On Thu, 11 Oct 2018 09:14:22 +0800
Zhenyu Wang <zhenyuw(a)linux.intel.com> wrote:
On 2018.10.10 23:22:20 +0000, Tian, Kevin wrote:
> > From: Alex Williamson
> > Sent: Thursday, October 11, 2018 4:39 AM
> >
> > On Tue, 9 Oct 2018 01:40:17 +0530
> > Kirti Wankhede <kwankhede(a)nvidia.com> wrote:
> >
> > > Generally a single instance of mdev device, a share of physical device,
is
> > > assigned to user space application or a VM. There are cases when
> > multiple
> > > instances of mdev devices of same or different types are required by user
> > > space application or VM. For example in case of vGPU, multiple mdev
> > devices
> > > of type which represents whole GPU can be assigned to one instance of
> > > application or VM.
> > >
> > > All types of mdev devices may not support assigning multiple mdev
> > devices
> > > to a user space application. In that case vendor driver can fail open()
> > > call of mdev device. But there is no way to know User space application
to
> > > about the configuration supported by vendor driver.
> > >
> > > To expose supported configuration, vendor driver should add
> > > 'single_usage_restriction' attribute to type-id directory.
Returning Y for
> > > this attribute indicates vendor driver has restriction of single mdev
> > > device of particular <type-id> assigned to one user space
application.
> > > Returning N indicates that multiple mdev devices of particular
<type-id>
> > > can be assigned to one user space application.
> > >
> > > User space application should read if 'single_usage_restriction'
attibute
> > > is present in <type-id> directory of all mdev devices which are
going to be
> > > used. If all read N then user space application can proceed with multiple
> > > mdev devices.
> > >
> > > This is optional and readonly attribute.
> > >
> > > Signed-off-by: Kirti Wankhede <kwankhede(a)nvidia.com>
> > > Reviewed-by: Neo Jia <cjia(a)nvidia.com>
> > > ---
> > > Documentation/ABI/testing/sysfs-bus-vfio-mdev | 16 ++++++++++++++++
> > > 1 file changed, 16 insertions(+)
> > >
> > > diff --git a/Documentation/ABI/testing/sysfs-bus-vfio-mdev
> > b/Documentation/ABI/testing/sysfs-bus-vfio-mdev
> > > index 452dbe39270e..3aca352a70e5 100644
> > > --- a/Documentation/ABI/testing/sysfs-bus-vfio-mdev
> > > +++ b/Documentation/ABI/testing/sysfs-bus-vfio-mdev
> > > @@ -85,6 +85,22 @@ Users:
> > > a particular <type-id> that can help in understanding the
> > > features provided by that type of mediated device.
> > >
> > > +What: /sys/.../mdev_supported_types/<type-
> > id>/single_usage_restriction
> > > +Date: October 2018
> > > +Contact: Kirti Wankhede <kwankhede(a)nvidia.com>
> > > +Description:
> > > + Reading this attribute will return Y or N. Returning Y
> > indicates
> > > + vendor driver has restriction of single mdev device of this
> > > + particular <type-id> assigned to one user space application.
> > > + Returning N indicates that multiple mdev devices of
> > particular
> > > + <type-id> can be assigned to one user space application.
> > > + This is optional and readonly attribute.
> > > +Users:
> > > + User space application should read if
> > 'single_usage_restriction'
> > > + attibute is present in <type-id> directory of all mdev devices
> > > + which are going to be used. If all read N then user space
> > > + application can proceed with multiple mdev devices.
> >
> > But we don't say what userspace should do when this optional attribute
> > is not present. Do we know of any cases other than the NVIDIA GRID
> > vGPU drivers that have this restriction? Intel folks, are multiple
> > GVT-g mdevs currently allowed in a VM? I don't think the libvirt
> > algorithm is going to be as simple as suggested here and we should
> > probably understand what it really needs to be.
>
> technically I don't see a restriction in GVT-g side, i.e. multiple GVT-g
> mdevs can be assigned to same VM. But the fact is that Intel GPU is
> integrated thus just one per platform. Then guest i915 driver may have
> problem to operate multiple vGPUs if with some assumption on integrated
> part. I don't think we verified such configuration. Zhenyu?
>
yeah, as current intel gpu is single device, even guest i915 driver could
operate but user space gfx driver is not ready for multiple devices, so won't
be useful now.
The restriction sounds different here though, in the case of NVIDIA
it's the mdev vendor driver with the restriction, with Intel I think
I'm hearing that it's the guest graphics driver that's the
restriction. I don't think we'd want to place an artificial
restriction at the mdev layer for a guest driver deficiency that could
theoretically be fixed independently from the host mdev vendor driver.
Thanks,
Alex