08:20vedranm: Ristovski: spent most of the weekend (1.5 hours per compile, then a few minutes transfer, reboot, check) and ended at the wrong commit
08:21vedranm: I'll retry next weekend or the one after
08:21vedranm: but I noticed that it is memory size dependent, it fails at 32 MB, but not at 256 MB VRAM
08:21vedranm: I now wonder if it affects chips other then Kabini
16:28Ristovski: vedranm: You might want to try setting up ccache - it might speed up kernel rebuilds (by how much varies on how many source files have changed in between compiles, should theoretically be faster the longer you bisect as ultimately the last several versions to test should differ less)
18:51vedranm: Ristovski: already did, this is socket 1366 Xeon so it is not very fast overall
19:08Ristovski: vedranm: Ah, I see
19:10vedranm: Ristovski: thanks for the tips, I plan to anyhow check if other chips than Kabini are affected when VRAM is reduced to 32 MB
19:11Ristovski: vedranm: Sadly I no longer have access to my HD 8850M ("Cape Verde"), else I would test that out
19:12Ristovski: Its one gen older than your card
19:56vedranm: Ristovski: yup, SI would be interesting to check
19:58vedranm: I have access to a Carrizo and a Ryzen with Vega, let's see
19:59Ristovski: I have access to a gfx90c/Renoir via Ryzen 5700G, currently set to 256MB VRAM
20:02vedranm: Ristovski: would you mind checking how low can it be set?
20:02Ristovski: My bios allows as low as 32MB, and I have once booted with that just fine
20:02vedranm: FWIW, Kabini with 32 MB and 5.15 kernel doesn't even put the picture on display, but firmware loads fine
20:02vedranm: with 6.1 fimware loading fails
20:03vedranm: if firmware loads fine for you, it is not affected
20:03Ristovski: I believe I upped it back to 256MB since I was experiencing some crashing, but it turned out to be something else
20:03Ristovski: hmm yeah it loaded and worked fine for me
20:56vedranm: Ristovski: just took a quick look and HP's Carrizo and Vega in 2200GE don't allow below 128 MB in BIOS
20:56Ristovski: interesting
20:56vedranm: what board does your 5700G have?
21:10Ristovski: vedranm: Motherboard? Asus B450M-A II
21:12Ristovski: It was the only decent B450 board I could find at the time which also had dedicated USB bios flashback (didn't support Zen3 out of the box, had to flash a new bios)
21:18vedranm: Ristovski: good choice
21:18vedranm: I remember those struggles with Zen 3 and old boards
21:18vedranm: can you really set it as low as 32 MB?
21:19Ristovski: it might have been 64MB actually, but definitely a two digit number
21:19Ristovski: I remember it was maxed out (and spilling over to GTT) on sway even with nothing else open
21:20vedranm: Ristovski: anyhow, if you get a chance to check, I would be grateful
21:20vedranm: any kernel >=5.17 should be affected
21:21vedranm: of course, spills are expected
21:21Ristovski: Will check when I can. I am on 6.10-rc5
21:30johnny0: vedranm: kaveri was also impacted by this change, required the same minimum 256MB UMA or disabling UVD via amdgpu.ip_block_mask
21:38Ristovski: johnny0: so its the UVD block that breaks actually? interesting
21:43agd5f: not sure what issue you are chasing, but on those old APUs like kabini and kaveri, both video encode/decode and display buffers had to be in UMA (carve out), so if you only have 32M, that's not much space, especially if you have multiple displays
21:44vedranm: Ristovski: loading UVD firmware breaks
21:45johnny0: Ristovski: https://lore.kernel.org/lkml/67e2904a-4b80-5173-2d8f-6a3d1f702fbb@amd.com/ is what I have saved in my notes when I hit the issue
21:46Ristovski: I see
21:47vedranm: johnny0: I see
21:48vedranm: how come this isn't an issue on radeon then?
21:49vedranm: and I don't get "Move buffer fallback to memcpy unavailable" and -19 error code
21:49vedranm: it's -22: https://gitlab.freedesktop.org/drm/amd/-/issues/3448
22:12johnny0: vedranm: Yeah, it was -22 for me too. That list message indicated a bigger UMA buffer would probably help though (and it did).
22:36johnny0: vedranm: as far as why radeon isn't impacted, IIRC the drm_buddy changes weren't backported to radeon