[Intel-gfx] [PATCH] drm/i915/evict: watch out for unevictable nodes

[Chris Wilson]

Quoting Matthew Auld (2020-04-08 18:04:56)
> In an address space there can be sprinkling of I915_COLOR_UNEVICTABLE
> nodes, which lack a parent vma. For platforms with cache coloring we
> might be very unlucky and abut with such a node thinking we can simply
> unbind the vma.

I did notice this myself recently (from observation); it's highly
unlikely to ever matter.

> Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> ---
>  drivers/gpu/drm/i915/i915_gem_evict.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c
> index 4518b9b35c3d..9e462c6a4c6a 100644
> --- a/drivers/gpu/drm/i915/i915_gem_evict.c
> +++ b/drivers/gpu/drm/i915/i915_gem_evict.c
> @@ -227,6 +227,10 @@ i915_gem_evict_something(struct i915_address_space *vm,
>         }
>  
>         while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
> +               /* If we find any non-objects (!vma), we cannot evict them */
> +               if (node->color == I915_COLOR_UNEVICTABLE)
> +                       return -ENOSPC;

Returning error immediately looks ok, I was worried as around here we
usually have lists to clean up, but this is after those. However, I
would suggest that setting ret = -ENOSPC would be more consistent with
the flow.

Fwiw, we can actually pull this logic into evict_for_node for a bit of
simplification.
-Chris