Re: [PATCH 3/3] drm/panthor: Rreset device and load FW after failed PM suspend

[Boris Brezillon <boris.brezillon@collabora.com>]

On Fri, 11 Oct 2024 23:57:01 +0100
Adrián Larumbe <adrian.larumbe@collabora.com> wrote:

> On rk3588 SoCs, during a runtime PM suspend, the transition to the
> lowest voltage/frequency pair might sometimes fail for reasons not yet
> understood. In that case, even a slow FW reset will fail, leaving the
> device's PM runtime status as unusuable.
> 
> When that happens, successive attempts to resume the device upon running
> a job will always fail.
> 
> Fix it by forcing a synchronous device reset, which will lead to a
> successful FW reload, and also reset the device's PM runtime error
> status before resuming it.
> 
> Signed-off-by: Adrián Larumbe <adrian.larumbe@collabora.com>
> ---
>  drivers/gpu/drm/panthor/panthor_device.c | 10 ++++++++++
>  drivers/gpu/drm/panthor/panthor_device.h |  2 ++
>  drivers/gpu/drm/panthor/panthor_sched.c  |  7 +++++++
>  3 files changed, 19 insertions(+)
> 
> diff --git a/drivers/gpu/drm/panthor/panthor_device.c b/drivers/gpu/drm/panthor/panthor_device.c
> index 5430557bd0b8..ec6fed5e996b 100644
> --- a/drivers/gpu/drm/panthor/panthor_device.c
> +++ b/drivers/gpu/drm/panthor/panthor_device.c
> @@ -105,6 +105,16 @@ static void panthor_device_reset_cleanup(struct drm_device *ddev, void *data)
>  	destroy_workqueue(ptdev->reset.wq);
>  }
>  
> +int panthor_device_reset_sync(struct panthor_device *ptdev)
> +{
> +	panthor_fw_pre_reset(ptdev, false);
> +	panthor_mmu_pre_reset(ptdev);
> +	panthor_gpu_soft_reset(ptdev);
> +	panthor_gpu_l2_power_on(ptdev);
> +	panthor_mmu_post_reset(ptdev);
> +	return panthor_fw_post_reset(ptdev);
> +}
> +
>  static void panthor_device_reset_work(struct work_struct *work)
>  {
>  	struct panthor_device *ptdev = container_of(work, struct panthor_device, reset.work);
> diff --git a/drivers/gpu/drm/panthor/panthor_device.h b/drivers/gpu/drm/panthor/panthor_device.h
> index 0e68f5a70d20..05a5a7233378 100644
> --- a/drivers/gpu/drm/panthor/panthor_device.h
> +++ b/drivers/gpu/drm/panthor/panthor_device.h
> @@ -217,6 +217,8 @@ struct panthor_file {
>  int panthor_device_init(struct panthor_device *ptdev);
>  void panthor_device_unplug(struct panthor_device *ptdev);
>  
> +int panthor_device_reset_sync(struct panthor_device *ptdev);
> +
>  /**
>   * panthor_device_schedule_reset() - Schedules a reset operation
>   */
> diff --git a/drivers/gpu/drm/panthor/panthor_sched.c b/drivers/gpu/drm/panthor/panthor_sched.c
> index c7b350fc3eba..9a854c8c5718 100644
> --- a/drivers/gpu/drm/panthor/panthor_sched.c
> +++ b/drivers/gpu/drm/panthor/panthor_sched.c
> @@ -3101,6 +3101,13 @@ queue_run_job(struct drm_sched_job *sched_job)
>  		return dma_fence_get(job->done_fence);
>  	}
>  
> +	if (ptdev->base.dev->power.runtime_error) {
> +		ret = panthor_device_reset_sync(ptdev);
> +		if (drm_WARN_ON(&ptdev->base, ret))
> +			return ERR_PTR(ret);
> +		drm_WARN_ON(&ptdev->base, pm_runtime_set_active(ptdev->base.dev));
> +	}

I'd rather pretend the suspend/resume worked (even if it didn't) and
deal with the consequences (force a slow reset on the next resume), than
spread the 'if-PM-op-failed-force-sync-reset' thing everywhere we do a
pm_runtime_resume_and_get(). Also not sure how resetting the GPU will
help fixing the OPP transition failure.

> +
>  	ret = pm_runtime_resume_and_get(ptdev->base.dev);
>  	if (drm_WARN_ON(&ptdev->base, ret))
>  		return ERR_PTR(ret);