Re: Include request for reset-rework branch v4

["Christian König" <deathsimple@vodafone.de>]

On 03.05.2012 19:20, Alex Deucher wrote:
> 2012/5/3 Jerome Glisse<j.glisse@gmail.com>:
>> On Thu, May 3, 2012 at 12:39 PM, Christian König
>> <deathsimple@vodafone.de>  wrote:
>>> On 03.05.2012 18:32, Jerome Glisse wrote:
>>>> On Thu, May 3, 2012 at 4:19 AM, Christian König<deathsimple@vodafone.de>
>>>>   wrote:
>>>>> On 02.05.2012 18:01, Jerome Glisse wrote:
>>>>>> On Wed, May 2, 2012 at 9:11 AM, Christian König<deathsimple@vodafone.de>
>>>>>>   wrote:
>>>>>>> Hi Dave,
>>>>>>>
>>>>>>> there still seems to be the need for some further discussion about the
>>>>>>> SA
>>>>>>> code,
>>>>>>> so I again split that out of the patchset and tested the result a bit.
>>>>>>>
>>>>>>> Most of the stuff still works fine without those offending changes, so
>>>>>>> to
>>>>>>> avoid
>>>>>>> mailing around unrelated and already reviewed patches, I request the
>>>>>>> include
>>>>>>> the following 17 patches into drm-next.
>>>>>>>
>>>>>>> If you prefer to merge they are also available from
>>>>>>> git://people.freedesktop.org/~deathsimple/linux branch reset-rework.
>>>>>>>
>>>>>>> Cheers,
>>>>>>> Christian.
>>>>>>>
>>>>>> I am ok with this 17 patchset, i just sent 3 patch on top of those 17
>>>>>> that
>>>>>> bring back some other of the previous cleanup.
>>>>> At least for now those three are NAK, cause I just realized we need to
>>>>> put
>>>>> those on top of a more sophisticated fence implementation.
>>>>>
>>>>> Your idea of not using a list, but 64 bit sequences instead actually
>>>>> sounds
>>>>> quite nifty to me. Going to hack something together in the next couple of
>>>>> hours.
>>>>>
>>>>> Christian.
>>>> Btw you said that you are having issue when using multiple ring, it
>>>> comes to my attention that you never sync with the GFX ring (unless
>>>> asked by userspace) that's wrong, before scheduling on another ring
>>>> than GFX index you need to emit semaphore to make the ring wait for
>>>> the last emited fence on the GFX ring because of ttm. What might
>>>> happen is that ttm scheduled bo move on the GFX ring and that the ring
>>>> you work on start using those bo at there soon to be GPU offset while
>>>> the bo data is not there yet.
>>> Yeah, already stumbled over that but IIRC Alex already solved that issue..
>>>
>>> We set the sync_obj to the fence of the move operation, so when there is a
>>> move operation in progress we sync to it correctly (at least I hope so).
>>>
>>> Christian.
>>>
>> Looking at code doesn't seems ok, yes you use the fence from the move
>> operation but you only emit wait if userspace ask for it, that's
>> wrong.
>>
>> if (!(p->relocs[i].flags&  RADEON_RELOC_DONT_SYNC)) {
>>
> The default is to sync all the rings.  We only skip the sync if the
> _DONT_ sync flag is set.
Yeah, but the _DONT_ sync flag is just an optimization, and we only want 
to not sync to things userspace has submitted.

  E.g. userspace tells us that two jobs can happily run on the same bo 
even if they are both writing to it....

So Jerome is completely right, userspace doesn't expect that TTM is 
jumping in between and moving the bo around, that is indeed a bug and 
another thing for the todo list.

Christian.