Re: [PATCH 6/8] drm/panthor: Implement XGS queues

[Simona Vetter <simona.vetter@ffwll.ch>]

On Wed, Aug 28, 2024 at 06:26:02PM +0100, Mihail Atanassov wrote:
> +int panthor_xgs_queue_create(struct panthor_file *pfile, u32 vm_id,
> +			     int eventfd_sync_update, u32 *handle)
> +{
> +	struct panthor_device *ptdev = pfile->ptdev;
> +	struct panthor_xgs_queue_pool *xgs_queue_pool = pfile->xgs_queues;
> +	struct panthor_xgs_queue *queue;
> +	struct drm_gpu_scheduler *drm_sched;
> +	int ret;
> +	int qid;
> +
> +	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
> +	if (!queue)
> +		return -ENOMEM;
> +
> +	kref_init(&queue->refcount);
> +	INIT_LIST_HEAD(&queue->fence_ctx.in_flight_jobs);
> +	INIT_WORK(&queue->release_work, xgs_queue_release_work);
> +	queue->ptdev = ptdev;
> +
> +	ret = drmm_mutex_init(&ptdev->base, &queue->lock);

This is guaranteed buggy, because you kzalloc queue, with it's own
refcount, but then tie the mutex cleanup to the entirely different
lifetime of the drm_device.

Just spotted this while reading around.
-Sima
-- 
Simona Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch