Hardware enablement record · Aug–Nov 2026
Everything that was dead, degraded or unclaimed on this machine, what it took to fix, and the traps worth knowing if you are doing the same thing. Secure Boot is on, booting a signed unified kernel image through a systemd-boot menu; the camera runs entirely on the GPU; fingerprints now survive a reboot; a freeze reboots itself instead of hanging; updates, Flatpaks and the boot chain look after themselves. 133 of 135 automated checks passing, and the two failures are an empty fingerprint enrolment.
Status reflects verified behaviour, not just installed packages — every row below was tested on the running system.
| Subsystem | Before | Now | Status |
|---|---|---|---|
| Fingerprint | No stack installed; sensor unsupported by stock libfprint. Then: enrolments wiped on every reboot, and matches reported as failures | Enrolled, matching, and persistent across reboots. Unlocks login, lock screen, sudo, polkit |
Fixed |
| Webcam | Sensor failed to probe; later, apps saw only raw Bayer nodes; then 1080p at 6 fps with second-long freezes | 1920×1088 at a steady 30 fps, hardware denoise, no pixel work on the CPU | Fixed |
| Camera effects | Nothing — no Linux equivalent of Windows Studio Effects | Background blur on the NPU at 11.3 ms/frame, offered as a second camera | Added |
| Boot menu | No OS chooser; firmware timeout 0 and a UKI has no menu | systemd-boot, signed, with Manjaro preselected and a working Windows entry | Fixed |
| Speakers | Silent — 3 of 4 amps never enumerated | All four MAX98390 amps bound and audible | Fixed |
| Video acceleration | None — only the i965 driver, which has no Arrow Lake support | 42 VA-API profiles incl. AV1 decode and encode | Fixed |
| GPU / NPU compute | NPU firmware loaded but no userspace; no OpenCL | OpenCL 3.0 (128 EUs), Level Zero on both GPU and NPU | New |
| Thermals | Firmware-only control; thermald absent |
thermald active and managing Intel DPTF | Improved |
| Power | One fixed profile regardless of power source | Auto-switches: performance on AC, balanced on battery | Improved |
| Memory | Disk swap only; 6 GB dirty-page ceiling | 7.7 GB zstd zram at priority 100; bounded writeback | Improved |
| OLED panel | 8-bit colour, SDR, fixed 120 Hz, manual brightness | 10-bit via DSC, HDR + wide gamut, adaptive sync, light-sensor brightness | Improved |
| Secure Boot | Off; enabling it dropped the machine to grub rescue | On, with custom keys. Microsoft and Samsung certificates preserved | Done |
| Boot chain | GRUB refuses the kernel under Secure Boot, unfixably | Signed Unified Kernel Image booted straight from firmware; GRUB kept as a Secure-Boot-off rescue | Replaced |
| Sleep | Never measured | 91% of a suspend cycle in S0ix, reaching the deepest substate | Verified |
| Fingerprint tooling | Command line only, opaque result codes | A packaged desktop app for enrolling, testing and removing fingers | New |
| Peripherals | Never audited | Fn keys, touchscreen, mic beamforming, BT codecs all verified working | Verified |
| Touchpad | Gestures fine; tap-to-click off | 5-finger capable; 2/3-finger tap buttons enabled | Improved |
| Snapshots | Pre-upgrade only; every time-based schedule off, and no cron entry to fire them | Boot and daily snapshots armed, BTRFS mode pinned so the GUI cannot switch it to rsync | Improved |
| Wi-Fi | Regulatory domain unset — country 00, the world fallback | Set to the local domain. 6 GHz stays firmware-blocked; see below | Partly |
| Stale drivers | Four obsolete Xorg DDX drivers installed | Removed — nothing depended on them | Improved |
The reader is an Egis / LighTuning 1c7a:05a5 (ETU906Axx-E). Stock libfprint 1.94.100 does not carry it — the runtime device table stops at 05a1 and 05ae. It needs the SDCP-capable fork, which lists this exact ID with a 15-stage enrolment profile.
Two packaging problems had to be solved on the way: the AUR PKGBUILD declared provides=(libfprint-2.so) without the version, so pacman refused to satisfy fprintd's dependency; and the driver itself had a genuine crash bug.
Upstream bug fixed locally. egismoc_close() never cleared self->task_ssm. Any verify that timed out left the pointer dangling, and the next open hit g_assert(self->task_ssm == NULL) — so every fingerprint prompt you ignored killed fprintd with SIGABRT. The build now clears it on close and downgrades the assert to a warning. This patch lives in the PKGBUILD's prepare() and must survive any re-fetch from the AUR.
Each is wired as -auth sufficient pam_fprintd.so. The leading dash means a missing module is skipped rather than fatal, and sufficient means failure falls straight through to your password — verified by testing all four stacks with correct and incorrect passwords against a throwaway account.
su was deliberately left alone: it authenticates as the target account, and root has no enrolled print, so it would fall back to a password anyway.
The AUR package declares a placeholder pkgver=0, and pacman ranks a numeric segment above an alphabetic one — so 0-1 outranks the real r1831.<hash> and every update check reported an upgrade that would have silently dropped both local patches. That was first hidden behind IgnorePkg; it is now fixed properly with epoch=1, so 1:r1831… outranks the placeholder unconditionally. The pin was removed as redundant.
Every reboot used to log Deleted stored finger 7 … as it is unknown to device. This is a match-on-chip reader — templates live on the sensor, bound to a host keypair. libfprint’s fpi_sdcp_device_init_claim(), caches that keypair to disk but scopes the whole cache to the kernel boot ID, so the next boot deletes it and takes the generate_new: path, minting a fresh random EC keypair. The sensor then sees a stranger, stops reporting the enrolment IDs, and fprintd prunes its own records to match. The same happens after the 24-hour claim expiry without any reboot at all.
That file conflates two things with very different lifetimes. The session secrets really are stale after a reboot; the keypair is a long-lived identity and there is no reason to discard it. The patch keeps the keypair, drops the secrets, and leaves is_connected false so a full SDCP Connect runs under the same pk_h. Verified across a real reboot: the claim from the previous boot was reused and the device returned enrolment ID E7:01:63:2A:0D:7A…, matching the host’s stored print byte for byte.
The host private key was already on disk in the clear before this change — the patch adds nothing new to the filesystem, but it does extend how long those bytes stay useful to anyone who can read them. The directory is 0700 root. Deleting the cache file forces a clean re-pair, at the cost of the enrolments.
With that fixed, the sensor recognised the finger and the login still failed. The journal told the whole story in two lines: report_verify_status: result verify-match, then verify_cb: result verify-no-match.
The driver reports the result, then advances to a housekeeping state to issue one more command. But reporting a match makes libfprint cancel whatever is still in flight, so that command dies with G_IO_ERROR_CANCELLED, the state machine is marked failed, and egismoc_task_ssm_done() hands it to fpi_device_action_error() — which discards the result already reported and never reaches the state that would have called fpi_device_verify_complete(). The fix jumps straight from the check callback to IDENTIFY_COMPLETE, skipping the trailing sensor reset; that is safe because every identify already begins with one.
Two ways this hides from you. fprintd-list reads fprintd’s own record, not the sensor — it will happily list fingers the device has already forgotten, because the two are only reconciled when a match is attempted and fails. And a successful verify logs nothing at all, so an empty log is not evidence of failure. Judge from report_verify_status, or from what the client actually shows.
Enrolment is the part the command line handles badly: bare result codes, no sense of progress. The reader reports both a step count and a reason for every rejected scan, which is enough to build something followable. Fingerprint is now installed as a packaged application — a progress ring counting real steps against the reader's own total, rejection reasons in plain language, and a timeout that says No finger detected rather than verify-unknown-error.
Live operations drive the fprintd-* helpers and parse their per-stage output; listing and deleting use synchronous D-Bus. That split avoids running a GLib main loop beside Qt's just to receive two signals.
fprintd-enroll -f left-index-finger
fprintd-list $USER # see what's enrolled
fprintd-delete $USER # start over
Rest the fingertip still for about a second per stage. Rapid tapping produces partial reads and verify-no-match.
The OV02E10 sensor was failing at probe with external clock 26000000 is not supported. The in-tree driver only accepts 19.2 MHz; this board feeds it 26 MHz. That is a per-board property, not a platform one.
A DKMS module built from the matching v7.1.8 driver source carries three changes:
SRGGB10 instead of SGRBG10. The sensor shifts its Bayer phase on one axis only when flipped, so the textbook derivation (SGBRG10) is wrong — it renders red hair as green. The correct order was found by testing all four and matching channel means against a known-good capture.Through PipeWire, applications see two real sources: Built-in Front Camera (libcamera directly) and Integrated Camera (the relay loopback), plus StudioEffects when the NPU blur is running. The ~48 raw ipu6 entries in wpctl are internal device objects that expose no sources, so no picker shows them.
Fixing the driver was only half of it. The IPU6 publishes about 48 raw Bayer capture nodes, and every one of them advertises itself as a video capture device. An application scanning /dev/video* picks one, starts the sensor — the privacy light comes on — and then renders nothing, because only libcamera can debayer that data. It looks exactly like a broken camera.
There are now three working routes, all verified:
v4l2-relayd, for apps that open a device path directly — Zoom, OBS, Discord. Find it by that card name: its /dev/videoN number changes between boots.The raw nodes are now hidden from application enumeration, so only the working device is offered. libcamera is unaffected: it walks the media graph rather than reading that metadata.
Two traps in the bridge, both silent. The shipped relay unit sets a DeviceAllow whitelist — and once that is set at all, anything absent is denied. /dev/dma_heap/system was absent, so libcamera could not get a dma-buf provider, quietly disabled software debayering, and emitted a flat grey frame. Separately, libcamera outputs 1928×1088 while the relay asked for 1280×720, and the stock pipeline has videoconvert but no scaler, so negotiation failed without an error. Both are fixed; both would have read as “the camera is broken”.
One app remains broken: pantheon-camera hangs. It enumerates through GStreamer's device monitor, which offers only the libcamera device, and it cannot drive it. Not worth chasing — use Snapshot.
Accepting the faster clock without re-timing the PLL meant the whole clock tree ran 26 ÷ 19.2 = 1.354× fast: 40.62 fps measured, from a mode named 30fps. Exposure is set in lines, so a shorter frame is a shorter maximum exposure — 24.6 ms instead of 33.3 ms. Indoors the auto-exposure ran out of exposure and bought the rest with gain. The proof was sitting in the controls:
exposure max=2242 value=2242 # pinned at maximum
analogue_gain max=248 value=193 # 12.1x, register is 4.4 fixed point
Rather than re-derive PLL dividers that are not public, add blanking: vts_def 2244 → 3039, which is 2244 × 48.75/36. vts_min stays at 2244 so the faster rate is still reachable.
| Before | After | |
|---|---|---|
| Frame rate | 40.62 fps | 30.01 fps |
| Max exposure | 2242 lines | 3037 lines |
| Analogue gain, same room | 193 (12.1×) | 144 (9.0×) |
| High-frequency energy | 6.14 | 4.70 |
Same light, 25% less gain — and the advertised 30 fps is now true. Note that vblank cannot be used as a runtime knob for this: the driver reapplies vts_def on every set_fmt, and libcamera caches the exposure range at configure time, so a mid-stream change is accepted, ignored, then reset.
