[PATCH 3/9] drm/amdkfd: Fix MQD updates

[Oded Gabbay]

On Wed, Sep 27, 2017 at 7:09 AM, Felix Kuehling <Felix.Kuehling at amd.com> wrote:
> When a queue is mapped, the MQD is owned by the FW. The FW overwrites
> the MQD on the next unmap operation. Therefore the queue must be
> unmapped before updating the MQD.
>
> For the non-HWS case, also fix disabling of queues and creation of
> queues in disabled state.
>
> Signed-off-by: Oak Zeng <Oak.Zeng at amd.com>
> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
> ---
>  .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c  | 84 ++++++++++++++++------
>  1 file changed, 62 insertions(+), 22 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 be925a4..dccb493 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> @@ -49,6 +49,8 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm,
>                                 enum kfd_unmap_queues_filter filter,
>                                 uint32_t filter_param);
>
> +static int map_queues_cpsch(struct device_queue_manager *dqm);
> +
>  static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
>                                         struct queue *q,
>                                         struct qcm_process_device *qpd);
> @@ -274,6 +276,9 @@ static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
>         dqm->dev->kfd2kgd->set_scratch_backing_va(
>                         dqm->dev->kgd, qpd->sh_hidden_private_base, qpd->vmid);
>
> +       if (!q->properties.is_active)
> +               return 0;
> +
>         retval = mqd->load_mqd(mqd, q->mqd, q->pipe, q->queue, &q->properties,
>                                q->process->mm);
>         if (retval)
> @@ -365,23 +370,50 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q)
>                 goto out_unlock;
>         }
>
> -       if (q->properties.is_active)
> -               prev_active = true;
> +       /* Save previous activity state for counters */
> +       prev_active = q->properties.is_active;
> +
> +       /* Make sure the queue is unmapped before updating the MQD */
> +       if (sched_policy != KFD_SCHED_POLICY_NO_HWS) {
> +               retval = unmap_queues_cpsch(dqm,
> +                               KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
> +               if (retval != 0) {
> +                       pr_err("unmap queue failed\n");
> +                       goto out_unlock;
> +               }
> +       } else if (sched_policy == KFD_SCHED_POLICY_NO_HWS &&
> +                  prev_active &&
> +                  (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
> +                   q->properties.type == KFD_QUEUE_TYPE_SDMA)) {
> +               retval = mqd->destroy_mqd(mqd, q->mqd,
> +                               KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
> +                               KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
> +               if (retval) {
> +                       pr_err("destroy mqd failed\n");
> +                       goto out_unlock;
> +               }
> +       }
> +
> +       retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
> +
> +       if (sched_policy != KFD_SCHED_POLICY_NO_HWS)
> +               retval = map_queues_cpsch(dqm);
> +       else if (sched_policy == KFD_SCHED_POLICY_NO_HWS &&
> +                q->properties.is_active &&
> +                (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
> +                 q->properties.type == KFD_QUEUE_TYPE_SDMA))
> +               retval = mqd->load_mqd(mqd, q->mqd, q->pipe, q->queue,
> +                                      &q->properties, q->process->mm);
>
>         /*
> -        *
>          * check active state vs. the previous state
>          * and modify counter accordingly
>          */
> -       retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
> -       if ((q->properties.is_active) && (!prev_active))
> +       if (q->properties.is_active && !prev_active)
>                 dqm->queue_count++;
>         else if (!q->properties.is_active && prev_active)
>                 dqm->queue_count--;
>
> -       if (sched_policy != KFD_SCHED_POLICY_NO_HWS)
> -               retval = execute_queues_cpsch(dqm);
> -
>  out_unlock:
>         mutex_unlock(&dqm->lock);
>         return retval;
> @@ -863,6 +895,27 @@ static int unmap_sdma_queues(struct device_queue_manager *dqm,
>  }
>
>  /* dqm->lock mutex has to be locked before calling this function */
> +static int map_queues_cpsch(struct device_queue_manager *dqm)
> +{
> +       int retval;
> +
> +       if (dqm->queue_count <= 0 || dqm->processes_count <= 0)
> +               return 0;
> +
> +       if (dqm->active_runlist)
> +               return 0;
> +
> +       retval = pm_send_runlist(&dqm->packets, &dqm->queues);
> +       if (retval) {
> +               pr_err("failed to execute runlist\n");
> +               return retval;
> +       }
> +       dqm->active_runlist = true;
> +
> +       return retval;
> +}
> +
> +/* dqm->lock mutex has to be locked before calling this function */
>  static int unmap_queues_cpsch(struct device_queue_manager *dqm,
>                                 enum kfd_unmap_queues_filter filter,
>                                 uint32_t filter_param)
> @@ -918,20 +971,7 @@ static int execute_queues_cpsch(struct device_queue_manager *dqm)
>                 return retval;
>         }
>
> -       if (dqm->queue_count <= 0 || dqm->processes_count <= 0)
> -               return 0;
> -
> -       if (dqm->active_runlist)
> -               return 0;
> -
> -       retval = pm_send_runlist(&dqm->packets, &dqm->queues);
> -       if (retval) {
> -               pr_err("failed to execute runlist");
> -               return retval;
> -       }
> -       dqm->active_runlist = true;
> -
> -       return retval;
> +       return map_queues_cpsch(dqm);
>  }
>
>  static int destroy_queue_cpsch(struct device_queue_manager *dqm,
> --
> 2.7.4
>

Applied to -next
Thanks,
Oded