[PATCH] dma-buf/fence-array: fix deadlock in fence-array

[Rob Clark]

On Mon, Oct 17, 2016 at 2:52 PM, Gustavo Padovan <gustavo at padovan.org> wrote:
> 2016-10-17 Rob Clark <robdclark at gmail.com>:
>
>> Currently with fence-array, we have a potential deadlock situation.  If we
>> fence_add_callback() on an array-fence, the array-fence's lock is acquired
>> first, and in it's ->enable_signaling() callback, it will install cb's on
>> it's array-member fences, so the array-member's lock is acquired second.
>>
>> But in the signal path, the array-member's lock is acquired first, and the
>> array-fence's lock acquired second.
>>
>> To solve that, always enabling signaling up-front (in the fence_array
>> constructor) without the fence_array's lock held.
>
> Do we always want to enable signaling for arrays? One of the things we
> removed from the Sync Framework was the need to enable signalling at
> creation time.
>
> Just merging fencing doesn't mean you want signaling, that is supposed
> to happen only when poll() is called on the sync file.

It was something Maarten suggested, as an alternative to introducing a
wq into the mix or worse hacks..
https://lists.freedesktop.org/archives/dri-devel/2016-October/120868.html

I think I agree with him that it is an optimization that is unlikely
to be useful in the case of fence-arrays.  If you need to wait on
multiple fences from different timelines, you probably aren't doing
that in hw.

BR,
-R

> Gustavo
>