drm/i915: Drop i915_gem_obj_is_pinned() from set-cache-level

Since the remove of the pin-ioctl, we only care about not changing the
cache level on buffers pinned to the hardware as indicated by
obj->pin_display. By knowing that only objects pinned to the hardware
will have an elevated vma->pin_count, so we can coallesce many of the
linear walks over the obj->vma_list.

v2: Try and retrospectively add comments explaining the steps in
rebinding the active VMA.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drivers/gpu/drm/i915/i915_gem.c

@@ -3657,50 +3657,103 @@ i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
 	return 0;
 }
 
+/**
+ * Changes the cache-level of an object across all VMA.
+ *
+ * After this function returns, the object will be in the new cache-level
+ * across all GTT and the contents of the backing storage will be coherent,
+ * with respect to the new cache-level. In order to keep the backing storage
+ * coherent for all users, we only allow a single cache level to be set
+ * globally on the object and prevent it from being changed whilst the
+ * hardware is reading from the object. That is if the object is currently
+ * on the scanout it will be set to uncached (or equivalent display
+ * cache coherency) and all non-MOCS GPU access will also be uncached so
+ * that all direct access to the scanout remains coherent.
+ */
 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;
+	bool bound = false;
 	int ret = 0;
 
 	if (obj->cache_level == cache_level)
 		goto out;
 
-	if (i915_gem_obj_is_pinned(obj)) {
+	/* Inspect the list of currently bound VMA and unbind any that would
+	 * be invalid given the new cache-level. This is principally to
+	 * 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, vma_link) {
+		if (!drm_mm_node_allocated(&vma->node))
+			continue;
+
+		if (vma->pin_count) {
 			DRM_DEBUG("can not change the cache level of pinned objects\n");
 			return -EBUSY;
 		}
 
-	list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
 		if (!i915_gem_valid_gtt_space(vma, cache_level)) {
 			ret = i915_vma_unbind(vma);
 			if (ret)
 				return ret;
-		}
+		} else
+			bound = true;
 	}
 
-	if (i915_gem_obj_bound_any(obj)) {
+	/* We can reuse the existing drm_mm nodes but need to change the
+	 * cache-level on the PTE. We could simply unbind them all and
+	 * rebind with the correct cache-level on next use. However since
+	 * we already have a valid slot, dma mapping, pages etc, we may as
+	 * rewrite the PTE in the belief that doing so tramples upon less
+	 * state and so involves less work.
+	 */
+	if (bound) {
+		/* 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.
+		 */
 		ret = i915_gem_object_wait_rendering(obj, false);
 		if (ret)
 			return ret;
 
-		i915_gem_object_finish_gtt(obj);
+		if (!HAS_LLC(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
+			 * 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_release_mmap(obj);
 
-		/* Before SandyBridge, you could not use tiling or fence
-		 * registers with snooped memory, so relinquish any fences
-		 * currently pointing to our region in the aperture.
+			/* 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
+			 * dropped the fence as all snoopable access is
+			 * supposed to be linear.
 			 */
-		if (INTEL_INFO(dev)->gen < 6) {
 			ret = i915_gem_object_put_fence(obj);
 			if (ret)
 				return ret;
+		} else {
+			/* 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
+			 * rewrite it without confusing the GPU or having
+			 * to force userspace to fault back in its mmaps.
+			 */
 		}
 
-		list_for_each_entry(vma, &obj->vma_list, vma_link)
-			if (drm_mm_node_allocated(&vma->node)) {
-				ret = i915_vma_bind(vma, cache_level,
-						    PIN_UPDATE);
+		list_for_each_entry(vma, &obj->vma_list, vma_link) {
+			if (!drm_mm_node_allocated(&vma->node))
+				continue;
+
+			ret = i915_vma_bind(vma, cache_level, PIN_UPDATE);
 			if (ret)
 				return ret;
 		}
@@ -3711,6 +3764,10 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 	obj->cache_level = cache_level;
 
 out:
+	/* Flush the dirty CPU caches to the backing storage so that the
+	 * object is now coherent at its new cache level (with respect
+	 * to the access domain).
+	 */
 	if (obj->cache_dirty &&
 	    obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
 	    cpu_write_needs_clflush(obj)) {