You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
M3 Air: qualify each shipped J613/J615 firmware profile separately #26
Scott confirms J613 exact26.6.2/.8 cold boot, native panel, active /opt/mesa-m3/25g83 and GPU checker pass with the published installer. J615 current14 and explicit native25 profiles are packaged; J615 physical acceptance is still pending. A system-firmware26.6.2 label alone does not select native25 GPU firmware.
Completion criteria: Retain a matrix for J613/J615 × current14/exact26.6.2 with actual stage1, stub, selected marker and packages. Check desktop/display, GL readback, correct legacy Vulkan or truthful native25-unavailable result, and error/recovery behavior. Do not repeat the already supplied J613 boot proof or copy calibration. J61525 must remain explicit; no automatic firmware/stage1 migration.
native, eDP-1 2560x1664 with notch, apple-panel-bl; J613/25G83 display handoff accepted
GPU
G16G: J613 power acknowledged; mesa-m3-session.statedecision=active, prefix=/opt/mesa-m3/25g83; aurora-m3-gpu-check --detailsresult=pass (native GL; Vulkan not claimed on this profile)
Wi-Fi
connected on boot, 5745 MHz VHT80 2×2, 585 Mbit/s
Logs
nothing new vs 20261010.8: the same 4 UATPageTable::prepare_map … already owned lines. The tas2764 … Failed to parse devicetree pair also appeared two boots ago on .8, so it isn't new either
ESP (#19):--archive-esp-history on .13 still refused with symlink in boot metadata: …/modules-7.1.12-2-12.0-sep-ARCH/build/include/dt-bindings/clock/qcom,dispcc-sm8150.h, as expected since #62 isn't released yet. I moved boot.bin.old and the limine_history UKI off the ESP by hand and the install ran. The ESP is back to 118 MB free, so this Air is ready to retest #62 in the next release.
Drafted with an AI assistant (Claude Code); reviewed before posting.
J613 on 2026.10.11.13, current14 / legacy profile: GPU FAIL (display fine). Fresh install with the release's install-aurora-sep.sh --read-only --m3-gpu. Auto-detect picked legacy: no GPU descriptor in the previous DT, stub and iBoot checks passed.
native, J613 display started: 2560x1664@60, 64 notch rows shown; one DPMS off/on cycle came back, no DCP fault
Wi-Fi
connected on boot (5240 MHz, 80 MHz)
GPU
fail: probe -110, no render node. mesa-m3-session.statedecision=inactive reason=no-gpu; aurora-m3-gpu-check --details: FAIL: Apple GPU is inactive (no-gpu)
GPU start, in order:
asahi 292400000.gpu: G15G: T8122 revision 2.0 identity admitted (10 active / 10 slots)
asahi 292400000.gpu: M3: GPU firmware identified as T8122 C0 firmware-compat 14.8.3 (start experiment)
asahi 292400000.gpu: M3 G15G start: performance cap 8 lowered to 2 (asahi.t8122_pstate_cap)
asahi 292400000.gpu: M3 G15G start: refused at the GPU registers and identity (EPROBE_DEFER); the firmware was not started (x3, then)
asahi 292400000.gpu: M3 G15G: PMP GPU device 5 power acknowledged
asahi 292400000.gpu: M3 G15G: power acknowledged, version=0x07022000 counts=0x0011010a core-mask=0x3ff, firmware=RTKit-2419.140.12.release
asahi 292400000.gpu: M3 G15G start: SGX setup write: SGX+0xd14000 = 0x70001 (asahi.t8122_sgx_setup=t6030)
asahi 292400000.gpu: M3 G15G: GPU handoff lock checked.
asahi 292400000.gpu: RTKit: Initializing (protocol version 12)
asahi 292400000.gpu: M3: InitData self-check passed: version 0xc08e21e83800490, 7 record pointers, 12 IO mappings, GPU region 0x103fffb8000
asahi 292400000.gpu: M3: publishing owned initdata 0xfffffc204062ff40
asahi 292400000.gpu: M3: device control 0x13 acknowledgement timed out, elapsed_ns=2000972146
asahi 292400000.gpu: M3 firmware readiness: magic=0x0 runtime=1 newA=1/1 newB=0
asahi 292400000.gpu: Asahi: sealed 587087-byte firmware fault snapshot offered to devcoredump (M3FWD001)
asahi 292400000.gpu: M3 G15G verdict: device-control-failed (ETIMEDOUT)
asahi 292400000.gpu: probe with driver asahi failed with error -110
This looks like the same current14 stall IcedTea363 and garymeyerca reported in chriskearney#35 (10 cores, mask 0x3ff, os-fw 14.7 / GPU firmware 14.8.3, system firmware 27.0), now reproduced on 2026.10.11.13 with the PMP power request in place. This Air is still on current14 and can test the DPE leakage candidate (HwDataA+0x418c..+0x41d8) or another build if that helps.
The one thing that differs from the stub-matched setup is the Mac's system firmware 27.0 (sepOS/iBoot1 from macOS 27). The installer says the system-firmware version doesn't select the profile; it may still matter for the GPU start. An --m3-report (taken just before the install, on the previous kernel) and the M3FWD001 devcoredump of this boot are saved and can be shared privately if useful. They are not attached here because they hold machine-specific data.
Not tested: Vulkan/GL (no GPU), sleep, external displays, Touch ID (read-only SEP).
Drafted with an AI assistant (Claude Code); reviewed before posting.
@satchlj, about the untested current14 candidate from chriskearney#35: the DPE leakage block at HwDataA+0x418c..+0x41d8 (20 words, nonzero on your J613 under macOS 26, zero in Aurora's current14 InitData).
This J613 (10 cores, mask 0x3ff, os-fw 14.7 / GPU firmware 14.8.3, system firmware 27.0) still stalls on current14 with 2026.10.11.13 (report above). It can test that block. Could you share the 20 words as your J613 had them, or the replacement branch once it exists? I would build the current14 kernel from the release's source and recipe with only that block set, cold-boot it as a separate GRUB entry, and report the G15G lines, aurora-m3-gpu-check and the readiness words from the M3FWD001 snapshot.
In the meantime I went through every combination of the three start knobs no report had varied yet: t8122_sgx_setup {t6030, none} × t8122_fender {rule, adt} × t8122_clkgen {e1c, e5c, none}, 12 cold boots on the same kernel (asahi.t8122_start=1 plus the knobs). The start lines confirm each setting took effect (no SGX setup write, fender=0x12c000, clkgen=e5c(...) etc.). All 12 end exactly like the default boot:
M3: device control 0x13 acknowledgement timed out, elapsed_ns=~2000000000
M3 firmware readiness: magic=0x0 runtime=1 newA=1/1 newB=0
M3 G15G verdict: device-control-failed (ETIMEDOUT)
probe with driver asahi failed with error -110
PMP power acknowledgement and the InitData self-check pass every time. So the stall doesn't depend on the SGX setup write, the Fender window or the clock-generator slot, alone or combined, which fits your reading that an InitData input is missing. All 12 dmesg logs and M3FWD001 snapshots are saved.
@GehrmannSolutions thanks for the 12-boot sweep. It's useful to have the SGX setup write, the Fender window and the clock-generator slot ruled out, alone and combined.
Here are the 20 words as macOS 26 writes them on our J613, so you can test the block on current14.
Source. macOS 26.6.2 running on a 10-core J613 (Mac15,12, core mask 0x3ff) under the m1n1 hypervisor. We read the HwData object back from guest memory at the end of each run. The block is byte-identical in all six captures (six separate boots). macOS writes the same 20 words into a second host object too.
Values (u32, little-endian, starting at HwDataA+0x418c):
They're integers, not floats. If your builder types this range as f32, write the raw u32 bits. Read as floats, these values are almost zero.
They aren't in the device tree. We searched both of this Mac's ADTs (14.x and 26.x stubs) in u32/u16/u64/float/ASCII encodings, with scalings. 12500 doesn't occur anywhere, and nothing under /arm-io/sgx or the PMGR/PMP nodes matches. The values also look like fixed configuration rather than per-machine calibration, but we've only measured one machine, so that isn't proven.
The 14.8.3 offset is our inference. The measurement is from macOS 26 firmware. We place the block at +0x418c in the current14 layout because the distance from the preceding block is the same in both layouts. Nothing shows the 14.8.3 firmware gives those bytes the same meaning.
Writer not confirmed. The captures were taken after the GPU firmware had run. The copy in a second host object suggests macOS writes the block, but that isn't confirmed.
For the A/B, setting only this block on top of your current default would be the cleanest test. The G15G lines, the 0x13 result and the readiness words (magic/runtime/newA/newB) from the M3FWD001 snapshot would show whether it changes anything.
Drafted with an AI assistant (Claude Code); reviewed before posting.
@satchljThe DPE block is it. With your 20 words at HwDataA+0x418c the current14 GPU starts on this J613, and both OpenGL and Vulkan pass.
Setup: J613, 10 cores (mask 0x3ff), os-fw 14.7 / GPU firmware 14.8.3 (RTKit-2419.140.12.release), system firmware 27.0, stage 1 v1.6.1. Kernel is 2026.10.11.13 (a85f8eee) built from the release source and recipe, plus a knob asahi.t8122_dpe_leak=1 that writes exactly the 20 u32 words (1, 500, 5, 12500 ×5, 3000 ×3, 12500 ×8, 1) at HwDataA+0x418c and nothing else. The P1 mirror stays zero. All other start knobs are at the defaults (sgx_setup=t6030, fender=rule, clkgen=e1c, pstate_cap=2). Cold boot.
M3: DPE leakage block set (HwDataA+0x418c, 20 words) (asahi.t8122_dpe_leak=1)
M3 G15G: power acknowledged, version=0x07022000 counts=0x0011010a core-mask=0x3ff, firmware=RTKit-2419.140.12.release
M3 firmware readiness: magic=0xabcdabcd runtime=1 newA=1/1 newB=0
M3 G15G verdict: firmware-running: the firmware accepted the InitData (version 0xc08e21e83800490) and the device controls; no job has run yet
M3 G15G verdict: job-completed: the first compute job finished, GPU timestamps 0x2ee76f7c..0x2ee7bf7b, 853291 ns of GPU time
M3 native render storage initialized, PB=256 MiB
M3 G15G verdict: job-completed: the first render job finished, GPU timestamps 0x2f488f07..0x2f491817, 1462000 ns of GPU time
Hyprland 0.56.2 renders on the GPU: Renderer: zink Vulkan 1.4(Apple M3 (G15G C0) (MESA_HONEYKRISP))
Same kernel, block + P1 mirror (dpe_leak=2, P1 +0x15c..+0x1a8 set as well)
also starts: magic=0xabcdabcd, compute and render jobs complete, OpenGL and Vulkan PASS
Same kernel, block off (dpe_leak=0)
the old stall: 0x13 ack timeout, magic=0x0 runtime=1 newA=1/1 newB=0, -110
So the block alone flips the result, on a machine whose system firmware is 27.0. That also suggests system firmware 27.0 isn't a blocker for current14.
One side note from the log: M3: thermal: no SoC die temperature zone (error -19); runtime cap at most 2 (618 MHz).
So HwDataA+0x418c alone is enough; the P1 mirror neither helps nor hurts the start. Not tested yet: longer runs, sleep, and whether the values are the same on an 8-core Air. The knob patch is small and can be shared if useful.
Posted by Claude (Claude Code), an AI assistant, on @satchlj's behalf.
The J613 stage 1 has a new pre-release: v1.6.1-m3air.stage1 on aurora-silicon/m1n1 (PR #31). It replaces j613-stage1-20261009.
Install steps are unchanged: fill the image with j613_stage1_fill.py, check it with --check, then run kmutil configure-boot from the paired 26.6.2 recoveryOS.
The version tag has changed. The stage 1 now reports v1.6.1-m3air.stage1 in /chosen/asahi,m1n1-stage1-version, so an installer that checks the version needs that tag listed.
The old j613-stage1-20261009 pre-release will be removed once the installer lists the new tag.
J615: the J615 image is experimental and hasn't been booted on a J615.
J613 current14 / legacy: GPU, OpenGL, Vulkan and accelerated Hyprland now work on my 8-core M3 Air (mask 0x37e). This follows my earlier failure report in chriskearney#35.
Published aurora-2026.10.11.13 source, a85f8eee2fd54ed9e7237f25d1f085e9d7848111, and release recipe, plus the local opt-in patch below
Installed test build
Separate linux-aurora-dpe 7.1.12.aurora2-20261011.131; running kernel 7.1.12-2-20261011.131-sep-ARCH, selected through Limine
Build tools
Clang / LLD 22.1.8, Rust 1.93.1
Desktop
Omarchy on Arch Linux ARM; hyprland 0.56.2-4; mesa-m3 26.1.4.m3.2-5, active prefix /opt/mesa-m3
The local patch independently reproduces the reported 20 raw little-endian u32 words at HwDataA+0x418c..+0x41d8: 1, 500, 5, 12500 ×5, 3000 ×3, 12500 ×8, 1. It is gated to T8122 / G15 / firmware 14.8.3, defaults off, and leaves the P1 mirror untouched. It is not a patch obtained from the earlier tester or a claim that the published release already contains the fix. Successful startup on this 8-core machine does not establish that these are its own measured macOS values or universally suitable calibration.
Other GPU start settings remained at the defaults: sgx_setup=t6030, fender=rule, clkgen=e1c, pstate_cap=2 (618 MHz). The first test boot kept mesa_m3=off for a standalone probe; the subsequent boot enabled the Mesa desktop session. No stage-1 or stub migration was performed; the machine remains on system firmware 27.0 with the current14 GPU profile. The ordinary GPU-off kernel entry remains available as the default fallback.
Check
Result
GPU startup
PASS:/dev/dri/renderD128 exists; firmware accepts InitData and device controls
Selected startup lines from the accelerated desktop boot:
M3: DPE leakage block set (HwDataA+0x418c, 20 words) (asahi.t8122_dpe_leak=1)
M3 firmware readiness: magic=0xabcdabcd runtime=1 newA=1/1 newB=0
M3 G15G verdict: firmware-running: the firmware accepted the InitData (version 0xc08e21e83800490) and the device controls; no job has run yet
The same boot then logged first compute and render job completion, with 867,291 ns and 1,247,666 ns of GPU time respectively. The standalone probe finished with gl.render=pass, gl.swap=pass, vk.job=pass words=65536 wrong=0, result=pass, exit=0, and loaded its Mesa implementation libraries from /opt/mesa-m3.
This is additional 8-core J613 current14 acceptance evidence with a local patch. We have not performed a fresh same-build DPE-off/on A/B, a long-duration stress test, suspend/resume, or external-display testing. The results above come from the standalone probe and desktop session/logs; they do not assert a separate aurora-m3-gpu-check --details run or completion of the whole J613/J615 qualification matrix.
Exact local opt-in patch
--- a/drivers/gpu/drm/asahi/asahi.rs+++ b/drivers/gpu/drm/asahi/asahi.rs@@ -130,6 +130,10 @@
description: "AGX GPU driver for Apple silicon SoCs",
license: "Dual MIT/GPL",
params: {
+ t8122_dpe_leak: u32 {+ default: 0,+ description: "Experimental current14 T8122 DPE block: 0 unchanged, 1 measured J613 values",+ },
g16_packet_commands: u32 {
default: 8,
description: "25G83 commands per scheduler packet (1..32, at most 16 per engine)",
--- a/drivers/gpu/drm/asahi/initdata.rs+++ b/drivers/gpu/drm/asahi/initdata.rs@@ -729,6 +729,46 @@
#[ver(V >= V14_8_3)]
for i in 0..tables.len {
raw.sram_k[i] = cfg.sram_k;
+ }++ // Opt-in reproduction of the J613 current14 experiment in+ // omacom/linux-aurora#26. These are raw u32 words, not floats.+ // No P1 mirror or other firmware layout is changed.+ #[ver(G == G15 && V == V14_8_3)]+ if cfg.chip_id == 0x8122 {+ match *module_parameters::t8122_dpe_leak.value() {+ 0 => (),+ 1 => {+ const OFFSET: usize = 0x418c;+ const WORDS: [u32; 20] = [+ 1, 500, 5,+ 12500, 12500, 12500, 12500, 12500,+ 3000, 3000, 3000,+ 12500, 12500, 12500, 12500,+ 12500, 12500, 12500, 12500, 1,+ ];+ kernel::static_assert!(+ core::mem::size_of::<raw::HwDataA::ver>() == 0x4360+ );+ kernel::static_assert!(OFFSET + 20 * 4 == 0x41dc);+ // SAFETY: raw exclusively borrows the initialized HwDataA.+ // Its exact G15/14.8.3 size is asserted above. All writes+ // lie within it; little-endian bytes permit unaligned+ // storage and preserve the integer bit patterns.+ unsafe {+ let dest = (raw as *mut raw::HwDataA::ver).cast::<u8>();+ for (i, word) in WORDS.iter().enumerate() {+ core::ptr::copy_nonoverlapping(+ word.to_le_bytes().as_ptr(),+ dest.add(OFFSET + i * 4),+ 4,+ );+ }+ }+ pr_info!("M3: DPE leakage block set (HwDataA+0x418c, 20 words) (asahi.t8122_dpe_leak=1)\n");+ }+ _ => return Err(EINVAL),+ }
}
// On G15 the G13 leak-coefficient arrays overlap the DPE leakage config
Posted by Claude (Claude Code), an AI assistant, on @satchlj's behalf.
Thanks @garymeyerca for the 8-core result. With the 10-core one, that is two J613s that start on 14.8.3 with the block
(the 10-core one also stalls again without it in a same-kernel A/B).
PR #95 writes the 20 words at HwDataA +0x418c..+0x41db by default, only on T8122 with the 14.8.3 GPU
firmware (G15). Other SoCs and the 26.x path are unchanged. The new parameter asahi.t8122_hwdata_418c turns it
off: =0 leaves the range at zero as before, for an A/B.
Could you confirm on that kernel?
Boot it as you boot current14 now, without any local patch for these words. You should see M3 G15G: HwDataA +0x418c..+0x41db: the 20 words measured on a J613 under macOS 26 ... and the firmware-running verdict.
Optionally boot once with asahi.t8122_hwdata_418c=0 and check that the device control 0x13 acknowledgement timed out line comes back.
Please include your core count and both firmware versions. We still don't know what the words mean, and they were
measured on one 10-core J613 only, so results from other machines (a J615 especially) are useful.
@satchlj: PASS on the original PR #95 head 3500d57cae7888ab5027eb10a5cfdc64f9dea9d3, without my previous local DPE patch. This answers the requested test in #26 (comment).
Prepared with Codex assistance from this machine's build logs and fresh current-boot checks.
The expected HwDataA log line appears, followed by:
M3 firmware readiness: magic=0xabcdabcd runtime=1 newA=1/1 newB=0
M3 G15G verdict: firmware-running: the firmware accepted the InitData (version 0xc08e21e83800490) and the device controls; no job has run yet
The first compute and render jobs then complete: 918,083 ns and 1,202,166 ns of GPU time. There is no device-control 0x13 acknowledgement timeout in the current boot's kernel log.
Honeykrisp Mesa 26.1.4, 65,536 words / 0 wrong, result=pass, exit=0
The existing cap remains 2 / 618 MHz. The kernel reports no SoC die temperature zone (error -19) and retains the cap.
Not tested: the optional same-kernel asahi.t8122_hwdata_418c=0 comparison, long-duration stress, suspend/resume, or external displays. The working locally patched kernel and normal GPU-off entry are retained.
I see the fix has since been carried through #112. This report verifies original PR #95 commit 3500d57…, not the rebased #112 commit or current integration/aurora. Their hardware qualification is still separate.
A data point for the rebased fix, with one caveat: not plain integration/aurora, but #114 on top of it.
Kernel:#114 (review/j613-external-dp at a372a728e), whose base is 1bf7fffcd (the merge of #112), plus six local patches for two external displays with the lid closed that touch only drivers/gpu/drm/apple, the display gate and the J613 DT, nothing in drm/asahi. No local GPU patch: our earlier DPE patch is gone.
Machine: J613 (MacBook Air 13" M3), GPU 10 cores (G15G: T8122 revision 2.0 identity admitted (10 active / 10 slots)). System firmware 27.0 (iBoot-10151.140.19.700.2), OS firmware 14.7. GPU firmware T8122 C0 firmware-compat 14.8.3.
Results:
M3 G15G: HwDataA +0x418c..+0x41db: wrote the 20 words (asahi.t8122_hwdata_418c=0 leaves them 0), then verdict: firmware-running, firmware ready, and job-completed for the first compute and the first render job. ✅
Accelerated desktop: Hyprland renders through zink on Honeykrisp (Apple M3 (G15G C0)); mesa-m3 session active. ✅
No GPU fault or timeout lines in any boot today.
Seven boots today on this base (J613 (M3 MacBook Air): external DisplayPort over USB-C, and T8122 PMP record requests #114 alone and with the six patches), with three s2idle suspend/resume cycles (lid open and lid closed) and up to two external displays (on dcpext and on the internal pipe). The GPU came back after each resume and the desktop kept rendering. The current boot has been up for three hours of normal desktop use, including the soak below.
32-minute GPU soak: llama.cpp (Vulkan on Honeykrisp, Mesa 26.1.4) with Qwen2.5-3B Q4_K_M, all layers on the GPU, 16 back-to-back llama-bench runs (pp512 + tg64), on battery. All runs completed; pp512 stayed at 32.30-32.36 t/s and tg64 at 1.82-1.86 t/s from the first run to the last, so no drift. No GPU fault, timeout or recovery lines; the only GPU lines were the periodic thermal summary (no SoC die temperature zone, runtime cap 2 (618 MHz)). The SMC temperatures we can read (NAND, Wi-Fi module, battery) stayed flat (<= 42 C), and battery draw averaged 9.7 W (max 16 W). For comparison, the same model on 4 CPU threads: pp128 22 t/s, tg32 10 t/s, so token generation on the capped GPU is about five times slower than on the CPU.
Not done: the asahi.t8122_hwdata_418c=0 A/B, explicit GL/Vulkan readback tests, and runs longer than three hours (this is now our daily kernel, so everyday use will add to it).
(Tests and this report by TARS, my Claude Code agent; reviewed by me.)
Drafted with an AI assistant (Claude Code); I reviewed it before posting.
A question about the M3 GPU on 7.2.10, from a J613 that now has the native display but renders in software.
Where the J613 is: current14 (14.x stub, Stage 1 m1n1 v1.6.1), linux-aurora 7.2.10.aurora1-1. Stage 2 is aura with the T8122 display handoff (aurora-silicon/aura#4). The panel comes up natively (J613 display started: 2560x1664@60, 64 notch rows shown). The GPU node stays disabled and the desktop renders in software, which is very slow at this resolution.
As I read #139 and 80797e4f, that's intended for now:
the M3 GPU runtime (firmware-event lifetime, GPUVM) is deferred to its maintainer's 7.2 redesign;
the carried backend is compile-only and refuses probe;
the commit message adds that T8122 "remains closed pending provenance decision on its register lists and InitData words".
So the 7.1.12 results above (#95/#112 on 8- and 10-core J613s) don't carry over to the current kernel yet.
Questions:
Who owns the 7.2 M3 GPU runtime, and is there a branch or issue to follow?
Is the provenance decision tracked anywhere? Is there anything testers can supply for it, or is it purely a maintainer call?
On 7.2, will drm/asahi still expect the boot loader to fill in the T8122 GPU node and enable it, as m1n1-aurora's G15 setup does on 7.1? Or will that move into the kernel? aura has no T8122 GPU setup today (ADT: GPU: unsupported chip!). I'm porting m1n1-aurora's as a draft so aura isn't the blocker, and I'd like to match whatever the 7.2 runtime will expect.
When there's something to test, this J613 (current14) is available, with aura and serial console.
G15G: T8122 revision 2.0 identity admitted (10 active / 10 slots) and the boot loader's power configuration were accepted (14 operating points).
The +0x418c words were written, and the GPU power-up was acknowledged: [drm] Initialized asahi 0.2.0.
mesa-m3 reports decision=active, and Hyprland renders on renderD128 with the native display and notch.
It runs at the default cap (state 2, 618 MHz).
So with aura as Stage 2, the boot loader's half of the 7.1 interface works on hardware. That covers the third question above for 7.1. Whatever the 7.2 M3 runtime ends up expecting from the boot loader, aura can be adapted to it, and this J613 can test it.
@scottjones replying to your 7.2 runtime questions: we have some related work underway on an 8-core J613, current14 / GPU firmware 14.8.3 (RTKit-2419.140.12.release).
We've built and installed a separate experimental kernel and matching Mesa for the G15 runtime and native AGX3 compiler. The pinned sources are aurora-silicon/linux at ee4f0f3ee2958ddc0243bb5e71a0bea88215d01b and aurora-silicon/mesa at c8947bf269511b94d4bf8d445518d7be27aa522d, with local changes. The resulting kernel is 7.1.12-ARCH-g15-j613-test1; this is not a working 7.2 port or the maintainer's announced redesign.
The candidate includes firmware identity checks, T8122 InitData integration, base-M3 shader scheduling, render/direct-compute command handling, receipt-based completion, and the existing thermal governor with state-2 fallback. Full ARM64 kernel/Mesa builds and host-side checks passed. Hardware qualification is still pending; no performance improvement has been established. The first test is configured for console-only operation at state 2 / 618 MHz, with thermal protection enabled. Our current desktop is still running the established runtime in a separate unified 7.1.12 kernel.
For your three questions:
I can't identify the official 7.2 runtime owner or development branch from our local work. This candidate is an independent experiment; no PR has been published for it.
We haven't resolved the provenance question. Successful builds or hardware tests would not settle that, and our experiment should not be read as provenance clearance. We have retained source revisions and local patches for review.
Our candidate still depends on the existing bootloader-supplied T8122 GPU setup, including board/power data, the PMP supplier link and logical device-5 power vote. We have left our shared m1n1/DTB unchanged. That describes our experiment's requirements, not a confirmed interface for upstream 7.2. We haven't tested it with aura or on a 10-core machine.
Separately, our higher-clock testing on the established runtime is tracked in #164. That evidence does not qualify this new runtime. Your aura/serial-console setup would provide a useful additional test configuration once there is a hardware-validated candidate to compare.
Prepared and posted with Codex assistance from the local build and validation records.
The isolated 7.1.12 test kernel booted cleanly again: the nine-class UAT check, firmware readiness, compute graph, and DRM initialization all passed, with no GPU MMU fault. The Mesa smoke test got past the shader-scheduling refusal from my previous update, then stopped during command-buffer finalization because its fail-closed check rejected the G15G wide-visibility CDM cache-flush barrier.
The tested Mesa source already documents Apple's G15 ComputeContext::encodeBarrier word (0x600215fe). G15G was admitted as a barrier target but missing from the wide-visibility allowlist. I added G15G to the allowlist and pack/decode coverage. Host checks passed (3/3), the direct asahi-agx3-packing test passed (1/1), and the ARM64 Mesa build passed. The build is installed only in the isolated test prefix, with the previous prefix backed up; the next hardware retest is pending.
There is still no successful GPU render/compute readback, so this is not hardware qualification or a performance result. It remains a local 7.1.12 experiment and does not resolve the 7.2 runtime or provenance questions.
Progress update from the 8-core J613 (current14 / GPU firmware 14.8.3) G15G experiment: the native runtime and AGX3 compiler have now completed bounded render and direct-compute readback on hardware.
Gary manually selected the isolated revision-4 test entry, kernel 7.1.12-ARCH-g15-j613-test1. It retained the conservative configuration: experimental runtime explicitly enabled, thermal control on, and ceiling/default/fallback all fixed at state 2 / 618 MHz. This remains a local 7.1.12 experiment based on kernel ee4f0f3ee2958ddc0243bb5e71a0bea88215d01b and Mesa c8947bf269511b94d4bf8d445518d7be27aa522d, with the previously described local prerequisites and fixes.
Startup passed the nine-class UAT leaf self-check, firmware identity and power admission, owned InitData/device-control acceptance, firmware READY transition, compute-graph construction, and asahi DRM initialization. The test used Mesa 26.3.0-devel through Zink/Honeykrisp on Apple M3 (G15G C0).
The bounded smoke check then passed twice in the same boot:
render: 4,096/4,096 pixels read back correctly on each run;
direct compute: 64/64 words correct on each run;
both processes exited 0;
post-test kernel logs contain no GPU fault, timeout, MMU fault, probe failure, receipt failure, warning, oops, hung-task, or over-cap indication.
The two passes establish bounded same-boot readback and repeatability for this exact revision. They do not establish a performance improvement, sustained-load stability, suspend/resume behavior, or higher-clock safety.
One important limitation remains: this candidate kernel does not expose macsmc_soc_die. The governor reports error -19 and stays fail-safe at state 2 / 618 MHz. No higher ceiling or sustained load was attempted. The next prerequisite for broader qualification is to restore and verify the J613 die-zone mapping on this candidate, while keeping the working unified kernel as the recovery default.
This result does not resolve the 7.2 runtime or provenance questions. Test execution and evidence review were performed with Codex assistance; Gary selected and unlocked the test boot.
Follow-up to the bounded state-2 readback result: the same 8-core J613 G15G experiment has now completed repeated bounded render/direct-compute testing at firmware state 8 / 1338 MHz.
The missing thermal input was the candidate kernel's macsmc-hwmon board table: M3 aggregation support was present, but apple,j613 had no default die-key entry. Revision 5 applies the exact 12-sensor J613 CPU/GPU die mapping from #146. On 7.1.12-ARCH-g15-j613-test2, all 12 sensors registered in macsmc_soc_die, the UAT self-check and firmware READY path passed, and DRM initialized normally.
The test entry retained state 2 / 618 MHz as its boot and fail-safe floor and set state 8 / 1338 MHz as the experimental ceiling. The governor raises one state at a time only on fresh cool readings, lowers at 75 C, and immediately returns to state 2 at 83 C or on loss of a usable sensor reading. The established unified kernel remains the persistent recovery default.
Results with Mesa 26.3.0-devel through Zink/Honeykrisp on Apple M3 (G15G C0):
baseline at cap 2: render 4,096/4,096 pixels and direct compute 64/64 words passed;
cap-8 qualification: 20 consecutive render/compute passes with exact readback;
tiny-workload continuity run: 2,643 consecutive passes in 120 seconds, all correct (the jobs were too small to drive firmware above state 3, so this is not counted as high-clock qualification);
sustained load, run 1: 60 seconds / 14,779 iterations, with 1,048,576 rendered pixels and 524,280 compute words verified exactly; firmware peak state 8, maximum die temperature 53.213 C;
independent sustained repeat: 60 seconds / 14,519 iterations, identical exact readbacks; firmware peak state 8, maximum die temperature 53.448 C.
The sustained compute shader performs 256 dependent integer rounds per each of 524,280 invocations, with every final value independently reproduced and checked on the CPU. Neither sustained run produced a GPU fault, MMU fault, timeout, receipt failure, probe failure, or over-cap indication.
One conservative behavior is visible at idle: if the aggregated die value remains bit-identical for 60 seconds, the policy treats it as a stuck reading and falls back to state 2. It automatically raises again once a changed reading arrives. This happened at roughly 40 C, not under thermal load; it is safe but the idle/quantization handling can be refined.
This establishes repeatable bounded state-8 operation for this exact local test2 revision. It is not yet a long-duration stability or performance result. Suspend/resume, external displays, longer stress, and the eventual 7.2 runtime remain unqualified, and this result does not resolve the provenance question.
Prepared and posted with Codex assistance from the saved build, boot, readback, temperature, and kernel-log evidence.
Follow-up to the state-8 qualification: I ran a bounded Mesa driver-optimization campaign on the same 8-core J613, current14 / GPU firmware 14.8.3, kernel 7.1.12-ARCH-g15-j613-test2, without changing clocks, voltages, thermal limits, or the recovery boot default.
The workload separately measures dependent integer compute, render-call throughput, and small direct-compute submission throughput, then combines them as a geometric-mean throughput index. Each full sample includes a 12-second warm-up and three 10-second phases. Every accepted sample had exact render/compute readback, no GPU/MMU/timeout/receipt/probe/over-cap fault, a usable die-temperature reading, and firmware state 8 / 1338 MHz reached. The initial three-sample median was 14,518.597; peak die temperature was 55.370 C.
Four G15G-only compiler variants were host-tested, cross-built on a separate Mac Pro, checksum-verified, installed into the isolated Mesa prefix, and measured on the Air:
Variant
Two-sample median vs initial baseline
Result
Enable the existing allowlisted loop-head hint for an iadd loop header
-1.95%
inconclusive; not retained
ILP scheduler ALU-latency model 3 instead of 4
-0.89%
inconclusive; not retained
ILP scheduler ALU-latency model 5 instead of 4
-1.46%
inconclusive; not retained
Keep the pressure schedule instead of ILP-rescheduling G15G loop blocks
-2.16%
inconclusive; not retained
The predeclared comparison threshold was 3%, so none of these changes established an improvement. All correctness and thermal gates passed, but no speculative change was merged. The campaign stopped at its four-iteration plateau rule.
I restored the original Mesa candidate and ran a fresh three-sample confirmation. Its median was 14,300.632, 1.50% below the initial snapshot and therefore still inside the same 3% variability band. All three final samples again passed exact readback and fault/thermal gates, reached state 8, and stayed at or below 52.245 C.
The useful result is negative but actionable: the current generic ILP latency of 4 remains the best measured setting among 3/4/5 for this workload, the existing G15G loop reschedule should remain enabled, and the unwitnessed iadd loop-head hint should not be enabled from this evidence. The remaining post-pack stall-padding idea was not attempted because the captured binary does not yet prove any stall is dead; weakening a live hazard would not be justified.
Testing and this update were performed with Codex assistance from the saved build, checksum, readback, temperature, and kernel-log records.
Follow-up qualification on the same 8-core J613 / G15G C0, current14 firmware profile, experimental kernel 7.1.12-ARCH-g15-j613-test2, and isolated Mesa 26.3.0-devel prefix:
One-hour state-8 mixed load: PASS
vkmark ran continuously on Honeykrisp while the guarded harness repeated dependent compute, rendering, and small direct-compute submission phases for 3,623.58 seconds.
85/85 samples: exact render and compute readback
85/85: fault-free, thermally safe, and firmware state 8 / 1338 MHz reached
14,359 independent die-temperature reads, zero read errors
maximum die temperature: 55.980 C
whole-window kernel scan: no GPU/MMU fault, timeout, receipt failure, recovery, over-cap, warning, oops, or hung-task signature
vkmark remained active, produced no stderr, and completed normally (final score 498)
The harness's median throughput index was 10,777.666 under deliberate contention with vkmark; this is not comparable to the standalone performance baseline.
Suspend/resume: FAIL
A fresh pre-suspend full sample passed exact readback, fault, thermal, and state-8 gates (throughput index 14,562.525; peak 48.635 C). The Air then entered s2idle and became unreachable.
RTC wake is not exposed on this kernel. Ethernet and Wi-Fi magic wake and their USB/PCI parent wake paths were enabled, with repeated wake packets sent from two wired hosts; they did not wake the machine. After a physical wake event, the Air performed a fresh boot rather than resuming:
boot ID changed;
last -x records the experimental-kernel session as crash;
the persistent unified recovery kernel booted automatically;
Btrfs performed tree-log replay, confirming an unclean restart;
no pstore record or GPU devcoredump survived.
The final previous-boot journal was unavailable after the unclean restart, so this evidence does not identify the failing component or prove the GPU caused the restart. It establishes only that this exact experimental-kernel boot does not survive the tested s2idle cycle. No post-resume GPU claim can be made.
The temporary wake settings were restored. The recovery system is writable and healthy, Btrfs device error counters are all zero, and the current kernel log has no GPU/MMU/Btrfs/oops/panic fault. Raw sample, temperature, vkmark, and kernel-window evidence plus checksums are preserved locally and off-machine.
Testing and evidence review were performed with Codex assistance at Gary's request.
Symptom: cursor never moves; the internal keyboard (same MTP dockchannel HID package) keeps typing. The HID device and its evdev node exist, but no touch reports ever flow — libinput never registers a pointer device, so the desktop never sees the pad.
Kernel signature: on every boot without exception, exactly 3 identical failure cycles, then it stops retrying until reset/reboot. Persistent journal (Oct 18 → Nov 3): 12 failure cycles across 4 boots, zero successes until one recovery (below):
rtkit-helper 2fa400000.mtp: RTKit: syslog message: boot.c:382: Calibration blob 'TCAL' updated
rtkit-helper 2fa400000.mtp: RTKit: syslog message: boot.c:382: Calibration blob 'FCAL' updated
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_bl_trp_hbpp.c:148: Failed to attach to AFE (-6)
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_bl_trp_hbpp.c:149: AFE Chip Boot Failure: No HINT_L
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_pwr.c:507: AFE bootload failed
rtkit-helper 2fa400000.mtp: RTKit: syslog message: touch_pwr_mgr.c:228: [TPM] Mac Power Manager State Machine Stopped
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_transport.c:206: afe_transport_stop already stopped
rtkit-helper 2fa400000.mtp: RTKit: syslog message: touch.c:88: Touch Reset Ext Request: OK
<same failure x2, then give up until next reset/reboot>
The coprocessor itself is otherwise healthy on those same boots: ST management ready, Keyboard ready, Actuator ready, Accel ready, TCAL/FCAL calibration loads fine. Only the touch AFE never attaches.
Recovery datapoint — by accident, while testing an in-place driver reset (first-ever clean bring-up in our journal):
The rebound firmware in that boot never came up — after RTKit: Initializing (protocol version 12) it emitted nothing for minutes and never attempted an AFE attach. The rebind alone did not restore the pad in place; it also dropped both HID interfaces (keyboard and multi-touch — both live on this coprocessor), leaving the user's internal inputs dead until the next reboot.
A full reboot after that reset did come up clean — first attempt, zero retries (12 failures → the journal's only success):
A second in-place reset within the same healthy boot crashed the rebound firmware outright — worse than the silent hang:
rtkit-helper 2fa400000.mtp: RTKit:Initializing (protocol version 12)
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_transport.c:220: Z2SPI AFE transport stopped
rtkit-helper 2fa400000.mtp: RTKit: syslog message: touch.c:88: Touch Reset Ext Request: OK
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_pwr.c:493: failed waiting on Touch reset completion event
rtkit-helper 2fa400000.mtp: RTKit: co-processor has crashed
rtkit-helper 2fa400000.mtp: RTKit: Version: RTKit-2419.140.12.release - Client: AppleMTPFirmwareMac-3300.7~470
rtkit-helper 2fa400000.mtp: PSR = 0x600012c5 PC = 0x1054560 ESR = 0xbe000000 FAR = 0x3fffffd33fc
rtkit-helper 2fa400000.mtp: RTKit: (full exception dump + mailbox history available on request)
Same consequence: both HID interfaces dropped; inputs only returned after a reboot. So the practical procedure is strictly "reset once, then reboot" — repeated in-place resets are actively harmful.
Working hypothesis: the touch AFE gets stuck in a latched bad power/boot state that ordinary power cycles do not clear (hence No HINT_L — the bootloader never raises its hint line), and an in-place MTP coprocessor reset — fragile as it is — appears to open a path to a clean attach on the next boot. Confidence caveat: one success immediately after four uniformly failing boots — strong correlation, not proof. Follow-up on persistence: since that one recovery the pad has stayed clean across the recovery boot, a full power-off overnight, and the next morning's cold boot — the wedge has not recurred so far (will keep watching).
Current workaround should the failure recur (run from a TTY; expect both inputs dark until the following reboot):
I can supply full boot logs (failing and successful bring-ups, incl. the crash dump), dmesg of a failing boot, devicetree dumps for 2fa400000.mtp / 2fab28000.input (and their mailbox/IOMMU nodes), or more cold-boot datapoints on request. If there's anything useful to test on a real j615 (e.g. the experimental v1.6.1-m3air.stage1 image referenced here), happy to.
Drafted with an AI assistant (Claude Code) at my request.
Thanks for such a careful report. The reset-then-reboot datapoint is especially useful: four boots failing the same way, then a clean first attempt right after one reset, is the clearest lead we have on this. Good to hear it has stayed clean since, cold boot included.
One request: please avoid repeated in-place resets of the MTP. The second one crashed the coprocessor and took the keyboard down with it, so it's not worth the risk.
We don't have a J615, so your reports are what we have to go on. A few questions:
Did the trackpad ever work on Linux on this Mac? If an earlier kernel worked, booting it (Limine entry or snapshot) the next time it fails would tell us whether aurora14 regressed. Which installer version and which kernel did you first boot?
Cold boot vs restart: does it fail the same way after a full shutdown with a minute's wait as after a restart, especially a restart straight from macOS? And with and without the charger connected?
Firmware: the output of cat /proc/device-tree/chosen/asahi,os-fw-version and cat /proc/device-tree/chosen/asahi,system-fw-version, plus your installer version.
One boot's kernel log with timestamps: journalctl -k -b | grep -E 'rtkit-helper|mtp|dockchannel|spi|AFE'
Does the trackpad work in macOS right now?
Is anything attached at boot (USB-C devices, external display, dock)?
OS-platform datapoint: this never appeared before the macOS "Liquid Glass" update — and it now reproduces on both macOS and Omarchy.
For the record (your account-holder's experience, j615 as described in my earlier comment):
Before updating macOS to the 26 "Liquid Glass" release, this trackpad never showed the problem — on either OS.
Since that update the failure happens on both macOS and Omarchy: the trackpad silently produces no pointer events in macOS too, matching the Linux-side "HID node exists, touch AFE never attaches" behavior.
That cross-OS reproduction makes the Linux bring-up story less likely as the primary cause, and fits what the logs already suggested: the touch AFE on the pad assembly ends up in a latched state where it never raises its hint line (No HINT_L), and that state survives ordinary power cycles. An in-place coprocessor reset + reboot (our recovery) appears to have cleared the persisted state once — and the pad has stayed clean since, across boots and a full power-off.
Questions others on affected hardware may be able to answer:
Does the same AFE wedge (No HINT_L after reset-loop) show up in macOS diagnostics/console logs on updates-era machines?
A/B check: does the Linux-side rtkit-helper rebind + reboot also fix behavior on the macOS side for anyone hit by this?
Did the 26.x update change anything trackpad-related (touch-assembly AFE, calibration, coprocessor firmware)? For comparison, under aurora the crashed dump identifies the MTP firmware as AppleMTPFirmwareMac-3300.7~470 — worth cross-checking against what updated macOS now provisions.
Will update this thread if the wedge recurs on our machine.
Update: the AFE wedge returned; the one-time reset recovery does not reproduce — j615 touch bring-up is effectively blocked
Follow-up to my comment above. Machine has been in near-continuous use for ~2 days since; here is the complete picture.
Unchanged failure signature — exactly 3 cycles on every failing boot:
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_bl_trp_hbpp.c:148: Failed to attach to AFE (-6)
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_bl_trp_hbpp.c:149: AFE Chip Boot Failure: No HINT_L
rtkit-helper 2fa400000.mtp: RTKit: syslog message: afe_pwr.c:507: AFE bootload failed
rtkit-helper 2fa400000.mtp: RTKit: syslog message: touch_pwr_mgr.c:228: [TPM] Mac Power Manager State Machine Stopped
Everything else on the MTP coprocessor is healthy on those same boots (ST management ready, Keyboard ready, Actuator ready, Accel ready, TCAL/FCAL loaded; keyboard HID works) — only the touch AFE never raises its hint line. Persistent journal totals: 11 of 13 boots wedged, exactly 2 clean boots ever — both immediately after the first-ever in-place rtkit-helper unbind/rebind on this unit (Nov 3 ~03:16), followed by a power-off.
All attempts to reproduce the recovery failed:
attempt
sequence
result
1
in-place rtkit reset → warm reboot 5 s later
wedge persists
2
in-place rtkit reset → full power-off → ~6 min off → power-on
One further data point: the second in-place reset ever attempted (during a healthy session) crashed the rebound firmware — RTKit: co-processor has crashed, ESR=0xbe000000 (SP alignment), PC 0x1054560, preceded by afe_pwr.c:493: failed waiting on Touch reset completion event and afe_transport.c:220: Z2SPI AFE transport stopped. Both HID interfaces vanished until the next reboot; full exception dump + mailbox history available on request.
Question for the folks working this profile-qualification issue: is the MTP/AFE bring-up tied to the j615 firmware-profile selection tracked here, and could this unit's v1.6.1-dirty stage1 chain be selecting a non-qualified Apple firmware set (vs. the shipped/qualified j615 image)? This machine is available as a test unit — happy to flash any official or experimental j615 image (including v1.6.1-m3air.stage1) and provide complete before/after journals, dmesg, and device-tree dumps. For now it runs with an external mouse.
Activity
scottjones commentedon Oct 11, 2026
Drafted with an AI assistant (Claude Code); I reviewed it before posting.
J613 on 2026.10.11.13: PASS, updated in place from
20261010.8with the omacom one-liner and--m3-profile=j613-25g83.v1.6.1-m3next.stage1linux-aurora 7.1.12.aurora2-20261011.13,m1n1-aurora 1.6.1.aurora17-5(stage 2v1.6.1-omarchy.aurora17),mesa-m3 26.1.4.m3.2-5apple-panel-bl;J613/25G83 display handoff acceptedG16G: J613 power acknowledged;mesa-m3-session.statedecision=active,prefix=/opt/mesa-m3/25g83;aurora-m3-gpu-check --detailsresult=pass(native GL; Vulkan not claimed on this profile)20261010.8: the same 4UATPageTable::prepare_map … already ownedlines. Thetas2764 … Failed to parse devicetreepair also appeared two boots ago on.8, so it isn't new eitherESP (#19):
--archive-esp-historyon.13still refused withsymlink in boot metadata: …/modules-7.1.12-2-12.0-sep-ARCH/build/include/dt-bindings/clock/qcom,dispcc-sm8150.h, as expected since #62 isn't released yet. I movedboot.bin.oldand thelimine_historyUKI off the ESP by hand and the install ran. The ESP is back to 118 MB free, so this Air is ready to retest #62 in the next release.GehrmannSolutions commentedon Oct 11, 2026
Drafted with an AI assistant (Claude Code); reviewed before posting.
J613 on 2026.10.11.13, current14 / legacy profile: GPU FAIL (display fine). Fresh install with the release's
install-aurora-sep.sh --read-only --m3-gpu. Auto-detect pickedlegacy: no GPU descriptor in the previous DT, stub and iBoot checks passed.stub_info.jsonProductVersion 14.8.3, iBoot2iBoot-10151.140.19.700.2, os-fw 14.7, system-fw 27.0, stage 1 cleanv1.6.1linux-aurora 7.1.12.aurora2-20261011.13(7.1.12-2-20261011.13-sep-ARCH),m1n1-aurora 1.6.1.aurora17-5(stage 2v1.6.1-omarchy.aurora17),mesa-m3 26.1.4.m3.2-5chosen.asahi,t8122-dcp=1,t8122-gpu=1,t8122-gpu-power-standin=1,t8122-gpu-fuse-leakage=1J613 display started: 2560x1664@60, 64 notch rows shown; one DPMS off/on cycle came back, no DCP faultmesa-m3-session.statedecision=inactive reason=no-gpu;aurora-m3-gpu-check --details:FAIL: Apple GPU is inactive (no-gpu)GPU start, in order:
This looks like the same current14 stall IcedTea363 and garymeyerca reported in chriskearney#35 (10 cores, mask
0x3ff, os-fw 14.7 / GPU firmware 14.8.3, system firmware 27.0), now reproduced on 2026.10.11.13 with the PMP power request in place. This Air is still on current14 and can test the DPE leakage candidate (HwDataA+0x418c..+0x41d8) or another build if that helps.The one thing that differs from the stub-matched setup is the Mac's system firmware 27.0 (sepOS/iBoot1 from macOS 27). The installer says the system-firmware version doesn't select the profile; it may still matter for the GPU start. An
--m3-report(taken just before the install, on the previous kernel) and the M3FWD001 devcoredump of this boot are saved and can be shared privately if useful. They are not attached here because they hold machine-specific data.Not tested: Vulkan/GL (no GPU), sleep, external displays, Touch ID (read-only SEP).
GehrmannSolutions commentedon Oct 11, 2026
Drafted with an AI assistant (Claude Code); reviewed before posting.
@satchlj, about the untested current14 candidate from chriskearney#35: the DPE leakage block at
HwDataA+0x418c..+0x41d8(20 words, nonzero on your J613 under macOS 26, zero in Aurora's current14 InitData).This J613 (10 cores, mask
0x3ff, os-fw 14.7 / GPU firmware 14.8.3, system firmware 27.0) still stalls on current14 with 2026.10.11.13 (report above). It can test that block. Could you share the 20 words as your J613 had them, or the replacement branch once it exists? I would build the current14 kernel from the release's source and recipe with only that block set, cold-boot it as a separate GRUB entry, and report theG15Glines,aurora-m3-gpu-checkand the readiness words from theM3FWD001snapshot.In the meantime I went through every combination of the three start knobs no report had varied yet:
t8122_sgx_setup{t6030,none} ×t8122_fender{rule,adt} ×t8122_clkgen{e1c,e5c,none}, 12 cold boots on the same kernel (asahi.t8122_start=1plus the knobs). The start lines confirm each setting took effect (no SGX setup write,fender=0x12c000,clkgen=e5c(...)etc.). All 12 end exactly like the default boot:PMP power acknowledgement and the InitData self-check pass every time. So the stall doesn't depend on the SGX setup write, the Fender window or the clock-generator slot, alone or combined, which fits your reading that an InitData input is missing. All 12 dmesg logs and M3FWD001 snapshots are saved.
satchlj commentedon Oct 11, 2026
@GehrmannSolutions thanks for the 12-boot sweep. It's useful to have the SGX setup write, the Fender window and the clock-generator slot ruled out, alone and combined.
Here are the 20 words as macOS 26 writes them on our J613, so you can test the block on current14.
Source. macOS 26.6.2 running on a 10-core J613 (Mac15,12, core mask
0x3ff) under the m1n1 hypervisor. We read the HwData object back from guest memory at the end of each run. The block is byte-identical in all six captures (six separate boots). macOS writes the same 20 words into a second host object too.Values (u32, little-endian, starting at
HwDataA+0x418c):i.e.
1, 500, 5, 12500 ×5, 3000 ×3, 12500 ×8, 1.Things to know before you test:
f32, write the rawu32bits. Read as floats, these values are almost zero./arm-io/sgxor the PMGR/PMP nodes matches. The values also look like fixed configuration rather than per-machine calibration, but we've only measured one machine, so that isn't proven.+0x418cin the current14 layout because the distance from the preceding block is the same in both layouts. Nothing shows the 14.8.3 firmware gives those bytes the same meaning.For the A/B, setting only this block on top of your current default would be the cleanest test. The
G15Glines, the0x13result and the readiness words (magic/runtime/newA/newB) from theM3FWD001snapshot would show whether it changes anything.🤖 Generated with Claude Code
GehrmannSolutions commentedon Oct 11, 2026
Drafted with an AI assistant (Claude Code); reviewed before posting.
@satchlj The DPE block is it. With your 20 words at
HwDataA+0x418cthe current14 GPU starts on this J613, and both OpenGL and Vulkan pass.Setup: J613, 10 cores (mask
0x3ff), os-fw 14.7 / GPU firmware 14.8.3 (RTKit-2419.140.12.release), system firmware 27.0, stage 1v1.6.1. Kernel is 2026.10.11.13 (a85f8eee) built from the release source and recipe, plus a knobasahi.t8122_dpe_leak=1that writes exactly the 20 u32 words (1, 500, 5, 12500 ×5, 3000 ×3, 12500 ×8, 1) atHwDataA+0x418cand nothing else. The P1 mirror stays zero. All other start knobs are at the defaults (sgx_setup=t6030,fender=rule,clkgen=e1c,pstate_cap=2). Cold boot.aurora-m3-gpu-checkmesa-m3-session.statedecision=active reason=activeRenderer: zink Vulkan 1.4(Apple M3 (G15G C0) (MESA_HONEYKRISP))dpe_leak=2, P1+0x15c..+0x1a8set as well)magic=0xabcdabcd, compute and render jobs complete, OpenGL and Vulkan PASSdpe_leak=0)0x13ack timeout,magic=0x0 runtime=1 newA=1/1 newB=0, -110So the block alone flips the result, on a machine whose system firmware is 27.0. That also suggests system firmware 27.0 isn't a blocker for current14.
One side note from the log:
M3: thermal: no SoC die temperature zone (error -19); runtime cap at most 2 (618 MHz).So
HwDataA+0x418calone is enough; the P1 mirror neither helps nor hurts the start. Not tested yet: longer runs, sleep, and whether the values are the same on an 8-core Air. The knob patch is small and can be shared if useful.satchlj commentedon Oct 12, 2026
Posted by Claude (Claude Code), an AI assistant, on @satchlj's behalf.
The J613 stage 1 has a new pre-release:
v1.6.1-m3air.stage1on aurora-silicon/m1n1 (PR #31). It replacesj613-stage1-20261009.j613_stage1_fill.py, check it with--check, then runkmutil configure-bootfrom the paired 26.6.2 recoveryOS.v1.6.1-m3air.stage1in/chosen/asahi,m1n1-stage1-version, so an installer that checks the version needs that tag listed.j613-stage1-20261009pre-release will be removed once the installer lists the new tag.garymeyerca commentedon Oct 13, 2026
J613 current14 / legacy: GPU, OpenGL, Vulkan and accelerated Hyprland now work on my 8-core M3 Air (mask
0x37e). This follows my earlier failure report in chriskearney#35.Posted with Codex assistance at my request, using this machine's build logs and runtime results. Thanks @satchlj for the DPE candidate and @GehrmannSolutions for the 10-core success report.
apple,j613,apple,t8122, C0 / revision 2.0; 8 active GPU cores / 10 slots, mask0x37eos-fw-version=14.7,system-fw-version=27.0; GPU firmware-compat 14.8.3,RTKit-2419.140.12.release; current14 / legacyv1.6.1-dirty/v1.6.1-omarchy.aurora17;m1n1-aurora 1.6.1.aurora17-5aurora-2026.10.11.13source,a85f8eee2fd54ed9e7237f25d1f085e9d7848111, and release recipe, plus the local opt-in patch belowlinux-aurora-dpe 7.1.12.aurora2-20261011.131; running kernel7.1.12-2-20261011.131-sep-ARCH, selected through Liminehyprland 0.56.2-4;mesa-m3 26.1.4.m3.2-5, active prefix/opt/mesa-m3The local patch independently reproduces the reported 20 raw little-endian u32 words at
HwDataA+0x418c..+0x41d8:1, 500, 5, 12500 ×5, 3000 ×3, 12500 ×8, 1. It is gated to T8122 / G15 / firmware 14.8.3, defaults off, and leaves the P1 mirror untouched. It is not a patch obtained from the earlier tester or a claim that the published release already contains the fix. Successful startup on this 8-core machine does not establish that these are its own measured macOS values or universally suitable calibration.Relevant boot options:
Other GPU start settings remained at the defaults:
sgx_setup=t6030,fender=rule,clkgen=e1c,pstate_cap=2(618 MHz). The first test boot keptmesa_m3=offfor a standalone probe; the subsequent boot enabled the Mesa desktop session. No stage-1 or stub migration was performed; the machine remains on system firmware 27.0 with the current14 GPU profile. The ordinary GPU-off kernel entry remains available as the default fallback./dev/dri/renderD128exists; firmware accepts InitData and device controls/opt/mesa-m3/bin/mesa-m3-probe, legacy ABIchip=0x8122 G15G rev=0x20 dies=1 clusters=1 USC3 HAL0 params=600 match=1glerror=0x0, window swap; GLES 3.2 Mesa 26.1.4decision=active reason=active fallback=none prefix=/opt/mesa-m3zink Vulkan 1.4(Apple M3 (G15G C0) (MESA_HONEYKRISP)); desktop operation confirmedSelected startup lines from the accelerated desktop boot:
The same boot then logged first compute and render job completion, with 867,291 ns and 1,247,666 ns of GPU time respectively. The standalone probe finished with
gl.render=pass,gl.swap=pass,vk.job=pass words=65536 wrong=0,result=pass,exit=0, and loaded its Mesa implementation libraries from/opt/mesa-m3.This is additional 8-core J613 current14 acceptance evidence with a local patch. We have not performed a fresh same-build DPE-off/on A/B, a long-duration stress test, suspend/resume, or external-display testing. The results above come from the standalone probe and desktop session/logs; they do not assert a separate
aurora-m3-gpu-check --detailsrun or completion of the whole J613/J615 qualification matrix.Exact local opt-in patch
satchlj commentedon Oct 13, 2026
Posted by Claude (Claude Code), an AI assistant, on @satchlj's behalf.
Thanks @garymeyerca for the 8-core result. With the 10-core one, that is two J613s that start on 14.8.3 with the block
(the 10-core one also stalls again without it in a same-kernel A/B).
PR #95 writes the 20 words at HwDataA +0x418c..+0x41db by default, only on T8122 with the 14.8.3 GPU
firmware (G15). Other SoCs and the 26.x path are unchanged. The new parameter
asahi.t8122_hwdata_418cturns itoff:
=0leaves the range at zero as before, for an A/B.Could you confirm on that kernel?
M3 G15G: HwDataA +0x418c..+0x41db: the 20 words measured on a J613 under macOS 26 ...and thefirmware-runningverdict.asahi.t8122_hwdata_418c=0and check that thedevice control 0x13 acknowledgement timed outline comes back.Please include your core count and both firmware versions. We still don't know what the words mean, and they were
measured on one 10-core J613 only, so results from other machines (a J615 especially) are useful.
garymeyerca commentedon Oct 13, 2026
@satchlj: PASS on the original PR #95 head
3500d57cae7888ab5027eb10a5cfdc64f9dea9d3, without my previous local DPE patch. This answers the requested test in #26 (comment).Prepared with Codex assistance from this machine's build logs and fresh current-boot checks.
0x37eRTKit-2419.140.12.release; current14 / legacyv1.6.1-dirty, stage 2v1.6.1-omarchy.aurora17; no stage-1 or stub migration7.1.12-2-20261013.95-sep-ARCH; separatelinux-aurora-pr95 7.1.12.aurora2-20261013.95package, installed and booted through Limineasahi.t8122_start=1; no oldasahi.t8122_dpe_leakoption and noasahi.t8122_hwdata_418coverride; sysfs confirms the upstream parameter defaults to 1mesa-m3 26.1.4.m3.2-5, prefix/opt/mesa-m3, Hyprland 0.56.2The expected HwDataA log line appears, followed by:
The first compute and render jobs then complete: 918,083 ns and 1,202,166 ns of GPU time. There is no device-control 0x13 acknowledgement timeout in the current boot's kernel log.
J613 display started: 2560x1664@60, 64 notch rows showndecision=active reason=active fallback=none prefix=/opt/mesa-m3zink Vulkan 1.4(Apple M3 (G15G C0) (MESA_HONEYKRISP))aurora-m3-gpu-check --detailsgl.render=pass, correct red/blue pixels,glerror=0x0,gl.swap=passresult=pass,exit=0The existing cap remains 2 / 618 MHz. The kernel reports no SoC die temperature zone (
error -19) and retains the cap.Not tested: the optional same-kernel
asahi.t8122_hwdata_418c=0comparison, long-duration stress, suspend/resume, or external displays. The working locally patched kernel and normal GPU-off entry are retained.I see the fix has since been carried through #112. This report verifies original PR #95 commit
3500d57…, not the rebased #112 commit or currentintegration/aurora. Their hardware qualification is still separate.1 remaining item
GehrmannSolutions commentedon Oct 14, 2026
A data point for the rebased fix, with one caveat: not plain
integration/aurora, but #114 on top of it.Kernel: #114 (
review/j613-external-dpata372a728e), whose base is1bf7fffcd(the merge of #112), plus six local patches for two external displays with the lid closed that touch onlydrivers/gpu/drm/apple, the display gate and the J613 DT, nothing indrm/asahi. No local GPU patch: our earlier DPE patch is gone.Machine: J613 (MacBook Air 13" M3), GPU 10 cores (
G15G: T8122 revision 2.0 identity admitted (10 active / 10 slots)). System firmware 27.0 (iBoot-10151.140.19.700.2), OS firmware 14.7. GPU firmwareT8122 C0 firmware-compat 14.8.3.Results:
M3 G15G: HwDataA +0x418c..+0x41db: wrote the 20 words (asahi.t8122_hwdata_418c=0 leaves them 0), thenverdict: firmware-running,firmware ready, andjob-completedfor the first compute and the first render job. ✅Apple M3 (G15G C0)); mesa-m3 session active. ✅llama-benchruns (pp512 + tg64), on battery. All runs completed; pp512 stayed at 32.30-32.36 t/s and tg64 at 1.82-1.86 t/s from the first run to the last, so no drift. No GPU fault, timeout or recovery lines; the only GPU lines were the periodic thermal summary (no SoC die temperature zone, runtime cap 2 (618 MHz)). The SMC temperatures we can read (NAND, Wi-Fi module, battery) stayed flat (<= 42 C), and battery draw averaged 9.7 W (max 16 W). For comparison, the same model on 4 CPU threads: pp128 22 t/s, tg32 10 t/s, so token generation on the capped GPU is about five times slower than on the CPU.Not done: the
asahi.t8122_hwdata_418c=0A/B, explicit GL/Vulkan readback tests, and runs longer than three hours (this is now our daily kernel, so everyday use will add to it).(Tests and this report by TARS, my Claude Code agent; reviewed by me.)
drm/asahi: m3: T8122: write the HwDataA +0x418c block on 14.8.3
drm/asahi: m3: T8122: write the HwDataA +0x418c block on 14.8.3
scottjones commentedon Oct 21, 2026
Drafted with an AI assistant (Claude Code); I reviewed it before posting.
A question about the M3 GPU on 7.2.10, from a J613 that now has the native display but renders in software.
Where the J613 is: current14 (14.x stub, Stage 1
m1n1 v1.6.1),linux-aurora 7.2.10.aurora1-1. Stage 2 is aura with the T8122 display handoff (aurora-silicon/aura#4). The panel comes up natively (J613 display started: 2560x1664@60, 64 notch rows shown). The GPU node stays disabled and the desktop renders in software, which is very slow at this resolution.As I read #139 and
80797e4f, that's intended for now:So the 7.1.12 results above (#95/#112 on 8- and 10-core J613s) don't carry over to the current kernel yet.
Questions:
ADT: GPU: unsupported chip!). I'm porting m1n1-aurora's as a draft so aura isn't the blocker, and I'd like to match whatever the 7.2 runtime will expect.When there's something to test, this J613 (current14) is available, with aura and serial console.
scottjones commentedon Oct 21, 2026
Drafted with an AI assistant (Claude Code); I reviewed it before posting.
A follow-up to my question above: the same J613 now runs its GPU, on 7.1.12 rather than 7.2.
0x3ff), GPU firmware 14.8.3 (current14), 14.x stub, Stage 1m1n1 v1.6.1c46d7dcd, 7.1.12) from the release recipe, as a side-by-side package;asahi.t8122_start=1on its entry only/etc/m1n1.conft8122-dcp=1,t8122-gpu=1,t8122-gpu-power-standin=1,t8122-gpu-fuse-leakage=1mesa-m3 26.1.4.m3.2-5,/etc/mesa-m3/t8122-gpu-experimentResult:
G15G: T8122 revision 2.0 identity admitted (10 active / 10 slots)and the boot loader's power configuration were accepted (14 operating points).[drm] Initialized asahi 0.2.0.decision=active, and Hyprland renders onrenderD128with the native display and notch.So with aura as Stage 2, the boot loader's half of the 7.1 interface works on hardware. That covers the third question above for 7.1. Whatever the 7.2 M3 runtime ends up expecting from the boot loader, aura can be adapted to it, and this J613 can test it.
Not tested yet: suspend/resume and long GPU runs.
garymeyerca commentedon Oct 21, 2026
@scottjones replying to your 7.2 runtime questions: we have some related work underway on an 8-core J613, current14 / GPU firmware 14.8.3 (
RTKit-2419.140.12.release).We've built and installed a separate experimental kernel and matching Mesa for the G15 runtime and native AGX3 compiler. The pinned sources are
aurora-silicon/linuxatee4f0f3ee2958ddc0243bb5e71a0bea88215d01bandaurora-silicon/mesaatc8947bf269511b94d4bf8d445518d7be27aa522d, with local changes. The resulting kernel is7.1.12-ARCH-g15-j613-test1; this is not a working 7.2 port or the maintainer's announced redesign.The candidate includes firmware identity checks, T8122 InitData integration, base-M3 shader scheduling, render/direct-compute command handling, receipt-based completion, and the existing thermal governor with state-2 fallback. Full ARM64 kernel/Mesa builds and host-side checks passed. Hardware qualification is still pending; no performance improvement has been established. The first test is configured for console-only operation at state 2 / 618 MHz, with thermal protection enabled. Our current desktop is still running the established runtime in a separate unified 7.1.12 kernel.
For your three questions:
Separately, our higher-clock testing on the established runtime is tracked in #164. That evidence does not qualify this new runtime. Your aura/serial-console setup would provide a useful additional test configuration once there is a hardware-validated candidate to compare.
Prepared and posted with Codex assistance from the local build and validation records.
garymeyerca commentedon Oct 21, 2026
Progress update from our 8-core J613 (current14 / GPU firmware 14.8.3) experiment:
The isolated 7.1.12 test kernel booted cleanly again: the nine-class UAT check, firmware readiness, compute graph, and DRM initialization all passed, with no GPU MMU fault. The Mesa smoke test got past the shader-scheduling refusal from my previous update, then stopped during command-buffer finalization because its fail-closed check rejected the G15G wide-visibility CDM cache-flush barrier.
The tested Mesa source already documents Apple's G15
ComputeContext::encodeBarrierword (0x600215fe). G15G was admitted as a barrier target but missing from the wide-visibility allowlist. I added G15G to the allowlist and pack/decode coverage. Host checks passed (3/3), the directasahi-agx3-packingtest passed (1/1), and the ARM64 Mesa build passed. The build is installed only in the isolated test prefix, with the previous prefix backed up; the next hardware retest is pending.There is still no successful GPU render/compute readback, so this is not hardware qualification or a performance result. It remains a local 7.1.12 experiment and does not resolve the 7.2 runtime or provenance questions.
garymeyerca commentedon Oct 22, 2026
Progress update from the 8-core J613 (current14 / GPU firmware 14.8.3) G15G experiment: the native runtime and AGX3 compiler have now completed bounded render and direct-compute readback on hardware.
Gary manually selected the isolated revision-4 test entry, kernel
7.1.12-ARCH-g15-j613-test1. It retained the conservative configuration: experimental runtime explicitly enabled, thermal control on, and ceiling/default/fallback all fixed at state 2 / 618 MHz. This remains a local 7.1.12 experiment based on kernelee4f0f3ee2958ddc0243bb5e71a0bea88215d01band Mesac8947bf269511b94d4bf8d445518d7be27aa522d, with the previously described local prerequisites and fixes.Startup passed the nine-class UAT leaf self-check, firmware identity and power admission, owned InitData/device-control acceptance, firmware READY transition, compute-graph construction, and asahi DRM initialization. The test used Mesa 26.3.0-devel through Zink/Honeykrisp on
Apple M3 (G15G C0).The bounded smoke check then passed twice in the same boot:
The two passes establish bounded same-boot readback and repeatability for this exact revision. They do not establish a performance improvement, sustained-load stability, suspend/resume behavior, or higher-clock safety.
One important limitation remains: this candidate kernel does not expose
macsmc_soc_die. The governor reports error-19and stays fail-safe at state 2 / 618 MHz. No higher ceiling or sustained load was attempted. The next prerequisite for broader qualification is to restore and verify the J613 die-zone mapping on this candidate, while keeping the working unified kernel as the recovery default.This result does not resolve the 7.2 runtime or provenance questions. Test execution and evidence review were performed with Codex assistance; Gary selected and unlocked the test boot.
garymeyerca commentedon Oct 22, 2026
Follow-up to the bounded state-2 readback result: the same 8-core J613 G15G experiment has now completed repeated bounded render/direct-compute testing at firmware state 8 / 1338 MHz.
The missing thermal input was the candidate kernel's
macsmc-hwmonboard table: M3 aggregation support was present, butapple,j613had no default die-key entry. Revision 5 applies the exact 12-sensor J613 CPU/GPU die mapping from #146. On7.1.12-ARCH-g15-j613-test2, all 12 sensors registered inmacsmc_soc_die, the UAT self-check and firmware READY path passed, and DRM initialized normally.The test entry retained state 2 / 618 MHz as its boot and fail-safe floor and set state 8 / 1338 MHz as the experimental ceiling. The governor raises one state at a time only on fresh cool readings, lowers at 75 C, and immediately returns to state 2 at 83 C or on loss of a usable sensor reading. The established unified kernel remains the persistent recovery default.
Results with Mesa 26.3.0-devel through Zink/Honeykrisp on
Apple M3 (G15G C0):The sustained compute shader performs 256 dependent integer rounds per each of 524,280 invocations, with every final value independently reproduced and checked on the CPU. Neither sustained run produced a GPU fault, MMU fault, timeout, receipt failure, probe failure, or over-cap indication.
One conservative behavior is visible at idle: if the aggregated die value remains bit-identical for 60 seconds, the policy treats it as a stuck reading and falls back to state 2. It automatically raises again once a changed reading arrives. This happened at roughly 40 C, not under thermal load; it is safe but the idle/quantization handling can be refined.
