[Intel-gfx] [PATCH v13 5/7] vfio: ABI for mdev display dma-buf operation
Alex Williamson
alex.williamson at redhat.com
Wed Aug 9 16:57:42 UTC 2017
On Wed, 9 Aug 2017 08:31:00 +0000
"Zhang, Tina" <tina.zhang at intel.com> wrote:
> > -----Original Message-----
> > From: intel-gvt-dev [mailto:intel-gvt-dev-bounces at lists.freedesktop.org] On
> > Behalf Of Alex Williamson
> > Sent: Tuesday, August 8, 2017 1:43 AM
> > To: Zhang, Tina <tina.zhang at intel.com>
> > Cc: Tian, Kevin <kevin.tian at intel.com>; intel-gfx at lists.freedesktop.org; dri-
> > devel at lists.freedesktop.org; kwankhede at nvidia.com; kraxel at redhat.com;
> > intel-gvt-dev at lists.freedesktop.org; Wang, Zhi A <zhi.a.wang at intel.com>; Lv,
> > Zhiyuan <zhiyuan.lv at intel.com>
> > Subject: Re: [PATCH v13 5/7] vfio: ABI for mdev display dma-buf operation
> >
> > On Mon, 7 Aug 2017 08:11:43 +0000
> > "Zhang, Tina" <tina.zhang at intel.com> wrote:
> >
> > > After going through the previous discussions, here are some summaries may
> > be related to the current discussion:
> > > 1. How does user mode figure the device capabilities between region and
> > dma-buf?
> > > VFIO_DEVICE_GET_REGION_INFO could tell if the mdev supports region case.
> > > Otherwise, the mdev supports dma-buf.
> >
> > Why do we need to make this assumption? What happens when dma-buf is
> > superseded? What happens if a device supports both dma-buf and regions?
> > We have a flags field in vfio_device_gfx_plane_info, doesn't it make sense to use
> > it to identify which field, between region_index and fd, is valid? We could even
> > put region_index and fd into a union with the flag bits indicating how to
> > interpret the union, but I'm not sure everyone was onboard with this idea.
> > Seems like a waste of 4 bytes not to do that though.
> It seems we discussed this idea before:
> https://lists.freedesktop.org/archives/intel-gvt-dev/2017-June/001304.html
> https://lists.freedesktop.org/archives/intel-gvt-dev/2017-June/001333.html
These are both from Gerd. Gerd, do you have any objection to using a
union to provide either the dmabuf fd or region index? It's
inefficient to provide separate fields when we can only provide one or
the other and I don't like the idea of using implicit values to
determine which is active.
> > Thinking further, is the user ever in a situation where they query the graphics
> > plane info and can handle either a dma-buf or a region? It seems more likely
> > that the user needs to know early on which is supported and would then require
> > that they continue to see compatible plane information... Should the user then
> > be able to specify whether they want a dma-buf or a region? Perhaps these flag
> > bits are actually input and the return should be -errno if the driver cannot
> > produce something compatible.
> From the previously discussion, it seems user space workflow will look quite different
> for these two cases. So once user space finds out which case is supported, it just uses
> that case, and won't change it.
And that's supported, the user tests whether a given interface type is
supported and continues to request updates for that interface type.
> Meanwhile, I'm not sure whether there will be a mdev would like to support both region
> and dma-buf cases. In my opinion, either region or dma-buf is supported by one mdev. (Yeah,
> agree, there may be other cases in future)
There's no requirement to support both.
> It's like we want to propose a general interface used to share guest's buffer with host. And the
> general interface, so far, has two choice: region and dma-buf. So each mdev likes this interface
> can implement one kind of it and gets the benefit from the general interface.
> So, if we think about this, the difference in user mode should be as little as possible.
The difference seems pretty minimal here, the user probes supported
interface types, and explicitly picks one by requesting updates using
that interface type. The difference is only in the interpretation of
one dword field. Furthermore, we're not limiting ourselves to these
two interface types, this same API could support dmabuf-v2 if we define
a flag bit for it and define the structure of the interface union.
