A hung user queue must not be replayed after reset because its context may be corrupted. so Skip guilty queues and queues whose VM generation has changed when mapping or restoring them.
Signed-off-by: Prike Liang <[email protected]> --- drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 19 +++++++++++++++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h | 2 ++ 2 files changed, 19 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c index 6c4743ddf04a..cc212226fb48 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c @@ -148,6 +148,8 @@ static void amdgpu_userq_hang_detect_work(struct work_struct *work) } } + WRITE_ONCE(queue->guilty, true); + if (amdgpu_userq_is_reset_type_supported(adev, queue->queue_type, AMDGPU_RESET_TYPE_PER_QUEUE)) { int r; @@ -174,7 +176,7 @@ static void amdgpu_userq_hang_detect_work(struct work_struct *work) * handler to another is illegal if you don't take extra precautions! */ if (gpu_reset) - amdgpu_userq_mgr_reset_work(&queue->userq_mgr->reset_work); + amdgpu_userq_mgr_reset_work(&uq_mgr->reset_work); } /* @@ -338,6 +340,10 @@ static int amdgpu_userq_restore_helper(struct amdgpu_usermode_queue *queue) adev->userq_funcs[queue->queue_type]; int r = 0; + if (READ_ONCE(queue->guilty) || + queue->generation != amdgpu_vm_generation(adev, queue->vm)) + return 0; + if (queue->state == AMDGPU_USERQ_STATE_PREEMPTED) { trace_amdgpu_userq_state_start(queue); @@ -388,6 +394,10 @@ static int amdgpu_userq_map_helper(struct amdgpu_usermode_queue *queue) adev->userq_funcs[queue->queue_type]; int r; + if (READ_ONCE(queue->guilty) || + queue->generation != amdgpu_vm_generation(adev, queue->vm)) + return 0; + if (queue->state == AMDGPU_USERQ_STATE_UNMAPPED) { trace_amdgpu_userq_state_start(queue); @@ -833,6 +843,7 @@ amdgpu_userq_query_status(struct drm_file *filp, union drm_amdgpu_userq *args) args->out_qs.flags = 0; if (queue->state == AMDGPU_USERQ_STATE_HUNG || queue->state == AMDGPU_USERQ_STATE_INVALID_VA || + READ_ONCE(queue->guilty) || queue->generation != amdgpu_vm_generation(adev, &fpriv->vm)) args->out_qs.flags |= AMDGPU_USERQ_QUERY_STATUS_FLAGS_BROKEN; if (amdgpu_in_reset(adev)) @@ -1598,7 +1609,11 @@ int amdgpu_userq_post_reset(struct amdgpu_device *adev, bool vram_lost) int ret = 0, r; xa_for_each(&adev->userq_doorbell_xa, queue_id, queue) { - if (queue->state == AMDGPU_USERQ_STATE_HUNG && !vram_lost) { + if (vram_lost || READ_ONCE(queue->guilty) || + queue->generation != amdgpu_vm_generation(adev, queue->vm)) + continue; + + if (queue->state == AMDGPU_USERQ_STATE_HUNG) { trace_amdgpu_userq_state_start(queue); userq_funcs = adev->userq_funcs[queue->queue_type]; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h index d1a2e2a45861..69d6ad4e6b11 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h @@ -51,6 +51,8 @@ struct amdgpu_userq_obj { struct amdgpu_usermode_queue { int queue_type; enum amdgpu_userq_state state; + /* Reset due to a hang; must not be re-mapped, userspace recreates. */ + bool guilty; uint64_t doorbell_handle; uint64_t doorbell_index; u32 doorbell_offset; -- 2.34.1
