[PATCH] dma-buf: keep the signaling time of merged fences v3

[Christian König]

Am 30.06.23 um 19:19 schrieb Luben Tuikov:
> On 2023-06-30 08:00, Christian König wrote:
> [SNIP]
> ---
> Silly question perhaps:
> 	Could we not have returned an existing (signalled) fence with
> the wanted timestamp (when count == 0), as opposed to allocating a stub? Maybe
> allocation should be avoided?

No a silly, but a very good question.

Answer is yes, that's what we have done previously. Problem with that 
approach is that each fence reference prevents the driver who originally 
issued this fence from being able to unload.

That's especially ugly in cases of hotplug because you have to keep a 
zombie driver instance around for potentially quite some time.

Regards,
Christian.