#asahi-dev · 2026-06-22 21:31
StanfordAppliedCyber[m]are the gxf instructions locked out of m3?#asahi-dev · 2026-06-20 02:06
StanfordAppliedCyber[m]ah, i see there's alrady some knowledge on these but these exact instructions are in the unknown block#asahi-dev · 2026-06-20 02:04
StanfordAppliedCyber[m]^ once we figured out what they did#asahi-dev · 2026-06-20 02:01
StanfordAppliedCyber[m]we'd then have to patch sptm or emulate the instruction(s)#asahi-dev · 2026-06-20 02:01
StanfordAppliedCyber[m]because if i were to take the "virtualize sptm" approach idk how i'd deal with this-- i have no clue what this instruction does, and it'd be hard to figure out because i don't have the m2 hardware to actually run it#asahi-dev · 2026-06-20 02:00
StanfordAppliedCyber[m]i know sven mentioned that someone got sptm working on m2 under m1n1, presumably m2 didn't have the plumbing that locks us out of gxf?#asahi-dev · 2026-06-20 01:59
StanfordAppliedCyber[m]0x00201328#asahi-dev · 2026-06-20 01:59
StanfordAppliedCyber[m]0x00201401#asahi-dev · 2026-06-20 01:58
StanfordAppliedCyber[m]there appear to be some gxf instructions that mess with cache coherency that are causing problems#asahi-dev · 2026-06-20 01:42
StanfordAppliedCyber[m]once we support exclaves none of this will be necessary and it'll all just work#asahi-dev · 2026-06-20 01:42
StanfordAppliedCyber[m]for other people to use the hypervisor this should likely be done when you first install + boot macos in the new volume just patch that file as one of the steps#asahi-dev · 2026-06-20 01:42
StanfordAppliedCyber[m]for context btw, the medina thing checks the chip id against a hardcoded list, the product id against a hardcoded value, and then the actual bplist flag. i figure for a normal vm boot either chip id or product id would change, but if i modify that other stuff breaks, so just patch the bplist#asahi-dev · 2026-06-20 01:10
StanfordAppliedCyber[m]right now boot deterministically reproduces the latest issue (some memory caching problem) whereas before we'd only make it to the latest issue 50% of the time, and the other 50% random nonsense would happen (sep timeout, spmi timeout, etc) and require a reboot#asahi-dev · 2026-06-20 01:09
StanfordAppliedCyber[m]I think i managed to fix the spmi issue, and make boot way more reliable, with some timer accounting stuff#asahi-dev · 2026-06-19 23:19
StanfordAppliedCyber[m]not quite at login window, but this feels good#asahi-dev · 2026-06-19 23:19
StanfordAppliedCyber[m]windowserver is alive, i just saw it take over the display, solution was just to disable medina by modifying the bplist for windowserver directly#asahi-dev · 2026-06-19 21:21
StanfordAppliedCyber[m]but i also don't really have anything concrete on it#asahi-dev · 2026-06-19 21:20
StanfordAppliedCyber[m]anyone have any idea what's up with this? it's an intermittent timeout, likely introduced due to timing side effects of the genter --> hvc patch i do, but anyone have any ideas on how to get rid of it? it's slowing down my iteration cycle significantly#asahi-dev · 2026-06-19 21:19
StanfordAppliedCyber[m]the following issue is some intermittent flake that occurs on roughly 40% of boots:... (full message at <https://matrix.org/oftc/media/v1/media/download/AalZDond9bsgr8hWXkhLqgiynIsXSdeKn1g-SceEMnrPPyVDREZmRbKAcyM_wWIZr63viZGNbNfYRomsXGPkeDBCefKdpI8AAG1hdHJpeC5vcmcvdHhMdm5sVWx0QmJnVVVHcnZiRU5NTkNh>)#asahi-dev · 2026-06-19 21:09
StanfordAppliedCyber[m]how about m5, has anyone looked at that?#asahi-dev · 2026-06-19 21:09
StanfordAppliedCyber[m]true, a18pro definitely works, m4 works (right?), m4 pro does not work and no one has checked max?#asahi-dev · 2026-06-19 21:01
StanfordAppliedCyber[m]that's why i'm doing it#asahi-dev · 2026-06-19 20:59
