#asahi-dev · 2026-09-28 08:36
jannauhuh, they seem to break later as well. I just had a pending smc mbox irq in addition to the uart one#asahi-dev · 2026-09-28 08:32
svenhuh, odd#asahi-dev · 2026-09-28 08:31
jannauother irqs (spmi and smc mbox) work#asahi-dev · 2026-09-28 08:30
jannauthe exact same kernel / mini / rootfs works on m3 max#asahi-dev · 2026-09-28 08:28
jannauuserspace is a minimal alpine rootfs which spawns getty#asahi-dev · 2026-09-28 08:28
svenhttps://github.com/AsahiLinux/m1n1/blob/c42cf43d03887208e70ec5dcc4da0b2af3699c54/src/chickens.c#L228 something weird like that?#asahi-dev · 2026-09-28 08:27
jannauthis is broken with and without the hypervisor#asahi-dev · 2026-09-28 08:27
jannauany ideas how irqs might break on m3 ultra? uart0 irqs are not read. This doesn't seem to be a hw problem. the irq is unmasked and pending (reading the event reg from the HV)#asahi-dev · 2026-09-27 20:49
svenpossibly, yeah#asahi-dev · 2026-09-27 20:49
chaos_princesscould RT in UNEXP_RT mean "realtime"?#asahi-dev · 2026-09-27 20:48
svenhttps://github.com/AsahiLinux/m1n1/commit/fabe27e3f14891674a2fc42e9fbd2b51542230cd https://github.com/AsahiLinux/m1n1/commit/8fa422c0d3c11b3cc09e758c80d88ad3f5bf9a6a that sounds similar, so i'm probably just missing some new carveouts or whatever that have to be mapped differently#asahi-dev · 2026-09-27 20:35
sventhat just leaves that weird AMCC panic then#asahi-dev · 2026-09-27 20:32
svenxnu thinks the cluster is offline while it's actually stuck in that wfi loop and then tries to bring it back up which can't happen because it's been powered on the entire time :D#asahi-dev · 2026-09-27 20:32
svenoh lol, i think the pmgr panic and the wfi loop are a symptom of the same thing#asahi-dev · 2026-09-27 19:28
svenit either dies with that PMGR thing, with that AMCC thing or it ends up trying to put all cores into deep sleep which makes them all get stuck in an endless 1: wfi; 1b loop#asahi-dev · 2026-09-27 19:26
svenit's still somewhat unstable but if it boots the login screen i usually have 1-2 minutes before it dies#asahi-dev · 2026-09-27 19:26
svenyuka: fwiw, i'll try to get this thing up as a PR in the next 1-2 weeks, let me know if you need something specific traced before that on my m4 pro#asahi-dev · 2026-09-27 18:27
svenhuh.. that's new#asahi-dev · 2026-09-27 18:27
svenpanic(cpu 2 caller 0xfffffe003a2fe8c4): AMCC1 PLANE2 UNEXP_RT_HIT_DIR error: INTSTS1 0x00000100 AFERRLOG0/1/2/3/4/5 0xc8000800/0xfe2902/0x11c1f5d0/0x800/0x4/0 ADDR 0x102c8000800 CMD/SIZE 0x5(ncrdincri)/0x7f AID/TID 0x1f5d0/0x11c/0 @AppleH16GFamilyPlatformErrorHandler.cpp:2915#asahi-dev · 2026-09-27 16:50
svennickchan: done#asahi-dev · 2026-09-27 16:06
jannauthe write to (cpu_start_base + 0x4) is clearly bad on SoCs with 6 core clusters. should strace it on M3*#asahi-dev · 2026-09-27 16:04
jannaubased on the pmgr ADT CPU start might have changed on T6050. T8152 still has misc-cores-offset#asahi-dev · 2026-09-27 16:03
jannauyuka: I'm suspicious because (0x280600000 + 0x88000) + [0, 4] both read 0x3fffe after reset. those read 0x0 on M3 Max (with the correct PMGR reg)#asahi-dev · 2026-09-27 15:37
sven+an#asahi-dev · 2026-09-27 15:37
svensecure kernel for m5+ is also going to be even more cursed with this, but that's issue for later i guess#asahi-dev · 2026-09-27 15:26
