[PATCH] drm/amdgpu: Fix recursive evict/restore worker on a multiple mapped bo

Shaoyun Liu Shaoyun.Liu at amd.com
Fri Aug 4 23:03:14 UTC 2017


When one buffer object is mapped muliple times, KFD will treat them as
different memory object but share the same ttm_bo with a single eviction
fence. Seperate the ttm_bo_wait and attach eviction fence code to avoid
ttm_bo_wait trigger the eviction worker on the bo that just be assigned
with the eviction fence

Signed-off-by: Shaoyun Liu <Shaoyun.Liu at amd.com>

Change-Id: I4a7dba32c21cbd5fc77512c0476b8d2b4b6ea155
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index a3ee5bf..8e03f44 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -2309,9 +2309,6 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info)
 		goto ttm_reserve_fail;
 	}
 
-	if (!list_empty(&duplicate_save))
-		pr_err("BUG: list of BOs to reserve has duplicates!\n");
-
 	amdgpu_sync_create(&sync_obj);
 	ctx.sync = &sync_obj;
 
@@ -2360,12 +2357,11 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info)
 
 	/* Wait for validate to finish and attach new eviction fence */
 	list_for_each_entry(mem, &process_info->kfd_bo_list,
-		validate_list.head) {
-		struct amdgpu_bo *bo = mem->bo;
-
-		ttm_bo_wait(&bo->tbo, false, false);
-		amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
-	}
+		validate_list.head)
+		ttm_bo_wait(&mem->bo->tbo, false, false);
+	list_for_each_entry(mem, &process_info->kfd_bo_list,
+		validate_list.head)
+		amdgpu_bo_fence(mem->bo, &process_info->eviction_fence->base, true);
 
 	/* Attach eviction fence to PD / PT BOs */
 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
-- 
1.9.1



More information about the amd-gfx mailing list