#
#asahi-dev · 2024-09-16 13:13
mazand as long as you don't need to refer to a MPIDR (no GIC, no problem), you should be OK.
#asahi-dev · 2024-09-16 13:12
mazright, that makes it much simpler. you still have the same number of CPUs (half of what the HW has) in the various cpumasks,
#asahi-dev · 2024-09-16 13:11
nickchanthere is only one performance controller
#asahi-dev · 2024-09-16 13:11
nickchanmaz: All CPUs switches at the same time
#asahi-dev · 2024-09-16 13:10
mazwhat if another CPU sends an IPI to the one that goes down? or are *all* the CPUs switching at the same time?
#asahi-dev · 2024-09-16 13:09
nickchannothing seems to explode that horribly on the iPad 7 where I did get cpufreq to work, so maybe I can just taint the kernel during init and ignore it for now...
#asahi-dev · 2024-09-16 13:09
mazthen the cpufreq magic is going to be interesting, because I don't think there is much infrastructure for surprise-removal of CPUs.
#asahi-dev · 2024-09-16 13:08
nickchanmaz: Correct
#asahi-dev · 2024-09-16 13:07
mazso this effectively appears as CPU hotplug on/off with affinity change?
#asahi-dev · 2024-09-16 13:06
nickchanWhat changes is the MPIDR
#asahi-dev · 2024-09-16 13:06
nickchanthere are two MIDR: one for A10/T2, another for A10X
#asahi-dev · 2024-09-16 13:06
nickchanmaz: MIDR is not MPIDR
#asahi-dev · 2024-09-16 13:06
mazand is it MIDR, not MPIDR? -- unless the affinity changes as well?
#asahi-dev · 2024-09-16 13:05
mazthe problem is that you now have multiple MIDRs for the same part -- how do you select an erratum?
#asahi-dev · 2024-09-16 13:03
nickchanbut now it looks like it maybe should end up in arch/arm64/kernel/cpu_errata.c...
#asahi-dev · 2024-09-16 13:02
nickchanso far this is in arch/arm64/kernel/cpufeature.c
#asahi-dev · 2024-09-16 13:00
jannaumy point was that either qemu or kvm actively tries to avoid the same situation for the guest kernel. I think the taint is justified
#asahi-dev · 2024-09-16 12:59
nickchanapple_fusion_cursed_fixup()
#asahi-dev · 2024-09-16 12:59
nickchanprobably should rename bad_aarch32_el0_fixup() to something more specific to a10 while doing this though
#asahi-dev · 2024-09-16 12:58
nickchanthere's quite a bit of places in the kernel that uses mpidr though, id rather taint the kernel first to let people know about this cursed CPU
#asahi-dev · 2024-09-16 12:58
mazit's not that KVM cares. it's luserspace responsibility to pin the threads to CPUs they care about.
#asahi-dev · 2024-09-16 12:57
nickchanA10 only has EL0 and EL1 so that's non issue
#asahi-dev · 2024-09-16 12:57
jannaukvm doesn't/didn't work across big and little cores to avoid changing MPIDR iirc
#asahi-dev · 2024-09-16 12:52
nickchan*super sure
#asahi-dev · 2024-09-16 12:52
nickchanprobably should ensure that it always enter in E-mode in the loader since the device tree is like that...
#asahi-dev · 2024-09-16 12:52
nickchanthough I am not super suer if a changing MPIDR warrents a taint
#asahi-dev · 2024-09-16 12:51
jannauvery limited testing from my side (j493 only)
#asahi-dev · 2024-09-16 12:50
jannauI've pushed asahi-6.11-rc and rebased asahi-wip and the bits/* branches onto v6.11
#asahi-dev · 2024-09-16 12:28
nickchanprobably right at where aarch32 el0 is disabled
#asahi-dev · 2024-09-16 12:26
nickchanI am seriously considering to cause an immediate TAINT_CPU_OUT_OF_SPEC when I get around to the cpufreq stuff
#asahi-dev · 2024-09-16 12:26
nickchanoh hell nah A10(X) MPIDR changes when with their p-state
#asahi-dev · 2024-09-16 11:04
nickchanrevision == 3 would mean boot cmdline length 1024
#asahi-dev · 2024-09-16 11:03
nickchanluckily for the newest increase in boot cmdline length you should be able to get around with checking the revision field of the BootArgs struct
#asahi-dev · 2024-09-16 11:02
nickchanthere are some even older ones that have device tree issue that I have not gotten around to fixing yet...
#asahi-dev · 2024-09-16 11:01
nickchanfor the BootArgs struct stuff because some very old firmwares like iBoot-2817 doesn't even have mem_size_actual I made some assumption and assumed the dram bottom to 0x800000000, then rounded things up to the nearest 256 MB
#asahi-dev · 2024-09-16 09:53
mazI wish the author had provided a script...
#asahi-dev · 2024-09-16 09:52
mazjannau: yeah, I had a fun time fixing those.
#asahi-dev · 2024-09-16 09:50
jannauoh, joy. changinging the platform_device.remove function pointer signature from returning int to void for ~20 drivers
#asahi-dev · 2024-09-16 09:43
emxlIf you dont get any output on your screen the culprit is probably a new version of the boot_args struct (if you have a recent macOS version). This moves mem_size_actual, breaking the mapping code and crashing m1n1 before it can initialize the screen console. I fixed this for me by increasing cmdline from 608 to 1024 bytes.
#asahi-dev · 2024-09-16 09:40
emxlkj1243: You asked 2 weeks ago if you could help develop for M3. I think it would be helpful if you got to work on whatever init code for your platform is missing in m1n1. You can build m1n1 for your platform, flash it to a new OS install (then enable the -v (for verbose) boot arg with nvram boot-args="-v") and see what breaks and what works.
#asahi-dev · 2024-09-14 22:03
chadmedthat said, we kinda boot way too fast for it to even matter :p
#asahi-dev · 2024-09-14 22:02
chadmedi dont think we can make it not suck for us because we reset dcp so we're always going to get some amount of flicker
#asahi-dev · 2024-09-14 22:01
chadmedit might need to be explicitly blocked...
#asahi-dev · 2024-09-14 22:00
chadmedjannau: plymouth just waits for _a_ drm driver, so simpledrm triggers it for us
#asahi-dev · 2024-09-14 18:00
nickchanthere is also a possibility that I am missing some code from the loader but idk.
#asahi-dev · 2024-09-14 17:15
nickchanespecially without the luxury of access tracing
#asahi-dev · 2024-09-14 17:15
nickchancan never know for sure if what you are doing is right when dealing with these hardware...
#asahi-dev · 2024-09-14 17:14
nickchanoh no I think everything cpufreq related had to be retracted for now
#asahi-dev · 2024-09-13 17:41
nicolas17goes back under his rock
#asahi-dev · 2024-09-13 17:40
nicolas17but where is 120-125ms pro

plus ›