diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 17 |
1 files changed, 17 insertions, 0 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index cba44ec5d2b0..37de451447b1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -1393,6 +1393,7 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, uint32_t filter_param) { int retval = 0; + struct mqd_manager *mqd_mgr; if (!dqm->sched_running) return 0; @@ -1424,6 +1425,22 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, return retval; } + /* In the current MEC firmware implementation, if compute queue + * doesn't response to the preemption request in time, HIQ will + * abandon the unmap request without returning any timeout error + * to driver. Instead, MEC firmware will log the doorbell of the + * unresponding compute queue to HIQ.MQD.queue_doorbell_id fields. + * To make sure the queue unmap was successful, driver need to + * check those fields + */ + mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]; + if (mqd_mgr->read_doorbell_id(dqm->packets.priv_queue->queue->mqd)) { + pr_err("HIQ MQD's queue_doorbell_id0 is not 0, Queue preemption time out\n"); + while (halt_if_hws_hang) + schedule(); + return -ETIME; + } + pm_release_ib(&dqm->packets); dqm->active_runlist = false; |