diff mbox series

[v3,6/7] drm/syncobj: Add a fast path to drm_syncobj_array_wait_timeout

Message ID 20250328164621.59150-7-tvrtko.ursulin@igalia.com (mailing list archive)
State New
Headers show
Series few drm_syncobj optimisations | expand

Commit Message

Tvrtko Ursulin March 28, 2025, 4:46 p.m. UTC 
Running the Cyberpunk 2077 benchmark we can observe that waiting on DRM
sycobjs is relatively hot, but the 96% of the calls are for a single
object. (~4% for two points, and never more than three points. While
a more trivial workload like vkmark under Plasma is even more skewed
to single point waits.)

Therefore lets add a fast path to bypass the kcalloc/kfree and use a pre-
allocated stack array for those cases.

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
Reviewed-by: Maíra Canal <mcanal@igalia.com>
---
v2:
 * Document rationale for stack array in a comment.

v3:
 * Added DRM_SYNCOBJ_FAST_PATH_ENTRIES to avoid hardcoding fast path array
   size.
---
 drivers/gpu/drm/drm_syncobj.c | 24 ++++++++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)
diff mbox series

Patch

diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c
index bf2fbe07add2..34a6dc50ec31 100644
--- a/drivers/gpu/drm/drm_syncobj.c
+++ b/drivers/gpu/drm/drm_syncobj.c
@@ -236,6 +236,14 @@  static void
 syncobj_eventfd_entry_func(struct drm_syncobj *syncobj,
 			   struct syncobj_eventfd_entry *entry);
 
+/*
+ * Empirically vast majority of ioctls pass in a single syncobj (96%) and never
+ * more than three points. Therefore implement a fast path with a small stack
+ * array to avoid going into the allocator sometimes several times per
+ * userspace rendered frame.
+ */
+#define DRM_SYNCOBJ_FAST_PATH_ENTRIES 4
+
 /**
  * drm_syncobj_find - lookup and reference a sync object.
  * @file_private: drm file private pointer
@@ -1035,6 +1043,7 @@  static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 						  uint32_t *idx,
 						  ktime_t *deadline)
 {
+	struct syncobj_wait_entry stack_entries[DRM_SYNCOBJ_FAST_PATH_ENTRIES];
 	struct syncobj_wait_entry *entries;
 	uint32_t signaled_count, i;
 	struct dma_fence *fence;
@@ -1049,9 +1058,14 @@  static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 	    !access_ok(user_points, count * sizeof(*user_points)))
 		return -EFAULT;
 
-	entries = kcalloc(count, sizeof(*entries), GFP_KERNEL);
-	if (!entries)
-		return -ENOMEM;
+	if (count > ARRAY_SIZE(stack_entries)) {
+		entries = kcalloc(count, sizeof(*entries), GFP_KERNEL);
+		if (!entries)
+			return -ENOMEM;
+	} else {
+		memset(stack_entries, 0, sizeof(stack_entries));
+		entries = stack_entries;
+	}
 
 	/* Walk the list of sync objects and initialize entries.  We do
 	 * this up-front so that we can properly return -EINVAL if there is
@@ -1174,7 +1188,9 @@  static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 						  &entries[i].fence_cb);
 		dma_fence_put(entries[i].fence);
 	}
-	kfree(entries);
+
+	if (entries != stack_entries)
+		kfree(entries);
 
 	return timeout;
 }