20:18 <marex> austriancoder: it seems I still keep running into

20:18 <marex> [ 480.305964] etnaviv-gpu 59000000.gpu: MMU fault status 0x00000002

20:18 <marex> [ 480.316294] etnaviv-gpu 59000000.gpu: MMU 0 fault addr 0xff680000

20:18 <marex> this crap

20:48 <marex> seems like this bug has been showing up a lot more and on various hardware too

20:48 <marex> https://lists.freedesktop.org/archives/etnaviv/2017-August/000890.html 2017, iMX6

20:49 <marex> https://source.puri.sm/Librem5/linux-emcraft/-/issues/13#note_13766 2019, iMX8MQ

20:49 <marex> https://oftc.irclog.whitequark.org/etnaviv/2021-06-15 2021, iMX8MQ

20:49 <marex> and obviously, stm32mp1, too

20:50 <marex> so that would imply something is actually botched in the mmuv2 code ?

20:50 <marex> because all those GPUs are MMUv2

20:50 <marex> mwalle: mntmn: laanwj: ^

20:52 <mntmn> we got mmu faults on ls1028a for sure

20:52 <marex> oh right, that one too

21:27 <mntmn> btw is there a reason why NIR is still not the default for etnaviv?

21:59 <marex> mntmn: ask austriancoder

21:59 <marex> mntmn: btw did you try enabling devcoredump on the LS ?

22:05 <mntmn> no, IIRC it was quite usable in the end

22:05 <mntmn> hope to get back to ls1028a shenanigans soon

22:06 <mntmn> marex: fwiw we discovered that qtwebengine runs very well now on imx8mq... on par with chromium/angle

22:07 <mntmn> also wayland native which is nice

22:09 <marex> mntmn: for me it fails in rather nasty ways

22:09 <marex> mntmn: every once in a while, there is texture corruption, which makes it unusable

22:09 <mntmn> on imx8mq?

22:09 <marex> mntmn: stm32mp1, mx6qp, and also on mx8mq

22:12 <marex> I wonder if the GPU might be fetching data which the CPU didn't write to DRAM yet

22:13 <mntmn> hmmm ok

22:14 <mntmn> i can tell you that in chromium/blink, it never happens

22:14 <mntmn> like, it's perfect

22:14 <mntmn> but i can try to use qtwebengine browser for a few days...

22:15 <marex> mntmn: stm32mp1 triggers it with stress-ng --cpu 2 in the background (it is dual-core CPU) , iperf3 -P 4, and glmark2-es2-wayland

22:19 <marex> mntmn: look at this

22:19 <marex> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c#n68

22:21 <mntmn> marex: what am i looking at

22:21 <marex> say ... if I call etnaviv_iommuv2_map with entry which is not in either MTLB and STLB

22:22 <marex> then etnaviv_iommuv2_ensure_stlb will allocate new STLB page, memset it, and plug it into MTLB

22:22 <marex> so now, MTLB is aware of this new page, which is all EXCEPTION

22:23 <marex> and only then the new STLB entry is inserted into that new STLB page

22:24 <marex> I wonder if the FE might suck in this intermediate state somehow, and that is what causes the exception

23:05 <mntmn> hm, i am not familiar with mmu workings, but if there is some non atomic creation of state surely it could cause weird effects if the timing is exactly right

23:12 <marex> mntmn: the code is rather cryptic