Re: [PATCH v2 2/8] drm/sched: Allow drivers to skip the reset and keep on running

[Philipp Stanner <phasta@mailbox.org>]

Hi,

thx for the update. Seems to be developing nicely. Some comments below.

On Fri, 2025-05-30 at 11:01 -0300, Maíra Canal wrote:
> When the DRM scheduler times out, it's possible that the GPU isn't
> hung;
> instead, a job may still be running, and there may be no valid reason
> to
> reset the hardware. This can occur in two situations:
> 
>   1. The GPU exposes some mechanism that ensures the GPU is still
> making
>      progress. By checking this mechanism, the driver can safely skip
> the
>      reset, re-arm the timeout, and allow the job to continue running
> until
>      completion. This is the case for v3d, Etnaviv, and Xe.
>   2. Timeout has fired before the free-job worker. Consequently, the
>      scheduler calls `timedout_job()` for a job that isn't timed out.
> 
> These two scenarios are problematic because the job was removed from
> the
> `sched->pending_list` before calling `sched->ops->timedout_job()`,
> which
> means that when the job finishes, it won't be freed by the scheduler
> though `sched->ops->free_job()`. As a result, the job and its
> resources
> won't be freed, leading to a memory leak.

nit: redundant "won't bee freed"

> 
> To resolve those scenarios, create a new `drm_gpu_sched_stat`, called

nit:
s/resolve those scenarios/solve those problems

> DRM_GPU_SCHED_STAT_NO_HANG, that allows a driver to skip the reset.
> The
> new status will indicate that the job should be reinserted into the

nit:
s/should/must

> pending list, and the hardware / driver is still responsible to
> signal job completion.

The driver is *always* responsible for signaling, well, "the hardware
fence". We could have a discussion about whether a job is "completed"
if the driver signals its hardware fence through the timedout_job()
callback, but I think it's safer to just change this sentence to:

"and the hardware / driver will still complete that job."

> 
> Signed-off-by: Maíra Canal <mcanal@igalia.com>
> ---
>  drivers/gpu/drm/scheduler/sched_main.c | 49
> ++++++++++++++++++++++++++++++++--
>  include/drm/gpu_scheduler.h            |  3 +++
>  2 files changed, 50 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/scheduler/sched_main.c
> b/drivers/gpu/drm/scheduler/sched_main.c
> index
> 3b0760dfa4fe2fc63e893cda733e78d08dd451d5..ddc53eadab7bb6a15109f43989a
> fa1f7a95a3b41 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -379,11 +379,16 @@ static void drm_sched_run_free_queue(struct
> drm_gpu_scheduler *sched)
>  {
>  	struct drm_sched_job *job;
>  
> -	spin_lock(&sched->job_list_lock);
>  	job = list_first_entry_or_null(&sched->pending_list,
>  				       struct drm_sched_job, list);
>  	if (job && dma_fence_is_signaled(&job->s_fence->finished))
>  		__drm_sched_run_free_queue(sched);
> +}
> +
> +static void drm_sched_run_free_queue_unlocked(struct
> drm_gpu_scheduler *sched)
> +{
> +	spin_lock(&sched->job_list_lock);
> +	drm_sched_run_free_queue(sched);
>  	spin_unlock(&sched->job_list_lock);
>  }

nit:
Took me a few seconds to realize why that's necessary. A sentence in
the commit message might have been good. But no big thing, up to you

>  
> @@ -536,6 +541,32 @@ static void drm_sched_job_begin(struct
> drm_sched_job *s_job)
>  	spin_unlock(&sched->job_list_lock);
>  }
>  
> +/**
> + * drm_sched_job_reinsert_on_false_timeout - Reinsert the job on a
> false timeout
> + *

Please remove this empty line. Our docu style in those files is not
consistent, and I'd like to move towards a more unified style.

