Userspace patches:

* radeonsi NV_timeline_semaphore
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37335

* RADV user queues
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/40808

Signed-off-by: David Rosca <[email protected]>
---
v2: don't use struct amdgpu_cs_post_dep
    fixed syncobj leak when dma_fence_chain_alloc fails

 .../gpu/drm/amd/amdgpu/amdgpu_userq_fence.c   | 50 ++++++++++++++++---
 include/uapi/drm/amdgpu_drm.h                 |  5 ++
 2 files changed, 47 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
index c0d68863fa17..f4366c5d1e35 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
@@ -460,9 +460,13 @@ int amdgpu_userq_signal_ioctl(struct drm_device *dev, void 
*data,
 
        struct drm_gem_object **gobj_write, **gobj_read;
        u32 *syncobj_handles, num_syncobj_handles;
+       u64 *syncobj_points = NULL;
        struct amdgpu_usermode_queue *queue;
        struct amdgpu_userq_fence *fence;
-       struct drm_syncobj **syncobj;
+       struct {
+               struct drm_syncobj *syncobj;
+               struct dma_fence_chain *chain;
+       } *syncobj;
        struct drm_exec exec;
        void __user *ptr;
        int r, i, entry;
@@ -482,19 +486,38 @@ int amdgpu_userq_signal_ioctl(struct drm_device *dev, 
void *data,
        if (IS_ERR(syncobj_handles))
                return PTR_ERR(syncobj_handles);
 
+       if (args->syncobj_points) {
+               ptr = u64_to_user_ptr(args->syncobj_points);
+               syncobj_points = memdup_array_user(ptr, num_syncobj_handles,
+                                           sizeof(u64));
+               if (IS_ERR(syncobj_points)) {
+                       r = PTR_ERR(syncobj_points);
+                       goto free_syncobj_handles;
+               }
+       }
+
        syncobj = kmalloc_array(num_syncobj_handles, sizeof(*syncobj),
                                GFP_KERNEL);
        if (!syncobj) {
                r = -ENOMEM;
-               goto free_syncobj_handles;
+               goto free_syncobj_points;
        }
 
        for (entry = 0; entry < num_syncobj_handles; entry++) {
-               syncobj[entry] = drm_syncobj_find(filp, syncobj_handles[entry]);
-               if (!syncobj[entry]) {
+               syncobj[entry].chain = NULL;
+               syncobj[entry].syncobj = drm_syncobj_find(filp, 
syncobj_handles[entry]);
+               if (!syncobj[entry].syncobj) {
                        r = -ENOENT;
                        goto free_syncobj;
                }
+               if (syncobj_points && syncobj_points[entry]) {
+                       syncobj[entry].chain = dma_fence_chain_alloc();
+                       if (!syncobj[entry].chain) {
+                               drm_syncobj_put(syncobj[entry].syncobj);
+                               r = -ENOMEM;
+                               goto free_syncobj;
+                       }
+               }
        }
 
        ptr = u64_to_user_ptr(args->bo_read_handles);
@@ -561,8 +584,15 @@ int amdgpu_userq_signal_ioctl(struct drm_device *dev, void 
*data,
                dma_resv_add_fence(gobj_write[i]->resv, &fence->base,
                                   DMA_RESV_USAGE_WRITE);
 
-       for (i = 0; i < num_syncobj_handles; i++)
-               drm_syncobj_replace_fence(syncobj[i], &fence->base);
+       for (i = 0; i < num_syncobj_handles; i++) {
+               if (syncobj[i].chain) {
+                       drm_syncobj_add_point(syncobj[i].syncobj, 
syncobj[i].chain,
+                                          &fence->base, syncobj_points[i]);
+                       syncobj[i].chain = NULL;
+               } else {
+                       drm_syncobj_replace_fence(syncobj[i].syncobj, 
&fence->base);
+               }
+       }
 
 exec_fini:
        /* drop the reference acquired in fence creation function */
@@ -580,9 +610,13 @@ int amdgpu_userq_signal_ioctl(struct drm_device *dev, void 
*data,
                drm_gem_object_put(gobj_read[i]);
        kvfree(gobj_read);
 free_syncobj:
-       while (entry-- > 0)
-               drm_syncobj_put(syncobj[entry]);
+       while (entry-- > 0) {
+               drm_syncobj_put(syncobj[entry].syncobj);
+               dma_fence_chain_free(syncobj[entry].chain);
+       }
        kfree(syncobj);
+free_syncobj_points:
+       kfree(syncobj_points);
 free_syncobj_handles:
        kfree(syncobj_handles);
 
diff --git a/include/uapi/drm/amdgpu_drm.h b/include/uapi/drm/amdgpu_drm.h
index 9f3090db2f16..ad643b41982c 100644
--- a/include/uapi/drm/amdgpu_drm.h
+++ b/include/uapi/drm/amdgpu_drm.h
@@ -502,6 +502,11 @@ struct drm_amdgpu_userq_signal {
         * @bo_write_handles.
         */
        __u32   num_bo_write_handles;
+       /**
+        * @syncobj_points: The list of syncobj points submitted by the user 
queue job
+        * for the corresponding @syncobj_handles.
+        */
+       __u64   syncobj_points;
 };
 
 struct drm_amdgpu_userq_fence_info {
-- 
2.43.0

Reply via email to