#asahi-dev · 2023-05-12 10:14
maznote that I'm not seeing any state corruption. only temporary lockups.##asahi-dev · 2023-05-12 10:06
mazthis is the only machine I had this on (the M1 mini only runs pure upstream, the studio is stuck on some ancient kernel). I should try and enable it on the M2 MBA to see.#asahi-dev · 2023-05-12 10:02
maz(and maybe I should move the M2 pro away from the Studio... that's a bad thermal sandwich.#asahi-dev · 2023-05-12 10:01
mazmarcan: yeah, it seems to come and go without a clear pattern. I wonder if I'm hitting some thermal limit that keeps the cluster in shutdown at times...#asahi-dev · 2023-05-12 09:42
mazmarcan: using your cpuidle driver, I get some pretty bad lockups (the machine freezes for about 10 seconds). this is specially visible at boot and shutdown times. haven't had a chance to work out which state trips the behaviour yet, but I suspect the cluster shutdown to be slow to wake-up on IPIs...#asahi-dev · 2023-05-11 08:39
mazmarcan: added the flag, no screaming anymore.#asahi-dev · 2023-05-11 06:08
mazanyhow, time to face the train. ttyl#asahi-dev · 2023-05-11 06:07
mazas soon as the build is complete, PD reconnects.#asahi-dev · 2023-05-11 06:07
mazitakes about 5 minutes, as the kernel compile is just over 6.#asahi-dev · 2023-05-11 06:07
maz(VM on 6 p-cores, nested VM cimpiling the kernel...)#asahi-dev · 2023-05-11 06:06
mazthat's my gut feeling. and it is reproducible like clockwork.#asahi-dev · 2023-05-11 06:05
mazthe fan is definitely spinning, and the box is really warm to the touch.#asahi-dev · 2023-05-11 06:03
mazwhat could possibly go wrong...#asahi-dev · 2023-05-11 06:03
mazmarcan: it strikes me that this box is almost half a studio, but with a quarter of the thermal capacity...#asahi-dev · 2023-05-11 06:01
mazI'll get to the cpuidle later today, need to get myself into London first...#asahi-dev · 2023-05-11 06:00
mazmarcan: the PD controller that I use for the serial port. it's probably in host mode by that time, but there is nothing connected to the USB lines. #asahi-dev · 2023-05-10 19:24
mazwhen pushing the M2 pro mini quite hard (all CPUs at freq, no cpuidle, kernel compilation in some NV VMs), the machine gets pretty hot and the PD controller disconnects. let it cool down a few seconds, and it reconnects...#asahi-dev · 2023-05-09 16:35
mazremoved the driver for the time being, I have other things to debug! ;-)#asahi-dev · 2023-05-09 16:34
mazyeah. I have a ton of things like this https://paste.debian.net/1279776/#asahi-dev · 2023-05-09 16:32
mazj`ey: # CONFIG_PROVE_RAW_LOCK_NESTING is not set#asahi-dev · 2023-05-09 16:25
mazmarcan: not sure you're already aware, but the apple cpuidle driver screams like mad when used with lockdep.#asahi-dev · 2023-05-08 08:49
maz(enables the feature requiring context switching in the guest, that is).#asahi-dev · 2023-05-08 08:48
mazso enabling it wouldn't be a big deal, and probably easy enough to track if the VMM enables it.#asahi-dev · 2023-05-08 08:48
mazthat's why.#asahi-dev · 2023-05-08 08:48
maznah, TACR.#asahi-dev · 2023-05-08 08:45
mazI wonder whether that's a regression.#asahi-dev · 2023-05-08 08:45
mazbut j`ey is correct, we seem to be missing ACTLR_EL1's save/restore, descpite what I said earlier.#asahi-dev · 2023-05-08 08:43
mazmarcan: I don't think actlr_el12 in the TIDCP range.#asahi-dev · 2023-05-08 08:41
mazj`ey: sure? thad'be a bug.#asahi-dev · 2023-05-08 08:40
mazwe already save/restore ACTLR_EL1, so this has zero impact on KVM.#asahi-dev · 2023-05-08 08:39
mazmarcan: you don't have to sell it to me, only to the arm64 maintainers. the one thing I don't want is to expose something that isn't architectural to *guests*.#asahi-dev · 2023-05-08 08:36
mazmarcan: as long as this stuff never gets near my machines and that I don't have to support it...#asahi-dev · 2023-05-08 08:36
mazmarcan: why should I care?#asahi-dev · 2023-05-05 11:07
mazwell, the obvious conclusion is that I must have screwed up. I'm shocked.#asahi-dev · 2023-05-05 11:01
mazmarcan: right. I must have messed up when I tried that months ago. TGE definitely is a good hint. thanks again.#asahi-dev · 2023-05-05 10:54
mazI'll get back to that at some point. I'm too deep into other sort of trap handling to context-switch now... thanks for giving it a go!#asahi-dev · 2023-05-05 10:52
mazthey avoid the whole save/restore madness, which is terribly crap.#asahi-dev · 2023-05-05 10:51
mazgotcha.#asahi-dev · 2023-05-05 10:50
mazmarcan: I have no idea what the Apple mode is though.#asahi-dev · 2023-05-05 10:50
mazmarcan: interesting. I wonder if I get caught by the PAC emulation happening once TGE has been switched back to 1. but the spec doesn't say anything about the value of TGE, and I guess it could use some tightening...#asahi-dev · 2023-05-05 10:25
mazone potential alternative would be to have a custom mini-guest that is used only for the purpose of computing the PAC for ERETA{A,B}, but that is incredibly overkill.#asahi-dev · 2023-05-05 10:23
mazit implies it, as ERETAA has to be emulatable by a higher EL. withtout that, you cannot emulate PAC correctly.#asahi-dev · 2023-05-05 10:22
mazI've vaguely heard of something like that.#asahi-dev · 2023-05-05 10:21
mazhowever, M2 seems to use different seeds for the authentication depending on the EL that performs it#asahi-dev · 2023-05-05 10:20
mazmarcan: with NV, we have to trap ERET between the virtual EL2 and EL1. which means that when you do an ERETAA/ERETAB, EL2 needs to emulate the authentication.#asahi-dev · 2023-05-05 08:30
mazmarcan: VMKEYLO_EL2 caught my eye... I wonder if that's the key to another problem NV I've been facing...#asahi-dev · 2023-05-04 15:46
maz(and no, I don't call myself a kernel developer anymore -- more like a part time tinkerer...)#asahi-dev · 2023-05-04 15:45
mazhey, I have 3 other M1/M2 to play with! :D#asahi-dev · 2023-05-04 15:44
maz 16:43:21 up 260 days, 23:40, 5 users, load average: 0.02, 0.08, 0.12#asahi-dev · 2023-05-04 15:44
mazuptime