#asahi-dev · 2026-05-06 18:47
yukaThis means basically removing the entire "Setting secondary CPU RVBARs...", right?#asahi-dev · 2026-05-06 18:47
jannaufor chainload.py you could skip setting rvbar if `rvbar == u.base`. it should not not differ unless someone plays with the entry point or changes how m1n1.macho is linked. In both cases it's fine if said person keeps the broken pieces#asahi-dev · 2026-05-06 18:42
yukaAlthough I would assume anyone working on that right now can also just comment out the smp_start_secondaries call in linux.py / chainload.py#asahi-dev · 2026-05-06 18:41
yukaI guess it could be reverted until we have fixed that#asahi-dev · 2026-05-06 18:41
yukaIf anything it might cause people to run into the SError#asahi-dev · 2026-05-06 18:40
yukais _not_ wrong?#asahi-dev · 2026-05-06 18:40
jannauyes, the t8132 CPU start offset is not wrong. I was wondering whether keeping it causes troubles#asahi-dev · 2026-05-06 18:40
yukaI don't think we have something like cpu_features->cyc_ovrd in the proxyclient side#asahi-dev · 2026-05-06 18:39
yukaFor chainload.py, is there any case where we are actually setting rvbar to something different?#asahi-dev · 2026-05-06 18:35
yukaimo there's nothing wrong with that one#asahi-dev · 2026-05-06 18:34
jannaushall I revert "smp: add CPU_START_OFF for T8132" as well?#asahi-dev · 2026-05-06 18:28
yukaI always applied this on my local checkout#asahi-dev · 2026-05-06 18:28
yukaAnd I think the switch boot core function is quite useful to have#asahi-dev · 2026-05-06 18:28
yukaI was going to ask: can you merge the smp.c bits from #563 after reverting #536? Then we have less rebase/merge pains later#asahi-dev · 2026-05-06 18:27
yuka536 vs 563#asahi-dev · 2026-05-06 18:27
yukaAhh#asahi-dev · 2026-05-06 18:27
yuka(number)#asahi-dev · 2026-05-06 18:27
yuka#536 is the wrong PR#asahi-dev · 2026-05-06 18:27
yukaSorry#asahi-dev · 2026-05-06 18:24
yuka2. the purpose of setting the cpu_impl_reg[0] is setting rvbar to the correct value (-> also wrong, the purpose is also to clear the lock bit so that the rvbar can be set again later once the core is started)#asahi-dev · 2026-05-06 18:23
yuka1. the lock bit means we can not write the rvbar (-> wrong, we can clear it, depending on whether this core is already running and what platform we are on)#asahi-dev · 2026-05-06 18:22
yuka#536 was fundamentally based on wrong assumptions:#asahi-dev · 2026-05-06 18:22
jannauand let's start fresh with what we learned#asahi-dev · 2026-05-06 18:22
yukathen I will start from a clean slate#asahi-dev · 2026-05-06 18:22
yukathat's ok#asahi-dev · 2026-05-06 18:22
yukaI think I now know the wrong assumptions and I will come up with something#asahi-dev · 2026-05-06 18:22
jannauI think I'll close my PR and revert the commits from #536#asahi-dev · 2026-05-06 18:18
yukaall P-Cores#asahi-dev · 2026-05-06 18:17
jannauall P-cores or just the boot CPU?#asahi-dev · 2026-05-06 18:16
jannautested on M1#asahi-dev · 2026-05-06 18:16
yukaP-Cores SError on write, E-Cores silently ignore the write#asahi-dev · 2026-05-06 18:16
yukaon M4, P-Cores and E-Cores behave differently#asahi-dev · 2026-05-06 18:16
jannauas suspected, BIT(0) is set on boot on all cores but clearing it is possible as long as the core is not running#asahi-dev · 2026-05-06 18:11
yuka> > > read64(0x210050000) = 0x10010002114001#asahi-dev · 2026-05-06 18:11
yuka> > > write64(0x210050000, 0x10010002114000)#asahi-dev · 2026-05-06 18:11
yuka> > > read64(0x210050000) = 0x10010002114001#asahi-dev · 2026-05-06 18:11
yukait doesn't SError but it also doesn't apply the new value#asahi-dev · 2026-05-06 18:11
yukaAh no#asahi-dev · 2026-05-06 18:10
yukawhy was the old code not working then...#asahi-dev · 2026-05-06 18:10
yukawhen they fixed the other smp stuff?#asahi-dev · 2026-05-06 18:10
yukapossibly this changed with mac os 26?#asahi-dev · 2026-05-06 18:10
yukaI can actually write it#asahi-dev · 2026-05-06 18:08
yukaeither way#asahi-dev · 2026-05-06 18:08
yukaso we need to take cpu_features into account#asahi-dev · 2026-05-06 18:08
yukaExactly#asahi-dev · 2026-05-06 18:08
jannaucompletely unrelated to its value