#
✕ nick:sven
#asahi-dev · 2026-06-07 14:18
sventrue
#asahi-dev · 2026-06-07 14:16
svenbtw. what exactly happens when the nvme driver SErrors because we try to access the pre-15 MMIO registers? does that only crash that one thread or does the kernel stop completely then?
#asahi-dev · 2026-06-07 14:15
svenlinear sq, nvmmu is mandatory on M2 iirc
#asahi-dev · 2026-06-07 14:15
svener, yeah, other way around
#asahi-dev · 2026-06-07 14:13
sven*then
#asahi-dev · 2026-06-07 14:13
svenand hen M2 dropped the linear sq, nvmmu stuff and apple still called that ans3, didn't they?
#asahi-dev · 2026-06-07 14:09
svendon't remember what ans2 was, maybe the t2 macs?
#asahi-dev · 2026-06-07 14:09
sveniirc last time we came to the conclusion that m1 was already ans3
#asahi-dev · 2026-06-07 13:03
svenother alternative would be to add apple,firmware-version but we also need to change the compatible because the old driver will SError very early
#asahi-dev · 2026-06-07 13:03
svenit's actually called ANS3, that ANS2 was wrong all along
#asahi-dev · 2026-06-07 12:34
sveni don't think so :(
#asahi-dev · 2026-06-07 11:40
svenhttps://git.kernel.org/pub/scm/linux/kernel/git/sven/linux.git/log/?h=b4/nvme-m3 something like that would be nvme support for macos 15+ if we go with the "new compatible" path
#asahi-dev · 2026-06-07 10:07
svenme neither
#asahi-dev · 2026-06-07 09:55
svenno idea, can't find anything by just reading the code :/
#asahi-dev · 2026-06-07 09:54
svenhmmmm
#asahi-dev · 2026-06-07 09:49
svenwould be interesting if this reproduces with lockdep with hopefully some splat as well
#asahi-dev · 2026-06-07 09:48
svenbeats me as well
#asahi-dev · 2026-06-07 09:45
svenand well... the atcphy struct would disappear, so that's probably not it because we'd just see a crash
#asahi-dev · 2026-06-07 09:45
svenhrm.. but the next probe would use a different mutex then as well and
#asahi-dev · 2026-06-07 09:45
svenmaybe some left-over state when atcphy probe fails?
#asahi-dev · 2026-06-07 09:42
sventrue
#asahi-dev · 2026-06-07 09:39
svenbut that would surprise me
#asahi-dev · 2026-06-07 09:39
svenbut maybe i overlooked something back then or something changed
#asahi-dev · 2026-06-07 09:39
sveni remember checking when i originally wrote the code
#asahi-dev · 2026-06-07 09:35
svenbut i'm pretty sure i guard against that by taking the mutex before registering all those devices
#asahi-dev · 2026-06-07 09:33
sventhe only thing that could happen was that e.g. the mux was registered and would already trigger a callback to atc from a different thread while atc was busy registering the phys
#asahi-dev · 2026-06-07 09:32
svenmaybe something changed?
#asahi-dev · 2026-06-07 09:32
svenhrm... i thought i was careful to avoid that
#asahi-dev · 2026-06-07 08:52
sventhough I guess that also happens with simple framebuffer
#asahi-dev · 2026-06-07 08:52
svenI just don’t wanna know how quick it burns through the battery; all cores were easily at 20-30% running llvmpipe
#asahi-dev · 2026-06-07 08:51
svengpu-less dcp was useable on m1 in December for me
#asahi-dev · 2026-06-06 17:10
svensure!
#asahi-dev · 2026-06-06 16:31
svenand tail = 2 with a queue depths of 2 sounds wrong
#asahi-dev · 2026-06-06 16:29
svenyeah, was just looking at that and that looks like it assumes all queues have the submission queue depth
#asahi-dev · 2026-06-06 16:27
svenso admin queue then, that's extra weird because its tags should still start at zero
#asahi-dev · 2026-06-06 16:26
svenis that error for the submission queue?
#asahi-dev · 2026-06-06 16:25
svenso the fix might be correct but the explanation isn't complete then
#asahi-dev · 2026-06-06 16:25
sventhen something else is going on, because that should prevent any tag collisions
#asahi-dev · 2026-06-06 16:12
svenreplied on-list as well
#asahi-dev · 2026-06-06 16:06
sventhat's quite weird
#asahi-dev · 2026-06-06 16:06
svennickchan: huh, that "reserve the first N tags" is how the pci driver handles the T2 macs. are you sure it doesn't work when talking to the controller directly?
#asahi-dev · 2026-06-05 17:07
svenlet's not do that
#asahi-dev · 2026-06-05 17:07
svenyeah no
#asahi-dev · 2026-06-05 16:44
svenyeah, as i said, i likely just read the specification wrong
#asahi-dev · 2026-06-05 16:02
sveni guess new compatible it is then
#asahi-dev · 2026-06-05 16:02
svenboth are too late for us to determine the right mmio offsets though
#asahi-dev · 2026-06-05 16:02
svensounds like either a regular nvme command or one of those special debug nvme commands then
#asahi-dev · 2026-06-05 16:01
svenand that's too late :(
#asahi-dev · 2026-06-05 15:59
sventhere was no way to tell the fw version from inside the nvme driver directly, was there?
#asahi-dev · 2026-06-05 15:58
svenokay, some real hw change then i guess

plus ›