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.ursu...@igalia.com>
Reviewed-by: Maíra Canal <mca...@igalia.com>
---
v2:
* Document rationale for stack array in a comment.
---
drivers/gpu/drm/drm_syncobj.c | 21 +++++++++++++++++----
1 file changed, 17 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c
index bf2fbe07add2..b906d6acb4ef 100644
--- a/drivers/gpu/drm/drm_syncobj.c
+++ b/drivers/gpu/drm/drm_syncobj.c
@@ -1035,6 +1035,12 @@ static signed long drm_syncobj_array_wait_timeout(struct
drm_syncobj **syncobjs,
uint32_t *idx,
ktime_t *deadline)
{
+ /*
+ * Empirically vast majority of calls here works with just a single
+ * point (96%) and never more than three points. Therefore a small stack
+ * array can cheaply avoid multiple per frame allocations.
+ */
+ struct syncobj_wait_entry stack_entries[4];
struct syncobj_wait_entry *entries;
uint32_t signaled_count, i;
struct dma_fence *fence;
@@ -1049,9 +1055,14 @@ 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 +1185,9 @@ 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;
}
--
2.48.0
