00:02 <robclark> I suppose that is plausible

00:28 <steev> yeah, seems like it's heavy cpu usage that does it.... maybe some bad interaction with the new lmh driver?

00:32 <robclark> I mean, I'd expect normal userspace cpu busyness shouldn't delay irq's, but possibly lots of file I/O plus badly behaved other drivers could

01:16 <steev> hm, let me try something

01:24 <steev> # Compiler tasks

01:24 <steev> "^(make|cargo|rustc|rls|cicc|gcc|git)" = { priority = 20, policy = "idle" }

01:24 <steev> so that's in my process scheduler config

01:24 <steev> and if i disable it, it compiles without running into the enc issue

01:25 <steev> unless

01:28 <steev> i wonder if something is screwy in the scheduler

01:28 <steev> kernel side, i mean

01:32 <robclark> well, I mean, approximately 99.9% of the perf issues we've been dealing with are scheduler (or schedutil/boost, bit+little, etc) related.. but you'd still kinda hope that userspace threads (espc non-rt ones) aren't going to be blocking irqs

01:33 <robclark> *big+little

02:15 <steev> the big issue with when it happens is, there's no kinda recovery

02:15 <steev> reboot is the only way

02:15 <steev> oh! speaking of which

02:15 <steev> bamse: so, i made a change to my systemd config -

02:15 <steev> didn't mean to send so early

02:17 <steev> https://github.com/puavo-org/puavo-os/commit/240a73ebc99c1708a29eb66b988f60ec650f0a6d - changing the default timeout for apps to die from 90s to 8s seems to help a lot with rebooting - even rebooting via ssh

02:19 <steev> also... i've been trying out budgie desktop, and, i just don't get what is so awesome about it

02:28 <robclark> steev: yeah, maybe there is some edge case which things don't recover from gracefully.. you should probably ask abhinav (who seems not to be here but is on #freedreno), since the odds of me having a chance to dig into that are low

02:30 <steev> okay let me find the paste so i can pass it along and i'll do that

07:11 <macc24> is there a way to make linux kernel not use big cores at all if load is low enough?

14:18 <robclark> macc24: you can offline individual cores, or use thread affinity.. and possibly there are other ways

14:19 <HdkR> I guess they'd more want a scheduler that lets you very heavily weight processes towards the little cores until you change the policy from userspace?

14:20 <HdkR> weigh

14:39 <robclark> yeah, apparently with schedutil you can do that.. although I'm not really too familiar with the details.. I know there has been a lot of perf work on "trogdor" devices to tune that, but usually the problem is the opposite (things running on little cores when we want to get them onto big cores)

14:39 <HdkR> I currently have that problem on the HDK888 sadly

14:40 <robclark> (or maybe it is combination of schedutil and cgroups)

16:17 <macc24> hmm