[bong-water-water-bong]

The appliance had no fan or power control at all (grep -rn 'fan|power_mode' os/ was empty), so on a
board with an embedded controller it ran on the firmware's fan curve, which is conservative, and reported no
fan readings or temperatures of its own. This brings the development box's setup into the image for the
boards it knows, and writes nothing anywhere else.

os/mini/board.sh (installed as /usr/sbin/1bit-board, BusyBox sh, no Python, no systemd):

• the EC fan curve (mode=curve plus the ramp-up/ramp-down trip points, which the EC applies on its own sensor
  with hysteresis);
• a predictive boost loop, 0.5 s poll: package power ≥ 35 W or GPU busy ≥ 40 % or Tctl ≥ 70 °C or a
  3 °C rise per poll → the CPU fans at level 4; back to the curve after 20 s quiet; level 5 only above 82 °C,
  held to 76 °C;
• the power profile: the EC power mode first (a mode write resets the SMU limits), then the SMU limits through
  ryzenadj, then the fabric/SoC clock pins — every value overridable from /etc/1bit/power, SKIP_POWER=1 to
  leave power alone;
• driver only: it never touches debug registers.

Board match is DMI based (BeyondMax Series / AXB35-02). On any other board every command exits without
writing a byte, status says fans: not controlled (this board is not one we know; the firmware curve stays),
and the boot banner says the same.

Deterministic by construction. os/mini/mkimage.sh builds ec_su_axb35 for the image kernel from the
source this repository ships (os/licenses/ec_su_axb35, GPL-2.0, verbatim upstream f62c2c22, built in a temp
copy so the tree stays clean) and builds ryzenadj from the pinned source in os/licenses/ryzenadj
(LGPL-3.0, tag v0.19.0, commit ea71739b) rather than depending on what the build machine has. Both source
trees and licences are installed into the image under /usr/share/licenses/1bit-os/, so every binary travels
with its corresponding source. A failed build fails the image build loudly; EC_MODULE=skip ships without the
driver.

Two Secure Boot traps, both handled:

1. Module signing. An out-of-tree module must be signed by a key the kernel trusts, and it has to be the
   MOK: the "Secure Boot Kernel" certificate carries no module-signing OID (verified on the development box —
   amdxdna.ko and ec_su_axb35.ko are both signed by strixhalo Secure Boot MOK, and only MOK.der has
   1.3.6.1.4.1.2312.16.1.2). mkimage.sh now signs the module with MOK.priv/MOK.der through the kernel's
   scripts/sign-file (MODULE_SIGN_KEY/MODULE_SIGN_CERT override) and warns loudly if it cannot.
2. The SMU cap and lockdown. ryzenadj writes the SMU mailbox through /dev/mem, which lockdown
   (integrity) refuses. On the development box that never bites because the kernel does not enter lockdown from
   Secure Boot: /sys/kernel/security/lockdown reads [none] and CONFIG_LOCK_DOWN_IN_EFI_SECURE_BOOT is not
   set in the shipped kernel config — the same kernel the image carries. If a machine ever does refuse the write
   (lockdown forced, no /dev/mem, or no tool), 1bit-board power now falls back to the EC's own balanced
   profile (85 W sustained, 120 W burst) instead of leaving the board in performance with no cap, and both
   the banner and power-status say which cap is in force (cap=90/90/105W … vs fallback=balanced …).

os/mini/init.sh: on a known board, load the EC driver, apply the curve and the power profile, start the
boost loop, and report [fans] …, [power] … and temps: Tctl …C EC …C in the banner and in the 1bit.check
self-test. Unknown board → one line, nothing written.

os/mini/release-check.sh: a fans row —

PASS  fans : controlled (fan1 4122 rpm, fan2 4091 rpm at level 4)
SKIP  fans : this board has no EC we drive: the firmware fan curve stays        (also SKIP in a VM)
FAIL  fans : the driver loaded but a fan reads 0 rpm at level 4 (…): the curve and the boost cannot cool this board

The FAIL case is deliberate: it is exactly how a fan failed on our own hardware.

/etc/1bit/power ships as a commented template whose defaults are the development box's profile
(performance mode, 90/90/105 W, level-5 trip at 82 °C unchanged — the appliance is the same board and pad class),
with the fan trip points marked as the first thing to revisit on a hotter unit.

os/LICENSES.md: the board driver (GPL-2.0) and the SMU tool (LGPL-3.0, pinned) with corresponding-source
entries.

