[Intel-gfx] [PATCH v2 3/3] drm/i915: Use page coloring to provide the guard page at the end of the GTT

[Matthew Auld]

On 6 February 2017 at 08:45, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> As we now mark the reserved hole (drm_mm.head_node) with the special
> UNEVICTABLE color, we can use the page coloring to avoid prefetching of
> the CS beyond the end of the GTT.
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
>  drivers/gpu/drm/i915/i915_gem_evict.c | 10 ++++++++--
>  drivers/gpu/drm/i915/i915_gem_gtt.c   | 19 ++++++++++++-------
>  2 files changed, 20 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c
> index c181b1bb3d2c..776c508cb9ae 100644
> --- a/drivers/gpu/drm/i915/i915_gem_evict.c
> +++ b/drivers/gpu/drm/i915/i915_gem_evict.c
> @@ -258,6 +258,9 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
>         int ret = 0;
>
>         lockdep_assert_held(&vm->i915->drm.struct_mutex);
> +       GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
> +       GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
> +
>         trace_i915_gem_evict_node(vm, target, flags);
>
>         /* Retire before we search the active list. Although we have
> @@ -273,9 +276,11 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
>                 /* Expand search to cover neighbouring guard pages (or lack!) */
>                 if (start > vm->start)
>                         start -= I915_GTT_PAGE_SIZE;
> -               if (end < vm->start + vm->total)
> -                       end += I915_GTT_PAGE_SIZE;
> +
> +               /* Always look at the page afterwards to avoid the end-of-GTT */
> +               end += I915_GTT_PAGE_SIZE;
>         }
> +       GEM_BUG_ON(start >= end);
>
>         drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
>                 /* If we find any non-objects (!vma), we cannot evict them */
> @@ -284,6 +289,7 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
>                         break;
>                 }
>
> +               GEM_BUG_ON(!node->allocated);
>                 vma = container_of(node, typeof(*vma), node);
>
>                 /* If we are using coloring to insert guard pages between
> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
> index 75cd2a56b8b4..959764394718 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.c
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
> @@ -2718,11 +2718,16 @@ static void i915_gtt_color_adjust(const struct drm_mm_node *node,
>                                   u64 *start,
>                                   u64 *end)
>  {
> -       if (node->color != color)
> +       if (node->allocated && node->color != color)
>                 *start += I915_GTT_PAGE_SIZE;
>
> +       /* Also leave a space between the unallocated reserved node after the
> +        * GTT and any objects within the GTT, i.e. we use the color adjustment
> +        * to insert a guard page to prevent prefetches crossing over the
> +        * GTT boundary.
> +        */
>         node = list_next_entry(node, node_list);
> -       if (node->allocated && node->color != color)
> +       if (node->color != color)
>                 *end -= I915_GTT_PAGE_SIZE;
>  }
>
> @@ -3241,14 +3246,14 @@ int i915_ggtt_init_hw(struct drm_i915_private *dev_priv)
>
>         INIT_LIST_HEAD(&dev_priv->vm_list);
>
> -       /* Subtract the guard page before address space initialization to
> -        * shrink the range used by drm_mm.
> +       /* Note that we use page colouring to enforce a guard page at the
> +        * end of the address space. This is required as the CS may prefetch
> +        * beyond the end of the batch buffer, across the page boundary,
> +        * and beyond the end of the GTT if we do provide a guard.
if we do not provide a guard?

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