05:11 <zhxt> hello, learned from the panfrost doc, I saw panfrost had already supported Mali G57 GPU now. I have a device with Mali G57 GPU running debian11 . Is there any guide/doc shows how to make the GPU work?

05:12 <zhxt> Any hints? thank a lot!

12:27 <CounterPillow> zhxt_: You need a more recent mesa than what Debian 11 ships iirc

12:28 <CounterPillow> furthermore you probably need to put something like governor_simpleondemand into your initramfs because Debian builds all of the devfreq drivers as modules which can make panfrost fail to load due to dependency issues when it's loaded

13:01 <alyssa> I was not expecting to find GenXML generated templates in someone else's blog: https://github.blog/2022-07-27-corrupting-memory-without-memory-corruption/

13:02 <alyssa> The funniest part to me is that kbase's memory aliasing ioctl is probably unused.

13:03 <alyssa> As I understand, it's a hack to emulate MRT on v4.

13:03 <alyssa> Mali T760 and newer shouldn't need it.

13:03 <alyssa> (Because Panfrost doesn't have this hack, Mali-T720 and older are capped at GL2.)

13:07 <alyssa> What's even funnier is that they use the immediate value selectors of write value jobs in their exploitation, but those are only introduced in Bifrost.

13:09 <alyssa> (That doesn't affect exploitability. As noted in the blog, they could have crafted a compute kernel to do the same work and that would've worked on the Mali-T600 class GPU the bug actually affects. But that's more annoying.)

13:22 <zhxt> CounterPillow, thank you, good to know. Is that enough? Do I need to do something with kernel? the kernel version is 5.10, which come from chromiumos-kernel.

13:23 <alyssa> zhxt: chromiumos builds the wrong GPU driver

13:23 <alyssa> You want a mainline kernel

13:23 <alyssa> (or at the very least, your own build of chromiumos-kernel with the MALI config options =n and CONFIG_DRM_PANFROST=y)

13:24 <CounterPillow> I'd use as recent of a kernel as I could if I was aiming for support of recent hardware.

13:29 <zhxt> alyssa: okay, thank you, I think I had understood now, :D

13:30 <zhxt> CounterPillow, the mainline/recent kernel may need more effort to run on the device, I think,

13:31 <CounterPillow> what device is it?

13:31 <CounterPillow> the SoC, specifically

13:31 <zhxt> mt8195 ,

13:31 <CounterPillow> I see, yeah I don't know anything about the mainline support of that

13:33 <zhxt> okay, maybe I can go and check the git logs.

13:52 <zhxt> seems alyssa was working on mt8195 several months ago :D https://gitlab.freedesktop.org/panfrost/linux/-/commits/mt8195-6

14:12 <alyssa> zhxt: That branch was a typo, it was my MT8192 branch

14:12 <alyssa> I don't have an MT8195 machine (and with the same GPU as MT8192 I can't really justify getting one)

14:13 <alyssa> MT819x mainlining is happening but not "there yet"

14:13 <alyssa> spearheaded by the good folks over in the kernel group at Collaboa

14:14 <alyssa> https://gitlab.collabora.com/google/chromeos-kernel/-/tree/mt8195-tracking-master-rolling

14:14 <alyssa> that's mainline + in-progress patches for MT8195

14:15 <alyssa> I'm not sure I'd recommend daily-driving it (IIRC it's based on linux-next, not a release)

14:15 <alyssa> but things should mostly 'just work' on that branch, including the GPU

14:15 <alyssa> (with appropriate firmware and recent Mesa)

15:03 <zhxt> alyssa: ah, thank you very much for the info, that helps me a lot :D