[RFC] dma-fence: dma-buf synchronization (v2)

[Rob Clark]

On Fri, Jul 13, 2012 at 4:44 PM, Maarten Lankhorst
<maarten.lankhorst at canonical.com> wrote:
> Hey,
>
> Op 13-07-12 20:52, Rob Clark schreef:
>> On Fri, Jul 13, 2012 at 12:35 PM, Tom Cooksey <tom.cooksey at arm.com> wrote:
>>> My other thought is around atomicity. Could this be extended to
>>> (safely) allow for hardware devices which might want to access
>>> multiple buffers simultaneously? I think it probably can with
>>> some tweaks to the interface? An atomic function which does
>>> something like "give me all the fences for all these buffers
>>> and add this fence to each instead/as-well-as"?
>> fwiw, what I'm leaning towards right now is combining dma-fence w/
>> Maarten's idea of dma-buf-mgr (not sure if you saw his patches?).  And
>> let dmabufmgr handle the multi-buffer reservation stuff.  And possibly
>> the read vs write access, although this I'm not 100% sure on... the
>> other option being the concept of read vs write (or
>> exclusive/non-exclusive) fences.
> Agreed, dmabufmgr is meant for reserving multiple buffers without deadlocks.
> The underlying mechanism for synchronization can be dma-fences, it wouldn't
> really change dmabufmgr much.
>> In the current state, the fence is quite simple, and doesn't care
>> *what* it is fencing, which seems advantageous when you get into
>> trying to deal with combinations of devices sharing buffers, some of
>> whom can do hw sync, and some who can't.  So having a bit of
>> partitioning from the code dealing w/ sequencing who can access the
>> buffers when and for what purpose seems like it might not be a bad
>> idea.  Although I'm still working through the different alternatives.
>>
> Yeah, I managed to get nouveau hooked up with generating irqs on
> completion today using an invalid command. It's also no longer a
> performance regression, so software syncing is no longer a problem
> for nouveau. i915 already generates irqs and r600 presumably too.
>
> Monday I'll take a better look at your patch, end of day now. :)

let me send you a slightly updated version.. I fixed locally some
locking fail in attach_fence() and get_fence() that I managed to
introduce when converting from global spinlock to using the
waitqueue's spinlock.

BR,
-R

> ~Maarten
> --
> To unsubscribe from this list: send the line "unsubscribe linux-media" in
> the body of a message to majordomo at vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html