> > Maybe we'd therefore define 3 flag bits: PROBE, DMABUF, REGION. In this
> > initial implementation, DMABUF or REGION would always be set by the user to
> > request that type of interface. Additionally, the QUERY bit could be set to probe
> > compatibility, thus if PROBE and REGION are set, the vendor driver would return
> > success only if it supports the region based interface. If PROBE and DMABUF are
> > set, the vendor driver returns success only if the dma-buf based interface is
> > supported. The value of the remainder of the structure is undefined for PROBE.
> > Additionally setting both DMABUF and REGION is invalid. Undefined flags bits
> > must be validated as zero by the drivers for future use (thus if we later define
> > DMABUFv2, an older driver should automatically return -errno when probed or
> > requested).
> >
> > It seems like this handles all the cases, the user can ask what's supported and
> > specifies the interface they want on every call. The user therefore can also
> > choose between region_index and fd and we can make that a union.
> Agree, that's a good proposal, which can handle all the cases.
> I'm just not sure about the usage case of "on every call". In previous discussion, it seems we think static is enough.
Is it somehow a problem for the user to set the type bit in the flag on
every call? This makes it explicit that the user is asking for an
update for a specific interface type, if the vendor driver doesn't
support it, return -errno. This makes it explicit whether the return
is a dmabuf fd or a region index and the user knows how to interpret
that union because they have asked for a specific type. As you've
likely gathered from my previous replies, I don't like implicit
interfaces. I don't want the interpretation of a field to depend on
something the user has done in the past and I see no reason to impose a
limitation on the vendor driver supporting only one of two pre-defined
interfaces types. Thanks,
Alex
> > > 2. For dma-buf, how to differentiate unsupported vs not initialized?
> > > For dma-buf, when the mdev doesn't support some arguments, -EINVAL will
> > be returned. And -errno will return when meeting other failures, like -ENOMEM.
> > > If the mdev is not initialized, there won't be any returned err. Just zero all the
> > fields in structure vfio_device_gfx_plane_info.
> >
> > So we're relying on special values again :-\ For which fields is zero not a valid
> > value? I prefer the probe interface above unless there are better ideas.
> >
> > > 3. The id field in structure vfio_device_gfx_plane_info So far we
> > > haven't figured out the usage of this field for dma-buf usage. So, this field is
> > changed to "region_index" and only used for region usage.
> > > In previous discussions, we thought this "id" field might be used for both dma-
> > buf and region usages.
> > > So, we proposed some ways, like adding flags field to the structure. Another
> > way to do it was to add this:
> > > enum vfio_device_gfx_type {
> > > VFIO_DEVICE_GFX_NONE,
> > > VFIO_DEVICE_GFX_DMABUF,
> > > VFIO_DEVICE_GFX_REGION,
> > > };
> > >
> > > struct vfio_device_gfx_query_caps {
> > > __u32 argsz;
> > > __u32 flags;
> > > enum vfio_device_gfx_type;
> > > };
> > > Obviously, we don't need to consider this again, unless we find the
> > "region_index" means something for dma-buf usage.
> >
> > The usefulness of this ioctl seems really limited and once again we get into a
> > situation where having two ioctls leaves doubt whether this is describing the
> > current plane state. Thanks,
> >
> > Alex
> >
> > > > > > > > > > include/uapi/linux/vfio.h | 28
> > > > > > > > > > ++++++++++++++++++++++++++++
> > > > > > > > > > 1 file changed, 28 insertions(+)
> > > > > > > > > >
> > > > > > > > > > diff --git a/include/uapi/linux/vfio.h
> > > > > > > > > > b/include/uapi/linux/vfio.h index ae46105..827a230
> > > > > > > > > > 100644
> > > > > > > > > > --- a/include/uapi/linux/vfio.h
> > > > > > > > > > +++ b/include/uapi/linux/vfio.h
> > > > > > > > > > @@ -502,6 +502,34 @@ struct vfio_pci_hot_reset {
> > > > > > > > > >
> > > > > > > > > > #define VFIO_DEVICE_PCI_HOT_RESET _IO(VFIO_TYPE,
> > > > VFIO_BASE +
> > > > > > > > > 13)
> > > > > > > > > >
> > > > > > > > > > +/**
> > > > > > > > > > + * VFIO_DEVICE_QUERY_GFX_PLANE - _IOW(VFIO_TYPE,
> > > > VFIO_BASE
> > > > > +
> > > > > > > 14,
> > > > > > > > > > +struct vfio_device_query_gfx_plane)
> > > > > > > > > > + *
> > > > > > > > > > + * Set the drm_plane_type and retrieve information
> > > > > > > > > > +about the gfx
> > > > > plane.
