[Nouveau] [PATCH 09/17] drm/radeon: use common fence implementation for fences

[Christian König]

Am 23.07.2014 10:54, schrieb Daniel Vetter:
> On Wed, Jul 23, 2014 at 10:46 AM, Christian König
> <deathsimple at vodafone.de> wrote:
>> Am 23.07.2014 10:42, schrieb Daniel Vetter:
>>
>>> On Wed, Jul 23, 2014 at 10:25 AM, Maarten Lankhorst
>>> <maarten.lankhorst at canonical.com> wrote:
>>>> In this case if the sync was to i915 the i915 lockup procedure would take
>>>> care of itself. It wouldn't fix radeon, but it would at least unblock your
>>>> intel card again. I haven't specifically added a special case to attempt to
>>>> unblock external fences, but I've considered it. :-)
>>> Actually the i915 reset stuff relies crucially on being able to kick
>>> all waiters holding driver locks. Since the current fence code only
>>> exposes an opaque wait function without exposing the underlying wait
>>> queue we won't be able to sleep on both the fence queue and the reset
>>> queue. So would pose a problem if we add fence_wait calls to our
>>> driver.
>>
>> And apart from that I really think that I misunderstood Maarten. But his
>> explanation sounds like i915 would do a reset because Radeon is locked up,
>> right?
>>
>> Well if that's really the case then I would question the interface even
>> more, cause that is really nonsense.
> I disagree - the entire point of fences is that we can do multi-gpu
> work asynchronously. So by the time we'll notice that radeon's dead we
> have accepted the batch from userspace already. The only way to get
> rid of it again is through our reset machinery, which also tells
> userspace that we couldn't execute the batch. Whether we actually need
> to do a hw reset depends upon whether we've committed the batch to the
> hw already. Atm that's always the case, but the scheduler will change
> that. So I have no issues with intel doing a reset when other drivers
> don't signal fences.

You submit a job to the hardware and then block the job to wait for 
radeon to be finished? Well than this would indeed require a hardware 
reset, but wouldn't that make the whole problem even worse?

I mean currently we block one userspace process to wait for other 
hardware to be finished with a buffer, but what you are describing here 
blocks the whole hardware to wait for other hardware which in the end 
blocks all userspace process accessing the hardware.

Talking about alternative approaches wouldn't it be simpler to just 
offload the waiting to a different kernel or userspace thread?

Christian.

>
> Also this isn't a problem with the interface really, but with the
> current implementation for radeon. And getting cross-driver reset
> notifications right will require more work either way.
> -Daniel