[PATCH] drm/radeon: fence virtual address and free it once idle v3

[Jerome Glisse]

On Mon, Aug 6, 2012 at 12:55 PM, Christian König
<deathsimple at vodafone.de> wrote:
> On 06.08.2012 18:30, Jerome Glisse wrote:
>>
>> On Mon, Aug 6, 2012 at 12:06 PM, Christian König
>> <deathsimple at vodafone.de> wrote:
>>>
>>> [SNIP]
>>>
>>> Additional to that patch we still need a minor fix to mesa (just move
>>> freeing the VM range after closing the handle). Going to send that in the
>>> next minute to the mesa-dev list.
>>>
>>> Otherwise the patch indeed fixes the problem, but also isn't the best
>>> idea
>>> for performance.
>>>
>>> Cheers,
>>> Christian.
>>>
>> This does not impact performance, it's all about destroying bo/va so
>> it's not a performance path. Or am i missing something here ?
>
>
> radeonsi currently allocates very small BOs (8-32 bytes) for T# descriptors,
> and throws them away immediately after drawing.
>
> That alone of course is insane under a performance aspect, but waiting for
> the last job to finish makes things completely worse.
>
> It just moves my priorities of hacking on radeonsi a bit around.
>
> Christian.

Well we could simply just not register any close callback for gem and
let the ttm bo delayed queue handle thing, that way userspace won't
stall.

Cheers,
Jerome