Re: [PATCH 17/20] drm/i915/uapi: add NEEDS_CPU_ACCESS hint

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On Thu, 2022-02-03 at 11:38 +0000, Matthew Auld wrote:
> On 03/02/2022 09:28, Thomas Hellström wrote:
> > 
> > On 1/26/22 16:21, Matthew Auld wrote:
> > > If set, force the allocation to be placed in the mappable portion
> > > of
> > > LMEM. One big restriction here is that system memory must be
> > > given as a
> > > potential placement for the object, that way we can always spill
> > > the
> > > object into system memory if we can't make space.
> > > 
> > > XXX: Still very much WIP and needs IGTs. Including now just for
> > > the sake
> > > of having more complete picture.
> > > 
> > > Signed-off-by: Matthew Auld <matthew.auld@intel.com>
> > > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> > > ---
> > >   drivers/gpu/drm/i915/gem/i915_gem_create.c | 28 ++++++++++++---
> > > ----
> > >   include/uapi/drm/i915_drm.h                | 31
> > > +++++++++++++++++++++-
> > >   2 files changed, 49 insertions(+), 10 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_create.c 
> > > b/drivers/gpu/drm/i915/gem/i915_gem_create.c
> > > index e7456443f163..98d63cb21e94 100644
> > > --- a/drivers/gpu/drm/i915/gem/i915_gem_create.c
> > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_create.c
> > > @@ -238,6 +238,7 @@ struct create_ext {
> > >       struct drm_i915_private *i915;
> > >       struct intel_memory_region
> > > *placements[INTEL_REGION_UNKNOWN];
> > >       unsigned int n_placements;
> > > +    unsigned int placement_mask;
> > >       unsigned long flags;
> > >   };
> > > @@ -334,6 +335,7 @@ static int set_placements(struct 
> > > drm_i915_gem_create_ext_memory_regions *args,
> > >       for (i = 0; i < args->num_regions; i++)
> > >           ext_data->placements[i] = placements[i];
> > > +    ext_data->placement_mask = mask;
> > >       return 0;
> > >   out_dump:
> > > @@ -408,7 +410,7 @@ i915_gem_create_ext_ioctl(struct drm_device
> > > *dev, 
> > > void *data,
> > >       struct drm_i915_gem_object *obj;
> > >       int ret;
> > > -    if (args->flags)
> > > +    if (args->flags &
> > > ~I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS)
> > >           return -EINVAL;
> > >       ret = i915_user_extensions(u64_to_user_ptr(args-
> > > >extensions),
> > > @@ -424,14 +426,22 @@ i915_gem_create_ext_ioctl(struct drm_device
> > > *dev, void *data,
> > >           ext_data.n_placements = 1;
> > >       }
> > > -    /*
> > > -     * TODO: add a userspace hint to force CPU_ACCESS for the
> > > object, 
> > > which
> > > -     * can override this.
> > > -     */
> > > -    if (!IS_DG1(i915) && (ext_data.n_placements > 1 ||
> > > -                  ext_data.placements[0]->type !=
> > > -                  INTEL_MEMORY_SYSTEM))
> > > -        ext_data.flags |= I915_BO_ALLOC_TOPDOWN;
> > > +    if (args->flags & I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS)
> > > {
> > > +        if (ext_data.n_placements == 1)
> > > +            return -EINVAL;
> > > +
> > > +        /*
> > > +         * We always need to be able to spill to system memory,
> > > if we
> > > +         * can't place in the mappable part of LMEM.
> > > +         */
> > > +        if (!(ext_data.placement_mask & BIT(INTEL_REGION_SMEM)))
> > > +            return -EINVAL;
> > > +    } else {
> > > +        if (!IS_DG1(i915) && (ext_data.n_placements > 1 ||
> > > +                      ext_data.placements[0]->type !=
> > > +                      INTEL_MEMORY_SYSTEM))
> > > +            ext_data.flags |= I915_BO_ALLOC_TOPDOWN;
> > > +    }
> > >       obj = __i915_gem_object_create_user_ext(i915, args->size,
> > >                           ext_data.placements,
> > > diff --git a/include/uapi/drm/i915_drm.h
> > > b/include/uapi/drm/i915_drm.h
> > > index 914ebd9290e5..ecfa805549a7 100644
> > > --- a/include/uapi/drm/i915_drm.h
> > > +++ b/include/uapi/drm/i915_drm.h
> > > @@ -3157,7 +3157,36 @@ struct drm_i915_gem_create_ext {
> > >        * Object handles are nonzero.
> > >        */
> > >       __u32 handle;
> > > -    /** @flags: MBZ */
> > > +    /**
> > > +     * @flags: Optional flags.
> > > +     *
> > > +     * Supported values:
> > > +     *
> > > +     * I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS - Signal to the
> > > kernel that
> > > +     * the object will need to be accessed via the CPU.
> > > +     *
> > > +     * Only valid when placing objects in
> > > I915_MEMORY_CLASS_DEVICE, and
> > > +     * only strictly required on platforms where only some of
> > > the device
> > > +     * memory is directly visible or mappable through the CPU,
> > > like 
> > > on DG2+.
> > > +     *
> > > +     * One of the placements MUST also be
> > > I915_MEMORY_CLASS_SYSTEM, to
> > > +     * ensure we can always spill the allocation to system
> > > memory, if we
> > > +     * can't place the object in the mappable part of
> > > +     * I915_MEMORY_CLASS_DEVICE.
> > > +     *
> > > +     * Note that buffers that need to be captured with 
> > > EXEC_OBJECT_CAPTURE,
> > > +     * will need to enable this hint, if the object can also be 
> > > placed in
> > > +     * I915_MEMORY_CLASS_DEVICE, starting from DG2+. The execbuf
> > > call 
> > > will
> > > +     * throw an error otherwise. This also means that such
> > > objects 
> > > will need
> > > +     * I915_MEMORY_CLASS_SYSTEM set as a possible placement.
> > > +     *
> > 
> > I wonder, should we try to migrate capture objects at execbuf time 
> > instead on an on-demand basis? If migration fails, then we just
> > skip 
> > capturing that object, similar to how the capture code handles
> > errors?
> 
> So IIUC if the object has been marked for capture, unmark the TOPDOWN
> annotation, if it has been set, to force allocating in the mappable 
> portion, or spill to system memory(if the placements allow it)? I
> think 
> that should work.

Yes that would temporarily mean drop the TOPDOWN flag and migrate the
object if needed (we can do that async from inside execbuf AFAICT).
We'd need to make the TOPDOWN flag mutable and part of object-
>mem_flags

>  Jon any thoughts

And on that subject, the TOPDOWN flag name is IMHO a bit misleading. It
really has a "GPU_ONLY" meaning, which translates to TOPDOWN
allocations in some memory regions only. For others (no small bar,
small bar + multi-tile, it might translate to nothing or to tile
selection)?

/Thomas