The GPU Scheduler's job initialization is split into two steps,
drm_sched_job_init() and drm_sched_job_arm(). One reason for this is
that actually arming a job results in the job's fences getting
initialized (armed).

Currently, the documentation does not explicitly state what
drm_sched_job_arm() does in this regard and which rules the API-User has
to follow once the function has been called.

Add a section to drm_sched_job_arm()'s docstring which details the
function's consequences regarding the job's fences.

Signed-off-by: Philipp Stanner <pstanner@redhat.com>
---
 drivers/gpu/drm/scheduler/sched_main.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
index 7e90c9f95611..e563eff4887c 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -831,6 +831,12 @@ EXPORT_SYMBOL(drm_sched_job_init);
  * Refer to drm_sched_entity_push_job() documentation for locking
  * considerations.
  *
+ * drm_sched_job_cleanup() can be used to disarm the job again - but only
+ * _before_ the job's fences have been published. Once a drm_sched_fence was
+ * published, the associated job needs to be submitted to and processed by the
+ * scheduler to avoid potential deadlocks on the DMA fences encapsulated by
+ * drm_sched_fence.
+ *
  * This can only be called if drm_sched_job_init() succeeded.
  */
 void drm_sched_job_arm(struct drm_sched_job *job)
-- 
2.46.0

On Mon, Aug 26, 2024 at 11:39:17AM +0200, Philipp Stanner wrote:
> The GPU Scheduler's job initialization is split into two steps,
> drm_sched_job_init() and drm_sched_job_arm(). One reason for this is
> that actually arming a job results in the job's fences getting
> initialized (armed).
> 
> Currently, the documentation does not explicitly state what
> drm_sched_job_arm() does in this regard and which rules the API-User has
> to follow once the function has been called.
> 
> Add a section to drm_sched_job_arm()'s docstring which details the
> function's consequences regarding the job's fences.
> 
> Signed-off-by: Philipp Stanner <pstanner@redhat.com>
> ---
>  drivers/gpu/drm/scheduler/sched_main.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> index 7e90c9f95611..e563eff4887c 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -831,6 +831,12 @@ EXPORT_SYMBOL(drm_sched_job_init);
>   * Refer to drm_sched_entity_push_job() documentation for locking
>   * considerations.
>   *
> + * drm_sched_job_cleanup() can be used to disarm the job again - but only
> + * _before_ the job's fences have been published. Once a drm_sched_fence was
> + * published, the associated job needs to be submitted to and processed by the
> + * scheduler to avoid potential deadlocks on the DMA fences encapsulated by
> + * drm_sched_fence.

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

> + *
>   * This can only be called if drm_sched_job_init() succeeded.
>   */
>  void drm_sched_job_arm(struct drm_sched_job *job)
> -- 
> 2.46.0
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch