[PATCH 15/87] drm/i915: Pack the partial view size and offset into a single u64
Tvrtko Ursulin
tvrtko.ursulin at linux.intel.com
Fri Jan 6 13:32:32 UTC 2017
On 05/01/2017 23:19, Chris Wilson wrote:
> Since the partial offset must be page aligned, we can use those low 12
> bits to encode the side of the partial view (which then cannot be larger
s/side/size/
> than 8MiB in pages).
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
> drivers/gpu/drm/i915/i915_gem.c | 22 +++++++++++++---------
> drivers/gpu/drm/i915/i915_gem_gtt.c | 4 ++--
> drivers/gpu/drm/i915/i915_gem_gtt.h | 25 +++++++++++++++++++++++--
> drivers/gpu/drm/i915/i915_vma.c | 9 ++++-----
> 4 files changed, 42 insertions(+), 18 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index f6f4ec894a7f..2c1631190a60 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -1698,6 +1698,9 @@ static unsigned int tile_row_pages(struct drm_i915_gem_object *obj)
> {
> u64 size;
>
> + BUILD_BUG_ON(ilog2(GEN7_FENCE_MAX_PITCH_VAL*128*32 >> PAGE_SHIFT) >
> + INTEL_PARTIAL_SIZE_BITS);
> +
> size = i915_gem_object_get_stride(obj);
> size *= i915_gem_object_get_tiling(obj) == I915_TILING_Y ? 32 : 8;
>
> @@ -1831,24 +1834,25 @@ int i915_gem_fault(struct vm_area_struct *area, struct vm_fault *vmf)
> vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, flags);
> if (IS_ERR(vma)) {
> struct i915_ggtt_view view;
> - unsigned int chunk_size;
> + unsigned int chunk;
>
> /* Use a partial view if it is bigger than available space */
> - chunk_size = MIN_CHUNK_PAGES;
> + chunk = MIN_CHUNK_PAGES;
> if (i915_gem_object_is_tiled(obj))
> - chunk_size = roundup(chunk_size, tile_row_pages(obj));
> + chunk = roundup(chunk, tile_row_pages(obj));
>
> memset(&view, 0, sizeof(view));
> view.type = I915_GGTT_VIEW_PARTIAL;
> - view.params.partial.offset = rounddown(page_offset, chunk_size);
> - view.params.partial.size =
> - min_t(unsigned int, chunk_size,
> - vma_pages(area) - view.params.partial.offset);
> + view.params.partial.offset_size = rounddown(page_offset, chunk);
> + view.params.partial.offset_size =
> + (view.params.partial.offset_size << INTEL_PARTIAL_SIZE_BITS) |
> + (min_t(unsigned int, chunk,
> + vma_pages(area) - view.params.partial.offset_size) - 1);
>
> /* If the partial covers the entire object, just create a
> * normal VMA.
> */
> - if (chunk_size >= obj->base.size >> PAGE_SHIFT)
> + if (chunk >= obj->base.size >> PAGE_SHIFT)
> view.type = I915_GGTT_VIEW_NORMAL;
>
> /* Userspace is now writing through an untracked VMA, abandon
> @@ -1878,7 +1882,7 @@ int i915_gem_fault(struct vm_area_struct *area, struct vm_fault *vmf)
>
> /* Finally, remap it using the new GTT offset */
> ret = remap_io_mapping(area,
> - area->vm_start + (vma->ggtt_view.params.partial.offset << PAGE_SHIFT),
> + area->vm_start + intel_partial_get_offset(&vma->ggtt_view.params.partial),
> (ggtt->mappable_base + vma->node.start) >> PAGE_SHIFT,
> min_t(u64, vma->size, area->vm_end - area->vm_start),
> &ggtt->mappable);
> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
> index f698006fe883..4e77baf7d652 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.c
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
> @@ -3492,7 +3492,7 @@ intel_partial_pages(const struct i915_ggtt_view *view,
> {
> struct sg_table *st;
> struct scatterlist *sg, *iter;
> - unsigned int count = view->params.partial.size;
> + unsigned int count = intel_partial_get_pages(&view->params.partial);
Thinking if get_pages could be improved since we use it elsewhere so
much - get_size_pages ? Or just the existing get_size() >> PAGE_SHIFT.
> unsigned int offset;
> int ret = -ENOMEM;
>
> @@ -3505,7 +3505,7 @@ intel_partial_pages(const struct i915_ggtt_view *view,
> goto err_sg_alloc;
>
> iter = i915_gem_object_get_sg(obj,
> - view->params.partial.offset,
> + intel_partial_get_page_offset(&view->params.partial),
> &offset);
> GEM_BUG_ON(!iter);
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h
> index 19ea4c942df4..023bf6ac3dc7 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.h
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h
> @@ -156,10 +156,31 @@ struct intel_rotation_info {
> };
>
> struct intel_partial_info {
> - u64 offset;
> - unsigned int size;
> + /* offset is page-aligned, leaving just enough bits for the size */
> +#define INTEL_PARTIAL_SIZE_BITS PAGE_SHIFT
> + u64 offset_size;
> };
>
> +static inline u32 intel_partial_get_pages(const struct intel_partial_info *pi)
> +{
> + return 1 + (pi->offset_size & GENMASK(INTEL_PARTIAL_SIZE_BITS-1, 0));
> +}
> +
> +static inline u32 intel_partial_get_size(const struct intel_partial_info *pi)
> +{
> + return intel_partial_get_pages(pi) << PAGE_SHIFT;
> +}
> +
> +static inline u64 intel_partial_get_offset(const struct intel_partial_info *pi)
> +{
> + return pi->offset_size & GENMASK(63, INTEL_PARTIAL_SIZE_BITS);
> +}
> +
> +static inline u64 intel_partial_get_page_offset(const struct intel_partial_info *pi)
> +{
> + return pi->offset_size >> INTEL_PARTIAL_SIZE_BITS;
> +}
> +
> struct i915_ggtt_view {
> enum i915_ggtt_view_type type;
>
> diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
> index 58f2483362ad..65770b7109c0 100644
> --- a/drivers/gpu/drm/i915/i915_vma.c
> +++ b/drivers/gpu/drm/i915/i915_vma.c
> @@ -96,11 +96,10 @@ __i915_vma_create(struct drm_i915_gem_object *obj,
> vma->ggtt_view = *view;
> if (view->type == I915_GGTT_VIEW_PARTIAL) {
> GEM_BUG_ON(range_overflows_t(u64,
> - view->params.partial.offset,
> - view->params.partial.size,
> - obj->base.size >> PAGE_SHIFT));
> - vma->size = view->params.partial.size;
> - vma->size <<= PAGE_SHIFT;
> + intel_partial_get_offset(&view->params.partial),
> + intel_partial_get_size(&view->params.partial),
> + obj->base.size));
> + vma->size = intel_partial_get_size(&view->params.partial);
> GEM_BUG_ON(vma->size >= obj->base.size);
> } else if (view->type == I915_GGTT_VIEW_ROTATED) {
> vma->size =
>
I am not sure the complication is worth the small saving. Don't we ever
expect to start creating >8MiB partials? If one day we get an efficient
way of querying for the biggest hole we could start doing that and then
this would have to be undone.
Regards,
Tvrtko
More information about the Intel-gfx-trybot
mailing list