#
#asahi-dev · 2026-10-10 10:39
yukain the guest
#asahi-dev · 2026-10-10 10:39
yuka> watchdog: BUG: soft lockup - CPU#1 stuck for 24s! [txg_sync:1092]
#asahi-dev · 2026-10-10 10:39
mazpromising... :-/
#asahi-dev · 2026-10-10 10:38
yukaI tested NV on M3 Max and it was broken (either host or guest hangs)
#asahi-dev · 2026-10-10 10:37
mazand while you're in KVM land, can you check whether NV1 is actually functional? M2 advertises it, but the page table walker is fsck'd. I suspect it has been fixed in M3/M4, but you might want to check.
#asahi-dev · 2026-10-10 10:36
mazah, that... yes, you need to keep that in sync with the rest. too bad we can't share it with the errata one.
#asahi-dev · 2026-10-10 10:35
MichaelReeves[m]yes
#asahi-dev · 2026-10-10 10:35
yukaand they automatically get impdef_pmuv3 and broken_gic_v3_seis
#asahi-dev · 2026-10-10 10:35
MichaelReeves[m]yes
#asahi-dev · 2026-10-10 10:35
yukaAh right, and now it's just apple_cpus
#asahi-dev · 2026-10-10 10:34
MichaelReeves[m]yuka: the other one was impdef PMUv3 workaround list
#asahi-dev · 2026-10-10 10:34
MichaelReeves[m]https://github.com/torvalds/linux/blob/3857c2fe5449541e24afc5efdb0f81a8a8f9a3a0/arch/arm64/kernel/cpufeature.c#L2347
#asahi-dev · 2026-10-10 10:34
MichaelReeves[m]7.3-rc6
#asahi-dev · 2026-10-10 10:33
mazMichaelReeves[m]: what kernel are you looking at?
#asahi-dev · 2026-10-10 10:32
MichaelReeves[m]to rephase/clarify, m1 and m2 are already in a list called has_vgic_v3[]. in v1 of the patch, I forgot to add m3 and m4 to it as well. because they aren't in that list, ARM64_HAS_ICH_HCR_EL2_TDIR remains unset and TC is set meaning we trap more than we need to. so in v2, I will add it to the list as well
#asahi-dev · 2026-10-10 10:32
yukathere were two lists which were at some point combined into one. one was the SEIS erratum and the other is...?
#asahi-dev · 2026-10-10 10:31
MichaelReeves[m]wait sorry I wrote my sentence the wrong way
#asahi-dev · 2026-10-10 10:30
MichaelReeves[m]maz: no we aren't avoiding setting TC. avoiding setting TC is what will happen if M3 and m4 not added to has_vgic_v3 (which I forgot in v1). m1 and m2 are added to has_vgic_v3 already as well
#asahi-dev · 2026-10-10 10:29
mazyup.
#asahi-dev · 2026-10-10 10:29
yukais the SEIS erratum the thing where some SErrors from the guest appear to the host when they should not?
#asahi-dev · 2026-10-10 10:28
yukaI still need to read up on these things to actually understand what I'm testing
#asahi-dev · 2026-10-10 10:28
yukagood morning
#asahi-dev · 2026-10-10 10:27
mazWhy do you need to avoid setting TC? since TDIR exists on this HW, TC is never needed, and doesn't get set (at least it doesn't on M1/M2).
#asahi-dev · 2026-10-10 10:25
MichaelReeves[m]oh good point, yuka is the one who has M4 so maybe we should wait for her
#asahi-dev · 2026-10-10 10:25
mazhave you tested M4 as well?
#asahi-dev · 2026-10-10 10:20
MichaelReeves[m]ok sounds good, yeah so you reply to v1 and then ill send v2
#asahi-dev · 2026-10-10 10:20
myriaYes, but I'll reply to the mailing list so there's a record of me testing it there
#asahi-dev · 2026-10-10 10:19
MichaelReeves[m]ok I'll send v2 now with has_vgic_v3 added and noted we tested that SEIS was needed. you used m3 max right?
#asahi-dev · 2026-10-10 10:19
MichaelReeves[m]perfect!!
#asahi-dev · 2026-10-10 10:19
myriaMichaelReeves[m]: yes, i had has_vgic_v3 patches
#asahi-dev · 2026-10-10 10:16
MichaelReeves[m]s/which/which avoids
#asahi-dev · 2026-10-10 10:16
MichaelReeves[m]yes, the SEIS workaround needs to trap DIR, I was talking about a seperate list has_vgic_v3[] I think which unneeded TC trap
#asahi-dev · 2026-10-10 10:06
mazand since DIR can generate an SError when deactivating an inactive interrupt, we must trap it to sanitise the access.
#asahi-dev · 2026-10-10 10:05
mazwhy not needed? we absolutely need DIR to work when the guest is using EOImode=1.
#asahi-dev · 2026-10-10 10:04
MichaelReeves[m]btw myria did you test with has_vgic_v3/can_trap_icv_dir_el1 as well? (that was the extra one that increases performance but isnt needed for it to work)
#asahi-dev · 2026-10-10 10:04
mazthanks.
#asahi-dev · 2026-10-10 10:03
MichaelReeves[m]thanks, yep will do next time
#asahi-dev · 2026-10-10 10:01
JamesCalligeros[m]You never know who’s following along at home
#asahi-dev · 2026-10-10 10:01
JamesCalligeros[m]Once things go on the list any important correspondence should go there
#asahi-dev · 2026-10-10 10:01
MichaelReeves[m]maz: sorry, didnt want to bother you and thought it was quicker to ask yuka
#asahi-dev · 2026-10-10 09:58
myriaSure thing
#asahi-dev · 2026-10-10 09:58
mazOK, thanks for testing. If you can report on the list what HW you've tested on, I'll take the patches after -rc1.
#asahi-dev · 2026-10-10 09:57
myriaThat one
#asahi-dev · 2026-10-10 09:57
myriahttp://www.loen.fr/tmp/QEMU_EFI.fd-bullseye
#asahi-dev · 2026-10-10 09:56
myriathe one sent to yuka, yes
#asahi-dev · 2026-10-10 09:56
mazmyria: did you test with the dreaded EFI binary?
#asahi-dev · 2026-10-10 09:55
myriaKernel without patches definitely serrors and causes a soc reset
#asahi-dev · 2026-10-10 09:53
mazMichaelReeves[m]: if you didn't get them, why didn't you say so over email? :(
#asahi-dev · 2026-10-10 09:26
myriaMichaelReeves[m]: Definitely get SErrors forwarded to host in some case, (not sure without patches or not, rebuilding to check), with patches it's definitely fine. Will let you know state without patches for sure when kernel finishes build
#asahi-dev · 2026-10-10 09:13
chadmedand ive been exploding nazis all day but wolfenstein is native vulkan so just goes through winevulkan

plus ›