Reviewed-by: Jacek Lawrynowicz <jacek.lawrynow...@linux.intel.com>
On 1/7/2025 6:32 PM, Maciej Falkowski wrote:
> From: Karol Wachowski <karol.wachow...@intel.com>
>
> Mark as invalid context of a job that returned HW context violation
> error and queue work that aborts jobs from faulty context.
> Add engine reset to the context abort thread handler to not only abort
> currently executing jobs but also to ensure NPU invalid state recovery.
>
> Signed-off-by: Karol Wachowski <karol.wachow...@intel.com>
> Signed-off-by: Maciej Falkowski <maciej.falkow...@linux.intel.com>
> ---
> drivers/accel/ivpu/ivpu_job.c | 25 +++++++++++++++++++++++++
> 1 file changed, 25 insertions(+)
>
> diff --git a/drivers/accel/ivpu/ivpu_job.c b/drivers/accel/ivpu/ivpu_job.c
> index c93ea37062d7..3c162ac41a1d 100644
> --- a/drivers/accel/ivpu/ivpu_job.c
> +++ b/drivers/accel/ivpu/ivpu_job.c
> @@ -533,6 +533,26 @@ static int ivpu_job_signal_and_destroy(struct
> ivpu_device *vdev, u32 job_id, u32
>
> lockdep_assert_held(&vdev->submitted_jobs_lock);
>
> + job = xa_load(&vdev->submitted_jobs_xa, job_id);
> + if (!job)
> + return -ENOENT;
> +
> + if (job_status == VPU_JSM_STATUS_MVNCI_CONTEXT_VIOLATION_HW) {
> + guard(mutex)(&job->file_priv->lock);
> +
> + if (job->file_priv->has_mmu_faults)
> + return 0;
> +
> + /*
> + * Mark context as faulty and defer destruction of the job to
> jobs abort thread
> + * handler to synchronize between both faults and jobs
> returning context violation
> + * status and ensure both are handled in the same way
> + */
> + job->file_priv->has_mmu_faults = true;
> + queue_work(system_wq, &vdev->context_abort_work);
> + return 0;
> + }
> +
> job = ivpu_job_remove_from_submitted_jobs(vdev, job_id);
> if (!job)
> return -ENOENT;
> @@ -946,6 +966,9 @@ void ivpu_context_abort_work_fn(struct work_struct *work)
> unsigned long ctx_id;
> unsigned long id;
>
> + if (vdev->fw->sched_mode == VPU_SCHEDULING_MODE_HW)
> + ivpu_jsm_reset_engine(vdev, 0);
> +
> mutex_lock(&vdev->context_list_lock);
> xa_for_each(&vdev->context_xa, ctx_id, file_priv) {
> if (!file_priv->has_mmu_faults || file_priv->aborted)
> @@ -959,6 +982,8 @@ void ivpu_context_abort_work_fn(struct work_struct *work)
>
> if (vdev->fw->sched_mode != VPU_SCHEDULING_MODE_HW)
> return;
> +
> + ivpu_jsm_hws_resume_engine(vdev, 0);
> /*
> * In hardware scheduling mode NPU already has stopped processing jobs
> * and won't send us any further notifications, thus we have to free
> job related resources