Verified on the development box (same board, same kernel): known, curve (identical trip points), status
with live rpms, boost raising the fans on load and releasing to the curve on TERM, power exercised end to
end against a fake EC — including the fallback path (ryzenadj forced to fail → EC mode becomes balanced,
banner says so) — so the real profile was never touched; the /etc/1bit/power parser (comments, spaces,
SKIP_POWER=1, FALLBACK_MODE); tools/copyright.py --check passes; sh -n clean on all four scripts.

Draft on purpose — image boot on hardware pending: the module build, the ryzenadj build and the VM boot test
run once the build lane opens (EC_MODULE=skip builds an image without the driver). No behaviour change on
machines without this board.

---

Verified by building the image and booting it in QEMU (24 Oct, build lane)

os/mini/mkimage.sh on this board, under a build hold (gpu-hold -l "build pi osfans" -t 30, nice -n 19 taskset -c 0-3), against ~/wt/release-rc/work/build:

step	result
board driver	ec_su_axb35.ko 33,846 bytes, in /lib/modules/<kver>/updates/, signed by MOK.der — modinfo -F signer says strixhalo Secure Boot MOK
SMU tool	ryzenadj v0.19.0 built from os/licenses/ryzenadj/src (68,184 bytes, 6 libs linked) → /opt/1bit/ryzenadj
initramfs contents	usr/sbin/1bit-board, etc/1bit/power, the signed .ko, opt/1bit/ryzenadj, and both licence trees with their full source (…/1bit-os/ec_su_axb35/src 31 files, …/1bit-os/ryzenadj/src 47 files)
image	1bit-os.img 2.3 GB, 1bit-os.efi 97 MB, initramfs.img 86 MB

os/mini/vmtest-release.sh (QEMU, KVM, OVMF, the real release check over SSH):

RESULT: PASS (16 checks, no failures)        # image with the driver
SKIP  fans : this board has no EC we drive: the firmware fan curve stays
RESULT: PASS (16 checks, no failures)        # EC_MODULE=skip image
SKIP  fans : this board has no EC we drive: the firmware fan curve stays

i.e. the SKIP path is exercised exactly as intended (no EC in a VM), the EC_MODULE=skip build still boots
and passes, and nothing else regressed (chat, bench, sealed, licences all PASS; chat-hrx, chat-npu,
identity SKIP for the documented VM reasons).

Three things the build lane taught this PR, all fixed here: the image kernel is clang-built, so the module
must build with LLVM=1 and the matching /usr/lib/llvm-<major>/bin on PATH (gcc chokes on
-mretpoline-external-thunk, and LLVM=1 alone dies on ld.lld: not found); the MOK private key is
0600 root, so signing goes through sudo -n when needed and always from a scratch copy (a root-owned
module inside the image tree broke the later chmod/cpio); $out/logs has to exist before the build logs
open.

Still pending on hardware: a boot on the real board with 1bit-release-check (the fans row should PASS
with live rpms there) — that is the owner's step, after the leaderboard is done.

[github-actions]

Preparing review...

[bong-water-water-bong]

Review (orchestrator): DMI-gated, driver-only, no byte written on an unknown board, GPL/LGPL sources shipped with the image, fans row with the dead-fan FAIL case; host checks against a fake EC are the right call while the lane is busy. Decisions: ship the SMU tool pinned (version + source commit recorded in os/, not 'if the builder has it') so the 90 W cap is deterministic; default power file on the known board = the dev box profile. Converted to draft until the module build + VMTEST (QEMU reaches the SKIP path, EC_MODULE=skip image still boots) run at 'lane shared'; the hardware boot + release check is the owner's step after 'leaderboard done'.

