Merge tag 'drm-next-2022-01-07' of git://anongit.freedesktop.org/drm/drm

[linux-2.6-microblaze.git] / drivers / gpu / drm / nouveau / nouveau_fence.c

diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c
index 26f9299..a3a04e0 100644 (file)
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -344,38 +344,48 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan,
 {
        struct nouveau_fence_chan *fctx = chan->fence;
        struct dma_resv *resv = nvbo->bo.base.resv;
-       struct dma_resv_iter cursor;
-       struct dma_fence *fence;
-       struct nouveau_fence *f;
-       int ret;
+       int i, ret;
 
        if (!exclusive) {
                ret = dma_resv_reserve_shared(resv, 1);
-
                if (ret)
                        return ret;
        }
 
-       dma_resv_for_each_fence(&cursor, resv, exclusive, fence) {
-               struct nouveau_channel *prev = NULL;
-               bool must_wait = true;
-
-               f = nouveau_local_fence(fence, chan->drm);
-               if (f) {
-                       rcu_read_lock();
-                       prev = rcu_dereference(f->channel);
-                       if (prev && (prev == chan ||
-                                    fctx->sync(f, prev, chan) == 0))
-                               must_wait = false;
-                       rcu_read_unlock();
-               }
+       /* Waiting for the exclusive fence first causes performance regressions
+        * under some circumstances. So manually wait for the shared ones first.
+        */
+       for (i = 0; i < 2; ++i) {
+               struct dma_resv_iter cursor;
+               struct dma_fence *fence;
+
+               dma_resv_for_each_fence(&cursor, resv, exclusive, fence) {
+                       struct nouveau_fence *f;
+
+                       if (i == 0 && dma_resv_iter_is_exclusive(&cursor))
+                               continue;
+
+                       f = nouveau_local_fence(fence, chan->drm);
+                       if (f) {
+                               struct nouveau_channel *prev;
+                               bool must_wait = true;
+
+                               rcu_read_lock();
+                               prev = rcu_dereference(f->channel);
+                               if (prev && (prev == chan ||
+                                            fctx->sync(f, prev, chan) == 0))
+                                       must_wait = false;
+                               rcu_read_unlock();
+                               if (!must_wait)
+                                       continue;
+                       }
 
-               if (must_wait) {
                        ret = dma_fence_wait(fence, intr);
                        if (ret)
                                return ret;
                }
        }
+
        return 0;
 }