drm/i915: Separate fence pin counting from normal bind pin counting

In order to correctly account for reserving space in the GTT and fences
for a batch buffer, we need to independently track whether the fence is
pinned due to a fenced GPU access in the batch or whether the buffer is
pinned in the aperture. Currently we count the fenced as pinned if the
buffer has already been seen in the execbuffer. This leads to a false
accounting of available fence registers, causing frequent mass evictions.
Worse, if coupled with the change to make i915_gem_object_get_fence()
report EDADLK upon fence starvation, the batchbuffer can fail with only
one fence required...

Fixes intel-gpu-tools/tests/gem_fenced_exec_thrash

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38735
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Tested-by: Paul Neumann <paul104x@yahoo.de>
[danvet: Resolve the functional conflict with Jesse Barnes sprite
patches, acked by Chris Wilson on irc.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

1 files changed, 90 insertions, 49 deletions

diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
index 49b3ebc0e7a6..4a43ef5dba31 100644
--- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
@@ -461,6 +461,54 @@ i915_gem_execbuffer_relocate(struct drm_device *dev,
 	return ret;
 }
 
+#define  __EXEC_OBJECT_HAS_FENCE (1<<31)
+
+static int
+pin_and_fence_object(struct drm_i915_gem_object *obj,
+		     struct intel_ring_buffer *ring)
+{
+	struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
+	bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
+	bool need_fence, need_mappable;
+	int ret;
+
+	need_fence =
+		has_fenced_gpu_access &&
+		entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
+		obj->tiling_mode != I915_TILING_NONE;
+	need_mappable =
+		entry->relocation_count ? true : need_fence;
+
+	ret = i915_gem_object_pin(obj, entry->alignment, need_mappable);
+	if (ret)
+		return ret;
+
+	if (has_fenced_gpu_access) {
+		if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
+			if (obj->tiling_mode) {
+				ret = i915_gem_object_get_fence(obj, ring);
+				if (ret)
+					goto err_unpin;
+
+				entry->flags |= __EXEC_OBJECT_HAS_FENCE;
+				i915_gem_object_pin_fence(obj);
+			} else {
+				ret = i915_gem_object_put_fence(obj);
+				if (ret)
+					goto err_unpin;
+			}
+		}
+		obj->pending_fenced_gpu_access = need_fence;
+	}
+
+	entry->offset = obj->gtt_offset;
+	return 0;
+
+err_unpin:
+	i915_gem_object_unpin(obj);
+	return ret;
+}
+
 static int
 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
 			    struct drm_file *file,
@@ -518,6 +566,7 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
 		list_for_each_entry(obj, objects, exec_list) {
 			struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
 			bool need_fence, need_mappable;
+
 			if (!obj->gtt_space)
 				continue;
 
@@ -532,58 +581,47 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
 			    (need_mappable && !obj->map_and_fenceable))
 				ret = i915_gem_object_unbind(obj);
 			else
-				ret = i915_gem_object_pin(obj,
-							  entry->alignment,
-							  need_mappable);
+				ret = pin_and_fence_object(obj, ring);
 			if (ret)
 				goto err;
-
-			entry++;
 		}
 
 		/* Bind fresh objects */
 		list_for_each_entry(obj, objects, exec_list) {
-			struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
-			bool need_fence;
-
-			need_fence =
-				has_fenced_gpu_access &&
-				entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
-				obj->tiling_mode != I915_TILING_NONE;
-
-			if (!obj->gtt_space) {
-				bool need_mappable =
-					entry->relocation_count ? true : need_fence;
-
-				ret = i915_gem_object_pin(obj,
-							  entry->alignment,
-							  need_mappable);
-				if (ret)
-					break;
-			}
+			if (obj->gtt_space)
+				continue;
 
-			if (has_fenced_gpu_access) {
-				if (need_fence) {
-					ret = i915_gem_object_get_fence(obj, ring);
-					if (ret)
-						break;
-				} else if (entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
-					   obj->tiling_mode == I915_TILING_NONE) {
-					/* XXX pipelined! */
-					ret = i915_gem_object_put_fence(obj);
-					if (ret)
-						break;
-				}
-				obj->pending_fenced_gpu_access = need_fence;
+			ret = pin_and_fence_object(obj, ring);
+			if (ret) {
+				int ret_ignore;
+
+				/* This can potentially raise a harmless
+				 * -EINVAL if we failed to bind in the above
+				 * call. It cannot raise -EINTR since we know
+				 * that the bo is freshly bound and so will
+				 * not need to be flushed or waited upon.
+				 */
+				ret_ignore = i915_gem_object_unbind(obj);
+				(void)ret_ignore;
+				WARN_ON(obj->gtt_space);
+				break;
 			}
-
-			entry->offset = obj->gtt_offset;
 		}
 
 		/* Decrement pin count for bound objects */
 		list_for_each_entry(obj, objects, exec_list) {
-			if (obj->gtt_space)
-				i915_gem_object_unpin(obj);
+			struct drm_i915_gem_exec_object2 *entry;
+
+			if (!obj->gtt_space)
+				continue;
+
+			entry = obj->exec_entry;
+			if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
+				i915_gem_object_unpin_fence(obj);
+				entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
+			}
+
+			i915_gem_object_unpin(obj);
 		}
 
 		if (ret != -ENOSPC || retry > 1)
@@ -600,16 +638,19 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
 	} while (1);
 
 err:
-	obj = list_entry(obj->exec_list.prev,
-			 struct drm_i915_gem_object,
-			 exec_list);
-	while (objects != &obj->exec_list) {
-		if (obj->gtt_space)
-			i915_gem_object_unpin(obj);
+	list_for_each_entry_continue_reverse(obj, objects, exec_list) {
+		struct drm_i915_gem_exec_object2 *entry;
+
+		if (!obj->gtt_space)
+			continue;
+
+		entry = obj->exec_entry;
+		if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
+			i915_gem_object_unpin_fence(obj);
+			entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
+		}
 
-		obj = list_entry(obj->exec_list.prev,
-				 struct drm_i915_gem_object,
-				 exec_list);
+		i915_gem_object_unpin(obj);
 	}
 
 	return ret;