Re: [PATCH v7 07/17] drm/panfrost: Stop all jobs before commencing device teardown

[Boris Brezillon]

On Fri, 28 Aug 2026 21:56:47 +0100
Adrián Larumbe <adrian.larumbe@xxxxxxxxxxxxx> wrote:

> We should make sure there are no inflight jobs at the time we finish driver
> structures and synchronise their IRQs, and even more so by the time clocks
> and power domains are disabled.
> 
> To that end, stop submissions from existing scheduler entities and force HW
> to stop all running jobs.
> 
> Signed-off-by: Adrián Larumbe <adrian.larumbe@xxxxxxxxxxxxx>
> ---
>  drivers/gpu/drm/panfrost/panfrost_device.c |  2 +
>  drivers/gpu/drm/panfrost/panfrost_job.c    | 71 ++++++++++++++++++------------
>  drivers/gpu/drm/panfrost/panfrost_job.h    |  1 +
>  3 files changed, 46 insertions(+), 28 deletions(-)
> 
> diff --git a/drivers/gpu/drm/panfrost/panfrost_device.c b/drivers/gpu/drm/panfrost/panfrost_device.c
> index 9549ddb7854a..9e02fb5f73c8 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_device.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_device.c
> @@ -341,6 +341,8 @@ void panfrost_device_fini(struct panfrost_device *pfdev)
>  	pm_runtime_get_sync(pfdev->base.dev);
>  	pm_runtime_disable(pfdev->base.dev);
>  
> +	panfrost_jm_stop_sched_jobs(pfdev);

Do we really want to do the SOFT_STOP+HARD_STOP dance in that case?
Can't we just SOFT_RESET the GPU and collect the pieces in the
panfrost_<component>_fini() functions?

> +
>  	panfrost_gem_fini(pfdev);
>  	panfrost_perfcnt_fini(pfdev);
>  	panfrost_jm_fini(pfdev);
> diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c b/drivers/gpu/drm/panfrost/panfrost_job.c
> index 5016d2d53da2..630298b7ea8a 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_job.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_job.c
> @@ -655,39 +655,12 @@ static u32 panfrost_active_slots(struct panfrost_device *pfdev,
>  	return js_state & *js_state_mask;
>  }
>  
> -static void
> -panfrost_reset(struct panfrost_device *pfdev,
> -	       struct drm_sched_job *bad)
> +static void panfrost_stop_jobs(struct panfrost_device *pfdev)
>  {
>  	u32 js_state, js_state_mask = 0xffffffff;
>  	unsigned int i, j;
> -	bool cookie;
>  	int ret;
>  
> -	if (!atomic_read(&pfdev->reset.pending))
> -		return;
> -
> -	/* Stop the schedulers.
> -	 *
> -	 * FIXME: We temporarily get out of the dma_fence_signalling section
> -	 * because the cleanup path generate lockdep splats when taking locks
> -	 * to release job resources. We should rework the code to follow this
> -	 * pattern:
> -	 *
> -	 *	try_lock
> -	 *	if (locked)
> -	 *		release
> -	 *	else
> -	 *		schedule_work_to_release_later
> -	 */
> -	for (i = 0; i < NUM_JOB_SLOTS; i++)
> -		drm_sched_stop(&pfdev->js->queue[i].sched, bad);
> -
> -	cookie = dma_fence_begin_signalling();
> -
> -	if (bad)
> -		drm_sched_increase_karma(bad);
> -
>  	/* Mask job interrupts and synchronize to make sure we won't be
>  	 * interrupted during our reset.
>  	 */
> @@ -730,6 +703,48 @@ panfrost_reset(struct panfrost_device *pfdev,
>  	}
>  	memset(pfdev->jobs, 0, sizeof(pfdev->jobs));
>  	spin_unlock(&pfdev->js->job_lock);
> +}
> +
> +void panfrost_jm_stop_sched_jobs(struct panfrost_device *pfdev)
> +{
> +	for (u32 i = 0; i < NUM_JOB_SLOTS; i++)
> +		drm_sched_wqueue_stop(&pfdev->js->queue[i].sched);

Do we have something cleaning up the pending list if this is stopped
while jobs are in-flight? If not, we probably need a ::cancel_job()
implementation.

> +
> +	panfrost_stop_jobs(pfdev);
> +}
> +
> +static void
> +panfrost_reset(struct panfrost_device *pfdev,
> +	       struct drm_sched_job *bad)
> +{
> +	unsigned int i;
> +	bool cookie;
> +
> +	if (!atomic_read(&pfdev->reset.pending))
> +		return;
> +
> +	/* Stop the schedulers.
> +	 *
> +	 * FIXME: We temporarily get out of the dma_fence_signalling section
> +	 * because the cleanup path generate lockdep splats when taking locks
> +	 * to release job resources. We should rework the code to follow this
> +	 * pattern:
> +	 *
> +	 *	try_lock
> +	 *	if (locked)
> +	 *		release
> +	 *	else
> +	 *		schedule_work_to_release_later
> +	 */
> +	for (i = 0; i < NUM_JOB_SLOTS; i++)
> +		drm_sched_stop(&pfdev->js->queue[i].sched, bad);
> +
> +	cookie = dma_fence_begin_signalling();
> +
> +	if (bad)
> +		drm_sched_increase_karma(bad);
> +
> +	panfrost_stop_jobs(pfdev);
>  
>  	/* Proceed with reset now. */
>  	panfrost_device_reset(pfdev, false);
> diff --git a/drivers/gpu/drm/panfrost/panfrost_job.h b/drivers/gpu/drm/panfrost/panfrost_job.h
> index 3e3b9717838f..dc9d3c9dc67b 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_job.h
> +++ b/drivers/gpu/drm/panfrost/panfrost_job.h
> @@ -73,6 +73,7 @@ int panfrost_jm_is_idle(struct panfrost_device *pfdev);
>  int panfrost_job_get_slot(struct panfrost_job *job);
>  int panfrost_job_push(struct panfrost_job *job);
>  void panfrost_job_put(struct panfrost_job *job);
> +void panfrost_jm_stop_sched_jobs(struct panfrost_device *pfdev);
>  
>  #ifdef CONFIG_DEBUG_FS
>  void panfrost_sched_debugfs_init(struct drm_minor *minor);
>