> + * @sched: scheduler instance
> + * @job: job to be reinserted on the pending list
> + *
> + * In the case of a "false timeout" - when a timeout occurs but the
> GPU isn't
> + * hung and the job is making progress, the scheduler must reinsert
> the job back
> + * into the pending list. Otherwise, the job and its resources won't
> be freed
> + * through the &drm_sched_backend_ops.free_job callback.
> + *
> + * Note that after reinserting the job, the scheduler enqueues the
> free-job
> + * work again if ready. Otherwise, a signaled job could be added to
> the pending
> + * list, but never freed.
> + *
> + * This function must be used in "false timeout" cases only.
> + */
> +static void drm_sched_job_reinsert_on_false_timeout(struct
> drm_gpu_scheduler *sched,
> +						    struct
> drm_sched_job *job)
> +{
> +	spin_lock(&sched->job_list_lock);
> +	list_add(&job->list, &sched->pending_list);
> +	drm_sched_run_free_queue(sched);
> +	spin_unlock(&sched->job_list_lock);
> +}
> +
>  static void drm_sched_job_timedout(struct work_struct *work)
>  {
>  	struct drm_gpu_scheduler *sched;
> @@ -569,6 +600,14 @@ static void drm_sched_job_timedout(struct
> work_struct *work)
>  			job->sched->ops->free_job(job);
>  			sched->free_guilty = false;
>  		}
> +
> +		/*
> +		 * If the driver indicated that the GPU is still
> running and wants
> +		 * to skip the reset, reinsert the job back into the
> pending list
> +		 * and re-arm the timeout.

Doesn't sound entirely correct to me – at this point, the driver itself
did already skip the reset. The scheduler has no control over that.

You might also just drop the comment, I think the function name and the
function's docstring make what's happening perfectly clear.

> +		 */
> +		if (status == DRM_GPU_SCHED_STAT_NO_HANG)
> +			drm_sched_job_reinsert_on_false_timeout(sche
> d, job);
>  	} else {
>  		spin_unlock(&sched->job_list_lock);
>  	}
> @@ -591,6 +630,9 @@ static void drm_sched_job_timedout(struct
> work_struct *work)
>   * This function is typically used for reset recovery (see the docu
> of
>   * drm_sched_backend_ops.timedout_job() for details). Do not call it
> for
>   * scheduler teardown, i.e., before calling drm_sched_fini().
> + *
> + * As it's used for reset recovery, drm_sched_stop() shouldn't be
> called
> + * if the driver skipped the timeout (DRM_GPU_SCHED_STAT_NO_HANG).

s/timeout/reset

>   */
>  void drm_sched_stop(struct drm_gpu_scheduler *sched, struct
> drm_sched_job *bad)
>  {
> @@ -676,6 +718,9 @@ EXPORT_SYMBOL(drm_sched_stop);
>   * drm_sched_backend_ops.timedout_job() for details). Do not call it
> for
>   * scheduler startup. The scheduler itself is fully operational
> after
>   * drm_sched_init() succeeded.
> + *
> + * As it's used for reset recovery, drm_sched_start() shouldn't be
> called
> + * if the driver skipped the timeout (DRM_GPU_SCHED_STAT_NO_HANG).

same

>   */
>  void drm_sched_start(struct drm_gpu_scheduler *sched, int errno)
>  {
> @@ -1197,7 +1242,7 @@ static void drm_sched_free_job_work(struct
> work_struct *w)
>  	if (job)
>  		sched->ops->free_job(job);
>  
> -	drm_sched_run_free_queue(sched);
> +	drm_sched_run_free_queue_unlocked(sched);
>  	drm_sched_run_job_queue(sched);
>  }
>  
> diff --git a/include/drm/gpu_scheduler.h
> b/include/drm/gpu_scheduler.h
> index
> 83e5c00d8dd9a83ab20547a93d6fc572de97616e..063c1915841aa54a0859bdccd3c
> 1ef6028105bec 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -393,11 +393,14 @@ struct drm_sched_job {
>   * @DRM_GPU_SCHED_STAT_NONE: Reserved. Do not use.
>   * @DRM_GPU_SCHED_STAT_RESET: The GPU hung and successfully reset.
>   * @DRM_GPU_SCHED_STAT_ENODEV: Error: Device is not available
> anymore.
> + * @DRM_GPU_SCHED_STAT_NO_HANG: Contrary to scheduler's belief, the
> GPU
> + * did not hang and it's operational.

s/it's/is

>   */
>  enum drm_gpu_sched_stat {
>  	DRM_GPU_SCHED_STAT_NONE,
>  	DRM_GPU_SCHED_STAT_RESET,
>  	DRM_GPU_SCHED_STAT_ENODEV,
> +	DRM_GPU_SCHED_STAT_NO_HANG,
>  };
>  
>  /**
> 

Thx, I'll look through the other ones soonish, too