10:09 <tomeu> lynxeye: just to be sure, you are waiting on me to reorder the commits?

10:31 <tomeu> nm, just went ahead and sent a v5

13:10 <marex> MESA: error: etna_cmd_stream_flush:238: submit failed: -28 (No space left on device)

13:10 <marex> hmmmm ... that looks familiar

13:21 <lynxeye> marex: GPU with MMUv2? If so see patches in drm-next.

13:22 <marex> lynxeye: stm32mp15xx , gc400

13:22 <marex> wouldn't surprise me something went wonky again in the last few releases

13:24 <lynxeye> marex: new Mesa got better in triggering a race in the kernel driver fixed by 5a40837debaa + 332f847212e4

13:27 <marex> I'm on 5.10 LTS here, so that shouldn't pose a problem

13:30 <lynxeye> marex: 5.10 implements softpin, which is what is being fixed here. The patches have not been backported to 5.10, as the whole context changed a lot, so backporting is considerable effort.

13:30 <marex> lynxeye: I know

13:30 <marex> lynxeye: is softpin used by the old GPUs too ?

13:31 <lynxeye> As the GC400 doesn't require softpin you could sidestep the issue by falling back to the old relocs in Mesa.

13:32 <lynxeye> Yes, it is used by default on all MMUv2 GPUs. It is only required for GPUs with texture descriptors, though.

13:32 <marex> that isn't GC400 though, is it ?

13:32 <marex> MMUv2 and texture descs

13:32 <lynxeye> GC400 in the STM32MP1 has MMUv2, but no texture descriptors.

13:33 <lynxeye> So softpin is used on this GPU, but you could hack it out in Mesa without much drawbacks.

13:34 <marex> lynxeye: just testing next, let's see

13:34 <marex> so yes, works, much better too

13:34 <marex> thanks

13:34 <marex> but then, I wonder, what to do with 5.10.y which is LTS

13:35 <marex> last time I did the backport, it was as you said, awful

13:35 <lynxeye> marex: If you really care about 5.10 then the only option is to backport all the softpin fixes :/

13:39 <marex> lynxeye: well, lets see, somehow I feel it would be good to have this fixed in LTS kernel releases

13:40 <marex> lynxeye: but thanks for the hint /wrt softpin, that seems to be the issue here