[Intel-gfx] [PATCH 1/2] intel: Recount fences after rewinding relocations

[Daniel Vetter]

On Wed, May 8, 2013 at 6:39 PM, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> On Wed, May 08, 2013 at 06:11:19PM +0200, Daniel Vetter wrote:
>> On Wed, May 08, 2013 at 04:33:09PM +0100, Chris Wilson wrote:
>> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=59771
>> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
>>
>> I guess we've walked past a "please stop and return" sign here a bit.
>> Originally this was just add for opportunistic state emission in mesa (and
>> still documents that check_aperture will return fobar). But then uxa
>> started to abuse it to reset batches completely in
>>
>> commit 441ef916ae6569c88b3d6abaf7fea4d69be49d76
>> Author: Chris Wilson <chris at chris-wilson.co.uk>
>> Date:   Thu Jan 10 19:14:21 2013 +0000
>>
>>     intel: Throttle harder
>>
>> which required some band-aids in libdrm to have some semblances of working
>>
>> commit fdda97007b1dbf95beb16a0e3510fd36c89e8c33
>> Author: Chris Wilson <chris at chris-wilson.co.uk>
>> Date:   Fri Jan 11 00:55:12 2013 +0000
>>
>>     intel: Remove the fence count contributions when clearing relocs
>>
>> Apparently not yet good enough. Personally I vote for reverting the libdrm
>> patch and doing a normal wait_rendering + unref/bo realloc for the uxa
>> throttling.
>
> The second patch should in theory be the fix for the root cause in the
> miscounting here. I left this here as it makes the count obviously
> correct irrespective of that patch (though it still grossly overcounts
> fences as there should be many relocations from the batch to the same
> tiled bo). The libdrm rewind interface is broken as designed and should
> be shot unless you actually want to fix it up.

The relocs rewind works insofar as it just removes the relocs, but
only if the very next thing you do after unwinding is submitting the
batch. Otherwise it's fireworks. Hence why I've suggested to also
revert the already committed duct-tape and use something else in uxa.

I've read through libdrm a bit more and I've thought that this should
be enough, but atm I'm too blind to find the bo_gem->reloc_tree_fences
reset on buffer reuse :(
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch