Re: [PATCH 07/10] dma-buf/resv: add new fences container implementation

["Koenig, Christian" <Christian.Koenig@amd.com>]

Am 22.08.19 um 15:02 schrieb Daniel Vetter:
> On Thu, Aug 22, 2019 at 10:23:29AM +0200, Christian König wrote:
>> Am 21.08.19 um 18:04 schrieb Daniel Vetter:
>>> On Wed, Aug 21, 2019 at 02:31:44PM +0200, Christian König wrote:
>>>> [SNIP]
>>>> +	/* Try to drop the last reference */
>>>> +	if (!dma_fence_array_recycle(staged))
>>> Without an rcu barrier here you're not syncing to new clients at all.
>>> I don't think this works, and I expect that once you've readded all the
>>> barriers and retry loops we're back to seqlocks.
>> The key difference is that RCU users now use dma_fence_get_rcu_safe() to
>> grab a reference to the current set of fences.
>>
>> In other words the whole array is reference counted and RCU protected
>> instead of each individual entry in the array.
>>
>> This way you don't need the sequence count any more because you grab a
>> reference to all of them at once and then can be sure that they don't
>> change.
> Hm yeah ... I think there's still some users left that have an open-coded
> rcu section though. But yeah if you can concince Chris that this is ok I
> think it makes sense as an overall cleanup of the hand-rolled fences array
> we have for shared fences. But I'd really like to untangle it from the
> entire semantics discussion, since that seems entirely unrelated.

Yeah, agree. To untangle that is a really good idea.

Going to send out the dma_fence_array as a replacement for shared fences 
separately first.

Christian.

> -Daniel