sveni think i can't get rid of the last one, but i should be able to virtualize those mmio regions and just report the correct values#asahi-dev · 2026-09-27 15:26
svenwith correct initial values for the CTRR sysregs that's even less binary patching for SPTM. what's missing is the mmio part of the locked memory regions and another patch where it fails if its base is somewhat above ram_start #asahi-dev · 2026-09-27 15:06
jannaugrub issue reported in https://gitlab.freedesktop.org/gnu-grub/grub/-/work_items/90#asahi-dev · 2026-09-27 14:51
yukaI think I tested many offsets and this one did not give SError or similar and it did result in the pmgr entries coming up#asahi-dev · 2026-09-27 14:49
jannausigh: "Try allocating in the lower 4GB first, and limit arm64 allocations to the 4GB boundary. (091e50562, 1cae86131)" from grub 2.16#asahi-dev · 2026-09-27 14:46
nickchansven: mind picking these aarch32 rtkit crashlog section additions for 7.4: https://lore.kernel.org/asahi/20260919-rtkit-armv7-crashlog-v1-0-f4a671ea453c@gmail.com/T#asahi-dev · 2026-09-27 14:42
svenit's luckily rare enough but that also means i have almost no motivation to debug it :D#asahi-dev · 2026-09-27 14:41
svennickchan: i dunno, but even the latest HEAD still has some cursed issue, probably something with caching around MMU enable, that i hit occasionally on M4 pro :(#asahi-dev · 2026-09-27 14:40
jannaunickchan: thanks. I meant to ping you in the PR and forgot#asahi-dev · 2026-09-27 14:39
nickchanWhy is starting secondary always cursed for everybody#asahi-dev · 2026-09-27 14:39
nickchanspeaking of which maybe I should try to fix the cursed concurrency issue with start A7-A9 CPUs for the third time (after stopping them for 2 times) that makes the secondary die when it tries to eret from EL3 into EL1 that has been there since the beginning#asahi-dev · 2026-09-27 14:38
nickchanvalidated the pre-M1 cpufreq changes #asahi-dev · 2026-09-27 14:36
yukaHmm#asahi-dev · 2026-09-27 14:34
jannaupmgr.misc-cores-offset#asahi-dev · 2026-09-27 14:33
jannauthe CPU start offset seems to be in pmgr.misc-cores, unfortunately no such property on t6050#asahi-dev · 2026-09-27 14:31
svennow let's see about that CTRR crap#asahi-dev · 2026-09-27 14:31
svenwith that mpidr thing i need one less binary patch to sptm \o/#asahi-dev · 2026-09-27 13:42
jannau"write32(cpu_start_base + 0x4, 1 << (4 * cluster + core));" looks very broken for clusters with 6 CPU cores#asahi-dev · 2026-09-27 13:42
jannauyuka: are you sure t6050 uses the same CPU start offset as t6031...? pmgr_reg + CPU_START_OFFSET + 4 does not read 0x0#asahi-dev · 2026-09-27 12:40
svenhrm, i guess i can just seed the CTRR regs which I emulate anyway with the correct values#asahi-dev · 2026-09-27 12:39
svenit checks if CTRR and the lockdown stuff in AMCC is configured correctly which it really isn't once I move SPTM 2GB up in RAM#asahi-dev · 2026-09-27 12:38
svenor at least move them to python so that they're done automatically#asahi-dev · 2026-09-27 12:38
svenyeah... i'll also see if i can get rid of the binary patches done to sptm now#asahi-dev · 2026-09-27 12:37
chadmedwe cant just move fast and break things all the time, thats how we end up with enterprise codebases#asahi-dev · 2026-09-27 12:36
svengetting this thing to work was clearly much more fun than now trying to understand wtf is actually required and what's just weird hacks that went nowhere :D