[Intel-gfx] [PATCH v10 22/23] drm/i915/vm_bind: Properly build persistent map sg table

[Matthew Auld]

On 18/01/2023 07:16, Niranjana Vishwanathapura wrote:
> Properly build the sg table for persistent mapping which can
> be partial map of the underlying object. Ensure the sg pages
> are properly set for page backed regions. The dump capture
> support requires this for page backed regions.
> 
> v2: Remove redundant sg_mark_end() call
> 
> Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura at intel.com>
> ---
>   drivers/gpu/drm/i915/i915_vma.c | 113 +++++++++++++++++++++++++++++++-
>   1 file changed, 112 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
> index 5b9ae5ebf55c..2f0994f0ed42 100644
> --- a/drivers/gpu/drm/i915/i915_vma.c
> +++ b/drivers/gpu/drm/i915/i915_vma.c
> @@ -1328,6 +1328,113 @@ intel_partial_pages(const struct i915_gtt_view *view,
>   	return ERR_PTR(ret);
>   }
>   
> +static unsigned int
> +intel_copy_dma_sg(struct sg_table *src_st, struct sg_table *dst_st,
> +		  u64 offset, u64 length, bool dry_run)
> +{
> +	struct scatterlist *dst_sg, *src_sg;
> +	unsigned int i, len, nents = 0;
> +
> +	dst_sg = dst_st->sgl;
> +	for_each_sgtable_dma_sg(src_st, src_sg, i) {
> +		if (sg_dma_len(src_sg) <= offset) {
> +			offset -= sg_dma_len(src_sg);
> +			continue;
> +		}
> +
> +		nents++;
> +		len = min(sg_dma_len(src_sg) - offset, length);
> +		if (!dry_run) {
> +			sg_dma_address(dst_sg) = sg_dma_address(src_sg) + offset;
> +			sg_dma_len(dst_sg) = len;
> +			dst_sg = sg_next(dst_sg);
> +		}
> +
> +		length -= len;
> +		offset = 0;
> +		if (!length)
> +			break;
> +	}
> +	WARN_ON_ONCE(length);
> +
> +	return nents;
> +}
> +
> +static unsigned int
> +intel_copy_sg(struct sg_table *src_st, struct sg_table *dst_st,
> +	      u64 offset, u64 length, bool dry_run)
> +{
> +	struct scatterlist *dst_sg, *src_sg;
> +	unsigned int i, len, nents = 0;
> +
> +	dst_sg = dst_st->sgl;
> +	for_each_sgtable_sg(src_st, src_sg, i) {
> +		if (src_sg->length <= offset) {
> +			offset -= src_sg->length;
> +			continue;
> +		}
> +
> +		nents++;
> +		len = min(src_sg->length - offset, length);
> +		if (!dry_run) {
> +			unsigned long pfn;
> +
> +			pfn = page_to_pfn(sg_page(src_sg)) + offset / PAGE_SIZE;
> +			sg_set_page(dst_sg, pfn_to_page(pfn), len, 0);
> +			dst_sg = sg_next(dst_sg);
> +		}
> +
> +		length -= len;
> +		offset = 0;
> +		if (!length)
> +			break;
> +	}
> +	WARN_ON_ONCE(length);
> +
> +	return nents;
> +}
> +
> +static noinline struct sg_table *

Not sure why this noinline is needed here?

Reviewed-by: Matthew Auld <matthew.auld at intel.com>

> +intel_persistent_partial_pages(const struct i915_gtt_view *view,
> +			       struct drm_i915_gem_object *obj)
> +{
> +	u64 offset = view->partial.offset << PAGE_SHIFT;
> +	struct sg_table *st, *obj_st = obj->mm.pages;
> +	u64 length = view->partial.size << PAGE_SHIFT;
> +	unsigned int nents;
> +	int ret = -ENOMEM;
> +
> +	st = kmalloc(sizeof(*st), GFP_KERNEL);
> +	if (!st)
> +		goto err_st_alloc;
> +
> +	/* Get required sg_table size */
> +	nents = intel_copy_dma_sg(obj_st, st, offset, length, true);
> +	if (i915_gem_object_has_struct_page(obj)) {
> +		unsigned int pg_nents;
> +
> +		pg_nents = intel_copy_sg(obj_st, st, offset, length, true);
> +		if (nents < pg_nents)
> +			nents = pg_nents;
> +	}
> +
> +	ret = sg_alloc_table(st, nents, GFP_KERNEL);
> +	if (ret)
> +		goto err_sg_alloc;
> +
> +	/* Build sg_table for specified <offset, length> section */
> +	intel_copy_dma_sg(obj_st, st, offset, length, false);
> +	if (i915_gem_object_has_struct_page(obj))
> +		intel_copy_sg(obj_st, st, offset, length, false);
> +
> +	return st;
> +
> +err_sg_alloc:
> +	kfree(st);
> +err_st_alloc:
> +	return ERR_PTR(ret);
> +}
> +
>   static int
>   __i915_vma_get_pages(struct i915_vma *vma)
>   {
> @@ -1360,7 +1467,11 @@ __i915_vma_get_pages(struct i915_vma *vma)
>   		break;
>   
>   	case I915_GTT_VIEW_PARTIAL:
> -		pages = intel_partial_pages(&vma->gtt_view, vma->obj);
> +		if (i915_vma_is_persistent(vma))
> +			pages = intel_persistent_partial_pages(&vma->gtt_view,
> +							       vma->obj);
> +		else
> +			pages = intel_partial_pages(&vma->gtt_view, vma->obj);
>   		break;
>   	}
>