[PATCH v4] dma-fence: Propagate errors to dma-fence-array container

[Koenig, Christian]

How about this instead:

Setting array->base.error = 1 during initialization.

Then cmpxchg(array->base.error, 1, error) whenever a fence in the array 
signals.

And then finally cmpxchg(array->base.error, 1, 0) when the array itself 
signals.

Christian.

Am 11.08.19 um 14:21 schrieb Chris Wilson:
> When one of the array of fences is signaled, propagate its errors to the
> parent fence-array (keeping the first error to be raised).
>
> v2: Opencode cmpxchg_local to avoid compiler freakout.
> v3: Be careful not to flag an error if we race against signal-on-any.
> v4: Same applies to installing the signal cb.
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Sumit Semwal <sumit.semwal at linaro.org>
> Cc: Gustavo Padovan <gustavo at padovan.org>
> Cc: Christian König <christian.koenig at amd.com>
> ---
>   drivers/dma-buf/dma-fence-array.c | 37 ++++++++++++++++++++++++++++++-
>   include/linux/dma-fence-array.h   |  2 ++
>   2 files changed, 38 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/dma-buf/dma-fence-array.c b/drivers/dma-buf/dma-fence-array.c
> index 12c6f64c0bc2..4d574dff0ba9 100644
> --- a/drivers/dma-buf/dma-fence-array.c
> +++ b/drivers/dma-buf/dma-fence-array.c
> @@ -23,10 +23,37 @@ static const char *dma_fence_array_get_timeline_name(struct dma_fence *fence)
>   	return "unbound";
>   }
>   
> +static void dma_fence_array_set_error(struct dma_fence_array *array)
> +{
> +	int error = READ_ONCE(array->pending_error);
> +
> +	if (!array->base.error && error)
> +		dma_fence_set_error(&array->base, error);
> +}
> +
> +static void dma_fence_array_set_pending_error(struct dma_fence_array *array,
> +					      int error)
> +{
> +	/*
> +	 * Propagate the first error reported by any of our fences, but only
> +	 * before we ourselves are signaled.
> +	 *
> +	 * Note that this may race with multiple fences completing
> +	 * simultaneously in error, but only one error will be kept, not
> +	 * necessarily the first. So long as we propagate an error if any
> +	 * fences were in error before we are signaled we should be telling
> +	 * an acceptable truth.
> +	 */
> +	if (error && !array->pending_error)
> +		WRITE_ONCE(array->pending_error, error);
> +}
> +
>   static void irq_dma_fence_array_work(struct irq_work *wrk)
>   {
>   	struct dma_fence_array *array = container_of(wrk, typeof(*array), work);
>   
> +	dma_fence_array_set_error(array);
> +
>   	dma_fence_signal(&array->base);
>   	dma_fence_put(&array->base);
>   }
> @@ -38,6 +65,8 @@ static void dma_fence_array_cb_func(struct dma_fence *f,
>   		container_of(cb, struct dma_fence_array_cb, cb);
>   	struct dma_fence_array *array = array_cb->array;
>   
> +	dma_fence_array_set_pending_error(array, f->error);
> +
>   	if (atomic_dec_and_test(&array->num_pending))
>   		irq_work_queue(&array->work);
>   	else
> @@ -63,9 +92,14 @@ static bool dma_fence_array_enable_signaling(struct dma_fence *fence)
>   		dma_fence_get(&array->base);
>   		if (dma_fence_add_callback(array->fences[i], &cb[i].cb,
>   					   dma_fence_array_cb_func)) {
> +			int error = array->fences[i]->error;
> +
> +			dma_fence_array_set_pending_error(array, error);
>   			dma_fence_put(&array->base);
> -			if (atomic_dec_and_test(&array->num_pending))
> +			if (atomic_dec_and_test(&array->num_pending)) {
> +				dma_fence_array_set_error(array);
>   				return false;
> +			}
>   		}
>   	}
>   
> @@ -141,6 +175,7 @@ struct dma_fence_array *dma_fence_array_create(int num_fences,
>   	array->num_fences = num_fences;
>   	atomic_set(&array->num_pending, signal_on_any ? 1 : num_fences);
>   	array->fences = fences;
> +	array->pending_error = 0;
>   
>   	return array;
>   }
> diff --git a/include/linux/dma-fence-array.h b/include/linux/dma-fence-array.h
> index 303dd712220f..faaf70c524ae 100644
> --- a/include/linux/dma-fence-array.h
> +++ b/include/linux/dma-fence-array.h
> @@ -42,6 +42,8 @@ struct dma_fence_array {
>   	atomic_t num_pending;
>   	struct dma_fence **fences;
>   
> +	int pending_error;
> +
>   	struct irq_work work;
>   };
>