[Intel-gfx] [PATCH 3/4] drm/i915: Add support for stealing purgable stolen pages

[Daniel Vetter]

On Wed, Jul 22, 2015 at 07:21:48PM +0530, ankitprasad.r.sharma at intel.com wrote:
> From: Chris Wilson <chris at chris-wilson.co.uk>
> 
> If we run out of stolen memory when trying to allocate an object, see if
> we can reap enough purgeable objects to free up enough contiguous free
> space for the allocation. This is in principle very much like evicting
> objects to free up enough contiguous space in the vma when binding
> a new object - and you will be forgiven for thinking that the code looks
> very similar.
> 
> At the moment, we do not allow userspace to allocate objects in stolen,
> so there is neither the memory pressure to trigger stolen eviction nor
> any purgeable objects inside the stolen arena. However, this will change
> in the near future, and so better management and defragmentation of
> stolen memory will become a real issue.
> 
> v2: Remember to remove the drm_mm_node.
> 
> v3: Rebased to the latest drm-intel-nightly (Ankit)
> 
> v4: correctedted if-else braces format (Tvrtko/kerneldoc)
> 
> Testcase: igt/gem_stolen
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> ---
>  drivers/gpu/drm/i915/i915_gem_stolen.c | 122 ++++++++++++++++++++++++++++++---
>  1 file changed, 111 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_gem_stolen.c b/drivers/gpu/drm/i915/i915_gem_stolen.c
> index 348ed5a..eaf0bdd 100644
> --- a/drivers/gpu/drm/i915/i915_gem_stolen.c
> +++ b/drivers/gpu/drm/i915/i915_gem_stolen.c
> @@ -430,18 +430,29 @@ cleanup:
>  	return NULL;
>  }
>  
> -struct drm_i915_gem_object *
> -i915_gem_object_create_stolen(struct drm_device *dev, u32 size)
> +static bool mark_free(struct drm_i915_gem_object *obj, struct list_head *unwind)
> +{
> +	if (obj->stolen == NULL)
> +		return false;
> +
> +	if (obj->madv != I915_MADV_DONTNEED)
> +		return false;
> +
> +	if (i915_gem_obj_is_pinned(obj))
> +		return false;
> +
> +	list_add(&obj->obj_exec_link, unwind);
> +	return drm_mm_scan_add_block(obj->stolen);
> +}
> +
> +static struct drm_mm_node *
> +stolen_alloc(struct drm_i915_private *dev_priv, u32 size)
>  {
> -	struct drm_i915_private *dev_priv = dev->dev_private;
> -	struct drm_i915_gem_object *obj;
>  	struct drm_mm_node *stolen;
> +	struct drm_i915_gem_object *obj;
> +	struct list_head unwind, evict;
>  	int ret;
>  
> -	if (!drm_mm_initialized(&dev_priv->mm.stolen))
> -		return NULL;
> -
> -	DRM_DEBUG_KMS("creating stolen object: size=%x\n", size);
>  	if (size == 0)
>  		return NULL;
>  
> @@ -451,11 +462,100 @@ i915_gem_object_create_stolen(struct drm_device *dev, u32 size)
>  
>  	ret = drm_mm_insert_node(&dev_priv->mm.stolen, stolen, size,
>  				 4096, DRM_MM_SEARCH_DEFAULT);
> -	if (ret) {
> -		kfree(stolen);
> -		return NULL;
> +	if (ret == 0)
> +		return stolen;
> +
> +	/* No more stolen memory available, or too fragmented.
> +	 * Try evicting purgeable objects and search again.
> +	 */
> +
> +	drm_mm_init_scan(&dev_priv->mm.stolen, size, 4096, 0);
> +	INIT_LIST_HEAD(&unwind);
> +
> +	list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
> +		if (mark_free(obj, &unwind))
> +			goto found;
> +
> +	list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
> +		if (mark_free(obj, &unwind))
> +			goto found;

Chris and I just discussed on irc that the bound_list isn't in a great LRU
order right now and Chris sent out a fix for that. But it only works if we
preferrentially shrink inactive objects first. Worth the bother or just a
FIXME? For the fb use-case alone it's not needed since we can't remove the
fb until it's no longer being displayed (otherwise the backwards-compat
code kicks in and synchronously kills the display at RMFB time), and that
pretty much means we can't put the underlying bo into any cache (and mark
it purgeable) either. But a FIXME comment here would be good for sure,
just in case this assumption ever gets broken.
-Daniel

> +
> +found:
> +	INIT_LIST_HEAD(&evict);
> +	while (!list_empty(&unwind)) {
> +		obj = list_first_entry(&unwind,
> +				       struct drm_i915_gem_object,
> +				       obj_exec_link);
> +		list_del_init(&obj->obj_exec_link);
> +
> +		if (drm_mm_scan_remove_block(obj->stolen)) {
> +			list_add(&obj->obj_exec_link, &evict);
> +			drm_gem_object_reference(&obj->base);
> +		}
>  	}
>  
> +	ret = 0;
> +	while (!list_empty(&evict)) {
> +		obj = list_first_entry(&evict,
> +				       struct drm_i915_gem_object,
> +				       obj_exec_link);
> +		list_del_init(&obj->obj_exec_link);
> +
> +		if (ret == 0) {
> +			struct i915_vma *vma, *vma_next;
> +
> +			list_for_each_entry_safe(vma, vma_next,
> +						 &obj->vma_list,
> +						 vma_link)
> +				if (i915_vma_unbind(vma))
> +					break;
> +
> +			/* Stolen pins its pages to prevent the
> +			 * normal shrinker from processing stolen
> +			 * objects.
> +			 */
> +			i915_gem_object_unpin_pages(obj);
> +
> +			ret = i915_gem_object_put_pages(obj);
> +			if (ret == 0) {
> +				i915_gem_object_release_stolen(obj);
> +				obj->madv = __I915_MADV_PURGED;
> +			} else {
> +				i915_gem_object_pin_pages(obj);
> +			}
> +		}
> +
> +		drm_gem_object_unreference(&obj->base);
> +	}
> +
> +	if (ret == 0)
> +		ret = drm_mm_insert_node(&dev_priv->mm.stolen, stolen, size,
> +					 4096, DRM_MM_SEARCH_DEFAULT);
> +	if (ret == 0)
> +		return stolen;
> +
> +	kfree(stolen);
> +	return NULL;
> +}
> +
> +struct drm_i915_gem_object *
> +i915_gem_object_create_stolen(struct drm_device *dev, u32 size)
> +{
> +	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct drm_i915_gem_object *obj;
> +	struct drm_mm_node *stolen;
> +
> +	lockdep_assert_held(&dev->struct_mutex);
> +
> +	if (!drm_mm_initialized(&dev_priv->mm.stolen))
> +		return NULL;
> +
> +	DRM_DEBUG_KMS("creating stolen object: size=%x\n", size);
> +
> +	stolen = stolen_alloc(dev_priv, size);
> +	if (stolen == NULL)
> +		return NULL;
> +
>  	obj = _i915_gem_object_create_stolen(dev, stolen);
>  	if (obj)
>  		return obj;
> -- 
> 1.9.1
> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch