Re: [PATCH v6 1/5] Introduce device group

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

On Thu, Aug 04, 2022 at 10:08:30AM +0800, Jason Wang wrote:
> On Wed, Aug 3, 2022 at 8:33 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> >
> > On Wed, Aug 03, 2022 at 04:04:50PM +0800, Jason Wang wrote:
> > >
> > > å 2022/8/3 14:10, Michael S. Tsirkin åé:
> > > > On Wed, Aug 03, 2022 at 12:44:38PM +0800, Jason Wang wrote:
> > > > > On Tue, Aug 2, 2022 at 9:42 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > > > > I feel some of my latest review opened some questions that I don't have
> > > > > > good answers for and might have felt a bit rambling.
> > > > > > So to focus the discussion:
> > > > > >
> > > > > > On Sun, Jul 31, 2022 at 06:43:50PM +0300, Max Gurtovoy wrote:
> > > > > > > +A device can be a member of one or more device groups.
> > > > > > Presumably this is so we can e.g. create subfunctions inside a VF.
> > > > > Then VF should have its own transport virtqueue. And subfunctions need
> > > > > to be created there. If we don't all thing in PF, we may end up with
> > > > > nesting issue when assign VF to the guest.
> > > > > > A VF now is a member of a SRIOV and SIOV type groups and we
> > > > > > can use type to distinguish between these.
> > > > > >
> > > > > > We should probably be explicit that each of these groups has to
> > > > > > have a distinct group type then.
> > > > > >
> > > > > > And this raises the question: different types have different
> > > > > > capabilities. So let's say admin queue is used to both
> > > > > > control features for SRIOV VFs and to create SIOV SFs.
> > > > > I don't get how the admin queue can be used to control VF features
> > > > > considering VF has its capabilities. (SR-IOV lacks the ability to
> > > > > provision a single VF).
> > > > Well look at latest proposal, last patch controls VF features from PF.
> > >
> > >
> > > Yes, so it works like previous MSI-X allocation which needs some care to
> > > prevent managed device from being probed before assigning features.
> > >
> > > This is technically possible, but I'm not sure it is a good design. For
> > > example, what happens if the management change the feature while the a
> > > driver is using the managed device.
> >
> > I think this should be prohibited in the spec.
> 
> Yes, but implementation wise, this needs to be considered.

Just check the DRIVER status bit, it's not difficult.

> >
> > It might be a good idea to have explicit commands that allow driver to
> > attach.
> >
> > For example the following might work for both VFs and SFs:
> >
> >
> > INIT
> >
> > configure
> >
> > ENABLE <- driver can attach now, configure is blocked
> >
> >
> > --- device can be used ---
> >
> > Note: some configs might be editable while device is in use.
> > E.g. enabling/disabling softmac dynamically.
> >
> > --- device can be used ---
> >
> > DISABLE -> takes control from driver. we can have a flag telling
> >            whether we want to be graceful about it and fail
> >            if driver is still attached or not
> >
> > configure - if we want to attach to another VM
> >
> > CLEANUP - release resources and forget config
> 
> Yes, but for SF it's not a must.
> 
> And should we add these states in the current state machine? If yes,
> it might complicate the migration compatibility.

Noidea what does it have to do with migration.

> >
> >
> >
> > >
> > > >
> > > > > > I guess we'll have a feature bit to say "command to create
> > > > > > SIOV SFs is supported" but how do we say that this command
> > > > > > is only supported for VFs not SFs?
> > > > > I think we should first answer if having VF and SF to be dealt with a
> > > > > single type of virtqueue is a good idea. They have something in common
> > > > > but they distinguish each other:
> > > > >
> > > > > - SF requires per virtual device lifecycle management
> > > > > - SF requires a transport other than PCI
> > > > > - SF requires more mediation in the software layer for presenting a
> > > > > virtual device
> > > > >
> > > > > Using a single type of virtqueue may end up with complex design.
> > > > > Having a dedicated queue for SF might be a better choice.
> > > > And dedicated feature bits for commands thereof?
> > >
> > >
> > > Only needed if we're using a single type of the queue.
> >
> >
> > Imagine a command only allowed for SFs not VFs. Does
> > the PF supporting SFs and VFs have the corresponding
> > feature bit or not?
> 
> I wonder if we can do:
> 
> 1) having two type of virtqueues
> 2) VFs goes to VF admin queue
> 3) SFs goes to SF transport queue
> 
> So if PF supports both SFs and VFs, it should have at least two feature bits.