> > > > > > > > > > + *
> > > > > > > > > > + * Return: 0 on success, -errno on failure.
> > > > > > > > > > + */
> > > > > > > > > > +struct vfio_device_gfx_plane_info {
> > > > > > > > > > + __u32 argsz;
> > > > > > > > > > + __u32 flags;
> > > > > > > > > > + /* in */
> > > > > > > > > > + __u32 drm_plane_type; /* type of plane: DRM_PLANE_TYPE_*
> > > > > */
> > > > > > > > > > + /* out */
> > > > > > > > > > + __u32 drm_format; /* drm format of plane */
> > > > > > > > > > + __u64 drm_format_mod; /* tiled mode */
> > > > > > > > > > + __u32 width; /* width of plane */
> > > > > > > > > > + __u32 height; /* height of plane */
> > > > > > > > > > + __u32 stride; /* stride of plane */
> > > > > > > > > > + __u32 size; /* size of plane in bytes, align on page*/
> > > > > > > > > > + __u32 x_pos; /* horizontal position of cursor plane, upper
> > > > > left corner
> > > > > > > > > in pixels */
> > > > > > > > > > + __u32 y_pos; /* vertical position of cursor plane, upper left
> > > > > corner in
> > > > > > > > > lines*/
> > > > > > > > > > + __u32 region_index;
> > > > > > > > > > + __s32 fd; /* dma-buf fd */
> > > > > > > > >
> > > > > > > > > How do I know which of these is valid, region_index or fd?
> > > > > > > > > Can I ask for one vs the other? What are the errno values
> > > > > > > > > to differentiate unsupported vs not initialized? Is there a "probe"
> > > > > > > > > flag that I can use to determine what the device supports
> > > > > > > > > without allocating
> > > > > > > those resources yet?
> > > > > > > > Dma-buf won't use region_index, which means region_index is
> > > > > > > > zero all the
> > > > > > > time for dma-buf usage.
> > > > > > > > As we discussed, there won't be errno if not initialized,
> > > > > > > > just keep all fields
> > > > > zero.
> > > > > > > > I will add the comments about these in the next version.
> > > > > > > > Thanks
> > > > > > >
> > > > > > > Zero is a valid region index.
> > > > > > If zero is valid, can we find out other invalid value? How about 0xffffffff?
> > > > >
> > > > > Unlikely, but also valid. Isn't this why we have a flags field in
> > > > > the structure, so we don't need to rely on implicit values as invalid.
> > > > > Also, all of the previously discussed usage models needs to be
> > > > > documented, either here in the header or in a Documentation/ file.
> > > > According to the previously discussion, we also have the following propose:
> > > > enum vfio_device_gfx_type {
> > > > VFIO_DEVICE_GFX_NONE,
> > > > VFIO_DEVICE_GFX_DMABUF,
> > > > VFIO_DEVICE_GFX_REGION,
> > > > };
> > > >
> > > > struct vfio_device_gfx_query_caps {
> > > > __u32 argsz;
> > > > __u32 flags;
> > > > enum vfio_device_gfx_type;
> > > > };
> > > > So, we may need to add one more ioctl, e.g.
> > VFIO_DEVICE_QUERY_GFX_CAPS.
> > > > User mode can query this before querying the plan info, and to see
> > > > which usage (dma-buf or region) is supported.
> > > > Does it still make sense?
> > > > Thanks.
> > > So, I think we can rely on VFIO_DEVICE_GET_REGION_INFO to tell user mode
> > whether the region case is using, instead of introducing a new ioctl.
> > > Thanks
> > >
> > > Tina
> > > >
> > > > Tina
> > > >
> > > >
> > > > > Thanks,
> > > > >
> > > > > Alex
> > > > > _______________________________________________
> > > > > dri-devel mailing list
> > > > > dri-devel at lists.freedesktop.org
> > > > > https://lists.freedesktop.org/mailman/listinfo/dri-devel
> >
> > _______________________________________________
> > intel-gvt-dev mailing list
> > intel-gvt-dev at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/intel-gvt-dev
More information about the Intel-gfx
mailing list