[PATCH 4/5] drm/amdgpu:cleanup job reset routine

[Liu, Monk]

> You need to walk the list of entities (with lock held) and check if 
entity->fence_context == job->s_fence->scheduled.context and only if you
found the right one set the guilty flag there.


Yeah you remind me, I did this in the old gpu reset patch, will add them back !

BR Monk
-----Original Message-----
From: Christian König [mailto:ckoenig.leichtzumerken at gmail.com] 
Sent: 2017年10月20日 17:07
To: Liu, Monk <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org
Subject: Re: [PATCH 4/5] drm/amdgpu:cleanup job reset routine

Am 20.10.2017 um 05:33 schrieb Monk Liu:
> merge the setting guilty on context into this function to avoid 
> implement extra routine.
>
> Change-Id: I7a0063464fdc85d5ac9080046380e745565ff540
> Signed-off-by: Monk Liu <Monk.Liu at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |  4 ++--
>   drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 13 ++++++++++++-
>   drivers/gpu/drm/amd/scheduler/gpu_scheduler.h |  2 +-
>   3 files changed, 15 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index dae1e5b..a07544d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -2868,7 +2868,7 @@ int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, struct amdgpu_job *job)
>   			amd_sched_job_kickout(&job->base);
>   
>   		/* only do job_reset on the hang ring if @job not NULL */
> -		amd_sched_hw_job_reset(&ring->sched);
> +		amd_sched_hw_job_reset(&ring->sched, NULL);
>   
>   		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
>   		amdgpu_fence_driver_force_completion(ring);
> @@ -2989,7 +2989,7 @@ int amdgpu_gpu_reset(struct amdgpu_device *adev)
>   		if (!ring || !ring->sched.thread)
>   			continue;
>   		kthread_park(ring->sched.thread);
> -		amd_sched_hw_job_reset(&ring->sched);
> +		amd_sched_hw_job_reset(&ring->sched, NULL);
>   		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
>   		amdgpu_fence_driver_force_completion(ring);
>   	}
> diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c 
> b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
> index e4d3b4e..322b6c1 100644
> --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
> +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
> @@ -442,7 +442,14 @@ static void amd_sched_job_timedout(struct work_struct *work)
>   	job->sched->ops->timedout_job(job);
>   }
>   
> -void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched)
> +static void amd_sched_set_guilty(struct amd_sched_job *s_job) {
> +	if (atomic_inc_return(&s_job->karma) > s_job->sched->hang_limit)
> +		if (s_job->s_entity->guilty)
> +			atomic_set(s_job->s_entity->guilty, 1);

The entity might be long freed at this point.

You need to walk the list of entities (with lock held) and check if 
entity->fence_context == job->s_fence->scheduled.context and only if you
found the right one set the guilty flag there.

Christian.

> +}
> +
> +void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched, struct amd_sched_job *bad)
>   {
>   	struct amd_sched_job *s_job;
>   
> @@ -455,6 +462,10 @@ void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched)
>   			s_job->s_fence->parent = NULL;
>   			atomic_dec(&sched->hw_rq_count);
>   		}
> +
> +		if (bad && bad == s_job) {
> +			amd_sched_set_guilty(s_job);
> +		}
>   	}
>   	spin_unlock(&sched->job_list_lock);
>   }
> diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
> index 2d59fc5..fff7cc7 100644
> --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
> +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
> @@ -172,7 +172,7 @@ int amd_sched_job_init(struct amd_sched_job *job,
>   		       struct amd_gpu_scheduler *sched,
>   		       struct amd_sched_entity *entity,
>   		       void *owner);
> -void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched);
> +void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched, struct amd_sched_job *job);
>   void amd_sched_job_recovery(struct amd_gpu_scheduler *sched);
>   bool amd_sched_dependency_optimized(struct dma_fence* fence,
>   				    struct amd_sched_entity *entity);