A denoise filter added against the remaining noise collapsed the bridge to 6.1 fps with inter-frame gaps of 466 ms at p90 and 1041 ms at worst. That number is not arbitrary. GstBaseSink carries an emergency escape hatch:
/* !!emergency!!, if we did not receive anything valid for more than a
second, render it anyway so the user sees something */
Every frame was late, so every frame was dropped — except one per second, which that clause renders anyway. One second rounded up to the next 33.3 ms frame is 1033 ms.
The frames were late because v4l2-relayd runs two separate pipelines and copies buffers between them carrying their original capture timestamps, while the output pipeline’s only source is a live appsrc with min-latency 0. The output side has no idea the input side exists, so every millisecond of capture is undeclared lateness against v4l2sink’s budget of roughly 38 ms. And relayd’s bus watch handles only ERROR and STATE_CHANGED, so the warning GStreamer posts about it is discarded: the whole thing fails silently.
Adding a queue makes it worse, not better. A queue adds delay, so every frame takes the emergency path — 1.4 fps — near enough one frame per emergency render. The instinct to buffer your way out of this is exactly wrong.
The iGPU’s VEBOX fixed-function unit does colour conversion, scaling and noise reduction at effectively no CPU cost. It needs the gst-plugin-va package, which was not installed.
VIDEOSRC="libcamerasrc ! vapostproc ! video/x-raw(memory:VAMemory),format=NV12 \
! vapostproc denoise=64 sharpen=0 disable-passthrough=true"
Two vapostproc stages, deliberately: VEBOX denoise is silently a no-op when the input surface is RGB, and libcamera’s software ISP hands out BGRx. Measured across denoise=0/16/32/48/64 on RGB input, spatial noise stayed flat at 5.33–5.35 — the filter was running and doing nothing. The first stage converts to NV12 on the GPU; only then does denoise engage.
Result over 300 consecutive frames: 30.00 fps, gap median 33.3 ms, p99 34.9 ms, worst 39.1 ms, zero gaps over 100 ms. Spatial noise about 10% below the CPU chain, at 20% of a core instead of 24%. Better picture, fewer stalls, less power.
The plain CPU chain was never safe either. It sat 0–2 ms inside that 38 ms budget, and a paired A/B caught it falling off — 11.85 fps, p99 1033.5 ms — in one round of three, with no filter present at all. A drop-in adding sync=false to v4l2sink removes the deadline rather than hoping to stay under it. Nothing paces the relay afterwards except the live 30 fps source, which is the correct pacing: this is a relay, not a player.
gst_parse_launch rejects a description ending in caps. relayd then calls gst_element_set_state() on NULL and only the splash frame ever reaches the loopback — a silent failure that looks like a dead camera.
Every one of the IPU6’s ~48 raw nodes reports V4L2_CAP_VIDEO_CAPTURE. The udev rule hides them from anything that consults udev, but browsers open device nodes directly and ignore it, so Firefox and Chromium both list 49 “cameras” of which 48 produce nothing. Routing them through PipeWire cuts the list to the two real sources: media.webrtc.camera.allow-pipewire for Firefox, --enable-features=WebRtcPipeWireCamera for Chromium and Brave.
Windows gets background blur from Samsung’s Studio Effects, running on the NPU. The Linux equivalent is linux-studio-effects: MODNet segmentation through OpenVINO on the NPU, composited and published as a second virtual camera.
MODNet FP16 at 256×256 runs 11.3 ms/frame on the NPU — half what the project’s README quotes. Compile takes 0.6 s.
30 fps steady, 31 ms of work per frame. Appears as StudioEffects alongside the unprocessed camera.
It did not work out of the box. Four defects, all fixed locally and worth sending upstream:
cap.read() taking 976 ms while the real work took 31 ms. A V4L2 device hands out a small buffer ring; read it in a tight loop and each read costs the true frame interval, but do 30 ms of work between reads and the reads themselves begin to block — 274 ms, then 977 ms, degrading as the consumer falls behind. Draining the camera in its own thread and taking the newest frame fixes it: 1.0 → 30 fps.cv2.VideoCapture(index) without a backend lets OpenCV 5 try GStreamer first, which fails on a loopback that already has a producer — and reports it as “camera busy/unavailable”, sending you after a conflict that does not exist.Two setup traps specific to this laptop. The project pins its virtual camera to /dev/video10, which here is Intel IPU6 ISYS Capture 9. And v4l2loopback was already loaded with one device for the camera bridge, so the setup script saw the module present, returned success, and created nothing. Load it with two devices and pin the effects one well clear of the IPU6’s range.
It is deliberately not a background service. Blur only matters during a call, and leaving it running keeps the camera powered and the privacy light on.
Four MAX98390 amplifiers sit on I²C bus 2 at 0x38, 0x39, 0x3c, 0x3d — two woofers, two tweeters. ACPI only enumerates the first, and no kernel quirk exists for this subsystem ID (0x144dc904), so nothing bound and the kernel reported speaker_outs=0. Audio routed to the ALC298 pin and died there.
A DKMS side-codec driver plus a boot service that creates the three missing I²C devices brings all four online.
The amps stay powered continuously. The HDA component-master quirk that would idle them does not exist upstream yet, so expect a small constant draw.
Thinner than Windows, which applies vendor DSP and speaker-protection tuning Linux has no equivalent for. EasyEffects closes much of the gap.
The only VA-API driver present was libva-intel-driver — the legacy i965 backend, which has no Arrow Lake support at all. Every video decode was running on the CPU. Replacing it with intel-media-driver yields 42 hardware profiles, including AV1 decode and encode. This is the single biggest change for battery life and fan noise during video playback.
