}
r = amdgpu_sync_fence(&p->gang_leader->sync, fence);
- dma_fence_put(fence);
+ if (r)
+ goto error;
+
+ /*
+ * When we have an explicit dependency it might be necessary to insert a
+ * pipeline sync to make sure that all caches etc are flushed and the
+ * next job actually sees the results from the previous one.
+ */
+ if (fence->context == p->gang_leader->base.entity->fence_context)
+ r = amdgpu_sync_fence(&p->gang_leader->explicit_sync, fence);
+error:
+ dma_fence_put(fence);
return r;
}
need_ctx_switch = ring->current_ctx != fence_ctx;
if (ring->funcs->emit_pipeline_sync && job &&
- ((tmp = amdgpu_sync_get_fence(&job->sched_sync)) ||
+ ((tmp = amdgpu_sync_get_fence(&job->explicit_sync)) ||
(amdgpu_sriov_vf(adev) && need_ctx_switch) ||
amdgpu_vm_need_pipeline_sync(ring, job))) {
need_pipe_sync = true;
(*job)->vm = vm;
amdgpu_sync_create(&(*job)->sync);
- amdgpu_sync_create(&(*job)->sched_sync);
+ amdgpu_sync_create(&(*job)->explicit_sync);
(*job)->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
(*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
drm_sched_job_cleanup(s_job);
amdgpu_sync_free(&job->sync);
- amdgpu_sync_free(&job->sched_sync);
+ amdgpu_sync_free(&job->explicit_sync);
dma_fence_put(&job->hw_fence);
}
amdgpu_job_free_resources(job);
amdgpu_sync_free(&job->sync);
- amdgpu_sync_free(&job->sched_sync);
+ amdgpu_sync_free(&job->explicit_sync);
if (job->gang_submit != &job->base.s_fence->scheduled)
dma_fence_put(job->gang_submit);
int r;
fence = amdgpu_sync_get_fence(&job->sync);
- if (fence && drm_sched_dependency_optimized(fence, s_entity)) {
- r = amdgpu_sync_fence(&job->sched_sync, fence);
- if (r)
- DRM_ERROR("Error adding fence (%d)\n", r);
- }
-
while (!fence && job->vm && !job->vmid) {
r = amdgpu_vmid_grab(job->vm, ring, job, &fence);
if (r)
struct drm_sched_job base;
struct amdgpu_vm *vm;
struct amdgpu_sync sync;
- struct amdgpu_sync sched_sync;
+ struct amdgpu_sync explicit_sync;
struct dma_fence hw_fence;
struct dma_fence *gang_submit;
uint32_t preamble_status;