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit f49e7fe)

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 4 🔵🔵🔵🔵⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Possible Issue The predictive boost loop in boost() function may not correctly handle the case where last temperature is not updated when load is 0. This can lead to incorrect boost behavior when the temperature does not change but other conditions (like power or GPU busy) are met. now=$(date +%s); load=0 if [ "$w" -ge "$BOOST_W" ] || [ "$busy" -ge "$BOOST_BUSY" ] || [ "$tc" -ge "$BOOST_HOT_C" ] \ || [ $(( tc - last )) -ge "$BOOST_SLOPE" ]; then load=1; fi Possible Issue The power() function uses a hardcoded sleep of 2 seconds after writing the EC power mode. This may not be sufficient for all hardware or may cause unnecessary delays. A more robust method for ensuring the EC has applied the mode should be considered. sleep 2 Possible Issue The build process for ec_su_axb35.ko may fail silently if the LLVM=1 flag is not properly set or if the required LLVM toolchain is not found. This could lead to an unsigned module being installed, which may not load correctly on systems with Secure Boot enabled. [ -x "$d/ld.lld" ] && [ -x "$d/clang" ] && { PATH="$d:$PATH"; export PATH; kmake="LLVM=1"; break; } done [ -n "$kmake" ] || echo " no LLVM toolchain (ld.lld) on this builder: the kernel is clang-built, the module will not build" fi

---

⚠️ Review coverage: The following files were not included in this review because of the token budget:

• os/licenses/ryzenadj/src/lib/api.c
• os/licenses/ec_su_axb35/src/src/ec_su_axb35.c
• os/licenses/ryzenadj/src/main.c
• os/licenses/ryzenadj/src/lib/win32/OlsApi.h
• os/licenses/ryzenadj/src/argparse.c
• os/licenses/ryzenadj/src/lib/ryzenadj.h
• os/licenses/ec_su_axb35/src/hwmon/ec-su_axb35-hwmon.c
• os/licenses/ryzenadj/src/lib/linux/osdep_linux_smu_kernel_module.c
• os/licenses/ryzenadj/src/lib/cpuid.c
• os/licenses/ryzenadj/src/argparse.h
• os/licenses/ryzenadj/src/lib/linux/osdep_linux.c
• os/licenses/ryzenadj/src/lib/nb_smu_ops.h
• os/licenses/ryzenadj/src/lib/win32/OlsDef.h
• os/licenses/ryzenadj/src/lib/ryzenadj_priv.h
• os/licenses/ryzenadj/src/lib/linux/osdep_linux_mem.h
• os/licenses/ryzenadj/src/lib/linux/osdep_linux_smu_kernel_module.h
• os/licenses/ryzenadj/src/win32/readjustService.ps1
• os/licenses/ryzenadj/src/win32/installServiceTask.bat
• os/licenses/ryzenadj/src/win32/uninstallServiceTask.bat
• os/licenses/ec_su_axb35/src/scripts/su_axb35_monitor
• os/licenses/ec_su_axb35/src/LICENSE
• os/licenses/ec_su_axb35/LICENSE
• os/licenses/ryzenadj/src/README.md
• os/licenses/ryzenadj/src/LICENSE
• os/licenses/ryzenadj/LICENSE
• os/licenses/ryzenadj/src/.github/workflows/release.yml
• os/licenses/ec_su_axb35/src/.clang-format
• os/LICENSES.md
• os/licenses/ryzenadj/src/CMakeLists.txt
• os/licenses/ryzenadj/src/.github/workflows/build.yml
• os/mini/power.conf.example
• os/licenses/ec_su_axb35/src/README.md
• os/licenses/ryzenadj/src/win32/RyzenAdjServiceTask.xml.template
• os/licenses/ec_su_axb35/src/Makefile
• os/licenses/ryzenadj/SOURCE
• os/licenses/ec_su_axb35/SOURCE
• os/licenses/ec_su_axb35/src/contrib/netdata/README.md
• os/licenses/ec_su_axb35/src/hwmon/Makefile
• os/licenses/ec_su_axb35/src/hwmon/README.md
• os/licenses/ec_su_axb35/src/python-gui/ec-fan-control.desktop
• os/licenses/ec_su_axb35/src/Kbuild

[github-advanced-security]

CodeQL found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

[github-actions]

Persistent review updated to latest commit 96bb144

[github-actions]

Preparing review...

[github-actions]

Preparing review...

[github-actions]

Persistent review updated to latest commit f49e7fe

[bong-water-water-bong]

Review (orchestrator): VMTEST PASS (16 checks) for both the driver image and the EC_MODULE=skip image, fans row SKIP in the VM as designed, module signed with the MOK that carries the module-signing attribute, SMU tool pinned and built from vendored source with the EC balanced-mode fallback when lockdown or the tool is missing. Merging; the hardware boot + 1bit-release-check (fans row PASS with live rpm, [power] cap=90/90/105W in the banner) is the owner's step after the leaderboard.

[github-actions]

Persistent review updated to latest commit f49e7fe