[Intel-gfx] [PATCH 47/64] drm/i915: Be more careful when unbinding vma

[Tvrtko Ursulin]

On 07/07/16 09:41, Chris Wilson wrote:
> When we call i915_vma_unbind(), we will wait upon outstanding rendering.
> This will also trigger a retirement phase, which may update the object
> lists. If, we extend request tracking to the VMA itself (rather than
> keep it at the encompassing object), then there is a potential that the
> obj->vma_list be modified for other elements upon i915_vma_unbind(). As
> a result, if we walk over the object list and call i915_vma_unbind(), we
> need to be prepared for that list to change.
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
>   drivers/gpu/drm/i915/i915_drv.h          |  2 ++
>   drivers/gpu/drm/i915/i915_gem.c          | 57 +++++++++++++++++++++++---------
>   drivers/gpu/drm/i915/i915_gem_shrinker.c |  7 +---
>   drivers/gpu/drm/i915/i915_gem_userptr.c  |  4 +--
>   4 files changed, 46 insertions(+), 24 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index 633585054669..27e1182544a2 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -3032,6 +3032,8 @@ int __must_check i915_vma_unbind(struct i915_vma *vma);
>    * _guarantee_ VMA in question is _not in use_ anywhere.
>    */
>   int __must_check __i915_vma_unbind_no_wait(struct i915_vma *vma);
> +
> +int i915_gem_object_unbind(struct drm_i915_gem_object *obj);
>   int i915_gem_object_put_pages(struct drm_i915_gem_object *obj);
>   void i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv);
>   void i915_gem_release_mmap(struct drm_i915_gem_object *obj);
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index c6816f9969d5..28a3079a7892 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -283,18 +283,38 @@ static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
>   	.release = i915_gem_object_release_phys,
>   };
>
> +int
> +i915_gem_object_unbind(struct drm_i915_gem_object *obj)
> +{
> +	struct i915_vma *vma;
> +	LIST_HEAD(still_in_list);
> +	int ret;
> +
> +	/* The vma will only be freed if it is marked as closed, and if we wait
> +	 * upon rendering to the vma, we may unbind anything in the list.
> +	 */
> +	while ((vma = list_first_entry_or_null(&obj->vma_list,
> +					       struct i915_vma,
> +					       obj_link))) {
> +		list_move_tail(&vma->obj_link, &still_in_list);
> +		ret = i915_vma_unbind(vma);
> +		if (ret)
> +			break;
> +	}
> +	list_splice(&still_in_list, &obj->vma_list);
> +
> +	return ret;
> +}
> +
>   static int
>   drop_pages(struct drm_i915_gem_object *obj)
>   {
> -	struct i915_vma *vma, *next;
>   	int ret;
>
>   	i915_gem_object_get(obj);
> -	list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link)
> -		if (i915_vma_unbind(vma))
> -			break;
> -
> -	ret = i915_gem_object_put_pages(obj);
> +	ret = i915_gem_object_unbind(obj);
> +	if (ret == 0)
> +		ret = i915_gem_object_put_pages(obj);
>   	i915_gem_object_put(obj);
>
>   	return ret;
> @@ -3442,8 +3462,7 @@ i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
>   int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
>   				    enum i915_cache_level cache_level)
>   {
> -	struct drm_device *dev = obj->base.dev;
> -	struct i915_vma *vma, *next;
> +	struct i915_vma *vma;
>   	int ret = 0;
>
>   	if (obj->cache_level == cache_level)
> @@ -3454,7 +3473,8 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
>   	 * catch the issue of the CS prefetch crossing page boundaries and
>   	 * reading an invalid PTE on older architectures.
>   	 */
> -	list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link) {
> +restart:
> +	list_for_each_entry(vma, &obj->vma_list, obj_link) {
>   		if (!drm_mm_node_allocated(&vma->node))
>   			continue;
>
> @@ -3463,11 +3483,18 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
>   			return -EBUSY;
>   		}
>
> -		if (!i915_gem_valid_gtt_space(vma, cache_level)) {
> -			ret = i915_vma_unbind(vma);
> -			if (ret)
> -				return ret;
> -		}
> +		if (i915_gem_valid_gtt_space(vma, cache_level))
> +			continue;
> +
> +		ret = i915_vma_unbind(vma);
> +		if (ret)
> +			return ret;
> +
> +		/* As unbinding may affect other elements in the
> +		 * obj->vma_list (due to side-effects from retiring
> +		 * an active vma), play safe and restart the iterator.
> +		 */
> +		goto restart;
>   	}

Does not look efficient for long lists but I don't see a solution right 
now. Any chance of this O(N^2) iteration hurting us in the real world?

>
>   	/* We can reuse the existing drm_mm nodes but need to change the
> @@ -3486,7 +3513,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
>   		if (ret)
>   			return ret;
>
> -		if (!HAS_LLC(dev) && cache_level != I915_CACHE_NONE) {
> +		if (!HAS_LLC(obj->base.dev) && cache_level != I915_CACHE_NONE) {
>   			/* Access to snoopable pages through the GTT is
>   			 * incoherent and on some machines causes a hard
>   			 * lockup. Relinquish the CPU mmaping to force
> diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> index 725a8c894517..8ad95695f80c 100644
> --- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
> +++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> @@ -155,7 +155,6 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
>   		INIT_LIST_HEAD(&still_in_list);
>   		while (count < target && !list_empty(phase->list)) {
>   			struct drm_i915_gem_object *obj;
> -			struct i915_vma *vma, *v;
>
>   			obj = list_first_entry(phase->list,
>   					       typeof(*obj), global_list);
> @@ -178,11 +177,7 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
>   			i915_gem_object_get(obj);
>
>   			/* For the unbound phase, this should be a no-op! */
> -			list_for_each_entry_safe(vma, v,
> -						 &obj->vma_list, obj_link)
> -				if (i915_vma_unbind(vma))
> -					break;
> -
> +			i915_gem_object_unbind(obj);
>   			if (i915_gem_object_put_pages(obj) == 0)
>   				count += obj->base.size >> PAGE_SHIFT;
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_userptr.c b/drivers/gpu/drm/i915/i915_gem_userptr.c
> index ca8b82ab93d6..e935b327f3f9 100644
> --- a/drivers/gpu/drm/i915/i915_gem_userptr.c
> +++ b/drivers/gpu/drm/i915/i915_gem_userptr.c
> @@ -104,7 +104,6 @@ static void cancel_userptr(struct work_struct *work)
>
>   	if (obj->pages != NULL) {
>   		struct drm_i915_private *dev_priv = to_i915(dev);
> -		struct i915_vma *vma, *tmp;
>   		bool was_interruptible;
>
>   		wait_rendering(obj);
> @@ -112,8 +111,7 @@ static void cancel_userptr(struct work_struct *work)
>   		was_interruptible = dev_priv->mm.interruptible;
>   		dev_priv->mm.interruptible = false;
>
> -		list_for_each_entry_safe(vma, tmp, &obj->vma_list, obj_link)
> -			WARN_ON(i915_vma_unbind(vma));
> +		WARN_ON(i915_gem_object_unbind(obj));
>   		WARN_ON(i915_gem_object_put_pages(obj));
>
>   		dev_priv->mm.interruptible = was_interruptible;
>

Regards,

Tvrtko