drm/i915: Replace i915_vma_put_fence()

Avoid calling i915_vma_put_fence() by using our alternate paths that
bind a secondary vma avoiding the original fenced vma. For the few
instances where we need to release the fence (i.e. on binding when the
GGTT range becomes invalid), replace the put_fence with a revoke_fence.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190822061557.18402-1-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/display/intel_overlay.c

@@ -770,10 +770,6 @@ static int intel_overlay_do_put_image(struct intel_overlay *overlay,
 	}
 	intel_frontbuffer_flush(new_bo->frontbuffer, ORIGIN_DIRTYFB);
 
-	ret = i915_vma_put_fence(vma);
-	if (ret)
-		goto out_unpin;
-
 	if (!overlay->active) {
 		u32 oconfig;
 

drivers/gpu/drm/i915/gem/i915_gem_domain.c

@@ -221,6 +221,8 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 	 * state and so involves less work.
 	 */
 	if (atomic_read(&obj->bind_count)) {
+		struct drm_i915_private *i915 = to_i915(obj->base.dev);
+
 		/* Before we change the PTE, the GPU must not be accessing it.
 		 * If we wait upon the object, we know that all the bound
 		 * VMA are no longer active.
@@ -232,18 +234,30 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 		if (ret)
 			return ret;
 
-		if (!HAS_LLC(to_i915(obj->base.dev)) &&
-		    cache_level != I915_CACHE_NONE) {
-			/* Access to snoopable pages through the GTT is
+		if (!HAS_LLC(i915) && cache_level != I915_CACHE_NONE) {
+			intel_wakeref_t wakeref =
+				intel_runtime_pm_get(&i915->runtime_pm);
+
+			/*
+			 * Access to snoopable pages through the GTT is
 			 * incoherent and on some machines causes a hard
 			 * lockup. Relinquish the CPU mmaping to force
 			 * userspace to refault in the pages and we can
 			 * then double check if the GTT mapping is still
 			 * valid for that pointer access.
 			 */
-			i915_gem_object_release_mmap(obj);
+			ret = mutex_lock_interruptible(&i915->ggtt.vm.mutex);
+			if (ret) {
+				intel_runtime_pm_put(&i915->runtime_pm,
+						     wakeref);
+				return ret;
+			}
 
-			/* As we no longer need a fence for GTT access,
+			if (obj->userfault_count)
+				__i915_gem_object_release_mmap(obj);
+
+			/*
+			 * As we no longer need a fence for GTT access,
 			 * we can relinquish it now (and so prevent having
 			 * to steal a fence from someone else on the next
 			 * fence request). Note GPU activity would have
@@ -251,12 +265,17 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 			 * supposed to be linear.
 			 */
 			for_each_ggtt_vma(vma, obj) {
-				ret = i915_vma_put_fence(vma);
+				ret = i915_vma_revoke_fence(vma);
 				if (ret)
-					return ret;
+					break;
 			}
+			mutex_unlock(&i915->ggtt.vm.mutex);
+			intel_runtime_pm_put(&i915->runtime_pm, wakeref);
+			if (ret)
+				return ret;
 		} else {
-			/* We either have incoherent backing store and
+			/*
+			 * We either have incoherent backing store and
 			 * so no GTT access or the architecture is fully
 			 * coherent. In such cases, existing GTT mmaps
 			 * ignore the cache bit in the PTE and we can

drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c

@@ -1024,6 +1024,9 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj,
 		struct i915_vma *vma;
 		int err;
 
+		if (i915_gem_object_is_tiled(obj))
+			return ERR_PTR(-EINVAL);
+
 		if (use_cpu_reloc(cache, obj))
 			return NULL;
 
@@ -1047,12 +1050,6 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj,
 			if (err) /* no inactive aperture space, use cpu reloc */
 				return NULL;
 		} else {
-			err = i915_vma_put_fence(vma);
-			if (err) {
-				i915_vma_unpin(vma);
-				return ERR_PTR(err);
-			}
-
 			cache->node.start = vma->node.start;
 			cache->node.mm = (void *)vma;
 		}

drivers/gpu/drm/i915/i915_gem.c

@@ -343,20 +343,16 @@ i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
 		return ret;
 
