[PATCH 7/7] drm/syncobj: use the timeline point in drm_syncobj_find_fence

Chunming Zhou zhoucm1 at amd.com
Fri Nov 23 13:15:03 UTC 2018 


在 2018/11/23 20:02, Koenig, Christian 写道:
> Am 23.11.18 um 12:03 schrieb Christian König:
>> Am 23.11.18 um 11:56 schrieb zhoucm1:
>>>
>>> On 2018年11月23日 18:10, Koenig, Christian wrote:
>>>> Am 23.11.18 um 03:36 schrieb zhoucm1:
>>>>> On 2018年11月22日 19:30, Christian König wrote:
>>>>>> Am 22.11.18 um 07:52 schrieb zhoucm1:
>>>>>>> On 2018年11月15日 19:12, Christian König wrote:
>>>>>>>> Implement finding the right timeline point in
>>>>>>>> drm_syncobj_find_fence.
>>>>>>>>
>>>>>>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>>>>>>> ---
>>>>>>>>     drivers/gpu/drm/drm_syncobj.c | 10 +++++++++-
>>>>>>>>     1 file changed, 9 insertions(+), 1 deletion(-)
>>>>>>>>
>>>>>>>> diff --git a/drivers/gpu/drm/drm_syncobj.c
>>>>>>>> b/drivers/gpu/drm/drm_syncobj.c
>>>>>>>> index 589d884ccd58..d42c51520da4 100644
>>>>>>>> --- a/drivers/gpu/drm/drm_syncobj.c
>>>>>>>> +++ b/drivers/gpu/drm/drm_syncobj.c
>>>>>>>> @@ -307,9 +307,17 @@ int drm_syncobj_find_fence(struct drm_file
>>>>>>>> *file_private,
>>>>>>>>             return -ENOENT;
>>>>>>>>           *fence = drm_syncobj_fence_get(syncobj);
>>>>>>>> -    if (!*fence) {
>>>>>>>> +    if (!*fence)
>>>>>>>>             ret = -EINVAL;
>>>>>>>> +
>>>>>>>> +    if (!ret && point) {
>>>>>>>> +        dma_fence_chain_for_each(*fence) {
>>>>>>>> +            if (!to_dma_fence_chain(*fence) ||
>>>>>>>> +                (*fence)->seqno <= point)
>>>>>>>> +                break;
>>>>>>> This condition isn't enough to find proper point.
>>>>>>> For two examples:
>>>>>>> a. No garbage collection happens, the points in chain are
>>>>>>> 1----3----6----9----12---18---20, if user wants to get point17, then
>>>>>>> we should return node 18.
>>>>>> And that is exactly what's wrong in the original logic. In this case
>>>>>> we need to return 12, not 18 because point 17 could have already been
>>>>>> garbage collected.
>>>>> I don't think so, the 'a' case I already assume there isn't garbage
>>>>> collection. If user wants to get point17, then we should return
>>>>> node 18.
>>>>> timeline means point[N]  must be signaled later than point[N-1].
>>>>> Point[12] just can make sure point[1] ~point[12] are signaled.
>>>>> Point[18] signal can make sure point[17] is signaled.
>>>>> So this case we need to return 18, not 12, which is key timeline
>>>>> concept.
>>>> No, exactly that's incorrect. When we ask for 17 and can't find it then
>>>> this means it either never existed or that it is signaled already.
>>>>
>>>> Returning a lower number in this case or even a stub fence is perfectly
>>>> fine since we only need to wait for that one in this case.
>>>>
>>>> If we return 18 in this case then we add incorrect synchronization when
>>>> there shouldn't be any.
>>> No, That will make timeline not work at all and break timeline
>>> semantics totally.
>>>
>>> If there aren't point18 and point20, the chain is
>>> 1----3----6----9----12, if user wants to get point 17, you also
>>> return 12? if yes, which absolutely is incorrect. The answer should
>>> be NO, right? point17 should be waited on there until a bigger point
>>> is coming.
>> Correct, but this is a different case. In this situation we either
>> return an error or wait for point 17 (or something >=17) to show up.
>>
>> The key difference is if point 17 shows up then we return point 17,
>> but if point 18 shows up then we need to return point 12.
>>
>>> For chain is 1----3----6----9----12---18---20, if user wants to wait
>>> on any one of points 13,14,15,16,17,18, we must wait for point 18,
>>> this is timeline semantic.
>> Ah, now I understand. You are still sticking with the assumption of a
>> fence number, right?
>>
>> In other words what you imply here is that we have the same semantic
>> as when somebody waits for a memory location to be written by number
>> 17, right? In this case the semantics you describe here indeed applies.
>>
>> But that is certainly not what we want to implement or otherwise we
>> will never be able to garbage collect the numbers in between.
>>
>> So if Vulkan has this requirement then we need to reject that.
> Backing of and reconsidering this I came to the conclusion that what you
> suggest here is actually the most defensive solution.
>
> In other words it is the solution where it's most likely that nothing
> goes wrong because the worst thing that can happen is that we
> synchronize to much, but never to less.
>
> Going to think about it how we can bring that into alignment with the
> proposed garbage collection.
Yeah, for garbae collection, I came up an idea this morning, we can pass 
the signaled stub fence when chain node is created, when you walk out 
all chain node, you can replace chain->fence with stub fence, that way, 
there is no redudant fence referenced in chain node, and we can keep the 
signal point in there.

Another use case, I'm not sure if you considered:
if chain is  1----3----6----9----12---18, a wait operation is on point 
17, then we return 18, another signal point 17 comes, then we still wait 
on 18(assume 18 takes very long time), that looks not reseonable, but 
this is just performance problem potientially. Seems the way of timeline 
sw_sync.c with comparing point for signal status can sovle it.

-David

>
> Thanks,
> Christian.
>
>> Regards,
>> Christian.
>>
>>> You can also check sw_sync.c for timeline meaning.
>>>
>>> -David
>>>> Christian.
>>>>
>>>>> -David
>>>>>>> b. garbage collection happens on point6, chain would be updated to
>>>>>>> 1---3---9---12---18---20, if user wants to get point5, then we
>>>>>>> should return node 3, but if user wants to get point 7, then we
>>>>>>> should return node 9.
>>>>>> Why? That doesn't seem to make any sense to me.
>>>>>>
>>>>>>> I still have no idea how to satisfy all these requirements with your
>>>>>>> current chain-fence. All these logic just are same we encountered
>>>>>>> before, we're walking them again. After solving these problems, I
>>>>>>> guess all design is similar as before.
>>>>>>>
>>>>>>> In fact, I don't know what problem previous design has, maybe there
>>>>>>> are some bugs, can't we fix these bugs by time going? Who can make
>>>>>>> sure his implementation never have bugs?
>>>>>> Well there where numerous problems with the original design. For
>>>>>> example we need to reject the requirement that timeline fences are in
>>>>>> order because that doesn't make sense in the kernel.
>>>>>>
>>>>>> When userspace does something like submitting fences in the order 1,
>>>>>> 5, 3 then it is broken and can keep the pieces. In other words the
>>>>>> kernel should not care about that, but rather make sure that it never
>>>>>> looses any synchronization no matter what.
>>>>>>
>>>>>> Regards,
>>>>>> Christian.
>>>>>>
>>>>>>> -David
>>>>>>>> +        }
>>>>>>>>         }
>>>>>>>> +
>>>>>>>>         drm_syncobj_put(syncobj);
>>>>>>>>         return ret;
>>>>>>>>     }
>>> _______________________________________________
>>> dri-devel mailing list
>>> dri-devel at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel

More information about the dri-devel mailing list