Re: [PATCH 2/6] drm/amdgpu: Mark contexts guilty for causing soft recoveries

[Christian König]

Am 11.07.23 um 23:34 schrieb André Almeida:
> If a DRM fence is set to -ENODATA, that means that this context was a
> cause of a soft reset, but is never marked as guilty. Flag it as guilty
> and log to user that this context won't accept more submissions.
>
> Signed-off-by: André Almeida <andrealmeid@xxxxxxxxxx>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 6 ++++++
>   1 file changed, 6 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> index 0dc9c655c4fb..fe8e47d063da 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> @@ -459,6 +459,12 @@ int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
>   	ctx_entity = &ctx->entities[hw_ip][ring]->entity;
>   	r = drm_sched_entity_error(ctx_entity);
>   	if (r) {
> +		if (r == -ENODATA) {
> +			DRM_ERROR("%s (%d) context caused a reset,"
> +				  "marking it guilty and refusing new submissions.\n",
> +				  current->comm, current->pid);
> +			atomic_set(&ctx->guilty, 1);
> +		}

I'm going back and forth with that as well.

Michel has a very good point that it often is sufficient to cancel just 
one rough shader to keep going.

But Marek has a very good point as well that when that happens multiple 
times we probably want to block the application from making further 
submissions.

Christian.

>   		DRM_DEBUG("error entity %p\n", ctx_entity);
>   		return r;
>   	}