On 13.06.25 23:20, Matthew Brost wrote:
> A small race exists between spsc_queue_push and the run-job worker, in
> which spsc_queue_push may return not-first while the run-job worker has
> already idled due to the job count being zero. If this race occurs, job
> scheduling stops, leading to hangs while waiting on the job’s DMA
> fences.
>
> Seal this race by incrementing the job count before appending to the
> SPSC queue.
>
> This race was observed on a drm-tip 6.16-rc1 build with the Xe driver in
> an SVM test case.
>
> Fixes: 1b1f42d8fde4 ("drm: move amd_gpu_scheduler into common location")
> Fixes: 27105db6c63a ("drm/amdgpu: Add SPSC queue to scheduler.")
> Cc: sta...@vger.kernel.org
> Signed-off-by: Matthew Brost <matthew.br...@intel.com>
Sorry for the late response, if it isn't already pushed to drm-misc-fixes then
feel free to add Reviewed-by: Christian König <christian.koe...@amd.com>
> ---
> include/drm/spsc_queue.h | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/include/drm/spsc_queue.h b/include/drm/spsc_queue.h
> index 125f096c88cb..ee9df8cc67b7 100644
> --- a/include/drm/spsc_queue.h
> +++ b/include/drm/spsc_queue.h
> @@ -70,9 +70,11 @@ static inline bool spsc_queue_push(struct spsc_queue
> *queue, struct spsc_node *n
>
> preempt_disable();
>
> + atomic_inc(&queue->job_count);
> + smp_mb__after_atomic();
> +
> tail = (struct spsc_node **)atomic_long_xchg(&queue->tail,
> (long)&node->next);
> WRITE_ONCE(*tail, node);
> - atomic_inc(&queue->job_count);
>
> /*
> * In case of first element verify new node will be visible to the
> consumer