Also removed: xf86-video-intel, -amdgpu, -ati and -nouveau. All deprecated, all useless under Wayland, and xf86-video-intel is actively harmful on modern Intel. Nothing depended on them.
Power management is deliberately split between two owners:
Do not install TLP, tuned or auto-cpufreq. They take over the same knobs power-profiles-daemon owns, and PPD is what KDE's battery applet drives. Adding one will produce contradictory, hard-to-debug behaviour.
The topology is 6 P-cores at 5.1 GHz (cpu0–5), 8 E-cores at 4.4 GHz (cpu6–13) and 2 low-power E-cores at 2.5 GHz with no L3 (cpu14–15). The kernel uses hybrid capacity scaling rather than ITMT, so /proc/sys/kernel/sched_itmt_enabled legitimately does not exist here. That is correct behaviour, not a missing setting — worth knowing before someone tries to “fix” it.
Measured on a real suspend cycle rather than inferred from settings — the residency counters only advance while actually suspended, so nothing here can be checked from a running desktop.
The only sleep state is s2idle. That is correct and not a limitation: Arrow Lake removed S3 entirely, so modern standby is the sole path, and its quality depends on whether the SoC actually reaches its low-power substates rather than merely idling.
| Measure | Result | Reading |
|---|---|---|
| Time suspended | 31.1 s | RTC-triggered wake, undisturbed |
| S0ix residency | 28.3 s — 91% | The remaining ~2.7 s is entry and exit overhead |
| S0i2.2 (deepest) | 27.3 s — 88% | Reaches the deepest substate, not just the shallow one |
| S0i2.1 | 1.0 s | Brief transit on the way down |
| Both PMC dies | agree | pmc0 and pmc1 report matching figures |
Everything came back clean: Wi-Fi reconnected and reachable, all four speaker amps still bound, camera capturing, PSR2 still enabled. The single dmesg warning on resume — intel_ipu6.psys: Failed to get runtime PM — concerns the IPU's processing-system block, which this camera pipeline does not use.
auto, which is what allows deep package idle at all — 97.5% while measuring, on an otherwise idle machine. That last condition matters: see the C-state note below.gb5-power-profile.sh re-asserts it each run, because a reassociation resets the driver setting.GB5_KEEP_120HZ=1 in /etc/default/gb5-power to keep 120 Hz everywhere.One device is pinned awake on purpose. The fingerprint reader is held at power/control = on. Autosuspend was tried and backed out: libfprint's own autosuspend database lists this ID only in its known-but-unsupported block, meaning upstream never validated it here. The saving would be a few milliwatts against a device you authenticate with all day.
Only three idle states are exposed, named C1_ACPI, C2_ACPI, C3_ACPI. Read literally, that says the machine is capped at C3. It is not. intel_idle has no native table for Arrow Lake — mainline covers Meteor Lake-L and even the newer Panther Lake-L, but not this — so it uses the firmware’s ACPI _CST, which names states by ACPI’s generic numbering rather than the hardware’s. The MWAIT hint gives the truth:
| Shown as | MWAIT hint | Actually is | Entered |
|---|---|---|---|
| C1_ACPI | 0x00 | C1 | 522,210× |
| C2_ACPI | 0x21 | C6 | 893,084× |
| C3_ACPI | 0x60 | C10 | 861,820× |
The cores reach C10 constantly — 40 minutes of a 55-minute uptime. Package C10 is a different thing: it needs every core in C10 and the uncore and GPU idle at once, so anything running at all keeps the package awake. Measured with a Node process, browsers and KDE polling active, the package spent 0.03 s of 10 s in any C-state.
A hardware-specific kernel build would gain nothing here. The only thing a native intel_idle table adds over this ACPI path is a nominal C1E state — and C1E auto-promotion is already enabled in hardware, with MSR_IA32_POWER_CTL bit 1 set on all 16 logical CPUs, so a C1 request is promoted to C1E by the CPU whatever the OS asks for. The idle path is not where this machine loses power.
Arrow Lake-H dropped hyper-threading. Sixteen logical CPUs are sixteen physical cores, every thread_siblings_list a single entry. It is very easy to assume six P-cores means CPUs 0–11 and get every per-core-type measurement wrong — that mistake was made here before the topology was checked.
| CPUs | Type | Max clock | cpu_capacity | Busy at idle |
|---|---|---|---|---|
| 0–5 | P-core, Lion Cove | 5100 MHz | 1024 | 14.5% |
| 6–13 | E-core, Skymont | 4400 MHz | 774 | 3.8% |
| 14–15 | LP-E, on the SoC tile | 2500 MHz | 312 | 1.1% |
Light background work lands mostly on the fast cores. That looks like a scheduling bug and mostly is not: Thread Director sends short latency-sensitive wakeups to a P-core at low clock, which finishes sooner and drops back to C10 faster than an LP-E core grinding through the same work would. Race-to-idle beats slow-and-steady at millisecond scale. It is only worth fighting for sustained low-priority load, and there the lever is cgroup placement for that one workload — not a daemon that repartitions the whole machine.
Intel’s low-power-mode daemon is meant to herd idle work onto the LP-E cores. It refuses to start here — Unsupported CPU model — because its table knows Arrow Lake as 0xb5 and this chip is family 6, model 0xc5, Arrow Lake-H: a different model number for the same architecture. One line makes it run, and it then reads the topology correctly (6 Pcores, 8 Ecores, 2 Lcores, TDP 28W). The patch is in the repo. The daemon is not installed, for two reasons.
It saves nothing. Back-to-back in the same minute, package RAPL: 7.94 W with it running, 7.87 W stopped, ~8.0 W after a full revert. One number, three times.
