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>

6 files changed, 129 insertions, 57 deletions

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 733f5f57bab..12e8cce7928 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -135,6 +135,7 @@ struct drm_i915_fence_reg {
 	struct list_head lru_list;
 	struct drm_i915_gem_object *obj;
 	uint32_t setup_seqno;
+	int pin_count;
 };
 
 struct sdvo_device_mapping {
@@ -1159,6 +1160,24 @@ int __must_check i915_gem_object_get_fence(struct drm_i915_gem_object *obj,
 					   struct intel_ring_buffer *pipelined);
 int __must_check i915_gem_object_put_fence(struct drm_i915_gem_object *obj);
 
+static inline void
+i915_gem_object_pin_fence(struct drm_i915_gem_object *obj)
+{
+	if (obj->fence_reg != I915_FENCE_REG_NONE) {
+		struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
+		dev_priv->fence_regs[obj->fence_reg].pin_count++;
+	}
+}
+
+static inline void
+i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj)
+{
+	if (obj->fence_reg != I915_FENCE_REG_NONE) {
+		struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
+		dev_priv->fence_regs[obj->fence_reg].pin_count--;
+	}
+}
+
 void i915_gem_retire_requests(struct drm_device *dev);
 void i915_gem_reset(struct drm_device *dev);
 void i915_gem_clflush_object(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 ff3066c4c76..c78930ed2e8 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2435,6 +2435,8 @@ i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
 
 	if (obj->fence_reg != I915_FENCE_REG_NONE) {
 		struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
+
+		WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count);
 		i915_gem_clear_fence_reg(obj->base.dev,
 					 &dev_priv->fence_regs[obj->fence_reg]);
 
@@ -2459,7 +2461,7 @@ i915_find_fence_reg(struct drm_device *dev,
 		if (!reg->obj)
 			return reg;
 
-		if (!reg->obj->pin_count)
+		if (!reg->pin_count)
 			avail = reg;
 	}
 
@@ -2469,7 +2471,7 @@ i915_find_fence_reg(struct drm_device *dev,
 	/* None available, try to steal one or wait for a user to finish */
 	avail = first = NULL;
 	list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
-		if (reg->obj->pin_count)
+		if (reg->pin_count)
 			continue;
 
 		if (first == NULL)
@@ -2664,6 +2666,7 @@ i915_gem_clear_fence_reg(struct drm_device *dev,
 	list_del_init(&reg->lru_list);
 	reg->obj = NULL;
 	reg->setup_seqno = 0;
+	reg->pin_count = 0;
 }
 
 /**
diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
index 49b3ebc0e7a..4a43ef5dba3 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;
diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
index 0770671fa8a..fc9bc19f6db 100644
--- a/drivers/gpu/drm/i915/intel_display.c
+++ b/drivers/gpu/drm/i915/intel_display.c
@@ -2041,6 +2041,8 @@ intel_pin_and_fence_fb_obj(struct drm_device *dev,
 		ret = i915_gem_object_get_fence(obj, pipelined);
 		if (ret)
 			goto err_unpin;
+
+		i915_gem_object_pin_fence(obj);
 	}
 
 	dev_priv->mm.interruptible = true;
@@ -2053,6 +2055,12 @@ err_interruptible:
 	return ret;
 }
 
+void intel_unpin_fb_obj(struct drm_i915_gem_object *obj)
+{
+	i915_gem_object_unpin_fence(obj);
+	i915_gem_object_unpin(obj);
+}
+
 static int i9xx_update_plane(struct drm_crtc *crtc, struct drm_framebuffer *fb,
 			     int x, int y)
 {
@@ -2284,7 +2292,7 @@ intel_pipe_set_base(struct drm_crtc *crtc, int x, int y,
 	ret = intel_pipe_set_base_atomic(crtc, crtc->fb, x, y,
 					 LEAVE_ATOMIC_MODE_SET);
 	if (ret) {
-		i915_gem_object_unpin(to_intel_framebuffer(crtc->fb)->obj);
+		intel_unpin_fb_obj(to_intel_framebuffer(crtc->fb)->obj);
 		mutex_unlock(&dev->struct_mutex);
 		DRM_ERROR("failed to update base address\n");
 		return ret;
@@ -2292,7 +2300,7 @@ intel_pipe_set_base(struct drm_crtc *crtc, int x, int y,
 
 	if (old_fb) {
 		intel_wait_for_vblank(dev, intel_crtc->pipe);
-		i915_gem_object_unpin(to_intel_framebuffer(old_fb)->obj);
+		intel_unpin_fb_obj(to_intel_framebuffer(old_fb)->obj);
 	}
 
 	mutex_unlock(&dev->struct_mutex);
@@ -3355,7 +3363,7 @@ static void intel_crtc_disable(struct drm_crtc *crtc)
 
 	if (crtc->fb) {
 		mutex_lock(&dev->struct_mutex);
-		i915_gem_object_unpin(to_intel_framebuffer(crtc->fb)->obj);
+		intel_unpin_fb_obj(to_intel_framebuffer(crtc->fb)->obj);
 		mutex_unlock(&dev->struct_mutex);
 	}
 }
@@ -7158,7 +7166,7 @@ static void intel_unpin_work_fn(struct work_struct *__work)
 		container_of(__work, struct intel_unpin_work, work);
 
 	mutex_lock(&work->dev->struct_mutex);
-	i915_gem_object_unpin(work->old_fb_obj);
+	intel_unpin_fb_obj(work->old_fb_obj);
 	drm_gem_object_unreference(&work->pending_flip_obj->base);
 	drm_gem_object_unreference(&work->old_fb_obj->base);
 
diff --git a/drivers/gpu/drm/i915/intel_drv.h b/drivers/gpu/drm/i915/intel_drv.h
index 1348705faf6..9cec6c3937f 100644
--- a/drivers/gpu/drm/i915/intel_drv.h
+++ b/drivers/gpu/drm/i915/intel_drv.h
@@ -374,6 +374,7 @@ extern void intel_init_emon(struct drm_device *dev);
 extern int intel_pin_and_fence_fb_obj(struct drm_device *dev,
 				      struct drm_i915_gem_object *obj,
 				      struct intel_ring_buffer *pipelined);
+extern void intel_unpin_fb_obj(struct drm_i915_gem_object *obj);
 
 extern int intel_framebuffer_init(struct drm_device *dev,
 				  struct intel_framebuffer *ifb,
diff --git a/drivers/gpu/drm/i915/intel_sprite.c b/drivers/gpu/drm/i915/intel_sprite.c
index d13989fda50..ad3bd929aec 100644
--- a/drivers/gpu/drm/i915/intel_sprite.c
+++ b/drivers/gpu/drm/i915/intel_sprite.c
@@ -503,7 +503,7 @@ intel_update_plane(struct drm_plane *plane, struct drm_crtc *crtc,
 			intel_wait_for_vblank(dev, to_intel_crtc(crtc)->pipe);
 			mutex_lock(&dev->struct_mutex);
 		}
-		i915_gem_object_unpin(old_obj);
+		intel_unpin_fb_obj(old_obj);
 	}
 
 out_unlock:
@@ -530,7 +530,7 @@ intel_disable_plane(struct drm_plane *plane)
 		goto out;
 
 	mutex_lock(&dev->struct_mutex);
-	i915_gem_object_unpin(intel_plane->obj);
+	intel_unpin_fb_obj(intel_plane->obj);
 	intel_plane->obj = NULL;
 	mutex_unlock(&dev->struct_mutex);
 out: