Linux 6.12 · iKKO MindOne

An unofficial forward-port to 6.12, done by hand — an independent research project

arm64 · 4 KiB pages · 6.12.92-android16-6-<stamp>-4k

---

⚠️ Experimental, research project, still moving

An independent, unofficial bring-up — not a vendor release, and built with no vendor support, assistance, or collaboration of any kind. It runs as a daily driver, but things change between commits, some hardware is unfinished, and nothing here is certified by anyone. Read Status before depending on it.

🎯 What this tree is for

The device shipped with a vendor kernel that stopped at Linux 5.10 (the android12-5.10 line, on Android 15). This tree carries it forward to a current, actively developed kernel line.

The kernel binary is Google's. This repository is the device half: the tree that describes the hardware, the configuration fragment that turns on what this board needs, and the 296 drivers it loads on top of a Generic Kernel Image.

kernel           //common:kernel_aarch64 — Google's GKI, unmodified gki_defconfig
this repository  device tree + mindone.fragment + 278 out-of-tree drivers + 18 in-tree platform drivers
contract         gki/aarch64/symbols/mindone — the symbols those drivers need from GKI
module_layout    0xf6779cbd

Status. The Kleaf GKI mixed build below is what runs on the device today, not merely the direction: kernel_aarch64 builds from an unmodified gki_defconfig + Google's own fragment, untouched by this tree, and the MediaTek platform drivers are either module_outs of a separate mindone kernel_build on top of it or ddk_module targets in mindone/modules/ — never inside kernel_aarch64 itself. Measured on the device (28.09): module_layout 0xf6779cbd, 296 modules loaded, and the running vmlinux has no KMI difference from Google's own ABI file for this base (0 of 776 documented imports missing). The standalone make path this document used to describe no longer builds a module set that matches the device — what that means for local iteration is in docs/BUILD.md.

No confidential vendor material went into this, and there was no vendor support or collaboration of any kind. The base is Google's Android Common Kernel. The device tree was reconstructed by hand from the stock DTB extracted off a retail unit — a compiled binary, not source, decoded and rebuilt from scratch as .dts/.dtsi. The out-of-tree drivers started from MediaTek's own published open-source kernel sources for this SoC family — the GPL sources MediaTek is required to release for its chips, not anything confidential or device-specific — then were fixed and adapted board by board; "What is actually ours" below says how much of that had to change.

📦 What is in here

Kernel	Google ACK android16-6.12-2026-09 (6.12.92), consumed as //common:kernel_aarch64 — this repository builds only the drivers on top of it
Device tree	mindone.dts + 18 mt6789-*.dtsi, rebuilt from the stock DTB, zero dtc warnings
Config	mindone/mindone.fragment on top of gki_defconfig, checked on every build by Kleaf's check_defconfig. arch/arm64/configs/mindone_defconfig remains for the legacy kernel-only make path, which no longer has a matching module build
Drivers	mindone/modules/ — 278 ddk_module targets, 3 257 C sources and headers
KMI	gki/aarch64/symbols/mindone — what those drivers need from the kernel, in the format Google's own tooling writes
Load order	mindone/vendor_boot.modules.load + mindone/vendor_dlkm.modules.load — the real 296, in the order the device loads them

🚫 No blobs, no firmware, no bootloader, no prebuilt binaries — verified by file content, not by extension. Proprietary userspace is extracted by each user from their own device.

🚀 Build

Android 16 dropped non-Bzlmod kernel builds, so Kleaf, the Bazel build Google ships with the ACK, is the supported entry point. The workspace is the ACK repo manifest with kernel/common replaced by this repository:

tools/bazel build //common/mindone:mindone_modules_install

A plain make build of the kernel image alone still works, without a full Kleaf workspace — see docs/BUILD.md for what it is still useful for and, more importantly, what it no longer builds:

TOOLCHAIN=/opt/toolchains/llvm-19.1.4-aarch64 ./build.sh ../k612-out

🔴 The clang version matters. clang decides symbol CRCs and module_layout. Build the kernel with one major version and the modules with another, and every module refuses to load without saying why — which is one reason the Kleaf path above, where both come from the same build, is the one that ships.

🔴 Memory. Generating BTF (pahole) peaks near 14.4 GiB. Under that ceiling the build dies with a bare Killed and FAILED: load BTF from vmlinux: Invalid argument, naming nothing. In a container or VM, check the limit that applies to the build, not the machine's total.

Modules, ABI checking, and the path to a flashable image → docs/BUILD.md.

🔧 What is actually ours

MediaTek keeps almost everything outside the kernel proper, and no device-specific material was available to start from. So most of this repository is work, not configuration.

most of 278 driver directories	carry local fixes: vendor code that was wrong for this board, targeted another SoC revision, or did not compile against a modern kernel. What changed and why is in the commit history for each driver.
2 modules written from scratch	for problems no vendor code solved — see the table below.
1 driver reconstructed	musb_hdrc_recon, a rebuilt MediaTek musb (built today as the `musb_hdrc` target).

The two written from scratch

mindone_thermal	thermal zones the stock tables never described
mindone_ufs_screen	storage behaviour specific to this board

🔌 The USB one is worth calling out. musb_hdrc_recon fixes a defect that made the device unable to sleep at all: on cable unplug the gadget teardown runs before the disconnect work, so the driver's own release path was unreachable and the "USB suspend lock" stayed held forever.

⬆️ Upstream work pulled back. UFS MediaTek clock-scaling and Vcore binding from v6.17 are merged into this 6.12 tree. A systematic diff of the MediaTek clock drivers against v6.18 was done and closed with no delta needed — a negative result worth stating, because it means the clock layer here is already current.

🖥 Hardware

✅ Works
Boot	~21 s to boot_completed
Audio	speaker, headset, Bluetooth, microphone; outgoing cellular-call microphone and earpiece/speaker switching verified with the matching userspace audio HAL
Cellular	calls, SMS, mobile data
Wireless	Wi-Fi, Bluetooth, NFC
Camera	both logical cameras of the flip module; the main IMX766 sensor's full 4096×3072 mode, brought up faster through burst I2C writes and chunked EEPROM calibration reads
Video	hardware decode and encode through the stateful V4L2 driver (mtk_vcodec), behind an open Codec2 HAL
Graphics	MDP
Security	fingerprint
System	USB (adb/MTP), charging, suspend/resume, 96 Hz display
Thermal	a kernel-owned skin-temperature zone at 46 °C, bound to CPU/GPU cooling in mindone_thermal

🚧 Unfinished
RPMB	hardware-backed key storage fails its MAC check; storage falls back safely, boot unaffected
60 Hz panel mode	exists on a branch, held back: the panel reports no physical size, so the second mode reaches apps at dpi 0
High-speed capture	120 fps not reached
HDR	not implemented
vSIM	parked deliberately, not attempted

🔁 Reproducible. Two builds from the same tree and toolchain now produce a byte-identical Image. build.sh pins the build user, host and SOURCE_DATE_EPOCH; before that the account name and the build moment were compiled in, and two builds differed by ~0.7 %. Verified on one machine into one output path — a different absolute build path is not separately tested.

Wi-Fi firmware pairing

IPv6-only address-buffer allocation and the firmware address-list encoding for IPv6 neighbor solicitation offload are fixed. RXFILTER now supports independent IPv4/IPv6 multicast policy while suspended, with capability probing and nonce-checked firmware readback. The accepted C37 driver/firmware pair passed multicast allow/block/restore and unicast controls on hardware; awake traffic retains the original policy. This is functional acceptance, not a measured battery-life improvement. APF remains disabled because the original firmware does not support its command interface.

RXFILTER requires the matching WIFI_RAM_CODE_soc2_2_1d_1.bin, SHA-256 bfb11a2b31fdd22aadb603b186e9ae0ebc98a476a7f7adf0f4a572d520abbfec. The companion device repository contains the firmware source tooling, release receipt and extraction checks. Firmware binaries are not distributed here; use your own extracted firmware and follow that repository's build instructions. An unsupported firmware query returns an error instead of reporting an applied RXFILTER policy. A successful ioctl alone does not establish wakeup or energy savings.

📊 Status

This is the kernel the device runs every day under LineageOS 23.2 (Android 16) — not a tree that merely compiles. But it is an independent bring-up by one person: expect rough edges, treat anything unusual as probably known, and keep a way to restore the stock firmware.

The battery driver still uses the userspace fuel gauge by default. The opt-in kernel algorithm has only a bounded powered handover/recovery check; complete discharge tracking, learned-state persistence and physical usable capacity have not been validated. Charger and GPU error-path fixtures do not establish hardware fault recovery or battery-life improvements.

Current focus is testing and optimisation rather than new features. Kernel 6.18 and Android 17 are where this goes next; both have been looked at, neither is committed to.

What has and has not been tested

Daily use	Yes — this is the author's phone, running this kernel every day
VTS / CTS	No. Neither has been run. No compatibility evidence exists, and none is claimed
Any certification	None, by anyone
Reproducible builds	Byte-identical Image on one machine into one output path. A different absolute build path is untested
Module ABI	Checked numerically on every build (module_layout, per-symbol CRC) — see docs/BUILD.md

📚 Documentation

docs/BUILD.md	Clean clone → kernel → modules → verified set
docs/MODULES.md	The driver trees: provenance, load order, gates
docs/HARDWARE.md	The device, and the behaviours that are invisible until they bite
docs/PORTING.md	Reusing this on another MT6789/MT8781 board

🕰 Where this came from

A forward-port. The device shipped on android12-5.10; the work went 5.10 → 6.1 → 6.12, and each step had to boot before the next could start.

This repository carries two branches. android16-6.12 is the one to build against -- that is where the drivers, the device tree, and all ongoing work live. android14-6.1 is the earlier step: the first forward-port of this device off its original vendor kernel, where the device tree and most of the driver work were first proven on real hardware. It is kept as history, unmaintained -- see its own README for what that means. The drivers are not duplicated there: mindone/modules/ lives only on android16-6.12. The current module build and acceptance target 6.12; the archived 6.1 branch is not a validated target for this module set.

🔎 Useful beyond this phone

Most of the work is SoC-level, not board-level, so it transfers to any MT6789 / MT8781 (Helio G99) bring-up on a modern kernel:

• 6.12 on a platform that shipped with 5.10 — the full path (android12-5.10 → android14-6.1 → android16-6.12-2026-09), including the API drift that actually hurts: scheduler, cpufreq and devfreq changes the MediaTek performance and DVFS helpers depend on.
• 278 out-of-tree MediaTek modules (275 loaded at boot, plus 17 more MediaTek platform drivers Kleaf compiles in-tree and 4 stock GKI modules — 296 in total) building against a current ACK — display, camera, audio DSP, Wi-Fi/BT, sensors, thermal, charging, UFS, USB, the CCCI modem interface, co-processors (sspm_v3, mcupm, vcp, adsp), memory and bus protection, and the GPU stack.
• A device tree reconstructed from a stock DTB to zero dtc warnings — 18 mt6789-*.dtsi covering clocks, pinctrl, regulators, display, camera, audio, thermal, reserved memory.
• SCMI/SSPM on MediaTek: upstream arm_scmi needs arm,scmi-shmem on the shared-memory nodes, which stock trees do not carry; the vendor tinysys protocol (id 0x80) needs a vendor_id on kernels ≥ 6.9.
• A reconstructed musb gadget/host driver, including the unplug path that otherwise holds a wakeup source and stops the device suspending entirely.

_{Search terms}

MT6789, MT8781, MT8781V-CA, Helio G99, MediaTek, mtk, android16-6.12, Android Common Kernel, ACK, LineageOS 23, Android 16, out-of-tree kernel modules, vendor modules, GKI, mtk-cmdq, wlan_drv_gen4m, eccci, ccci_mdinit, hf_manager, soc_temp_lvts, ufs-mediatek, sspm, mcupm, tinysys, device_apc_mt6789, imgsensor, mt6366, snd_soc_mt6789_afe, musb_hdrc, gen4m, MOLY modem firmware, md1img.

---

GPL-2.0, as upstream Linux — see COPYING and LICENSES/. The reconstructed device tree and everything in mindone/modules/ are under the same terms. Files carrying an upstream MediaTek dual GPL-2.0/BSD header keep it unchanged.