Repository navigation
Description
Naeem is helping bring up an M3 Max. This is the single tracking issue for his machine: establish its exact boot/firmware identity, preserve a working kernel-only baseline, then implement native display and G15C GPU support. Please keep reports and follow-up builds here so he does not have to repeat the same collection across tickets.
Priority: P1 · Owner: kernel/bootloader integration · Status: J516C identity recorded and its boot loader device tree reviewed; test release aurora-2026.10.11.911 (one-command M3 Max test kit) published for Naeem; waiting for its results.
What works in the current release
Aurora 2026.10.11.15 treats T6031/T6034 M3 Max as kernel-only and preserves the existing boot.bin. Software rendering on the boot framebuffer is expected; this is not an accelerated Max profile. M3 Air G15G and M3 Pro G15S acceptance does not establish G15C compatibility. No native Max display, OpenGL or Vulkan acceptance has been recorded here.
Naeem: the next step
Run the one-command test kit described in this comment and upload the one results file it produces. It restarts the Mac several times by itself, keeps every test boot one-shot, and can be stopped or undone at any time with sudo aurora-m3max-kit --stop or --restore.
For additional reports:
If you are already running Aurora, start with the report; reinstalling is not required just to collect it. In a writable folder, run:
curl -fsSL https://github.com/omacom/linux-aurora/releases/latest/download/install-aurora-sep.sh | bash -s -- --m3-reportAttach the resulting aurora-m3-report-<board>-<date>.tgz here, plus:
- Screen size, CPU/GPU core counts, RAM, and
uname -r. - Whether the built-in picture and desktop work, and any display/dock attached.
- Whether Wi-Fi, keyboard and trackpad work; name the current blocker if one prevents using the machine.
The collector does not install a kernel or rewrite the bootloader. It uses sudo, briefly loads phram when needed to read the bootloader's ADT through a read-only device, and removes a module it loaded. It masks common machine identifiers; inspect the archive before posting. If collection fails, paste the exact error instead—no raw ADT or calibration dump is needed.
If Aurora has not been installed yet, the kernel-only installer is:
curl -fsSL https://github.com/omacom/linux-aurora/releases/latest/download/install-aurora-sep.sh | bashBefore rebooting, retain the install summary and confirm it names the kernel-only Max path and preserved bootloader. After reboot, record uname -r, check the desktop, then collect the report above. If the installer refuses, stop and attach its message. Do not select an Air/Pro GPU profile for this Max.
Maintainer implementation milestones
- Record the actual board/SoC (J514C/J516C with T6031, or J514M/J516M with T6034), both OS and system firmware versions, stub/iBoot, stage1/stage2 identity, and booted kernel. Do not infer the firmware ABI from the installed macOS version.
- Naeem's Mac: J516C / T6031 (16 CPU, 40 GPU cores, 48 GB); system firmware 27.0; stub 14.8.3; iBoot
iBoot-10151.140.19.700.2; stage 1v1.6.1-dirty, stage 2v1.6.1-omarchy.aurora12; kernel7.1.12-2-20261011.15-sep-ARCH(kernel-only).
- Naeem's Mac: J516C / T6031 (16 CPU, 40 GPU cores, 48 GB); system firmware 27.0; stub 14.8.3; iBoot
- Review the existing report once; extract this Mac's GPU topology, address windows, interrupts, power domains, firmware segments, display/DART and PMP facts. Request only specific missing fields. Preserve its own calibration and power data.
- Done from the second report (2026-10-11): GPU windows, interrupts, power and clock gates, performance states (4 clusters × 10 cores), firmware segments and UAT regions; display, DART and power-management processor layout; all 258 power domains match the device tree. Firmware image IDs are collected by the test kit.
- Implement the exact board/firmware bootloader and device-tree handoff, including firmware-memory reservations and display ownership. Keep a tested restoration path and preserve kernel-only operation until the new profile is explicitly selected.
- Add the G15C runtime configuration and matching Mesa admission. Confirm its firmware ABI, memory mappings and queue/completion behavior; do not enable it by copying G15G/G15S identity tables or changing an allowlist alone.
- Pass source/build and adversarial failure/rollback controls, then a coordinated hardware boot: internal picture, responsive input, a correctly identified hardware renderer, pixel readback and repeatable submitted-job completion without GPU faults. Record OpenGL and Vulkan separately; unsupported capability must remain explicit.
- Deliver an admitted matched release and a single final install command for Naeem. Establish cold-boot recovery before adding external-display or suspend qualification.
Source and related work
Pinned current release source: f240960cc84cfda6f73e188050ccadb9435e938f.
- Installer Max kernel-only selection and report collector:
is_m3_kernel_only_chip,m3_next_steps,m3_report. - T6031 device tree and T6034 variant.
- AGX driver admission: T6031 is diagnostic, not a production GPU profile.
- Production M3 SoC configurations: runtime table contains T6030 and T8122, not a Max configuration.
- Bootloader integration.
- Per-Mac provisioning: document and test stock-install firmware and calibration #46 covers shared per-Mac provisioning; M3 Pro clean stage1: qualify explicit handoff admission on the supported stub #47 is the separate M3 Pro clean-stage1 case; M3 Air: qualify each shipped J613/J615 firmware profile separately #26 is Air qualification. None substitutes for Max acceptance. Track prioritization in Aurora roadmap: one kernel for M1/M2, M3 and Neo #56.
Completion: Naeem can install a documented matched Max profile, cold-boot a usable accelerated desktop, verify its supported graphics capabilities, and restore the previous working boot. Each result must identify the exact release, board and firmware. This issue remains open until that end-to-end path is demonstrated.
Launch placement: Next-stage accelerated M3 Max delivery in the same unified kernel and installer. Preserve the current kernel-only path; supported Air/Pro/Neo delivery need not wait for Max acceleration. Roadmap: #56.
Activity
malik-na commentedon Oct 11, 2026
Initial report: Mac15,9
Machine
uname -r:7.1.12-2-12.0-sep-ARCH(linux-aurora 7.1.12.aurora2-12.0)v1.6.1-dirty, stage2v1.6.1-omarchy.aurora12(m1n1-aurora 1.6.1.aurora12-1)What works on the current kernel
simple-framebuffer, 3456x2160. Nothing external attached.The report (
--m3-report, taken on the kernel above) is attached below:aurora-m3-report-j516c-20261012-014726.tgz. I checked it for identifiers before posting.Installer refusal. Earlier I ran the installer from the old
iconidentify/aurora-linuxrelease (aurora-2026.10.10.12). It stopped with nothing installed:Installed m1n1 is
1.6.1.aurora12-1, and theomarchyrepo only hasm1n1-aurora 1.6.1.aurora1-3, so a full update couldn't satisfy it. A fullomarchy updatealso breaks things on this machine. I haven't tried the 2026.10.11.13 installer from this repo yet. I'll wait for your go-ahead.The power survey hasn't been run yet.
chriskearney commentedon Oct 11, 2026
Thanks, Naeem—this confirms the installer/package mismatch on our side. The kernel-only Max path should accept your existing aurora12 and preserve
boot.bin; asking you to run a full updater was incorrect. The correction is in #77, and a fresh package is building. Please keep your working installation until the corrected release is available.I can see the report filename in your comment, but GitHub currently renders it as plain text with no attachment link. Please attach the existing
.tgzwhen convenient; there is no need to collect it again or run the power survey for this packaging fix.malik-na commentedon Oct 11, 2026
Here: https://www.transfernow.net/dl/20261011Ekk1ESHU
chriskearney commentedon Oct 11, 2026
The corrected release is live: Aurora 2026.10.11.15.
curl -fsSL https://github.com/omacom/linux-aurora/releases/latest/download/install-aurora-sep.sh | bashNaeem, the kernel-only path now accepts your existing aurora12 and preserves
boot.bin. A full Omarchy update is not required. Before rebooting, confirm the summary names the kernel-only Max path and preserved bootloader. After reboot, please reply withuname -rand whether the desktop works. Max GPU acceleration remains tracked here; this release fixes the installation blocker. Thanks for the report link—no need to collect it again.malik-na commentedon Oct 11, 2026
Installed 2026.10.11.15, but it took a workaround. The first run failed:
Cause (I think): this image has
IgnorePkg = *in/etc/pacman.conf. The helper correctly getslibfprintfrom the local candidate (libfprint-1.94.100-1.1) and only asks the repos forfprintdandlibgusb. But the-Sp/-Swcalls (around line 3946 ofinstall-aurora-sep.sh) still check dependencies, so pacman tries to pullextra/libfprint. That package is ignored, and--noconfirmdeclines it. The candidate isn't part of that transaction, so pacman never sees it. Passing-ddon those two calls should fix it, since the helper has already resolved dependencies and the final-Upcheck covers the full set.Workaround: install the two missing packages by hand with a copy of
pacman.confthat leaves out only theIgnorePkg = *line, then re-run the installer:The second run went through. The installer replaced
libfprintwith its own1.94.100-1.1, took the kernel-only Max path, and kept the bootloader as it was.After reboot:
uname -r:7.1.12-2-20261011.15-sep-ARCHsimpledrm, 3456x2160linux-aurora 7.1.12.aurora2-20261011.15,m1n1-aurora 1.6.1.aurora12-1(unchanged),aurora-touchid 20261003-2,libfprint 1.94.100-1.1,fprintd 1.94.5-2,libgusb 0.4.9-2malik-na commentedon Oct 11, 2026
https://www.transfernow.net/dl/20261011FPhvzsso
chriskearney commentedon Oct 11, 2026
Thanks, Naeem. Both results arrived, and the second report has what we needed.
