#
✕ nick:maz
#asahi-dev · 2026-05-19 10:23
mazyuka: it is ^_+?
#asahi-dev · 2026-05-14 07:36
mazinterrupt acknowledgement, for the most common example.
#asahi-dev · 2026-05-14 07:35
maza lot of MMIO registers have read side effects, and not only in the Apple world. that's a very common thing to do.
#asahi-dev · 2026-05-10 16:02
mazit is odd that macvdmtool doesn't work, since that's one of CS's ancestors...
#asahi-dev · 2026-05-10 16:01
mazah, true. that's set 2 (can't remember what the code is like)
#asahi-dev · 2026-05-10 15:54
mazyuka: CS defaults to using the SBU lines to serial. I suspect that macdvmtool is using the secondary set of USB2 lines.
#asahi-dev · 2026-04-28 13:18
mazyeah, that's even more hackish. I'd rather keep it obvious, even if that's ugly.
#asahi-dev · 2026-04-28 13:09
mazjannau: the problem is that as long as you don't have an irq, irqchip and irq_domain are unrelated. chicken and egg.
#asahi-dev · 2026-04-28 12:43
mazwe could match on the compatible string of the parent. bit ugly...
#asahi-dev · 2026-04-28 12:42
mazjannau: the problem I can see is that at the point where we initialise the MSI domain, we only have access to the parent domain, not to the parent irqchip.
#asahi-dev · 2026-04-26 16:26
mazsomething that the "managed MSIs" framework allows on other designs.
#asahi-dev · 2026-04-26 16:26
mazand that breaks any SW that uses per-CPU queues and using per-CPU data...
#asahi-dev · 2026-04-26 16:25
mazwhat's even worse with the AICv2+ is that not only you can't set the affinity, but you can't really influence where the affinity lands at all.
#asahi-dev · 2026-04-26 16:24
maz(where 1:N distribution is that the implementation picks an affinity based on a pre-defined population).
#asahi-dev · 2026-04-26 16:23
mazthe only vaguely similar thing is GICv3 1:N, which Linux doesn't support for related reasons.
#asahi-dev · 2026-04-26 16:23
mazjannau: no, walking the hierarchy is the only way. Linux is very ill-equipped for systems that can't set the affinity at the root controller level (Apple is the only one we have, I expect).
#asahi-dev · 2026-04-01 08:49
mazI don't think PMU events are a strong indication of a similar feature set, because the event coverage is on the pipeline and the memory subsystem, rarely at the ISA level.
#asahi-dev · 2026-04-01 07:18
mazjannau: it was never 100% clear to me that the A and M CPUs had the exact same feature set.
#asahi-dev · 2026-03-31 16:43
mazhas anyone managed to get the CPU id registers on one of the new macbook neo devices?
#asahi-dev · 2026-03-04 18:28
mazbasically an ipad with a keyboard.
#asahi-dev · 2026-02-11 17:40
mazwe used to do that in KVM, but that quickly led to some horrors, and we switched to the GIC active state, which is what the Apple EL2 register is replacing.
#asahi-dev · 2026-02-11 17:40
mazbut you rely on the guest only checking for the STATUS bit, and not taking IMASK into account.
#asahi-dev · 2026-02-11 17:39
mazsven: then you should be able to hack something by setting the IMASK bit in the CNT{P,V}_CTL_EL0 registers every time you have a pending timer interrupt.
#asahi-dev · 2026-02-11 16:51
mazeither way, this is shit.
#asahi-dev · 2026-02-11 16:51
mazsven: without this register, you can't mask the timer registers in a way that is invisible to the guest. so either you corrupt the guest state by masking the timer in the EL1 register, or you end-up with a screaming interrupt that prevents you from making forward progress.
#asahi-dev · 2026-02-06 11:03
mazarryn: in my modest experience, code is far less ambiguous then English. but if you think this is useful, go forth and write things up!
#asahi-dev · 2026-02-06 10:51
mazarryn: incomplete because things need to be reverse engineered?
#asahi-dev · 2026-02-04 13:11
mazfor those interested, I've just pushed a CS SW update, moving to a newer SDK, which fixes a break handling issue.
#asahi-dev · 2026-01-27 14:47
mazhuh. Elfring in the loop. no wonder you got annoyed...
#asahi-dev · 2026-01-27 14:41
mazwonder why that wasn't sent as a fix to Linus.
#asahi-dev · 2026-01-27 14:39
mazsven: ah, that'd be it!
#asahi-dev · 2026-01-27 14:39
mazI'll investigate once I find the regression I have introduced in NV...
#asahi-dev · 2026-01-27 14:39
mazcoming from apple_smc_probe(). that's on M2pro mini.
#asahi-dev · 2026-01-27 14:38
mazis anyone running upstream with lockdep on? I'm seeing a splat at boot time coming from the SMC driver (DEBUG_LOCKS_WARN_ON(lock->magic != lock)) which is a bit worrying.
#asahi-dev · 2026-01-25 13:07
maz(no cover subject)
#asahi-dev · 2026-01-25 12:09
mazmassive!
#asahi-dev · 2026-01-25 12:07
mazsven: a real UART then. wish we knew that 5 years ago, would have saved me some effort! :)
#asahi-dev · 2026-01-25 12:04
mazI hope there's an interrupt associated with the RX side.
#asahi-dev · 2026-01-25 12:03
mazis this supported all the way to M1?
#asahi-dev · 2026-01-25 12:02
mazit'd be awesome to be able to remove the need for extra HW (other mac or CS) for people to do early bringup, or even normal serial console.
#asahi-dev · 2026-01-24 11:11
mazoh no! happened to me a couple of times. extremely annoying.
#asahi-dev · 2026-01-24 10:55
maz+1. I don't think being proactive buys us anything here. only extra merging pain.
#asahi-dev · 2026-01-24 10:45
mazwe really could do without another binding IMO
#asahi-dev · 2026-01-24 10:42
mazfl0_id: the trick is to never use the /dev/ttyACM* names. Use the /dev/serial/by-id/ names, which are stable, irrespective of the discovery/plugging order.
#asahi-dev · 2026-01-23 14:27
mazs/the the/for the/
#asahi-dev · 2026-01-23 14:27
mazchaos_princess: control the interrupt masks the the EL1 timers, amongst other things.
#asahi-dev · 2026-01-23 14:24
mazyuka: it's only used with KVM to context-switch the guest timers. nothing will break until then.
#asahi-dev · 2026-01-23 13:53
mazbut the MSR UNDEFs.
#asahi-dev · 2026-01-23 13:53
mazactually, the register exists, as we do an MRS from the same one 3 instructions before.
#asahi-dev · 2026-01-23 13:50
mazeither they now completely prevent accessing this (which can be sidetracked but will also kill KVM on this HW), or they have moved the functionnality somewhere else.

plus ›