Re: [virtio-comment] [PATCH V2] virtio-net: introduce admin control virtqueue

[Haozhong Zhang <haozhong@zhan9.net>]

On 01/25/21 14:45, Jason Wang wrote:
> 
> On 2021/1/25 äå2:31, Haozhong Zhang wrote:
> > On 01/25/21 13:52, Jason Wang wrote:
> > > When implementing virtual devices like SR-IOV or sub-function. We're
> > > suffering from several issues:
> > > 
> > > - There's no management interface for management device to
> > >    configure features, attributes for a virtual device
> > > - Per virtual device control virtqueue could be very expensive as the
> > >    number of virtual devices could be very large
> > > - Virtualize per virtual device's control virtqueue could be very
> > >    challenge as we need the support of DMA isolation at queue level
> > > 
> > > So this patch introduces the feature of VIRTIO_NET_CTRL_ADMIN_VQ. This
> > > allows the device to implement a single admin control virtqueue to
> > > manage the features and attributes for a specific virtual device.
> > > 
> > > The idea is simple, a new virtual device id is introduced on top of
> > > the existing virtio_net_ctrl structure. This id is transport or device
> > > specific to uniquely identify a management or virtual device.
> > > 
> > > With this, we get a way of using management device (PF) to configure
> > > per virtual device features and attributes. And since the admin
> > > control virtqueue belongs to management device (PF), the DMA is
> > > naturally isolated at device level instead of the queue level for per
> > > virtual device control vq.
> > > 
> > > When the admin cvq is offered by management device and normal cvq is
> > > offered by virtual device. A new command class is introduced decide
> > > whether or not to accept commands from normal cvq for a virtual
> > > device.
> > > 
> > > Signed-off-by: Jason Wang <jasowang@redhat.com>
> > > ---
> > > Changes from V1:
> > > - use 'virtual device' instead of 'function'
> > > - introuce trust command
> > > - clairfy that the admin cvq could be used to configure management
> > >    device itself
> > > ---
> > >   content.tex | 59 ++++++++++++++++++++++++++++++++++++++++++++++++++---
> > >   1 file changed, 56 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/content.tex b/content.tex
> > > index 620c0e2..989b4f6 100644
> > > --- a/content.tex
> > > +++ b/content.tex
> > > @@ -2940,6 +2940,9 @@ \subsection{Feature bits}\label{sec:Device Types / Network Device / Feature bits
> > >   \item[VIRTIO_NET_F_CTRL_MAC_ADDR(23)] Set MAC address through control
> > >       channel.
> > > +\item[VIRTIO_NET_F_ADMIN_CTRL_VQ(56)] Admin control channel is
> > > +    available.
> > > +
> > >   \item[VIRTIO_NET_F_HASH_REPORT(57)] Device can report per-packet hash
> > >       value and a type of calculated hash.
> > > @@ -3840,11 +3843,12 @@ \subsubsection{Processing of Incoming Packets}\label{sec:Device Types / Network
> > >   \subsubsection{Control Virtqueue}\label{sec:Device Types / Network Device / Device Operation / Control Virtqueue}
> > >   The driver uses the control virtqueue (if VIRTIO_NET_F_CTRL_VQ is
> > > -negotiated) to send commands to manipulate various features of
> > > -the device which would not easily map into the configuration
> > > -space.
> > > +negotiated but VIRTIO_NET_F_ADMIN_CTRL_VQ is not negotiated) to send
> > > +commands to manipulate various features of the device which would not
> > > +easily map into the configuration space.
> > >   All commands are of the following form:
> > > +is not negotiated:
> > >   \begin{lstlisting}
> > >   struct virtio_net_ctrl {
> > > @@ -3864,6 +3868,29 @@ \subsubsection{Control Virtqueue}\label{sec:Device Types / Network Device / Devi
> > >   do except issue a diagnostic if \field{ack} is not
> > >   VIRTIO_NET_OK.
> > > +When both VIRTIO_NET_F_CTRL_VQ and VIRTIO_NET_F_ADMIN_CTRL_VQ are
> > > +negotiated, the driver can use the admin control virtqueue of the
> > > +management device to manipulate features of individual virtual devices
> > > +where the control virtqueue is not easily implemented. The definition
> > > +of management device and virtual device is transport or device
> > > +specific. E.g in the case of PCI SR-IOV, the management device is
> > > +implemented via the physical function (PF), then the virtual device is
> > > +the virtual function (VF) in this case.
> > > +
> > > +All commands are of the following form:
> > > +
> > > +\begin{lstlisting}
> > > +struct virtio_net_admin_ctrl {
> > > +        u32 virtual_device_id;
> > > +        struct virtio_net_ctrl ctrl;
> > > +};
> > > +\end{lstlisting}
> > > +
> > > +The \field{virtual_device_id} is an unique transport or device specific
> > > +identifier for a virtual device or management device. E.g for the case
> > > +of PCI SR-IOV, it's the PCI function id. Management device MUST
> > > +reserve 0 for \field{virtual_device_id} to identify itself.
> > > +
> > Hi Jason,
> > 
> > One question about the valid scope of virtual_device_id. Consider a
> > case, that
> > 
> > 0. A VF with PCI address "xx:yy.z" exists
> > 1. The virtio driver commits an admin command with virtual_device_id
> >     "xx:yy.z" admin queue
> > 2. Before the device processes the command, the virtio driver disables
> >     PF SRIOV, then re-enables it, and then creates a VF with the same
> >     address as "xx:yy.z".
> > 
> > Now, should the command committed in step 1 still be considered as
> > valid after step 2? If not, shall there be a flush or abort operation
> > to allow the driver to drop commands between step 1 and 2? Or simply
> > let the driver just wait until all admin commands have been processed.
> 
> 
> Good point. My understanding is:
> 
> From device level: the effect of the command ties to the lifecycle of the
> device (e.g VF). So if virtual device (VF) is destroyed (e.g SR-IOV is
> disabled), the command won't be effective any more even if it is re-created
> later. And if the virtual device is destroyed during processing the admin
> command, the management device (PF) must fail the command.
> 
> From driver level: The driver should synchronize between the admin control
> vq processing and virtual devices create/destroy to avoid unexpected
> results.
>
> Does this make sense?
> 
> I think we can clarify this in the spec.

I guess the above virtual device destroy also includes the case that
the virtual device is already in NEEDS_RESET status. If so, all make
sense. I think the driver can synchronize easily, for example by just
waiting all commands targeted to that VF failed.

Haozhong

Attachment: signature.asc
Description: PGP signature