The near-miss worth repeating. An earlier comparison seemed to show 3.73 W baseline against 7.17 W with lpmd — the daemon doubling power consumption. That was not a result. The baseline had been captured half an hour earlier on a quieter machine. Two RAPL readings taken far apart on a desktop with browsers and a dev server running measure the browsers, not the change. Both arms have to run in the same minute or the number is fiction — and the fiction is convincing, because it points somewhere plausible.
And it leaves the machine restricted after it exits. lpmd entered DEFAULT_HFI at startup and pinned everything to CPUs 0–13 — excluding 14 and 15, the two LP-E cores it had just identified as the low-power ones. Killing the daemon does not undo it.
The obvious fix does not stick, and that is worth knowing. Writing the cgroup file directly appears to work and then silently reverts — here, about an hour later. lpmd does not write cpuset.cpus; it sets a systemd property over D-Bus, leaving a runtime drop-in at /run/systemd/system.control/system.slice.d/50-AllowedCPUs.conf. systemd re-applies that on the next daemon-reload, which can be triggered by installing any unrelated service.
Clear the property, not the file:
for c in system.slice user.slice machine.slice; do
systemctl set-property --runtime "$c" AllowedCPUs=
done
systemctl daemon-reload
bash -c 'taskset -pc $$' # expect 0-15
When it is genuinely unset, systemd drops the cpuset controller from those slices, so cpuset.cpus.effective stops existing. That absence is the success condition.
The service is masked rather than disabled, so nothing can pull it in as a dependency.
Every short power measurement in this project failed the same way: on a live desktop the noise floor is several watts, so a twenty-second sample says more about what the desktop happened to be doing than about the change under test. The fix was to log continuously and select quiet windows afterwards. 139 undisturbed thirty-second samples on battery, filtered to the 113 where both P-cores and E-cores were under 2% busy:
| Measure | Idle on battery |
|---|---|
| Whole system, quiet moments | 4.86 W — a genuine idle floor, but not a runtime |
| Whole system, actual sustained drain | 6.19 W — 11.9 h ± 0.4, by coulomb count over 33 gauge steps |
| SoC package | 3.33 W |
| CPU cores | 0.23 W |
| GPU RC6 | 95.2% |
| Package C-states | C6 23.5% · C8 0.54% · C10 1.06% |
Spread: min 3.86, p25 4.54, median 4.86, p75 5.17, max 6.44 W, holding steady as the pack drains.
That idle figure is not a runtime, and treating it as one was a mistake worth describing. The battery’s coulomb counter is ground truth — capacity here is exactly charge_now / charge_full, so a percentage drop converts to energy linearly. Over 1.69 h of undisturbed operation it recorded 6.30 W, against 5.02 W for the median of instantaneous readings and 5.23 W for their mean. The snapshots under-report by 17–20%.
The reason is sampling. current_now × voltage_now is an instantaneous value, and at a 30 s interval it misses short bursts completely — the ten largest samples in that window ran up to 21.2 W. Averaging snapshots of a bursty load systematically under-counts energy. If you want a runtime, integrate the charge counter; if you want an idle floor, the snapshots are fine.
RAPL is not the system. intel-rapl measures the SoC package and nothing else — not the panel, not Wi-Fi, not the SSD, not the board. At idle here the package is 3.34 W of a 4.82 W machine, so tuning by RAPL alone optimises two thirds of the problem and misses the part that actually varies. Whole-machine draw comes from the battery, which means the numbers that decide battery life are only available while discharging.
Full-screen black against full-screen white, 45 s samples, order alternated, taking the quietest quartile of each:
| Screen | Whole-system draw |
|---|---|
| Black | ~5 W |
| White | ~15 W |
Content costs about 10 W — three times the draw, and more than the entire SoC package. The gap between a dark screen and a bright one is larger than the gap between idling and compiling. On an emissive panel this is expected: power follows light actually emitted, which is why a dark colour scheme is a real saving here and pure folklore on the LCD in a MacBook Air.
The brightness numbers from this machine are all void, and the reason is the interesting part. A run gave white at 100% costing 13.6 W, at 30% costing 7.0 W, and at 10% costing 13.0 W. Ten percent cannot cost more than thirty, so the method was wrong. The cause: this laptop has an ambient light sensor and KDE auto-brightness is on (automaticBrightness: true, PowerDevil driving the panel through iio-sensor-proxy). Every intel_backlight write was overridden within seconds — the tell was the panel finishing on 1126, a value nothing in the test had written. It is also a feedback loop: a white test screen lights the room, the sensor reads brighter, auto-brightness raises the panel.
/sys/kernel/debug/pmc_core/ltr_show reports the latency tolerance each IP block advertises, and the tightest one sets the floor for how deeply the package may sleep. Every block read zero except the USB controller, which was demanding 128 µs — and package C8 and C10 read exactly zero.
The cause was a TP-Link UE300 gigabit Ethernet adapter sitting at power/control = on. Unplugged, the USB controller’s tolerance relaxed and C8/C10 began accumulating for the first time. The tightest constraint is now ISH at 10.5 ms — three orders of magnitude looser.
Worth knowing generally: one stubborn device can forfeit the platform’s deep idle states, and nothing in the usual tooling points at it. ltr_show names the culprit directly.
Half of an earlier alarm here was architectural. Pkg%pc7 and Pkg%pc9 cannot be reported on this family — turbostat maps Arrow Lake to adl_features, whose supported set omits them, with an upstream comment saying so. Only pc2, pc3, pc6, pc8 and pc10 exist. So “C7 through C10 read zero” was two real zeros and two columns that were never going to print.
charge_now does not update continuously. It holds a value and jumps, in steps of exactly 49 mAh, roughly every 445 seconds. Sampled every 5 s it reads perfectly flat for a minute while about 6 mAh is actually drawn.
A quiet-versus-balanced comparison was run here as four alternating 12-minute arms and produced 98, 98, 98 and 49 mAh. Read as step counts those are 2, 2, 2, 1 — there is no sub-step information in them at all. The apparent halving in the last arm is one step that had not landed yet.
So the result is “not measurable”, which is not the same claim as “no difference”. The experiment could not have detected a difference had one existed. Count the gauge’s steps before believing any battery A/B: under about eight per arm you are measuring the gauge, not the machine.
The runtime figure is unaffected, because it spans far more steps. Extended to 3.93 h and 33 steps it converges to 6.19 W, 11.9 h ± 0.4 — quantisation down to ±3.0%. It settled as data accumulated, with overlapping error bars throughout: 12.1 ± 0.9, 12.1 ± 0.6, 11.7 ± 0.4, 11.9 ± 0.4.
Idle depth is fine: C10 residency 91.6%, package 3.48 W. The wakeup rate is not — 2208 interrupts per second at 3% busy, 86% of them local timer interrupts, and over half of all C10 entries waking before their residency target is repaid.
powertop’s software section came back entirely empty, so attribution came from per-process voluntary context switches instead:
| wakeups/s | process |
|---|---|
| 96.8 | the agent taking the measurement |
| 75.0 | rcu_preempt |
| 63.3 | kwin_wayland |
| 50.6 | irq/193-iwlwifi |
| 26.8 | plasma-systemmonitor |
| 14.0 | ksystemstats |
The largest single wakeup source during the idle measurements was the process doing the measuring. That does not invalidate the figures, it locates them: they describe a machine with something light running on it, which is a fair model of real use. But the true floor, with a logged-in desktop and genuinely nothing happening, is lower than anything recorded here.
The one actionable item: plasma-systemmonitor and ksystemstats spend about 41 wakeups per second keeping monitor widgets updated. Pure overhead when nobody is looking at them.
Worth ruling out explicitly, because 1908 timer interrupts per second invites the conclusion that dynticks is broken: it is not. The kernel has CONFIG_NO_HZ_FULL=y with no nohz_full= on the command line, which behaves as dynticks-idle, and CONFIG_HZ is 1000. A genuinely periodic tick across 16 CPUs would be 16000/s.
When the battery runs out this machine hibernates cleanly rather than losing power mid-write — CanHibernate yes, resume= set, and UPower’s CriticalPowerAction=Auto resolving to HybridSleep at 2%.
But hibernation must fit used RAM into disk swap, and zram does not count. There is 8.80 GiB of disk swap against 30 GiB of RAM. At a light desktop 7.44 GiB was in use, leaving 1.36 GiB. A large build, a VM or a heavy browser session crosses that, hibernation fails, and AllowRiskyCriticalPowerAction=false means UPower declines to try anything else — so it becomes a hard power-off with whatever was unsaved.
The tuning script had been switching the panel to 2880x1800@60 on battery since it was written, and logging success every time. This panel has no 60 Hz mode at native resolution. kscreen-doctor lists exactly one 2880×1800 mode and it is 120 Hz; every 60 Hz mode it offers is at a lower resolution.
The reason it looked fine: kscreen-doctor exits 0 even when it prints “Output mode not found”, so an && guard logged a success that never occurred. Verify what applied by reading the state back, never by checking the exit code.
Out of the box the kernel sizes the hibernation image up to image_size (about 2/5 of RAM) and the stock swap partition is 8.8 GiB. On a 31 GiB machine that is never enough under load — with an Android build running, RAM in use was 17.7 GiB.
An aborted hibernate is not harmless. One on 2026-08-22 reset the USB bus and wedged the Samsung SCAI/EC interface, silently killing fingerprint authentication, the keyboard backlight and power_on_lid_open for nine hours before anyone noticed. Each failed in a way that looks like something else: the fingerprint just asks for a password instead, and the backlight’s sysfs node accepts writes it cannot honour.
Capping image_size is not a fix. It was tried first, at 7 GiB. With 17.7 GiB in use the machine was still short by nearly nine gigabytes; the cap only meant the kernel gave up earlier. image_size is a best-effort upper limit — the kernel frees memory to shrink the image toward it — so it helps only when the shortfall is small.
The fix is a 20 GiB swapfile on its own btrfs subvolume:
mount -o subvolid=5 /dev/nvme0n1p2 /mnt/btrfs-top
btrfs subvolume create /mnt/btrfs-top/@swap # sibling of @, so Timeshift never snapshots it
mount -o subvol=@swap,noatime /dev/nvme0n1p2 /swap
btrfs filesystem mkswapfile --size 20g /swap/swapfile
swapon /swap/swapfile
btrfs inspect-internal map-swapfile -r /swap/swapfile # -> resume_offset
A successful swapon is itself the proof it is right: btrfs refuses to swap on anything compressed, copy-on-write, or spanning multiple devices, and the root here is mounted compress=zstd:1.
/sys/power/resume and resume_offset are writable at runtime. The whole fix can be applied and tested with zero boot risk — set the offset first, then the device. Only making it survive a reboot needs the kernel command line, and therefore a UKI regeneration under Secure Boot. That distinction is worth knowing before talking yourself out of the fix, as happened here for most of a day.
A UKI’s .cmdline section is UTF-16. Grepping the raw bytes reports every key as missing even when the file is perfectly correct — which nearly caused a good change to be reverted here. Strip the nulls first:
objcopy -O binary --only-section=.cmdline \
/boot/efi/EFI/Linux/manjaro-7.1-x86_64.efi /tmp/c.txt
tr -d '\0' < /tmp/c.txt
Keep the previous UKI beside the new one, signed, as a fallback: it appears as its own systemd-boot entry and still carries the old command line. On this machine the reboot also swapped the NVMe enumeration order — the Linux disk became nvme1n1 — and nothing broke only because every reference is by UUID. A /dev/-based resume= would not have booted.
Verified with a real S4 cycle: same boot_id, uptime continued 527 s to 781 s. Two things still need repairing on resume, and they are handled by a systemd-sleep hook rather than left to be rediscovered.
| After resume | State |
|---|---|
| kbd_backlight | broken — every CSFI call answers 0xff |
| power_on_lid_open | broken — reads empty, and each read costs a CSFI failure |
| usb_charging | works |
| platform_profile | works |
Only partly wedged, which is what makes it easy to miss — the profile still switches, so power management looks healthy while the keyboard backlight is dead. Reloading samsung_galaxybook re-runs its ACPI init and clears it.
S4 also resets the fingerprint reader six times, destroying its SDCP session while the kernel boot_id is unchanged and the cached claim still looks current. The driver is patched to treat the resulting AuthorizedIdentity MAC failure as a reconnect trigger, keeping the host keypair so the enrolments stay valid:
touch #1 Enrollment ID identified by the device
SDCP AuthorizedIdentity verification failed
"Invalidated the cached SDCP session..." <- the patch
next claim SDCP ConnectResponse verified successfully
touch #2 verify-match
The first touch after every hibernate still fails. The driver only learns the session is stale by failing, so it cannot repair mid-operation. Expect one rejected touch after a resume, then normal behaviour.
One more thing S4 exposed: the machine resumed straight into an unlocked desktop. Locking on the way down beats locking on the way up, because a pre-sleep lock cannot lose the race.
powertop run on battery for a few minutes.A second pass over everything the enablement work never touched. Each line below was checked against live output, not assumed — and two of the answers are “this hardware does not exist”, which is worth knowing before someone spends an evening trying to configure it.
bluez-utils was missing and is now installed.The pad handles five simultaneous contacts, so the hardware is not the limit. Plasma is. Probing KWin's scripting API directly rather than guessing:
registerTouchpadSwipeShortcut = MISSING
registerTouchpadPinchShortcut = MISSING
registerScreenEdge = global
registerShortcut = global
The C++ API exists in KWin but is not exposed to scripts, so on Plasma 6.7 there is no supported way to bind your own touchpad gestures. What you get is the built-in set — three and four finger swipes for Overview, the grid and desktop switching — plus two and three finger tap buttons, which are enabled.
Going further means an external daemon reading libinput events and injecting keystrokes (fusuma plus ydotool). That is a background process, membership of the input group and a uinput device: real complexity and a privilege change in exchange for convenience. Deliberately not done.
Clamshell chassis, so there is nothing to rotate against. KWin reports auto-rotate as incapable. Nothing to enable.
The digitizer's HID report descriptor carries only touch and vendor usages, with no Pen usage anywhere. A pen will not work on this panel.
xe driver loads but binds nothing; i915 owns the GPU. Harmless.The card is an Intel Wi-Fi 7 BE201 with 59 channels in the 6 GHz band. When this was first investigated all 59 reported (disabled); 35 still do, and the remaining 24 are (no IR) — present but passive-only, which is the normal regime for a 6 GHz client rather than a fault. Either way the obvious fix is not the fix.
The regulatory domain was unset, so the kernel used country 00 — the world fallback, which has no 6 GHz allocation at all. Setting it to the local domain is correct and worth doing: the global table gains 5925–6425 MHz @ 320 MHz. It also does not enable 6 GHz here, because this PHY is self-managed: the firmware owns its regulatory domain and iw reg set cannot override it.
What actually decides is an ACPI method. iwlwifi queries _DSM UUID f21202bf-8f78-4dc6-a5b3-1f738e285ade, function 3 — enable 6E. Decompiling this machine's DSDT shows where that lands:
If ((Arg2 == 0x03))
{
Local0 = UHBS /* \UHBS */
Return (Local0)
}
UHBS is a 32-bit field in the ACPI NVS region, written by the BIOS at boot. There is no Linux-side override: lar_disable was removed in kernel 5.4, and neither iwlwifi nor iwlmld exposes a parameter for it. The only route from Linux is a DSDT override loaded through the initramfs, which on this machine also means rebuilding and re-signing the UKI.
Check the access point before blaming the laptop. When this was investigated, the local AP was visible only on 2.4 GHz at −34 dBm — no 5 GHz sibling, no 6 GHz, and no Reduced Neighbor Report in its beacon. RNR is how a 6 GHz BSS is discovered, because a client whose 6 GHz channels are disabled cannot scan for them directly. A working card would have had nothing to join.
6 GHz channels start at 5955 MHz, not 6000. A pattern matching 6[0-9][0-9][0-9] MHz silently skips the first several and reports zero channels where 59 exist. That mistake was made here and produced a wrong conclusion before the count was redone properly.
A kernel update that outruns DKMS takes out the camera, the speakers and the fingerprint reader at once. This is the part that makes the rest recoverable.
Pre-upgrade snapshots were already working — a pacman PreTransaction hook has been creating them all along. Three things were missing or misleading:
timeshift --check runs with schedules on — editing the config alone does nothing. Running it created both entries.The config file is not authoritative — timeshift --list is. A default /etc/timeshift/timeshift.json reading "btrfs_mode": "false" looks like proof that snapshots are doing rsync backups. It is not; that file had never been written to. Reading it and concluding the worst wasted real time here.
grub-btrfs was removed along with os-prober: it generated snapshot entries for a GRUB menu this machine no longer boots. The consequence is worth stating — there is no boot-menu rollback any more. Restoring happens from Timeshift itself, which works fine, but recovering a system that will not boot at all now needs live media rather than a menu entry.
