On 09.12.21 11:23, Christian König wrote:
Always waiting for the exclusive fence resulted on some performance
regressions. So try to wait for the shared fences first, then the
exclusive fence should always be signaled already.
v2: fix incorrectly placed "(", add some comment why we do this.
Signed-off-by: Christian König <christian.koenig@xxxxxxx>
Tested-by: Stefan Fritsch <sf@xxxxxxxxxxx>
Please also add a cc for linux-stable, so that this is fixed in 5.15.x
Cheers,
Stefan
---
drivers/gpu/drm/nouveau/nouveau_fence.c | 28 +++++++++++++------------
1 file changed, 15 insertions(+), 13 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c
index 05d0b3eb3690..0ae416aa76dc 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -353,15 +353,22 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
if (ret)
return ret;
- }
- fobj = dma_resv_shared_list(resv);
- fence = dma_resv_excl_fence(resv);
+ fobj = NULL;
+ } else {
+ fobj = dma_resv_shared_list(resv);
+ }
- if (fence) {
+ /* Waiting for the exclusive fence first causes performance regressions
+ * under some circumstances. So manually wait for the shared ones first.
+ */
+ for (i = 0; i < (fobj ? fobj->shared_count : 0) && !ret; ++i) {
struct nouveau_channel *prev = NULL;
bool must_wait = true;
+ fence = rcu_dereference_protected(fobj->shared[i],
+ dma_resv_held(resv));
+
f = nouveau_local_fence(fence, chan->drm);
if (f) {
rcu_read_lock();
@@ -373,20 +380,13 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
if (must_wait)
ret = dma_fence_wait(fence, intr);
-
- return ret;
}
- if (!exclusive || !fobj)
- return ret;
-
- for (i = 0; i < fobj->shared_count && !ret; ++i) {
+ fence = dma_resv_excl_fence(resv);
+ if (fence) {
struct nouveau_channel *prev = NULL;
bool must_wait = true;
- fence = rcu_dereference_protected(fobj->shared[i],
- dma_resv_held(resv));
-
f = nouveau_local_fence(fence, chan->drm);
if (f) {
rcu_read_lock();
@@ -398,6 +398,8 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
if (must_wait)
ret = dma_fence_wait(fence, intr);
+
+ return ret;
}
return ret;
