10:59 <mntmn> problems on the ls1028a version of gc7000 are mostly caused by something in xwayland/glamor i think

10:59 <mntmn> [ 1063.420138] etnaviv-gpu f0c0000.gpu: AXI bus error

10:59 <mntmn> [ 1063.424952] etnaviv-gpu f0c0000.gpu: MMU fault status 0x00000002

10:59 <mntmn> [ 1063.430979] etnaviv-gpu f0c0000.gpu: MMU 0 fault addr 0xfc0acec0

10:59 <mntmn> axi bus error is a new one!

11:00 <mntmn> because apps, even big ones like firefox, that use wayland directly don't seem to cause such problems

11:02 <lynxeye> mntmn: AXI bus error most likely hints to a pagetable being programmed incorrectly. I guess some >4GB physical address issue, as that is something I couldn't test properly, yet.

11:03 <mntmn> lynxeye: i see. the rdb has only 4GB but part of it including the cma area is >4GB i guess

11:04 <mntmn> lynxeye: btw your malidp sync fix worked, the only downside is that it cuts framerate in half for any bigger paints (roughly speaking).

11:04 <mntmn> but it makes everything much more usable

11:06 <lynxeye> mntmn: Yea, I guess the DRAM isn't starting at MMIO address 0, right? On i.MX8MQ I can't test this, as the GPU doesn't have access to any DRAM above the 4GB mark, as the interconnect only has 32 address bits, so the upper 1GB in a 4GB system are inaccessible to the GPU. :/

11:07 <lynxeye> mntmn: Broken sync is good for benchmarks, but not so much for real usage. ;)

11:07 <mntmn> haha

11:07 <lynxeye> Perf should also get better once the missing supertiling bits are figured out.

11:07 <mntmn> yes, i think so too

11:08 <mntmn> there is a dramatic speed difference between running quake at 720p vs 1080p so some kind of bandwidth bottleneck (perhaps because of no tiling)

11:10 <pcercuei> if you use glquake, it's not really suited for today's GPUs

11:10 <lynxeye> The malidp doesn't understand the Vivante tiling I guess. So we are back to blitting from tiled render target to linear scanout buffer, which is very unfortunate, as it's a huge waste of DRAM bandwidth.

11:12 <mntmn> if i'm not mistaken, the dram starts at 0x80000000, but only the first ~2GB, and the rest is at 0x2080000000

11:24 <mwalle> mntmn: thats correct

11:24 <mwalle> only the lower 2GB are mapped <32bits

11:25 <mntmn> tried now placing cma @ 0xc0000000, not sure if that helps.

11:25 <mwalle> btw i'm just cleanup up my branch a bit

12:05 <mwalle> mntmn: https://github.com/mwalle/linux/tree/ls1028-dp-cleanup << if you're interested

12:06 <mwalle> keep in mind that is wip and right now, just compile time tested, although it used to work before the commit history cleanup

12:06 <mwalle> *it is

12:06 <mntmn> mwalle: thanks a lot, i will take a look

12:07 <mntmn> i am certain now that hangs and messed up state are caused by something in xwayland/glamor, because chromium 93 with ozone works perfectly fine, but running it through xwayland causes hangs regularly