drm/amdgpu: recover vram bo from shadow after gpu reset V2
authorChunming Zhou <David1.Zhou@amd.com>
Thu, 21 Jul 2016 09:20:52 +0000 (17:20 +0800)
committerAlex Deucher <alexander.deucher@amd.com>
Mon, 22 Aug 2016 17:47:21 +0000 (13:47 -0400)
V2:
1. don't directly submit to many jobs at the same time.
2. delete unrelated printk.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

index 76fd2dec7b42033196d2a787c6d2b1723a9f18c5..c38dc47cd767ab0530884cf166bb7989b3356494 100644 (file)
@@ -2060,6 +2060,35 @@ bool amdgpu_need_backup(struct amdgpu_device *adev)
        return amdgpu_lockup_timeout > 0 ? true : false;
 }
 
+static int amdgpu_recover_vram_from_shadow(struct amdgpu_device *adev,
+                                          struct amdgpu_ring *ring,
+                                          struct amdgpu_bo *bo,
+                                          struct fence **fence)
+{
+       uint32_t domain;
+       int r;
+
+       if (!bo->shadow)
+               return 0;
+
+       r = amdgpu_bo_reserve(bo, false);
+       if (r)
+               return r;
+       domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
+       /* if bo has been evicted, then no need to recover */
+       if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
+               r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
+                                                NULL, fence, true);
+               if (r) {
+                       DRM_ERROR("recover page table failed!\n");
+                       goto err;
+               }
+       }
+err:
+       amdgpu_bo_unreserve(bo);
+       return r;
+}
+
 /**
  * amdgpu_gpu_reset - reset the asic
  *
@@ -2138,13 +2167,46 @@ retry:
                if (r) {
                        dev_err(adev->dev, "ib ring test failed (%d).\n", r);
                        r = amdgpu_suspend(adev);
+                       need_full_reset = true;
                        goto retry;
                }
+               /**
+                * recovery vm page tables, since we cannot depend on VRAM is
+                * consistent after gpu full reset.
+                */
+               if (need_full_reset && amdgpu_need_backup(adev)) {
+                       struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
+                       struct amdgpu_bo *bo, *tmp;
+                       struct fence *fence = NULL, *next = NULL;
+
+                       DRM_INFO("recover vram bo from shadow\n");
+                       mutex_lock(&adev->shadow_list_lock);
+                       list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
+                               amdgpu_recover_vram_from_shadow(adev, ring, bo, &next);
+                               if (fence) {
+                                       r = fence_wait(fence, false);
+                                       if (r) {
+                                               WARN(r, "recovery from shadow isn't comleted\n");
+                                               break;
+                                       }
+                               }
 
+                               fence_put(fence);
+                               fence = next;
+                       }
+                       mutex_unlock(&adev->shadow_list_lock);
+                       if (fence) {
+                               r = fence_wait(fence, false);
+                               if (r)
+                                       WARN(r, "recovery from shadow isn't comleted\n");
+                       }
+                       fence_put(fence);
+               }
                for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
                        struct amdgpu_ring *ring = adev->rings[i];
                        if (!ring)
                                continue;
+
                        amd_sched_job_recovery(&ring->sched);
                        kthread_unpark(ring->sched.thread);
                }