[Mesa-dev] [PATCH v2 15/25] winsys/amdgpu: extend amdgpu_add_fence to allow adding multiple fences

[Nicolai Hähnle]

From: Nicolai Hähnle <nicolai.haehnle at amd.com>

---
 src/gallium/winsys/amdgpu/drm/amdgpu_cs.c | 35 +++++++++++++++++++++----------
 src/gallium/winsys/amdgpu/drm/amdgpu_cs.h |  3 +++
 2 files changed, 27 insertions(+), 11 deletions(-)

diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
index e2d3a45..bffa725 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
@@ -945,59 +945,72 @@ static void amdgpu_add_fence_dependency(struct amdgpu_cs *acs,
       dep = &cs->request.dependencies[idx];
       memcpy(dep, &bo_fence->fence, sizeof(*dep));
    }
 
    for (unsigned j = new_num_fences; j < bo->num_fences; ++j)
       amdgpu_fence_reference(&bo->fences[j], NULL);
 
    bo->num_fences = new_num_fences;
 }
 
-static void amdgpu_add_fence(struct amdgpu_winsys_bo *bo,
-                             struct pipe_fence_handle *fence)
+/* Add the given list of fences to the buffer's fence list.
+ *
+ * Must be called with the winsys bo_fence_lock held.
+ */
+void amdgpu_add_fences(struct amdgpu_winsys_bo *bo,
+                       unsigned num_fences,
+                       struct pipe_fence_handle **fences)
 {
-   if (bo->num_fences >= bo->max_fences) {
-      unsigned new_max_fences = MAX2(1, bo->max_fences * 2);
+   if (bo->num_fences + num_fences > bo->max_fences) {
+      unsigned new_max_fences = MAX2(bo->num_fences + num_fences, bo->max_fences * 2);
       struct pipe_fence_handle **new_fences =
          REALLOC(bo->fences,
                  bo->num_fences * sizeof(*new_fences),
                  new_max_fences * sizeof(*new_fences));
-      if (new_fences) {
+      if (likely(new_fences)) {
          bo->fences = new_fences;
          bo->max_fences = new_max_fences;
       } else {
-         fprintf(stderr, "amdgpu_add_fence: allocation failure, dropping fence\n");
+         unsigned drop;
+
+         fprintf(stderr, "amdgpu_add_fences: allocation failure, dropping fence(s)\n");
          if (!bo->num_fences)
             return;
 
-         bo->num_fences--; /* prefer to keep a more recent fence if possible */
+         bo->num_fences--; /* prefer to keep the most recent fence if possible */
          amdgpu_fence_reference(&bo->fences[bo->num_fences], NULL);
+
+         drop = bo->num_fences + num_fences - bo->max_fences;
+         num_fences -= drop;
+         fences += drop;
       }
    }
 
-   bo->fences[bo->num_fences] = NULL;
-   amdgpu_fence_reference(&bo->fences[bo->num_fences], fence);
-   bo->num_fences++;
+   for (unsigned i = 0; i < num_fences; ++i) {
+      bo->fences[bo->num_fences] = NULL;
+      amdgpu_fence_reference(&bo->fences[bo->num_fences], fences[i]);
+      bo->num_fences++;
+   }
 }
 
 static void amdgpu_add_fence_dependencies_list(struct amdgpu_cs *acs,
                                                struct pipe_fence_handle *fence,
                                                unsigned num_buffers,
                                                struct amdgpu_cs_buffer *buffers)
 {
    for (unsigned i = 0; i < num_buffers; i++) {
       struct amdgpu_cs_buffer *buffer = &buffers[i];
       struct amdgpu_winsys_bo *bo = buffer->bo;
 
       amdgpu_add_fence_dependency(acs, buffer);
       p_atomic_inc(&bo->num_active_ioctls);
-      amdgpu_add_fence(bo, fence);
+      amdgpu_add_fences(bo, 1, &fence);
    }
 }
 
 /* Since the kernel driver doesn't synchronize execution between different
  * rings automatically, we have to add fence dependencies manually.
  */
 static void amdgpu_add_fence_dependencies(struct amdgpu_cs *acs)
 {
    struct amdgpu_cs_context *cs = acs->csc;
 
diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.h b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.h
index bdf7cb2..242410f 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.h
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.h
@@ -233,15 +233,18 @@ amdgpu_bo_is_referenced_by_cs_with_usage(struct amdgpu_cs *cs,
 }
 
 static inline bool
 amdgpu_bo_is_referenced_by_any_cs(struct amdgpu_winsys_bo *bo)
 {
    return bo->num_cs_references != 0;
 }
 
 bool amdgpu_fence_wait(struct pipe_fence_handle *fence, uint64_t timeout,
                        bool absolute);
+void amdgpu_add_fences(struct amdgpu_winsys_bo *bo,
+                       unsigned num_fences,
+                       struct pipe_fence_handle **fences);
 void amdgpu_cs_sync_flush(struct radeon_winsys_cs *rcs);
 void amdgpu_cs_init_functions(struct amdgpu_winsys *ws);
 void amdgpu_cs_submit_ib(void *job, int thread_index);
 
 #endif
-- 
2.9.3