[PATCH 8/8] amdgpu: use sync file for shared semaphores (v3)

[Christian König]

Am 12.04.2017 um 12:18 schrieb Chris Wilson:
> On Wed, Apr 12, 2017 at 11:55:24AM +0200, Christian König wrote:
>> Am 12.04.2017 um 06:57 schrieb Dave Airlie:
>>> +static int amdgpu_sem_lookup_and_remove(struct amdgpu_cs_parser *p,
>>> +				      uint32_t handle)
>>> +{
>>> +	int r;
>>> +	struct dma_fence *old_fence;
>>> +
>>> +	r = drm_syncobj_replace_fence(p->filp, handle, NULL);
>>> +	if (r)
>>> +		return r;
>>> +	dma_fence_put(old_fence);
>> Am I wrong or is old_fence not initialized here?
>>
>> Additional to that what happens when the fence in the sync object
>> was changed while we do the CS? Or even worse the handle got
>> assigned to a new sync object.
> We either ww_mutex the lot, or regard that as a userspace race where the
> order between the two concurrent CS emits is undefined and who gets the
> in-semaphore is happenstance.

My thinking to solve this was rather a) resolve the handle to an object 
only once and b) replace the old fence with NULL only if it didn't changed.

Letting userspace race here is also an option, but I would prefer it 
another solution.

Christian.

> -Chris
>