On Wed, Feb 12, 2014 at 07:18:40PM +0000, Chris Wilson wrote:
> For stolen pages, since it is verboten to access them directly on many
> architectures, we have to read them through the GTT aperture. If they
> are not accessible through the aperture, then we have to abort.
>
> This was complicated by
>
> commit 8b6124a633d8095b0c8364f585edff9c59568a96
> Author: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> Date: Thu Jan 30 14:38:16 2014 +0000
>
> drm/i915: Don't access snooped pages through the GTT (even for error capture)
>
> and the desire to use stolen memory for ringbuffers, contexts and
> batches in the future.
I am somewhat unclear as to whether we want to prefer the aperture for
reading back objects which may be mapped in multiple address spaces.
>
> Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> ---
> drivers/gpu/drm/i915/i915_gpu_error.c | 50 ++++++++++++++++++++++-------------
> 1 file changed, 31 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gpu_error.c b/drivers/gpu/drm/i915/i915_gpu_error.c
> index 0e1f7b691082..a2c3a639c3cc 100644
> --- a/drivers/gpu/drm/i915/i915_gpu_error.c
> +++ b/drivers/gpu/drm/i915/i915_gpu_error.c
> @@ -542,10 +542,11 @@ static struct drm_i915_error_object *
> i915_error_object_create_sized(struct drm_i915_private *dev_priv,
> struct drm_i915_gem_object *src,
> struct i915_address_space *vm,
> - const int num_pages)
> + int num_pages)
> {
> struct drm_i915_error_object *dst;
> - int i;
> + bool use_ggtt;
bool use_aperture;
> + int i = 0;
> u32 reloc_offset;
>
> if (src == NULL || src->pages == NULL)
> @@ -555,8 +556,32 @@ i915_error_object_create_sized(struct drm_i915_private *dev_priv,
> if (dst == NULL)
> return NULL;
>
> - reloc_offset = dst->gtt_offset = i915_gem_obj_offset(src, vm);
> - for (i = 0; i < num_pages; i++) {
> + dst->gtt_offset = i915_gem_obj_offset(src, vm);
> +
> + reloc_offset = dst->gtt_offset;
> + use_ggtt = (src->cache_level == I915_CACHE_NONE &&
> + i915_is_ggtt(vm) &&
> + src->has_global_gtt_mapping &&
> + reloc_offset + num_pages * PAGE_SIZE <= dev_priv->gtt.mappable_end);
> +
You should probably do this as two patches. Pull out the use_aperture
variable first, followed by the stolen check.
"drm/i915: Don't check each page on error object capture"
> + /* Cannot access stolen address directly, try to use the aperture */
> + if (src->stolen) {
> + use_ggtt = true;
> +
> + if (!src->has_global_gtt_mapping)
This is BUG() or at least WARN() isn't it? Hmm, is there some
possibility we get here after an object is unmapped but before it's
taken away?
> + goto unwind;
> +
> + reloc_offset = i915_gem_obj_ggtt_offset(src);
> + if (reloc_offset + num_pages * PAGE_SIZE > dev_priv->gtt.mappable_end)
This could probably be a separate patch since it changes the old behavior
that may get some pages from CPU, and others from aperture for a given
object.
> + goto unwind;
> + }
> +
> + /* Cannot access snooped pages through the aperture */
> + if (use_ggtt && src->cache_level != I915_CACHE_NONE && !HAS_LLC(dev_priv->dev))
> + goto unwind;
> +
Some nitpicks from the previous code maybe worth fixing in the future:
/* Prefer the aperture? */
use_aperture = src->has_global_gtt_mapping &&
i915_gem_obj_ggtt_offset(src) + num_pages * PAGE_SIZE <= dev_priv->gtt.mappable_end)
if (src->stolen && !use_aperture)
goto unwind;
if (use_aperture)
reloc_offset = i915_gem_obj_ggtt_offset(src)
else
reloc_offset = dst->gtt_offset;
I don't really understand why the current code has
use_ggtt = (src->cache_level == I915_CACHE_NONE...
That seems irrelevant to me for this decision.
> + dst->page_count = num_pages;
> + while (num_pages--) {
> unsigned long flags;
> void *d;
>
> @@ -565,10 +590,7 @@ i915_error_object_create_sized(struct drm_i915_private *dev_priv,
> goto unwind;
>
> local_irq_save(flags);
> - if (src->cache_level == I915_CACHE_NONE &&
> - reloc_offset < dev_priv->gtt.mappable_end &&
> - src->has_global_gtt_mapping &&
> - i915_is_ggtt(vm)) {
> + if (use_ggtt) {
> void __iomem *s;
>
> /* Simply ignore tiling or any overlapping fence.
> @@ -580,14 +602,6 @@ i915_error_object_create_sized(struct drm_i915_private *dev_priv,
> reloc_offset);
> memcpy_fromio(d, s, PAGE_SIZE);
> io_mapping_unmap_atomic(s);
> - } else if (src->stolen) {
> - unsigned long offset;
> -
> - offset = dev_priv->mm.stolen_base;
> - offset += src->stolen->start;
> - offset += i << PAGE_SHIFT;
> -
> - memcpy_fromio(d, (void __iomem *) offset, PAGE_SIZE);
> } else {
> struct page *page;
> void *s;
> @@ -604,11 +618,9 @@ i915_error_object_create_sized(struct drm_i915_private *dev_priv,
> }
> local_irq_restore(flags);
>
> - dst->pages[i] = d;
> -
> + dst->pages[i++] = d;
You could capture the object backwards and get rid of i if you felt like
a good time.
> reloc_offset += PAGE_SIZE;
> }
> - dst->page_count = num_pages;
>
> return dst;
>
I couldn't spot anything actually wrong, but I really do want use_ggtt
changed to something better.
With that
Reviewed-by: Ben Widawsky <ben@xxxxxxxxxxxx>
If you fix any of my other suggestions, I will re-review (please CC me).
--
Ben Widawsky, Intel Open Source Technology Center
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