StanfordAppliedCyber[m]asking because i don't actually want to spend all my time writing tooling, i want to actually get linux to work on this thing, but if it's not a suitable target i won't spend as much time#asahi-dev · 2026-06-19 20:59
StanfordAppliedCyber[m]assuming the aic register works and wfi still loses state on macos 26.6, would this be a find fw revision to target for linux?#asahi-dev · 2026-06-19 20:28
StanfordAppliedCyber[m]back to where i was on 26.5#asahi-dev · 2026-06-19 20:23
StanfordAppliedCyber[m]i jumped into this project without knowing the whole "modern apple devices lock out apple sysregs" thing so i just found ways to make do#asahi-dev · 2026-06-19 20:23
StanfordAppliedCyber[m]i have some hacks around interrupt handling, but tbh it's not bad#asahi-dev · 2026-06-19 20:23
StanfordAppliedCyber[m]not#asahi-dev · 2026-06-19 05:40
StanfordAppliedCyber[m]oh and the debugusb issue that is still unresolved, but i've found ways to workaround that one#asahi-dev · 2026-06-19 05:39
StanfordAppliedCyber[m]but it's as good as you can ask for for a new macos version#asahi-dev · 2026-06-19 05:39
StanfordAppliedCyber[m]it's not just plug and play#asahi-dev · 2026-06-19 05:39
StanfordAppliedCyber[m]26.6 did change some stuff both sptm and nonsptm related#asahi-dev · 2026-06-19 05:38
StanfordAppliedCyber[m]i'm somewhere in late iokit init right now#asahi-dev · 2026-06-19 05:38
StanfordAppliedCyber[m]nah fixing kernel regressions#asahi-dev · 2026-06-19 05:34
StanfordAppliedCyber[m]sep was annoying but i got lucky in that it was self inflicted#asahi-dev · 2026-06-19 05:34
StanfordAppliedCyber[m]which thankfully appears to be a me problem that no one else working on the neo has hit#asahi-dev · 2026-06-19 05:34
StanfordAppliedCyber[m]the only real bad unfixable problem i hit was the smp thing#asahi-dev · 2026-06-19 05:33
StanfordAppliedCyber[m]i just like to complain a lot :)#asahi-dev · 2026-06-19 05:27
StanfordAppliedCyber[m]also "nontrivial" likely means ~50 total lines of code across the entire emlator code base in this case, the fix i just made was literally 2 lines of code in the dart emulator#asahi-dev · 2026-06-19 05:26
StanfordAppliedCyber[m]but we just pin a fw version that we want to support, get that sptm emulation to work, and then we can ignroe the drift, just as we have been#asahi-dev · 2026-06-19 05:26
StanfordAppliedCyber[m]i agree it's instability is a virtual guarantee#asahi-dev · 2026-06-19 05:25
StanfordAppliedCyber[m]well, it's just another firmware ABI#asahi-dev · 2026-06-19 05:08
StanfordAppliedCyber[m]oh boy, apparently 26.6 did change sptm nontrivially#asahi-dev · 2026-06-19 04:54
StanfordAppliedCyber[m]XNU exclusively uses the 24mhz timer internally, only userspace as far as i can tell consumes the 1ghz timer, which conveniently can be turned off via the commpage#asahi-dev · 2026-06-19 04:29
StanfordAppliedCyber[m]and this would be implemented in m1n1 running in el2 presumably? would this work with kvm? this whole thing feels rather cursed but I suppose that's why we didn't do psci to start with#asahi-dev · 2026-06-19 01:44
StanfordAppliedCyber[m]what will be the path forward if we can't prevent the wfi loses register state thing? real psci?#asahi-dev · 2026-06-19 00:19
StanfordAppliedCyber[m]no sysreg regressions like in the 26.4.1 --> 26.5 either, ie, 26.5 locked a bunch of sysregs that were writable in 26.4.1 (notably, all the pac registers)#asahi-dev · 2026-06-19 00:16
StanfordAppliedCyber[m]knock wood this hypervisor works by end of this week, sptm doesn't have many changes from a quick look#asahi-dev · 2026-06-19 00:14
StanfordAppliedCyber[m]okay, lesson learned, under no circumstances touch this 1tr object ever again#asahi-dev · 2026-06-19 00:14
StanfordAppliedCyber[m]SMP WORKS AGAIN ON 26.6 BETA 2 :)