Re: [RFC] Host1x/TegraDRM UAPI (drm_tegra_channel_map)

[Thierry Reding <thierry.reding-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 4951 bytes --]

On Fri, Jun 26, 2020 at 01:47:46AM +0300, Dmitry Osipenko wrote:
> 23.06.2020 15:09, Mikko Perttunen пишет:
> > ### DRM_TEGRA_CHANNEL_MAP
> > 
> > Make memory accessible by the engine while executing work on the channel.
> > 
> > ```
> > #define DRM_TEGRA_CHANNEL_MAP_READWRITE        (1<<0)
> > 
> > struct drm_tegra_channel_map {
> >         /*
> >          * [in] ID of the channel for which to map memory to.
> >          */
> >         __u32 channel_id;
> >         /*
> >          * [in] GEM handle of the memory to map.
> >          */
> >         __u32 handle;
> > 
> >         /*
> >          * [in] Offset in GEM handle of the memory area to map.
> >          *
> >          * Must be aligned by 4K.
> >          */
> >         __u64 offset;
> 
> Could you please give a use-case example for this partial mapping?
> 
> I vaguely recalling that maybe it was me who suggested this in the past..
> 
> I kinda see that this could be useful for a case where userspace
> allocates a large chunk of memory and then performs sub-allocations in
> the userspace driver. But do we have a real-world example for this right
> now?

I think the main point about this IOCTL was to make mapping/unmapping
more efficient and avoid relocations for situations where we know it is
safe to do so.

The fact that this can be used to create partial mappings is mostly just
an added bonus, in my opinion. Doing this doesn't create much complexity
but in turn gives us a lot more flexibility.

A couple of places where I think this could be useful are OpenGL and
Vulkan, both of which support buffer suballocation. This has a couple of
advantages on modern GPUs where sometimes you want to use very large
allocation granularity, etc.

Now, I don't think that we'll see much of that in Tegra DRM directly,
although grate could certainly make use of this, I suspect. However, I
think for interoperation of dGPU and Tegra DRM (with VIC for post-
processing, or hopefully some of the other hardware acceleration
engines at some point), this might come in handy.

There are other possible use-cases within just Tegra DRM as well. We may
want to only partially map planar buffers for video post-processing, for
example. Or map each plane separately.

> Please see more below.
> 
> >         /*
> >          * [in] Length of memory area to map in bytes.
> >          *
> >          * Must be aligned by 4K.
> >          */
> >         __u64 length;
> > 
> >         /*
> >          * [out] IOVA of mapped memory. Userspace can use this IOVA
> >          *   directly to refer to the memory to skip using relocations.
> >          *   Only available if hardware memory isolation is enabled.
> >          *
> >          *   Will be set to 0xffff_ffff_ffff_ffff if unavailable.
> >          */
> >         __u64 iova;
> > 
> >         /*
> >          * [out] ID corresponding to the mapped memory to be used for
> >          *   relocations or unmapping.
> >          */
> >         __u32 mapping_id;
> >         /*
> >          * [in] Flags.
> >          */
> >         __u32 flags;
> > 
> >         __u32 reserved[6];
> > };
> 
> It looks to me that we actually need a bit different thing here.
> 
> This MAP IOCTL maps a portion of a GEM and then returns the mapping_id.
> And I think we need something more flexible that will allow us to use
> GEM handles for the relocation IDs, which should fit nicely with the
> DMA-reservations.
> 
> What about an IOCTL that wraps GEM into another GEM? We could wrap a
> portion of GEM_A into a GEM_B, and then map the GEM_B using the MAP IOCTL.
> 
> It could be something like this:
> 
> ### DRM_TEGRA_BO_WRAP
> 
> struct drm_tegra_wrap_bo {
> 	__u32 bo_handle_wrapped; // out
> 	__u32 bo_handle;         // in
> 	__u64 offset;
> 	__u64 length;
> };
> 
> ### DRM_TEGRA_CHANNEL_MAP
> 
> struct drm_tegra_channel_map {
>         __u32 channels_mask;
> 	__u32 mapping_id;
>         __u32 bo_handle;
>         __u32 flags;
>         __u64 iova;
> };
> 
> ===
> 
> This allows multiple mapping_ids to have the same backing GEM, so the
> mapping_id could be resolved into a BO during of job's submission for
> the DMA-reservations handling.

That's pretty much what we have already above, isn't it? Just because we
call the field "mapping_id" doesn't mean that in the background we can't
create a GEM object and return it's handle as "mapping_id".

One advantage of Mikko's proposal is that we have a single IOCTL rather
than two to create the mapping, making it a bit more lightweight.

Thierry

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]