Re: [Qemu-devel] [RFC PATCH v1 1/1] vGPU core driver : to provide common interface for vGPU.

[Neo Jia]

On Tue, Feb 16, 2016 at 07:40:47AM +0000, Tian, Kevin wrote:
> > From: Neo Jia [mailto:address@hidden
> > Sent: Tuesday, February 16, 2016 3:37 PM
> > 
> > On Tue, Feb 16, 2016 at 07:27:09AM +0000, Tian, Kevin wrote:
> > > > From: Neo Jia [mailto:address@hidden
> > > > Sent: Tuesday, February 16, 2016 3:13 PM
> > > >
> > > > On Tue, Feb 16, 2016 at 06:49:30AM +0000, Tian, Kevin wrote:
> > > > > > From: Alex Williamson [mailto:address@hidden
> > > > > > Sent: Thursday, February 04, 2016 3:33 AM
> > > > > >
> > > > > > On Wed, 2016-02-03 at 09:28 +0100, Gerd Hoffmann wrote:
> > > > > > >   Hi,
> > > > > > >
> > > > > > > > Actually I have a long puzzle in this area. Definitely libvirt 
> > > > > > > > will use UUID to
> > > > > > > > mark a VM. And obviously UUID is not recorded within KVM. Then 
> > > > > > > > how does
> > > > > > > > libvirt talk to KVM based on UUID? It could be a good reference 
> > > > > > > > to this design.
> > > > > > >
> > > > > > > libvirt keeps track which qemu instance belongs to which vm.
> > > > > > > qemu also gets started with "-uuid ...", so one can query qemu via
> > > > > > > monitor ("info uuid") to figure what the uuid is.  It is also in 
> > > > > > > the
> > > > > > > smbios tables so the guest can see it in the system information 
> > > > > > > table.
> > > > > > >
> > > > > > > The uuid is not visible to the kernel though, the kvm kernel 
> > > > > > > driver
> > > > > > > doesn't know what the uuid is (and neither does vfio).  qemu uses 
> > > > > > > file
> > > > > > > handles to talk to both kvm and vfio.  qemu notifies both kvm and 
> > > > > > > vfio
> > > > > > > about anything relevant events (guest address space changes etc) 
> > > > > > > and
> > > > > > > connects file descriptors (eventfd -> irqfd).
> > > > > >
> > > > > > I think the original link to using a VM UUID for the vGPU comes from
> > > > > > NVIDIA having a userspace component which might get launched from a 
> > > > > > udev
> > > > > > event as the vGPU is created or the set of vGPUs within that UUID is
> > > > > > started.  Using the VM UUID then gives them a way to associate that
> > > > > > userspace process with a VM instance.  Maybe it could register with
> > > > > > libvirt for some sort of service provided for the VM, I don't know.
> > > > >
> > > > > Intel doesn't have this requirement. It should be enough as long as
> > > > > libvirt maintains which sysfs vgpu node is associated to a VM UUID.
> > > > >
> > > > > >
> > > > > > > qemu needs a sysfs node as handle to the vfio device, something
> > > > > > > like /sys/devices/virtual/vgpu/<name>.  <name> can be a uuid if 
> > > > > > > you want
> > > > > > > have it that way, but it could be pretty much anything.  The 
> > > > > > > sysfs node
> > > > > > > will probably show up as-is in the libvirt xml when assign a vgpu 
> > > > > > > to a
> > > > > > > vm.  So the name should be something stable (i.e. when using a 
> > > > > > > uuid as
> > > > > > > name you should better not generate a new one on each boot).
> > > > > >
> > > > > > Actually I don't think there's really a persistent naming issue, 
> > > > > > that's
> > > > > > probably where we diverge from the SR-IOV model.  SR-IOV cannot
> > > > > > dynamically add a new VF, it needs to reset the number of VFs to 
> > > > > > zero,
> > > > > > then re-allocate all of them up to the new desired count.  That has 
> > > > > > some
> > > > > > obvious implications.  I think with both vendors here, we can
> > > > > > dynamically allocate new vGPUs, so I would expect that libvirt would
> > > > > > create each vGPU instance as it's needed.  None would be created by
> > > > > > default without user interaction.
> > > > > >
> > > > > > Personally I think using a UUID makes sense, but it needs to be
> > > > > > userspace policy whether that UUID has any implicit meaning like
> > > > > > matching the VM UUID.  Having an index within a UUID bothers me a 
> > > > > > bit,
> > > > > > but it doesn't seem like too much of a concession to enable the use 
> > > > > > case
> > > > > > that NVIDIA is trying to achieve.  Thanks,
> > > > > >
> > > > >
> > > > > I would prefer to making UUID an optional parameter, while not tieing
> > > > > sysfs vgpu naming to UUID. This would be more flexible to different
> > > > > scenarios where UUID might not be required.
> > > >
> > > > Hi Kevin,
> > > >
> > > > Happy Chinese New Year!
> > > >
> > > > I think having UUID as the vgpu device name will allow us to have an 
> > > > gpu vendor
> > > > agnostic solution for the upper layer software stack such as QEMU, who 
> > > > is
> > > > supposed to open the device.
> > > >
> > >
> > > Qemu can use whatever sysfs path provided to open the device, regardless
> > > of whether there is an UUID within the path...
> > >
> > 
> > Hi Kevin,
> > 
> > Then it will provide even more benefit of using UUID as libvirt can be
> > implemented as gpu vendor agnostic, right? :-)
> > 
> > The UUID can be VM UUID or vGPU group object UUID which really depends on 
> > the
> > high level software stack, again the benefit is gpu vendor agnostic.
> > 
> 
> There is case where libvirt is not used while another mgmt. stack doesn't use
> UUID, e.g. in some Xen scenarios. So it's not about GPU vendor agnostic. It's
> about high level mgmt. stack agnostic. That's why we need make UUID as
> optional in this vGPU-core framework.

Hi Kevin,

As long as you have to create an object to represent vGPU or vGPU group, you
will have UUID, no matter which management stack you are going to use.

UUID is the most agnostic way to represent an object, I think.

(a bit off topic since we are supposed to focus on VFIO on KVM)

Since now you are talking about Xen, I am very happy to discuss that with you.
You can check how Xen has managed its object via UUID in xapi.

Thanks,
Neo

> 
> Thanks
> Kevin