Re: [Intel-gfx] [PATCH v9 5/7] vfio: Define vfio based dma-buf operations

[Alex Williamson <alex.williamson@redhat.com>]

On Tue, 20 Jun 2017 12:57:36 +0200
Gerd Hoffmann <kraxel@redhat.com> wrote:

> On Tue, 2017-06-20 at 08:41 +0000, Zhang, Tina wrote:
> > Hi,
> > 
> > Thanks for all the comments. Here are the summaries:
> > 
> > 1. Modify the structures to make it more general.
> > struct vfio_device_gfx_plane_info {
> > 	__u64 start;
> > 	__u64 drm_format_mod;
> > 	__u32 drm_format;
> > 	__u32 width;
> > 	__u32 height;
> > 	__u32 stride;
> > 	__u32 size;
> > 	__u32 x_pos;
> > 	__u32 y_pos;
> > 	__u32 generation;
> > };  
> 
> Looks good to me.
> 
> > struct vfio_device_query_gfx_plane {
> > 	__u32 argsz;
> > 	__u32 flags;
> > #define VFIO_GFX_PLANE_FLAGS_REGION_ID		(1 << 0)
> > #define VFIO_GFX_PLANE_FLAGS_PLANE_ID		(1 << 1)
> > 	struct vfio_device_gfx_plane_info plane_info;
> > 	__u32 id; 
> > };  
> 
> I'm not convinced the flags are a great idea.  Whenever dmabufs or a
> region is used is a static property of the device, not of each
> individual plane.
> 
> 
> I think we should have this for userspace to figure:
> 
> 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;
> };

We already have VFIO_DEVICE_GET_INFO which returns:

struct vfio_device_info {
        __u32   argsz;
        __u32   flags;
#define VFIO_DEVICE_FLAGS_RESET (1 << 0)        /* Device supports reset */
#define VFIO_DEVICE_FLAGS_PCI   (1 << 1)        /* vfio-pci device */
#define VFIO_DEVICE_FLAGS_PLATFORM (1 << 2)     /* vfio-platform device */
#define VFIO_DEVICE_FLAGS_AMBA  (1 << 3)        /* vfio-amba device */
#define VFIO_DEVICE_FLAGS_CCW   (1 << 4)        /* vfio-ccw device */
        __u32   num_regions;    /* Max region index + 1 */
        __u32   num_irqs;       /* Max IRQ index + 1 */
};

We could use two flag bits to indicate dmabuf or graphics region
support.  vfio_device_gfx_query_caps seems to imply a new ioctl, which
would be unnecessary.
 
> Then this to query the plane:
> 
> struct vfio_device_gfx_query_plane {
>         __u32 argsz;
>         __u32 flags;
>         struct vfio_device_gfx_plane_info plane_info;  /* out */
>         __u32 plane_type;                              /* in  */
> };

I'm not sure why we're using an enum for something that can currently
be defined with 2 bits, seems like this would be another good use of
flags.  We could even embed an enum into the flags if we want to
leave some expansion room, 4 bits maybe?  Also, I was imagining that a
device could support multiple graphics regions, that's where specifying
the "id" as a region index seemed useful.  We lose that ability here
unless we go back to defining a flag bit to specify how to interpret
this last field.

> 2. Remove dmabuf mgr fd and add these two ioctl commands to the vfio
> device fd.
> > VFIO_DEVICE_QUERY_GFX_PLANE : used to query
> > vfio_device_gfx_plane_info.  
> 
> Yes.
> 
> > VFIO_DEVICE_GET_DMABUF_FD: used to create and return the dmabuf fd.  

I'm not convinced this adds value, but I'll list it as an option:

VFIO_DEVICE_QUERY(VFIO_DEVICE_GFX_PLANE)
VFIO_DEVICE_GET_FD(VFIO_DEVICE_GFX_DMABUF_FD)

The benefit is that it might help to avoid a proliferation of ioctls on
the device the pain is that we need to either define a field or section
of flags which identify what is being queried or what type of device fd
is being requested.

> Yes.  The plane might have changed between query-plane and get-dmabuf
> ioctl calls though, we must make sure we handle that somehow.  Current
> patches return plane_info on get-dmabuf ioctl too, so userspace can see
> what it actually got.
> 
> With the generation we can also do something different:  Pass in
> plane_type and generation, and have VFIO_DEVICE_GET_DMABUF_FD return
> an error in case the generation doesn't match.  In that case it doesn't
> make much sense any more to have a separate plane_info struct, which
> was added so we don't have to duplicate things in query-plane and get-
> dmabuf ioctl structs.

I'm not sure I understand how this works for a region, the region is
always the current generation, how can the user ever be sure the
plane_info matches what is exposed in the region?  Thanks,

Alex