00:01 <alyssa> note to self: the infra for backwards opt passes has a serious footgun -- if you grab sources from the later instruction, it'll fuse them in earlier than the instructions producing those sources

00:02 <alyssa> propagating clamps and conversions doesn't hit this, but a new rule I tried does

00:02 <alyssa> the obvious way of fixing this may hurt scheduling

06:42 <Elemental> I have a Lenovo Duet with a Mali G72. I installed Linux on it using the Cadmium project (Debian based), Kernel 5.16.0-rc3, and am currently using Gnome. I'm getting several lockups per day. Some of them are hard lockups and others "soft" where I can still SSH in. Lots of panfrost errors in the Kernel log: gpu sched timeout, AS_ACTIVE bit stuck, page faults, etc. Any tips on troubleshooting this would be appreciated. Trying to figure out if it is a h

06:44 <Elemental> Trying to figure out if it is a hardware problem or not, since I just got it recently and can still return it. I did get at least one hard freeze in ChromeOS prior to installing Linux.

10:04 <tomeu> Elemental: hmm, not sure, we do test G72 in CI, but might not hit the same codepaths as gnome-shell and some apps

10:04 <tomeu> maybe macc24 will know something about it

11:05 <hexdump0815> i can confirm the observations of Elemental - i real world usage i also see quite a few panfrost related issues

11:06 <hexdump0815> this is with debian bullseye and a 5.15 kernel with mostly the cadmium mt8183 patches and mesa 21.3.2 - will have to move to sid and mesa head to see if this changes things a bit

11:06 <macc24> hmm

11:07 <hexdump0815> can it maybe be realted to the gpu frequency scaling?

11:07 <macc24> no clue

11:07 <macc24> i've been running with gpu devfreq on mt8183 for some time and never seen issues

11:08 <macc24> next week i'll be able to test panfrost from mesa-main on my krane

11:08 <hexdump0815> i see the exact same kinds of errors in the logs like Elemental and in my case its in xorg xfce (wm compositing is disabled) and some light opengl apps now and then

11:09 <macc24> >xorg

11:10 <hexdump0815> macc24: my plan is to move to sid and mesa head and maybe play with setting min=max freq for the gpu and see if i can find anything reproducable

11:11 <macc24> hexdump0815: check if cci devfreq impacts it too

11:11 <macc24> since that's not in ci too

11:11 <hexdump0815> macc24: it should be possible to build a kernel without the cci patches - right?

11:11 <macc24> hexdump0815: just set the governor to powersave/performance and it'll be fine

11:12 <hexdump0815> ok will try

11:15 <macc24> bbl

16:08 <Elemental> I woke up this morning thinking along the same lines, in regards to potential issues with frequency scaling or temperature.

16:08 <Elemental> One of the logged messages on boot is "[drm:panfrost_devfreq_init] Failed to register cooling device". There isn't a fan, but maybe there is some sort of other cooling method which isn't being used? Is there a temperature sensor I could monitor?

16:10 <Elemental> I'll try playing with the min/max frequency scaling as well (first limit it to minimum only) as hexdump0815 suggested.

17:55 <hexdump0815> Elemental: i think its not a temperature issue - as i understand it there is maybe sometimes a moment of instability when switching between frequencies if not everything is setup perfectly which might result in those timeouts, errors etc.

18:08 <Elemental> hexdump0815: I think you're right about that. It has been running fine now for a while with max_freq set to min_freq (300 MHz).

18:10 <Elemental> If it is still working fine by later today, I'll try setting min_freq and max_freq to 800 MHz, to see if it is related to the higher speed or the act of switching, as you suggest.

21:44 <alyssa> interesting...

21:44 <alyssa> soft_reset() works the first time it's used, but not the second

21:44 <alyssa> (Did the GPU hang in between?)

21:45 <macc24> alyssa: poke the gpu to go back into the reset vector, 0xffff f000

21:45 <macc24> eh, 0xffff fff0

21:46 <urja> -.-

21:46 <alyssa> Wondering if this is where the dummy job workaround comes in

21:52 <macc24> oh

21:52 <macc24> valhall

21:56 <alyssa> Yes

21:57 <alyssa> Assuming this is the dummy job thing .. guess I have a pile of code to write then