[PATCH] drm/amdgpu: skip scheduling IBs when GPU recovery

[Christian König]

Am 21.08.20 um 10:57 schrieb Dennis Li:
> If GPU begin to do recovery, skip scheduling IBs. Otherwise
> GPU recovery randomly fail.

NAK, this is prevented by stopping the scheduler during GPU recovery.

If you see jobs submitted to the hardware while we recover the GPU then 
the scheduler is re-started to early.

Regards,
Christian.

>
> Signed-off-by: Dennis Li <Dennis.Li at amd.com>
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> index dcfe8a3b03ff..054d7b0357fd 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> @@ -212,6 +212,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
>   	struct dma_fence *fence = NULL, *finished;
>   	struct amdgpu_job *job;
>   	int r = 0;
> +	int locked;
>   
>   	job = to_amdgpu_job(sched_job);
>   	finished = &job->base.s_fence->finished;
> @@ -220,6 +221,10 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
>   
>   	trace_amdgpu_sched_run_job(job);
>   
> +	locked = down_read_trylock(&ring->adev->reset_sem);
> +	if (!locked)
> +		dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if GPU recovery */
> +
>   	if (job->vram_lost_counter != atomic_read(&ring->adev->vram_lost_counter))
>   		dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */
>   
> @@ -231,6 +236,10 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
>   		if (r)
>   			DRM_ERROR("Error scheduling IBs (%d)\n", r);
>   	}
> +
> +	if (locked)
> +		up_read(&ring->adev->reset_sem);
> +
>   	/* if gpu reset, hw fence will be replaced here */
>   	dma_fence_put(job->fence);
>   	job->fence = dma_fence_get(fence);