diff options
Diffstat (limited to 'meta-amd-bsp/recipes-kernel/linux/linux-yocto-4.14.71/2131-drm-amdkfd-fix-hang-issue-if-KFD-memory-failed-to-re.patch')
| -rw-r--r-- | meta-amd-bsp/recipes-kernel/linux/linux-yocto-4.14.71/2131-drm-amdkfd-fix-hang-issue-if-KFD-memory-failed-to-re.patch | 82 |
1 files changed, 82 insertions, 0 deletions
diff --git a/meta-amd-bsp/recipes-kernel/linux/linux-yocto-4.14.71/2131-drm-amdkfd-fix-hang-issue-if-KFD-memory-failed-to-re.patch b/meta-amd-bsp/recipes-kernel/linux/linux-yocto-4.14.71/2131-drm-amdkfd-fix-hang-issue-if-KFD-memory-failed-to-re.patch new file mode 100644 index 00000000..bf2ad659 --- /dev/null +++ b/meta-amd-bsp/recipes-kernel/linux/linux-yocto-4.14.71/2131-drm-amdkfd-fix-hang-issue-if-KFD-memory-failed-to-re.patch @@ -0,0 +1,82 @@ +From b7cfd306079e1746315d6d2ad5c837623a07cbaf Mon Sep 17 00:00:00 2001 +From: Philip Yang <Philip.Yang@amd.com> +Date: Fri, 20 Oct 2017 14:32:27 -0400 +Subject: [PATCH 2131/4131] drm/amdkfd: fix hang issue if KFD memory failed to + restore + +In current logic, if restore failed, the restore work is rescheduled, +and new fence is created for process->eviction_fence, but the new fence +will attach to BOs created before the next restore, as a result, next +restore will start evict work because new fence is not signaled while +ttm_bo_wait on those new BOs. The eviction count can detect and skip +eviction but it schedules restore work again, this is recursive +eviction/restore. + +The fix is to create new fence only after restore is finished completely. + +BUG: SWDEV-134782 + +Change-Id: Ib2f25f8a50a5eb8247572a105085482dc7326733 +Signed-off-by: Philip Yang <Philip.Yang@amd.com> +--- + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 30 +++++++++++++----------- + 1 file changed, 16 insertions(+), 14 deletions(-) + +diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +index 900019d..5de8398 100644 +--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c ++++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +@@ -2287,7 +2287,7 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info) + struct amdkfd_vm *peer_vm; + struct kgd_mem *mem; + struct bo_vm_reservation_context ctx; +- struct amdgpu_amdkfd_fence *old_fence; ++ struct amdgpu_amdkfd_fence *new_fence; + int ret = 0, i; + struct list_head duplicate_save; + struct amdgpu_sync sync_obj; +@@ -2302,19 +2302,6 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info) + if (pd_bo_list == NULL) + return -ENOMEM; + +- /* Release old eviction fence and create new one. Use context and mm +- * from the old fence. +- */ +- old_fence = process_info->eviction_fence; +- process_info->eviction_fence = +- amdgpu_amdkfd_fence_create(old_fence->base.context, +- old_fence->mm); +- dma_fence_put(&old_fence->base); +- if (!process_info->eviction_fence) { +- pr_err("Failed to create eviction fence\n"); +- goto evict_fence_fail; +- } +- + i = 0; + mutex_lock(&process_info->lock); + list_for_each_entry(peer_vm, &process_info->vm_list_head, +@@ -2386,6 +2373,21 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info) + + amdgpu_sync_wait(ctx.sync, false); + ++ /* Release old eviction fence and create new one, because fence only ++ * goes from unsignaled to signaled, fence cannot be reused. ++ * Use context and mm from the old fence. ++ */ ++ new_fence = amdgpu_amdkfd_fence_create( ++ process_info->eviction_fence->base.context, ++ process_info->eviction_fence->mm); ++ if (!new_fence) { ++ pr_err("Failed to create eviction fence\n"); ++ ret = -ENOMEM; ++ goto validate_map_fail; ++ } ++ dma_fence_put(&process_info->eviction_fence->base); ++ process_info->eviction_fence = new_fence; ++ + /* Wait for validate to finish and attach new eviction fence */ + list_for_each_entry(mem, &process_info->kfd_bo_list, + validate_list.head) +-- +2.7.4 + |