[Intel-gfx] [PATCH v2 5/5] drm/i915: Move fence cancellation to runtime suspend

[Chris Wilson]

At the moment, we have dependency on the RPM as a barrier itself in both
i915_gem_release_all_mmaps() and i915_gem_restore_fences().
i915_gem_restore_fences() is also called along !runtime pm paths, but we
can move the markup of lost fences alongside releasing the mmaps into a
common i915_gem_runtime_suspend(). This has the advantage of locating
all the tricky barrier dependencies into one location.

Suggested-by: Daniel Vetter <daniel.vetter at ffwll.ch>
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
---
 drivers/gpu/drm/i915/i915_drv.c       |  6 ++----
 drivers/gpu/drm/i915/i915_drv.h       |  3 ++-
 drivers/gpu/drm/i915/i915_gem.c       | 25 +++++++++++++++++++++++--
 drivers/gpu/drm/i915/i915_gem_fence.c | 12 +++++-------
 4 files changed, 32 insertions(+), 14 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index d08d6617e3b6..6473bf58fbe5 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -2267,10 +2267,8 @@ static int vlv_resume_prepare(struct drm_i915_private *dev_priv,
 
 	vlv_check_no_gt_access(dev_priv);
 
-	if (rpm_resume) {
+	if (rpm_resume)
 		intel_init_clock_gating(dev);
-		i915_gem_restore_fences(dev);
-	}
 
 	return ret;
 }
@@ -2296,7 +2294,7 @@ static int intel_runtime_suspend(struct device *kdev)
 	 * We are safe here against re-faults, since the fault handler takes
 	 * an RPM reference.
 	 */
-	i915_gem_release_all_mmaps(dev_priv);
+	i915_gem_runtime_suspend(dev_priv);
 
 	intel_guc_suspend(dev);
 
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 7c686145861b..cc703c38b397 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -3125,9 +3125,10 @@ void i915_vma_destroy(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);
 
+void i915_gem_runtime_suspend(struct drm_i915_private *dev_priv);
+
 int __must_check i915_gem_object_get_pages(struct drm_i915_gem_object *obj);
 
 static inline int __sg_page_count(struct scatterlist *sg)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 8819fafbcc7a..d904e5f5d146 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -1954,10 +1954,10 @@ out:
 	intel_runtime_pm_put(i915);
 }
 
-void
-i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
+void i915_gem_runtime_suspend(struct drm_i915_private *dev_priv)
 {
 	struct drm_i915_gem_object *obj, *on;
+	int i;
 
 	/*
 	 * Only called during RPM suspend. All users of the userfault_list
@@ -1972,6 +1972,27 @@ i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
 		drm_vma_node_unmap(&obj->base.vma_node,
 				   obj->base.dev->anon_inode->i_mapping);
 	}
+
+	/* The fence will be lost when the device powers down. If any were
+	 * in use by hardware (i.e. they are pinned), we should not be powering
+	 * down! All other fences will be reacquired by the user upon waking.
+	 */
+	for (i = 0; i < dev_priv->num_fence_regs; i++) {
+		struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
+		struct i915_vma *vma;
+
+		if (WARN_ON(reg->pin_count))
+			continue;
+
+		vma = fetch_and_zero(&reg->vma);
+		if (!vma)
+			continue;
+
+		GEM_BUG_ON(!list_empty(&vma->obj->userfault_link));
+
+		list_move(&reg->link, &dev_priv->mm.fence_list);
+		vma->fence = NULL;
+	}
 }
 
 /**
diff --git a/drivers/gpu/drm/i915/i915_gem_fence.c b/drivers/gpu/drm/i915/i915_gem_fence.c
index f7081f4b5d22..230edc821823 100644
--- a/drivers/gpu/drm/i915/i915_gem_fence.c
+++ b/drivers/gpu/drm/i915/i915_gem_fence.c
@@ -343,6 +343,9 @@ i915_vma_get_fence(struct i915_vma *vma)
 	struct drm_i915_fence_reg *fence;
 	struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL;
 
+	/* Note that we revoke fences on runtime suspend. Therefore the user
+	 * must keep the device awake whilst using the fence.
+	 */
 	assert_rpm_wakelock_held(to_i915(vma->vm->dev));
 
 	/* Just update our place in the LRU if our fence is getting reused. */
@@ -368,19 +371,14 @@ i915_vma_get_fence(struct i915_vma *vma)
  * @dev: DRM device
  *
  * Restore the hw fence state to match the software tracking again, to be called
- * after a gpu reset and on resume.
+ * after a gpu reset and on resume. Note that on runtime suspend we only cancel
+ * the fences, to be reacquired by the user later.
  */
 void i915_gem_restore_fences(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = to_i915(dev);
 	int i;
 
-	/* Note that this may be called outside of struct_mutex, by
-	 * runtime suspend/resume. The barrier we require is enforced by
-	 * rpm itself - all access to fences/GTT are only within an rpm
-	 * wakeref, and to acquire that wakeref you must pass through here.
-	 */
-
 	for (i = 0; i < dev_priv->num_fence_regs; i++) {
 		struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
 		struct i915_vma *vma = reg->vma;
-- 
2.9.3