Re: [PATCH v2 1/4] Add virtio Admin Virtqueue

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

On Wed, Feb 09, 2022 at 10:27:25AM +0800, Jason Wang wrote:
> 
> 在 2022/1/30 下午11:30, Michael S. Tsirkin 写道:
> > On Sun, Jan 30, 2022 at 05:12:46PM +0200, Max Gurtovoy wrote:
> > > On 1/30/2022 4:41 PM, Michael S. Tsirkin wrote:
> > > > On Sun, Jan 30, 2022 at 11:56:30AM +0200, Max Gurtovoy wrote:
> > > > > On 1/30/2022 11:40 AM, Michael S. Tsirkin wrote:
> > > > > > On Sun, Jan 30, 2022 at 11:13:38AM +0200, Max Gurtovoy wrote:
> > > > > > > On 1/29/2022 5:53 AM, Jason Wang wrote:
> > > > > > > > On Fri, Jan 28, 2022 at 11:52 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > > > > > > > On Fri, Jan 28, 2022 at 04:49:34PM +0100, Cornelia Huck wrote:
> > > > > > > > > > On Fri, Jan 28 2022, "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > > > > > > > > > 
> > > > > > > > > > > On Fri, Jan 28, 2022 at 01:14:14PM +0100, Cornelia Huck wrote:
> > > > > > > > > > > > On Mon, Jan 24 2022, Max Gurtovoy <mgurtovoy@nvidia.com> wrote:
> > > > > > > > > > > > > +\section{Admin Virtqueues}\label{sec:Basic Facilities of a Virtio Device / Admin Virtqueues}
> > > > > > > > > > > > > +
> > > > > > > > > > > > > +Admin virtqueue is used to send administrative commands to manipulate
> > > > > > > > > > > > > +various features of the device and/or to manipulate various features,
> > > > > > > > > > > > > +if possible, of another device within the same group (e.g. PCI VFs of
> > > > > > > > > > > > > +a parent PCI PF device are grouped together. These devices can be
> > > > > > > > > > > > > +optionally managed by its parent PCI PF using its admin virtqueue.).
> > > > > > > > > > > > > +
> > > > > > > > > > > > > +Use of Admin virtqueue is negotiated by the VIRTIO_F_ADMIN_VQ
> > > > > > > > > > > > > +feature bit.
> > > > > > > > > > > > > +
> > > > > > > > > > > > > +Admin virtqueue index may vary among different device types.
> > > > > > > > > > > > So, my understanding is:
> > > > > > > > > > > > - any device type may or may not support the admin vq
> > > > > > > > > > > > - if the device type wants to be able to accommodate the admin vq, it
> > > > > > > > > > > >       also needs to specify where it shows up when the feature is negotiated
> > > > > > > > > > > > 
> > > > > > > > > > > > Do we expect that eventually all device types will need to support the
> > > > > > > > > > > > admin vq (if some use case comes along that will require all devices to
> > > > > > > > > > > > participate, for example?)
> > > > > > > > > > > I suspect yes. And that's one of the reasons why I'd rather we had a
> > > > > > > > > > > device independent way to locate the admin queue. There are less
> > > > > > > > > > > transports than device types.
> > > > > > > > > > So, do we want to bite the bullet now and simply say that every device
> > > > > > > > > > type has the admin vq as the last vq if the feature is negotiated?
> > > > > > > > > > Should be straightforward for the device types that have a fixed number
> > > > > > > > > > of vqs, and doable for those that have a variable amount (two device
> > > > > > > > > > types are covered by this series anyway.) I think we need to put it with
> > > > > > > > > > the device types, as otherwise the numbering of virtqueues could change
> > > > > > > > > > in unpredictable ways with the admin vq off/on.
> > > > > > > > > Well that only works once. The next thing we'll need we won't be able to
> > > > > > > > > make the last one ;) So I am inclined to add a per-transport field that
> > > > > > > > > gives the admin queue number.
> > > > > > > > Technically, there's no need to use the same namespace for admin
> > > > > > > > virtqueue if it has a dedicated notification area. If we go this way,
> > > > > > > > we can simply use 0 as queue index for admin virtqueue.
> > > > > > > Or we can use index 0xFFFF for admin virtqueue for compatibility.
> > > > > > I think I'd prefer a register with the #. For example we might want
> > > > > > to limit the # of VQs in order to pass extra data with the kick write.
> > > > > So you are suggesting adding a new cfg_type (#define
> > > > > VIRTIO_PCI_CAP_ADMIN_CFG 10) ?
> > > > > 
> > > > > that will look something like:
> > > > > 
> > > > > struct virtio_pci_admin_cfg {
> > > > > 
> > > > >       le32 queue_index; /* read only for the driver */
> > > > > 
> > > > >       le16 queue_size; /* read-write */
> > > > >       le16 queue_msix_vector; /* read-write */
> > > > >       le16 queue_enable; /* read-write */
> > > > >       le16 queue_notify_off; /* read-only for driver */
> > > > >       le64 queue_desc; /* read-write */
> > > > >       le64 queue_driver; /* read-write */
> > > > >       le64 queue_device; /* read-write */
> > > > >       le16 queue_notify_data; /* read-only for driver */
> > > > >       le16 queue_reset; /* read-write */
> > > > > 
> > > > > };
> > > > > 
> > > > > instead of re-using the struct virtio_pci_common_cfg ?
> > > > > 
> > > > > 
> > > > > or do you prefer extending the struct virtio_pci_common_cfg with "le16
> > > > > admin_queue_index; /* read only for the driver */ ?
> > > > The later. Other transports will need this too.
> > > > 
> > > > 
> > > > Cornelia has another idea which is that instead of
> > > > adding just the admin queue register to all transports,
> > > > we instead add a misc_config structure to all
> > > > transports. Working basically like device specific config,
> > > > but being device independent. For now it will only have
> > > > a single le16 admin_queue_index register.
> > > > 
> > > > For PCI we would thus add it with VIRTIO_PCI_CAP_MISC_CFG
> > > > 
> > > > The point here is that we are making it easier to add
> > > > more fields just like admin queue index in the future.
> > > OK.
> > > 
> > > #define VIRTIO_PCI_CAP_MISC_CFG 10
> > > 
> > > and
> > > 
> > > struct virtio_pci_misc_cfg {
> > >      le16 admin_queue_index; /* read-only for driver */
> > > };
> > > 
> > > Is agreed by all for V3 ? instead of the net and blk AQ index definitions.
> > We need to add it to MMIO and CCW I guess too.
> 
> 
> I wonder how much useful is this.
> 
> E.g for PCI we have an equation to calculate the queue notify address, if
> device choose to use dedicated notify for each queue it will probably end up
> with the last queue.
> 
> And I think the admin_queue_index should be stable regardless of the feature
> that has been negotiated?
> 
> Thanks


there's only one last queue though. I wouldn't just use it and hope
we don't need it for anything else.

> 
> > 
> > This is Cornelia's idea, we'll need her response.
> > 
> > 
> > > > 
> > > > > > > > Thanks
> > > > > > > > 
> > > > > > > > > Another advantage to this approach is that
> > > > > > > > > we can make sure admin queue gets a page by itself (which can be good if
> > > > > > > > > we want to allow access to regular vqs but not to the admin queue to
> > > > > > > > > guest) even if regular vqs share a page. Will help devices use less
> > > > > > > > > memory space.
> > > > > > > > > 
> > > > > > > > > --
> > > > > > > > > MST
> > > > > > > > >