Re: [RFC PATCH] dma-fence: Fix races of fence callbacks versus destructors by locking

[Christian König]

On 6/9/26 15:57, Philipp Stanner wrote:
> On Tue, 2026-06-09 at 14:36 +0100, Tvrtko Ursulin wrote:
>> I wanted to ask something else.
>> What happened to the idea to remove opportunistic signalling from 
>> dma_fence_is_signaled?
> 
> AFAIAR that was declared impossible because some system setups like
> mobile devices don't signal the fences with an interrupt handler, but
> have some sort of daemon (in userspace) who peridiocally wakes up to
> update the graphics output. It wakes up every T milliseconds and
> opportunistically asks through dma_fence_is_signaled() (and that
> through ops->signaled() which fences are already signaled.

That's a pretty good summary we should probably add it to the documentation.

Just one additional note: The userspace deamon is usually the compositor who is in control of the CRTCs and the interrupt is the periodical refresh timer, usually vblank.

Christian.

> 
> P.