This does not answer the question.  Let's say we have command X. We
would normally have feature COMMAND_X.  How do we communicate which of
the VQs support which command?


> >
> > >
> > > >    For example, I imagine
> > > > we could have commands to control the MAC of the group member. That is
> > > > the same for SF and VF, yes? How do we avoid duplication for that?
> > >
> > >
> > > In the transport vq, all configs (include mtu and features) were specified
> > > during the device creating command. It is not allowed to change mac
> > > afterwards. (If we need, the SF needs to be destroyed and created again with
> > > different configs).
> >
> > It was just an example. Are you implying SFs and VFs have completely
> > different needs with no overlap then?
> 
> There indeed overlaps, e.g the provisions of the configs. Other than
> these, there should be no other.

Yes provision but my point is that it is not just the config space.
Here's a better example of a resource which is not in device config: MAC
table size.  And one of the issues is that this is also something that can be
changed transparently as device is running.  So we could have a separate
command to provision it both for admin queue and for transport vq, and a
separate command to change it later, but it seems inelegant.


> The idea of the transport virtqueue
> is mainly for having a new transport. This is different from what I
> understand for the admin virtqueue.

Hmm my point was that a transport, or a bus, is in fact a way to address
a group of devices.  Which is exactly what admin queue does.

> > It seems weird since
> > fundamentally they look the same at a lot of levels.
> 
> Yes but only from the view of the functionality.

The concept of a device addressing other devices is what unifies them.
The patch dealing with device groups was developed in response to this.

> >
> >
> >
> > >
> > > >
> > > > > > Do we just make features list a superset of what is supported and simply
> > > > > > say in the spec which commands are legal with which group types?
> > > > > >
> > > > > >
> > > > > > Jason Cornelia what do you think?
> > > > > It looks to me it would be much more simpler if we use separated
> > > > > virtqueues for SRIOV and SIOV.
> > > > >
> > > > > Thanks
> > > > Then is it still helpful that we have the generic group type concept?
> > >
> > >
> > > Not sure, I wonder if the implicit group can do here. E.g _F_SRIOV with
> > > _F_ADMIN_VQ menast SR-IOV group.
> >
> >
> > I don't see how. PF can have SFs right?
> 
> Yes, but technically, we have capabilities then we know which
> virtqueue is doing PF and which is doing SF.

What are capabilities? An older proposal from nvidia had them but
it was dropped. Do you propose bringing them back?

> >
> >
> > >
> > > > I was hoping it will work so the same command can be used for VFs
> > > > and SFs.
> > >
> > >
> > > Yes, but the transport vq ties the mac and other configuration with the
> > > device creating. Not sure we can easily do the same for SR-IOV.
> > >
> > > Thanks
> >
> > We can if we either split SF out or artificially add creation to VFs.
> 
> I agree. But the artificial creation for VF requires more work.

Well we need to specify what can be changed when otherwise it's
a free for all, drivers will go crazy changing random fields
at random time and then we need to support this mess.
Just look at featuresm config fields and FEATURES_OK mess -
the spec said don't do it but implementations did not bother
checking and now we wasted man months already trying to fix it properly.

> >
> >
> >
> > But I expect more command will be exactly the same. Live migration?
> 
> My understanding is that the live migration is a basic facility like
> device status. It means it needs to be transport independent.
>
> That means the function could be accessed via PCI capability/MMIO and
> other transport so it does not look like an issue specific to admin
> virtqueue or transport virtqueue. We can define a common data
> structure then it can be mapped to the same or different commands in
> each type of transport(or virtqueue)?
> 
> Thanks
> 
> Thanks

I think I now understand that you would add capability to have admin
queue inside the vq transport. It does address some issues though not
all. I will need to ponder this.


> >
> >
> >
> > >
> > > >
> > > >
> > > > > >
> > > > > >
> > > > > > > +\item Self type (group identifier = 0) - this group has only one device in the group. Each virtio device is a member of at least one device group, the Self type group.
> > > > > > Presumably, this is here so we can send commands that refer to the
> > > > > > device itself as opposed to a group member (e.g. to
> > > > > > PF as opposed to VF). Is that right?
> > > > > >
> > > > > > It's handy but again the problem here is, this refers to
> > > > > > device as part of which group? Let's just drop this type?
> > > > > >
> > > > > >
> > > > > > --
> > > > > > MST
> > > > > >
> >