Release
.15:7.1.12-2-20261011.15-sep-ARCHwith the desktop on the boot framebuffer and your existing bootloader is the intended kernel-only state for this Max. Your manual step left the system as the installer would have, so nothing needs undoing.Installer failure: your diagnosis is correct. With
IgnorePkg = *, the missing-dependency step asks the repositories forfprintdandlibgusb. pacman then tries to pull the heldextra/libfprintinstead of the installer's ownlibfprint, and the step refuses. This is not specific to the Max: any held image withoutfprintdhits it. A fix is in progress for the next release; I'll link it here.Second report (
aurora-m3-report-j516c-20261012-023953.tgz): unlike the first, which ran on the older kernel without the reader, it includes the bootloader's device-tree data. That covers the GPU, display, DART and power-management layout we need for the next milestone, and we're working from it now. No further collection is needed for now.One question, whenever convenient: you mentioned that a full
omarchy updatebreaks things on this machine. What breaks (boot, display, a package conflict)? The error text, or the packages it tries to change, would help. Please don't run it again just to find out.I'll post here when there is a build for you to try.
9 remaining items
chriskearney commentedon Oct 13, 2026
To clarify: you haven't opened a PR, and you don't need to yet. That line is about future PRs. Your branches are built on our
.911code, which wasn't inintegration/aurorauntil #104. Now it is, so whenever you open a PR from one of your branches, it will show only your own commits, not ours mixed in.malik-na commentedon Oct 13, 2026
I understand now, Chris. Thanks for explaining 🤝
chriskearney commentedon Oct 13, 2026
Release aurora-2026.10.13.18 is out and is now Latest. It carries #78 (the
IgnorePkg = *dependency fix you diagnosed) and #104 (the M3 Max groundwork your branches are built on).On your J516C, the one-liner will stop before changing anything. Your Mac still has the kit's boot loader recorded (
t6031-gpu-off-1), and this release ships the common m1n1, which has no T6031 handoff. You'll see "this release's m1n1 has no M3 Max boot loader variant … Nothing was installed". That's expected, so there's no need to report it. Your current setup stays as it is.If you want this release's kernel on that Mac anyway,
linux-aurora-7.1.12.aurora2-20261013.18-aarch64.pkg.tar.zstis attached to the release, with its checksum inSHA256SUMS.malik-na commentedon Oct 13, 2026
On the
disp_fequestion: measured on the J516C with the display up. The display gate raises the minimum state ondisp_sys,disp_feanddisp_cpu; afterwards onlydisp_cpuholds it (register min field 0xf),disp_fereads 0, as on T6030. The PMP started on the DCP's request 3.5 s into the boot and the DCP stayed up (eDP-1 connected,disp_festill active).disp_feis a child ofdisp_cpuin the T6031 PMGR tree, so thedisp_cpufloor the report driver checks is the one that matters. No change needed for the J516C.malik-na commentedon Oct 13, 2026
M3 Max (J516C): the first compute job never starts on the four-cluster G15C, one question
Since the 13 Oct update: the host-side search is finished (boots t47–t76). The command, the StartCL microsequence, the queue, the pool manager and its firmware-control flush, the run-queue entry and the device controls now match what the J516C's own driver sends for its first compute job, field by field, as traced from a macOS guest under the m1n1 hypervisor on this Mac. Every job is still consumed by the firmware and never started by the hardware: no start timestamp, all-zero output, 0 ns of GPU time, no fault in any cluster's bank, and in the per-core snapshot (t50/t51) CDM activity on cluster 0's ten cores only.
