[Intel-gfx] [PATCH] drm/i915: Skip waking the device to service pwrite

[Chris Wilson]

If the device is in runtime suspend, resuming takes time and reduces our
powersaving. If this was for a small write into an object, that resume
will take longer than any savings in using the indirect GGTT access to
avoid the cpu cache.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_gem.c | 21 ++++++++++++++++++---
 1 file changed, 18 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 93dfa793975a..8940a6873ca5 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -1229,7 +1229,21 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
 	if (ret)
 		return ret;
 
-	intel_runtime_pm_get(i915);
+	if (i915_gem_object_has_struct_page(obj)) {
+		/* Avoid waking the device up if we can fallback, as
+		 * waking/resuming is very slow (10-100 ms depending
+		 * on PCI sleeps and our own resume time). This easily
+		 * dwarfs any performance advantage from using the
+		 * cache bypass of indirect GGTT access.
+		 */
+		if (!intel_runtime_pm_get_if_in_use(i915)) {
+			ret = -EFAULT;
+			goto out_unlock;
+		}
+	} else {
+		intel_runtime_pm_get(i915);
+	}
+
 	vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
 				       PIN_MAPPABLE | PIN_NONBLOCK);
 	if (!IS_ERR(vma)) {
@@ -1244,7 +1258,7 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
 	if (IS_ERR(vma)) {
 		ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
 		if (ret)
-			goto out_unlock;
+			goto out_rpm;
 		GEM_BUG_ON(!node.allocated);
 	}
 
@@ -1307,8 +1321,9 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
 	} else {
 		i915_vma_unpin(vma);
 	}
-out_unlock:
+out_rpm:
 	intel_runtime_pm_put(i915);
+out_unlock:
 	mutex_unlock(&i915->drm.struct_mutex);
 	return ret;
 }
-- 
2.14.1