 	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
-	vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
-				       PIN_MAPPABLE |
-				       PIN_NONBLOCK /* NOWARN */ |
-				       PIN_NOEVICT);
+	vma = ERR_PTR(-ENODEV);
+	if (!i915_gem_object_is_tiled(obj))
+		vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
+					       PIN_MAPPABLE |
+					       PIN_NONBLOCK /* NOWARN */ |
+					       PIN_NOEVICT);
 	if (!IS_ERR(vma)) {
 		node.start = i915_ggtt_offset(vma);
 		node.allocated = false;
-		ret = i915_vma_put_fence(vma);
-		if (ret) {
-			i915_vma_unpin(vma);
-			vma = ERR_PTR(ret);
-		}
-	}
-	if (IS_ERR(vma)) {
+	} else {
 		ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
 		if (ret)
 			goto out_unlock;
@@ -557,20 +553,16 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
 		wakeref = intel_runtime_pm_get(rpm);
 	}
 
-	vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
-				       PIN_MAPPABLE |
-				       PIN_NONBLOCK /* NOWARN */ |
-				       PIN_NOEVICT);
+	vma = ERR_PTR(-ENODEV);
+	if (!i915_gem_object_is_tiled(obj))
+		vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
+					       PIN_MAPPABLE |
+					       PIN_NONBLOCK /* NOWARN */ |
+					       PIN_NOEVICT);
 	if (!IS_ERR(vma)) {
 		node.start = i915_ggtt_offset(vma);
 		node.allocated = false;
-		ret = i915_vma_put_fence(vma);
-		if (ret) {
-			i915_vma_unpin(vma);
-			vma = ERR_PTR(ret);
-		}
-	}
-	if (IS_ERR(vma)) {
+	} else {
 		ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
 		if (ret)
 			goto out_rpm;

drivers/gpu/drm/i915/i915_gem_fence_reg.c

@@ -287,7 +287,7 @@ static int fence_update(struct i915_fence_reg *fence,
 }
 
 /**
- * i915_vma_put_fence - force-remove fence for a VMA
+ * i915_vma_revoke_fence - force-remove fence for a VMA
  * @vma: vma to map linearly (not through a fence reg)
  *
  * This function force-removes any fence from the given object, which is useful
@@ -297,26 +297,18 @@ static int fence_update(struct i915_fence_reg *fence,
  *
  * 0 on success, negative error code on failure.
  */
-int i915_vma_put_fence(struct i915_vma *vma)
+int i915_vma_revoke_fence(struct i915_vma *vma)
 {
-	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vma->vm);
 	struct i915_fence_reg *fence = vma->fence;
-	int err;
 
+	lockdep_assert_held(&vma->vm->mutex);
 	if (!fence)
 		return 0;
 
 	if (atomic_read(&fence->pin_count))
 		return -EBUSY;
 
-	err = mutex_lock_interruptible(&ggtt->vm.mutex);
-	if (err)
-		return err;
-
-	err = fence_update(fence, NULL);
-	mutex_unlock(&ggtt->vm.mutex);
-
-	return err;
+	return fence_update(fence, NULL);
 }
 
 static struct i915_fence_reg *fence_find(struct drm_i915_private *i915)

drivers/gpu/drm/i915/i915_vma.c

@@ -982,7 +982,9 @@ int i915_vma_unbind(struct i915_vma *vma)
 		GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
 
 		/* release the fence reg _after_ flushing */
-		ret = i915_vma_put_fence(vma);
+		mutex_lock(&vma->vm->mutex);
+		ret = i915_vma_revoke_fence(vma);
+		mutex_unlock(&vma->vm->mutex);
 		if (ret)
 			return ret;
 

drivers/gpu/drm/i915/i915_vma.h

@@ -421,8 +421,8 @@ static inline struct page *i915_vma_first_page(struct i915_vma *vma)
  *
  * True if the vma has a fence, false otherwise.
  */
-int i915_vma_pin_fence(struct i915_vma *vma);
-int __must_check i915_vma_put_fence(struct i915_vma *vma);
+int __must_check i915_vma_pin_fence(struct i915_vma *vma);
+int __must_check i915_vma_revoke_fence(struct i915_vma *vma);
 
 static inline void __i915_vma_unpin_fence(struct i915_vma *vma)
 {