The only values that still change anything are the two cluster-scaled compute registers, UMA PTC (
0xa599) and DUPM (0xd411):0x13200400020,0x200000001) the firmware waits out its progress check (2 x 250 ms) and halts the job;0x6000400020,0x3c000000140) it halts the job 2 ms after the kick, with the same recovery info (0x27), whether the preemption tail pointers are zero or not (t76).Answering the halt with a resume lets the firmware retire the job, so nothing is wedged; the dispatch just never runs.
Question: on the M3 Pro and M3 Air runtimes, is there anything a multi-cluster part needs before or with a compute kick that a one- or two-cluster part doesn't, such as a per-cluster enable, a cluster mask in the command or the queue, or a firmware setting? If you've seen the firmware give up on a dispatch within a couple of milliseconds, I'd like to know what it was.
Records:
agent-docswork/t6031-kit-refusals/results/on malik-na/linux-aurora (t47–t76); map malik-na#18, ticket malik-na#46.chriskearney commentedon Oct 13, 2026
@malik-na, the M3 Mesa sources are now public, so you can work on the G15C side directly. Everything is in iconidentify/mesa-omacom:
m3max/g15c(ff120e88): the Mesa in the.911kit'smesa-m3-g15c. It's four commits on the legacy tree: the G15C admission behindASAHI_M3_G15C=1, the G15S runtime choices for G15C, the T6031 VA-range table, and a drm-shim G15C profile.m3/legacy(contains7ff89a4d): the commonmesa-m3tree it builds on.The package recipes are in
integration/aurora(#116):tools/aurora-sep/recipes/mesa-m3-g15cbuilds the kit's package.tools/aurora-sep/recipes/MESA-SOURCES.mdmaps each package to its branch and shows how to make the source tarballs.Send Mesa changes as PRs against
m3max/g15c, with what was tested on the J516C.malik-na commentedon Oct 13, 2026
M3 Max (J516C / T6031): the first compute job runs; my question above is answered
My 14 Oct question (why the first dispatch never starts) is answered: it does start, and the shader cores fault.
With the J516C's own compute register values, the USC reads unmapped pages in the client's USC window: base + 0x4000 and + 0x4180 (USC L1 units 8 and 9, in the client's VM), then base + 0x48180 once the first of those pages is backed. Each read lands 0x0 to 0x180 past the end of a mapped object, in the single 16 KiB guard page the Mesa VA allocator leaves unbound after every allocation. The kit's Mesa lays the window out as the M3 Pro line does: the 2^40 base page with the container lead, objects allocated upward, one guard after each.
As an experiment only, not for a PR: with read-only kernel pages mapped into the four guard pages below the top object, every compute job completes (about 1.07 ms of GPU time) and compute-check gets 61440 of 65536 words right. The 4096 wrong words read zero, the first at index 512; I'm on those now.
Ruled out: the G15C helper-table count (the kit's Mesa already writes ten entries for G15C), and the soft-fault setting (the J516C's own value was in place for these runs).
Three questions, mainly for Chris:
max/g15c-mesa-20261011(the kit'sff120e88) toiconidentify/mesa-omacom? I'd like to propose a G15C-only change on it that backs each USC heap guard page with the object's own extra page instead of leaving it unbound, with G15S and the kernel interface unchanged. Would that be acceptable there?0x3c000000140, matches your selftest's32N | 96N << 32with N = 10. Do you know what these two values size?Records:
agent-docswork/t6031-kit-refusals/results/(t77 to t80) on malik-na/linux-aurora; map malik-na#18, tickets malik-na#47 to #51.chriskearney commentedon Oct 13, 2026
@malik-na, answers to your three questions:
m3max/g15coniconidentify/mesa-omacomis exactly the kit'sff120e88(see my comment just above yours). A G15C-only change there is welcome as a PR againstm3max/g15c, with G15S, G15G and the kernel interface unchanged.drivers/gpu/drm/asahi/g15_selftest.rs) names itDupmMinCount | DupmMaxCount << 32 = 32N | 96N << 32, with N the maximum number of active USCs per MGPU. That's 9 on the J516S, because its core mask enables 9 of 10 cores per cluster. Its comment notes it hasn't been checked against another core count, so your N = 10 match on the J516C is the first confirmation we have.chriskearney commentedon Oct 13, 2026
@malik-na, a follow-up to question 2: the M3 Pro GPU owner hasn't recorded a G15S USC read into the guard page after an allocation in any M3 Pro run. The overread you see on the G15C is new to us, and we have no per-USC table-size evidence to share. Whatever you find is the first data on it.
paganotoni commentedon Oct 16, 2026
Kernel-only report: Mac15,8 (J514C)
Machine
uname -r:7.1.12-2-11-ARCH(linux-aurora 7.1.12.aurora2-11)v1.6.1-dirty, stage2v1.6.1-omarchy.aurora1(m1n1-aurora 1.6.1.aurora1-3)What works
simple-framebuffer, 3024x1890. Nothing external attached.GPU / power
--m3max-kit.Report (
--m3-report):a67d12fdb87d1bad2f3f75bb951d173b76edb13f4356cb35ab2beb5450b6e098malik-na commentedon Oct 18, 2026
M3 Max (J516C / T6031): the GPU runs on the G15 Rust driver, with compute and render right
Following up on my 14 Oct update, there are two results.
