#asahi-dev · 2023-05-04 11:31
marcankettenis: should I be removing that from linux in our update script?#asahi-dev · 2023-05-04 11:27
marcanwe know there's weird magic involved#asahi-dev · 2023-05-04 11:27
marcanyup, but we also know Apple have a magic IMPDEF we need to poke to make AIC IRQ delivery work *at all* so...#asahi-dev · 2023-05-04 11:25
marcanbut if I can't repro in m1n1 we'll have to look at the latter option :p#asahi-dev · 2023-05-04 11:25
marcanI'll have to write a m1n1 test case first to prove this is the hardware and not KVM being terminally broken#asahi-dev · 2023-05-04 11:24
marcanI wonder if there's a magic bit we need to set somewhere to fix that...#asahi-dev · 2023-05-04 11:24
marcanyup#asahi-dev · 2023-05-04 11:23
marcanin particular, if the guest is running on p-cluster 0 and I put a host-side CPU burner on p-cluster 1, it gets better randomly (p-cluster IRQ delivery is random). if I put the burner on the e-cluster, everything goes back to normal since (in the default AIC config) IRQ delivery prioritizes the e-cluster (but prioritizes not waking up asleep CPUs first)#asahi-dev · 2023-05-04 11:22
marcanthe symptoms are consistent with host AIC IRQ delivery being broken or severely delayed while running the guest#asahi-dev · 2023-05-04 11:22
marcanmaz: repro'd the KVM pathological behavior. it's really obvious if you have a GUI.#asahi-dev · 2023-05-03 04:57
marcanso basically once it's coming you'll have plenty of notice#asahi-dev · 2023-05-03 04:57
marcankettenis, maz: eventually yes (to support "true" external installs), but not until someone writes an embeddable USB stack in Rust#asahi-dev · 2023-05-02 14:26
marcanchadmed: do you want me to grab impulses?#asahi-dev · 2023-05-02 14:26
marcanyeah#asahi-dev · 2023-05-02 14:26
marcanpovik: I can merge it into the next tree and give it a spin#asahi-dev · 2023-05-02 14:25
marcan:D#asahi-dev · 2023-05-02 13:33
marcanyeah, but you can always delete boot.bin to effectively turn the UEFI thing into a forced proxy, or use the csrutil backdoor to get both :)#asahi-dev · 2023-05-02 13:31
marcanmaz: this is just for asmedia though, and you know the installer will create all that for you right? ;)#asahi-dev · 2023-05-02 13:20
marcancould just use that for self-contained builds#asahi-dev · 2023-05-02 13:20
marcandoesn't u-boot have some kind of embedded filesystem anyway?#asahi-dev · 2023-05-02 10:12
marcankettenis: yup, works for me!#asahi-dev · 2023-05-02 06:38
marcanthat driver is unmaintained AIUI#asahi-dev · 2023-05-02 06:38
marcanjannau: I don't think xf86-input-evdev works properly at all with Apple trackpads, they are out of luck anyway#asahi-dev · 2023-05-02 06:37
marcankettenis: I think I prefer it in vendorfw so we can keep all firmware in the same place (even if there are several subsets there)#asahi-dev · 2023-05-01 14:52
marcancould also be we screwed something up in lower core machines, though I haven't found anything yet#asahi-dev · 2023-05-01 14:50
marcanreason being that I believe AIC2 will always try to deliver IRQs to CPUs that are alive instead of waking up CPUs that are asleep, and then will prefer e-clusters by default or if all CPUs are asleep, and it has a fallback where if the IRQ is not acked in a while it broadcasts to all CPUs, which would explain lag if a guest is busy but its CPUs are not properly handling host IRQs#asahi-dev · 2023-05-01 14:49
marcanbut it could also be something completely unrelated#asahi-dev · 2023-05-01 14:49
marcanI *think* those symptoms could be explained by some kind of pathological IRQ routing in KVM, like as if host IRQs don't get handled when they arrive at CPUs that are in the guest#asahi-dev · 2023-05-01 14:48
marcanmaz: so I saw this on reddit and I'm wondering whether it rings any bells: https://www.reddit.com/r/AsahiLinux/comments/134g00u/strange_scheduling_latency_on_the_host_when_kvm/#asahi-dev · 2023-05-01 14:36
marcan(hope we never have to deal with the complex firmware trees in uboot, those need hard/symlinks :p)#asahi-dev · 2023-05-01 14:36
marcanjust to make it clear this is for uboot and not the kernel#asahi-dev · 2023-05-01 14:36
marcanI think as a one-off we can do that, maybe <esp>/vendorfw/uboot/<filename>?#asahi-dev · 2023-05-01 14:35
marcancan you load it from the vendorfw.cpio directly? or do you want it standalone?#asahi-dev · 2023-05-01 14:35
marcanI figured an initramfs would be easier but if you think all that can work... :p#asahi-dev · 2023-05-01 14:34
marcankettenis: works for me, but then we need the magic ESP selection stuff upstream too, and all that has to run before PCIe/USB init which means NVMe has to come first#asahi-dev · 2023-05-01 14:15
marcanjannau: ok, nothing unexpected there#asahi-dev · 2023-05-01 14:01
marcancan you throw the ADT somewhere? at least /cpus#asahi-dev · 2023-05-01 13:47
marcanthough I guess I could probably emulate it on my j314 with ADT hacks#asahi-dev · 2023-05-01 13:47
marcancould be completely unrelated of course, but might as well check...#asahi-dev · 2023-05-01 13:46
marcansince that kind of thing could be explained by IRQs getting delivered to nonexistent CPUs first (AIC2 will fallback to all CPUs after a timeout if no CPU acks the IRQ)#asahi-dev · 2023-05-01 13:46
marcanthere's a reddit post about bad interactivity on an 8-core j314 and now I'm wondering if we're doing something wrong with the AIC2 or cpustart setup...#asahi-dev · 2023-05-01 13:45
marcandid anyone around here have a CPU-downbinned machine?#asahi-dev · 2023-05-01 11:34
marcankettenis: might as well, for consistency#asahi-dev · 2023-05-01 10:59
marcanso I guess the last few commits just made m1n1 slightly larger enough that the kernel lost something more critical :p#asahi-dev · 2023-05-01 10:58
marcanyeah, for the HV that is somewhat recent#asahi-dev · 2023-05-01 10:56
marcan(please don't use mach-o any more)#asahi-dev · 2023-05-01 10:56
marcankernel's probably missing the last few megabytes or something#asahi-dev · 2023-05-01 10:56
marcana 68MB mach-o when the max mach-o payload size is 64MB in the linker script#asahi-dev · 2023-05-01 10:55
marcanwell there's your problem#asahi-dev · 2023-05-01 10:52
marcanokay but I need the binary if I'm going to try to disassemble it to figure out what it's doing :p