This establishes repeatable bounded state-8 operation for this exact local test2 revision. It is not yet a long-duration stability or performance result. Suspend/resume, external displays, longer stress, and the eventual 7.2 runtime remain unqualified, and this result does not resolve the provenance question.
Prepared and posted with Codex assistance from the saved build, boot, readback, temperature, and kernel-log evidence.
garymeyerca commentedon Oct 22, 2026
Follow-up to the state-8 qualification: I ran a bounded Mesa driver-optimization campaign on the same 8-core J613, current14 / GPU firmware 14.8.3, kernel
7.1.12-ARCH-g15-j613-test2, without changing clocks, voltages, thermal limits, or the recovery boot default.The workload separately measures dependent integer compute, render-call throughput, and small direct-compute submission throughput, then combines them as a geometric-mean throughput index. Each full sample includes a 12-second warm-up and three 10-second phases. Every accepted sample had exact render/compute readback, no GPU/MMU/timeout/receipt/probe/over-cap fault, a usable die-temperature reading, and firmware state 8 / 1338 MHz reached. The initial three-sample median was 14,518.597; peak die temperature was 55.370 C.
Four G15G-only compiler variants were host-tested, cross-built on a separate Mac Pro, checksum-verified, installed into the isolated Mesa prefix, and measured on the Air:
iaddloop headerThe predeclared comparison threshold was 3%, so none of these changes established an improvement. All correctness and thermal gates passed, but no speculative change was merged. The campaign stopped at its four-iteration plateau rule.
I restored the original Mesa candidate and ran a fresh three-sample confirmation. Its median was 14,300.632, 1.50% below the initial snapshot and therefore still inside the same 3% variability band. All three final samples again passed exact readback and fault/thermal gates, reached state 8, and stayed at or below 52.245 C.
The useful result is negative but actionable: the current generic ILP latency of 4 remains the best measured setting among 3/4/5 for this workload, the existing G15G loop reschedule should remain enabled, and the unwitnessed
iaddloop-head hint should not be enabled from this evidence. The remaining post-pack stall-padding idea was not attempted because the captured binary does not yet prove any stall is dead; weakening a live hazard would not be justified.Testing and this update were performed with Codex assistance from the saved build, checksum, readback, temperature, and kernel-log records.
garymeyerca commentedon Oct 25, 2026
Follow-up qualification on the same 8-core J613 / G15G C0, current14 firmware profile, experimental kernel
7.1.12-ARCH-g15-j613-test2, and isolated Mesa 26.3.0-devel prefix:One-hour state-8 mixed load: PASS
vkmarkran continuously on Honeykrisp while the guarded harness repeated dependent compute, rendering, and small direct-compute submission phases for 3,623.58 seconds.vkmarkremained active, produced no stderr, and completed normally (final score 498)The harness's median throughput index was 10,777.666 under deliberate contention with
vkmark; this is not comparable to the standalone performance baseline.Suspend/resume: FAIL
A fresh pre-suspend full sample passed exact readback, fault, thermal, and state-8 gates (throughput index 14,562.525; peak 48.635 C). The Air then entered s2idle and became unreachable.
RTC wake is not exposed on this kernel. Ethernet and Wi-Fi magic wake and their USB/PCI parent wake paths were enabled, with repeated wake packets sent from two wired hosts; they did not wake the machine. After a physical wake event, the Air performed a fresh boot rather than resuming:
last -xrecords the experimental-kernel session ascrash;The final previous-boot journal was unavailable after the unclean restart, so this evidence does not identify the failing component or prove the GPU caused the restart. It establishes only that this exact experimental-kernel boot does not survive the tested s2idle cycle. No post-resume GPU claim can be made.
The temporary wake settings were restored. The recovery system is writable and healthy, Btrfs device error counters are all zero, and the current kernel log has no GPU/MMU/Btrfs/oops/panic fault. Raw sample, temperature,
vkmark, and kernel-window evidence plus checksums are preserved locally and off-machine.Testing and evidence review were performed with Codex assistance at Gary's request.
M4Marvin commentedon Nov 3, 2026
Machine: MacBook Air 15" M3 2024 — DMI
Mac15,13, DT j615 (t8122). Touch pad HID Uniq:HXMHB3000ZD0000F9A+B5LNon2fab14000.fifo/2fab28000.input, dockchannel-hid05AC:0358with two HID interfaces: multi-touch.0002/input/input0and keyboard.0003/input/input1.Software: linux-aurora
7.2.11.aurora14-1(7.2.11-14-1-ARCH), m1n1-aurora1.6.1.aurora22-1— stage1 self-reportsv1.6.1-dirty, stage2bc79acc4-dirty; omarchy-mac0.2.3-2, Hyprland 0.56.2.Symptom: cursor never moves; the internal keyboard (same MTP dockchannel HID package) keeps typing. The HID device and its evdev node exist, but no touch reports ever flow — libinput never registers a pointer device, so the desktop never sees the pad.
Kernel signature: on every boot without exception, exactly 3 identical failure cycles, then it stops retrying until reset/reboot. Persistent journal (Oct 18 → Nov 3): 12 failure cycles across 4 boots, zero successes until one recovery (below):
The coprocessor itself is otherwise healthy on those same boots:
ST management ready,Keyboard ready,Actuator ready,Accel ready, TCAL/FCAL calibration loads fine. Only the touch AFE never attaches.Recovery datapoint — by accident, while testing an in-place driver reset (first-ever clean bring-up in our journal):
What that taught, in three observations:
RTKit: Initializing (protocol version 12)it emitted nothing for minutes and never attempted an AFE attach. The rebind alone did not restore the pad in place; it also dropped both HID interfaces (keyboard and multi-touch — both live on this coprocessor), leaving the user's internal inputs dead until the next reboot.Working hypothesis: the touch AFE gets stuck in a latched bad power/boot state that ordinary power cycles do not clear (hence
No HINT_L— the bootloader never raises its hint line), and an in-place MTP coprocessor reset — fragile as it is — appears to open a path to a clean attach on the next boot. Confidence caveat: one success immediately after four uniformly failing boots — strong correlation, not proof. Follow-up on persistence: since that one recovery the pad has stayed clean across the recovery boot, a full power-off overnight, and the next morning's cold boot — the wedge has not recurred so far (will keep watching).Current workaround should the failure recur (run from a TTY; expect both inputs dark until the following reboot):
I can supply full boot logs (failing and successful bring-ups, incl. the crash dump),
dmesgof a failing boot, devicetree dumps for2fa400000.mtp/2fab28000.input(and their mailbox/IOMMU nodes), or more cold-boot datapoints on request. If there's anything useful to test on a real j615 (e.g. the experimentalv1.6.1-m3air.stage1image referenced here), happy to.scottjones commentedon Nov 3, 2026
Drafted with an AI assistant (Claude Code) at my request.
Thanks for such a careful report. The reset-then-reboot datapoint is especially useful: four boots failing the same way, then a clean first attempt right after one reset, is the clearest lead we have on this. Good to hear it has stayed clean since, cold boot included.
One request: please avoid repeated in-place resets of the MTP. The second one crashed the coprocessor and took the keyboard down with it, so it's not worth the risk.
We don't have a J615, so your reports are what we have to go on. A few questions:
cat /proc/device-tree/chosen/asahi,os-fw-versionandcat /proc/device-tree/chosen/asahi,system-fw-version, plus your installer version.journalctl -k -b | grep -E 'rtkit-helper|mtp|dockchannel|spi|AFE'M4Marvin commentedon Nov 3, 2026
OS-platform datapoint: this never appeared before the macOS "Liquid Glass" update — and it now reproduces on both macOS and Omarchy.
For the record (your account-holder's experience, j615 as described in my earlier comment):
No HINT_L), and that state survives ordinary power cycles. An in-place coprocessor reset + reboot (our recovery) appears to have cleared the persisted state once — and the pad has stayed clean since, across boots and a full power-off.No HINT_Lafter reset-loop) show up in macOS diagnostics/console logs on updates-era machines?rtkit-helperrebind + reboot also fix behavior on the macOS side for anyone hit by this?AppleMTPFirmwareMac-3300.7~470— worth cross-checking against what updated macOS now provisions.M4Marvin commentedon Nov 3, 2026
Update: the AFE wedge returned; the one-time reset recovery does not reproduce — j615 touch bring-up is effectively blocked
Follow-up to my comment above. Machine has been in near-continuous use for ~2 days since; here is the complete picture.
Unchanged failure signature — exactly 3 cycles on every failing boot:
Everything else on the MTP coprocessor is healthy on those same boots (
ST management ready,Keyboard ready,Actuator ready,Accel ready, TCAL/FCAL loaded; keyboard HID works) — only the touch AFE never raises its hint line. Persistent journal totals: 11 of 13 boots wedged, exactly 2 clean boots ever — both immediately after the first-ever in-placertkit-helperunbind/rebind on this unit (Nov 3 ~03:16), followed by a power-off.All attempts to reproduce the recovery failed:
One further data point: the second in-place reset ever attempted (during a healthy session) crashed the rebound firmware —
RTKit: co-processor has crashed,ESR=0xbe000000(SP alignment), PC0x1054560, preceded byafe_pwr.c:493: failed waiting on Touch reset completion eventandafe_transport.c:220: Z2SPI AFE transport stopped. Both HID interfaces vanished until the next reboot; full exception dump + mailbox history available on request.Versions (unchanged throughout): linux-aurora
7.2.11.aurora14-1(source commit4fdd0fc5c931ef9f14474942b37c3d451e97fc04), m1n1-aurora1.6.1.aurora22-1— stage1 self-reportsv1.6.1-dirty, stage2bc79acc4-dirty. Crash banner from the reset experiment:RTKit: RTKit-2419.140.12.release - Client: AppleMTPFirmwareMac-3300.7~470.Question for the folks working this profile-qualification issue: is the MTP/AFE bring-up tied to the j615 firmware-profile selection tracked here, and could this unit's
v1.6.1-dirtystage1 chain be selecting a non-qualified Apple firmware set (vs. the shipped/qualified j615 image)? This machine is available as a test unit — happy to flash any official or experimental j615 image (includingv1.6.1-m3air.stage1) and provide complete before/after journals,dmesg, and device-tree dumps. For now it runs with an external mouse.