1. The half-wrong output was a Mesa bug, now fixed. When a program has any uniform range, its state-loader program
runs the "profile" prefix after LoadShader. The prefix ends in a computed helper call into the USC window base page:
base + 0x100 + 16 x a per-core value. On the G15C that value is out of range for some cores, past the two ten-entry
helper tables. Those calls never returned:
reported.
constants or descriptors.
It has the same signature as the G15S six-to-ten-slot helper-table fix (kit Mesa commit 38e10535). The fix puts a return at every slot
after the tables, on G15C only. No kernel-side page backing is needed, so the experimental USC window fill stays out
of the kernel. Credit to Neo's agx3b: reading its state-loader code against the kit's is what found the call.
2. The M3 Max now runs on Eryk's stack. I ported aurora-silicon/linux#358 (G15 Rust driver) and aurora-silicon/mesa#23 (G15S native path) to the J516C as
a T6031 row and a G15C admission, with the fix above:
With both, on the J516C (40 cores), the firmware reaches READY, and in 5 s each:
That is about twice the throughput of the omacom kernel with the kit Mesa. There was no fault or hang. The kernel PR
lists its gaps: the J516C board data comes from a pinned firmware file for now, the 14.8.x firmware build still needs a
pin, and the render register values are still G15S's.
Questions:
entirely also gives correct results here. I kept it so every core still runs it.
beyond G15C.
Records: agent-docs
work/t6031-kit-refusals/results/(t83, t84, t87) on malik-na/linux-aurora; mapmalik-na#18, tickets malik-na#50 and malik-na#52.
asahi: size the G17P USC helper tables by cores per cluster
jbnunn commentedon Nov 6, 2026
Kernel-only report: Mac15,11 (J516M / T6034)
Machine
uname -r:7.2.11-14-1-ARCH(linux-aurora 7.2.11.aurora14-1)v1.6.1-dirty, stage2bc79acc4-dirty(m1n1-aurora 1.6.1.aurora22-1)What works
simple-framebuffer, 3456x2160 (x2r10g10b10). Nothing external attached.What doesn't: built-in panel brightness
The brightness keys do nothing.
/sys/class/backlightis empty, soomarchy-brightness-displayfinds no device and exits 1. I know this is expected on the kernel-only Max path. I'm adding one J516M-specific detail in case it helps with native display:t6034-j516m.dtb(andt6034-j514m.dtb) inlinux-aurora 7.2.11.aurora14-1has no DCP or display-subsystem nodes at all, while the other M3 boards in the same package do:t6030-j516sapple,t6030-dcp,apple,t6030-dcpext,apple,t6030-display-subsystemt6031-j516capple,t6031-dcp@386c00000,apple,t6031-display-subsystemt6034-j516mThis affects the 14" T6034 in #333 (J514M) the same way.
appledrm.koin this kernel has aliases fort6030/t6031DCP but none fort6034. Even with the DT nodes added, the driver would still need T6034 support (or confirmation that T6031's applies).Did not run
--m3max-kit. I'm happy to test a J516M display build.paganotoni commentedon Nov 6, 2026
Updated kernel-only
--m3-reportfrom this 14-inch M3 Max (J514C / T6031), still on the stock Max path (Hyprland onsimpledrm, no GPU). Did not run--m3max-kitor arm GPU.7.2.11-14-1-ARCH(linux-aurora 7.2.11.aurora14-1)d9861670e44dd7226cc27cddf6b696e66b2383f85fd086f1cbc888cd49f8cfc2SMC keys are missing from this capture (no sudo/debugfs). Hypervisor board-word capture is still pending tracer steps.
paganotoni commentedon Nov 6, 2026
Sudo
--m3-reportfrom the same J514C boot, with SMC keys included. Prefer this over the earlier file.91aba325881aab4f8d78ac1e317715f24cf9b0b538a4c19d4a47b4245b0bb0d1