e.g dirty page tracking.
> > > > Good point. It's better to reject these ioctls for now.
> > > >
> > > > PS. One thing I may need to clarify is that, we need the
> > > > VHOST_SET_OWNER ioctl to get the vq->handle_kick to work.
> > > > So if we don't call vhost_dev_ioctl(), we will need to
> > > > call vhost_dev_set_owner() directly.
> > I may miss something, it looks to me the there's no owner check in
> > vhost_vring_ioctl() and the vhost_poll_start() can make sure handle_kick
> > works?
> Yeah, there is no owner check in vhost_vring_ioctl().
> IIUC, vhost_poll_start() will start polling the file. And when
> event arrives, vhost_poll_wakeup() will be called, and it will
> queue work to work_list and wakeup worker to finish the work.
> And the worker is created by vhost_dev_set_owner().

Right, rethink about this. It looks to me we need:

- Keep VHOST_SET_OWNER, this could be used for future control vq where 
it needs a kthread to access the userspace memory

- Temporarily filter SET_LOG_BASE and SET_LOG_FD until we finalize the 
API for dirty page tracking.

- For kick through kthread, it looks sub-optimal but we can address this 
in the future, e.g call handle_vq_kick directly in vhost_poll_queue 
(probably a flag for vhost_poll) and deal with the synchronization in 
vhost_poll_flush carefully.

Thanks