All three cost hours because they present as something else entirely.
nvme0n1 was the Windows disk on one boot and the Linux disk on the next; the Linux ESP moved from nvme1n1p1 to nvme0n1p1 across a single reboot. Anything that hardcodes /dev/nvmeXn1pY is a coin flip — two scripts here had exactly that bug. Use PARTUUID, or look the device up at runtime with findmnt.
Windows fast startup is a partial hibernate, not a shutdown. It leaves a hibernation image on the NTFS volume, and a filesystem in that state cannot be safely written from Linux — mount it read-write anyway and you corrupt it. What you see instead is a volume that mounts read-only, or a copy that silently fails, which reads as a permissions problem rather than “Windows never actually shut down”.
Ask before writing, rather than finding out afterwards:
ntfs-3g.probe --readwrite /dev/<the windows partition>
0 safe
14 hibernated — fast startup, or a real hibernate
15 dirty — needs chkdsk from Windows
Only rc 0 is safe to write to. The durable fix is to turn fast startup off in Windows; until then, treat the volume as read-only from Linux. Note this machine has no ntfs3 in the kernel — writing goes through ntfs-3g over FUSE, so a mount -t ntfs3 fails with a misleading Can’t open blockdev.
Linux reads that clock as UTC, so every Windows → Linux boot starts wrong by the UTC offset until NTP corrects it. That is not cosmetic: the first seconds of the journal are stamped in the future, and journalctl --since keeps matching them forever afterwards — which silently corrupts any log window scoped by time, and produced a false test result here. Use --after-cursor when it matters, and set RealTimeIsUniversal=1 on the Windows side — config/windows/windows-use-utc-clock.reg in this record does it, so the fix ships as a part rather than a description.
Restarting the camera relay drops its node from PipeWire, so applications lose the bridge device until it is re-probed. Restart wireplumber alone to do that — restarting pipewire drops every client’s audio connection and will cost a running browser its microphone, which reads as “the mic broke” rather than “something restarted”.
These parts of the system are no longer stock. This is the part worth re-reading after a kernel bump — though since November the weekly updater and the boot guard do the re-checking themselves.
ov02e10-26mhz, max98390-hda and v4l2loopback all rebuild automatically. If a kernel update outpaces them, the camera or speakers go quiet.
libfprint-egismoc-sdcp-git conflicts with and provides libfprint, so pacman -Syu leaves it alone. Do not let anything pull stock libfprint back.
gb5-boot-guard.sh runs after every update and keeps the last-resort loader, the fallback image and the signing database right. Manjaro’s GRUB hook is disabled on purpose. (This card used to warn that hibernation was undersized; that was fixed with a 20 GiB swapfile on 23 August.)
/usr/local/sbin/gb5-verify.sh runs 135 checks across every area on this page and prints PASS / FAIL / NOTE with the evidence inline. A companion service runs it once, 45 seconds into the next graphical session, writes /var/log/gb5-verify.log, then disables itself — so there is a cold-boot record without it becoming a permanent background job.
sudo gb5-verify.sh # to stdout
sudo gb5-verify.sh /tmp/r.log # and to a file
# after every update, before rebooting — the signing hook can fail silently
sbctl verify
sudo gb5-boot-guard.sh --check # silent means the boot chain is right
Two of the checks depend on the hardware being free: if a camera application is open, the camera and microphone probes report contention rather than a fault. Several encode a trap that cost real time to find — that the relay pipeline carries no CPU filter, that GPU denoise is actually engaged rather than silently idling on an RGB surface, that no emergency frame drops appear in the journal, and that the sensor is still retimed for the 26 MHz clock.
Do not hardcode /dev/videoN anywhere. The v4l2loopback bridge takes whatever number is free at probe time — it was video48 one boot and video0 the next. Look it up by card name instead.
dkms status # all three built for the new kernel?
dmesg | grep -i ov02e10 # no "external clock" error
cam -l # camera enumerates
lsmod | grep max98390 # amps loaded
fprintd-verify -f right-index-finger # NOT bare: it tests only the FIRST finger
vainfo | grep -c VAProfile # expect 42
# sleep quality — only meaningful after an actual suspend
cat /sys/kernel/debug/pmc_core/slp_s0_residency_usec
cat /sys/kernel/debug/pmc_core/substate_residencies
dkms remove -m ov02e10-26mhz -v 1.0 --all restores the in-tree driver (and the dead camera).dkms remove -m max98390-hda -v 1.0 --all, then systemctl disable --now max98390-hda-i2c-setup. The speakers go silent again.pacman -S libfprint reverts to stock; the reader stops working. PAM stays safe either way — the - prefix skips a missing module./etc/udev/rules.d/99-gb5-power.rules and disable gb5-power-tune.service./etc/sysctl.d/99-gb5-tuning.conf and /etc/systemd/zram-generator.conf..bak-2026-08-19 copies sit beside every edited file in /etc/pam.d/.The working build tree is kept at ~/.cache/gb5-build (47 MB) so the fingerprint package can be rebuilt without re-deriving the two PKGBUILD patches.
Everything above is a description. A description is not a way to do it again — so the record now carries the parts, not just the account of them.
sudo ./setup.sh # guided, asks before each step
sudo ./setup.sh --list # what the steps are
Five steps — camera, speakers, fingerprint, system tuning, Secure Boot — each idempotent, so re-running one is the normal way to repair something after a kernel update rather than an exceptional act. A precheck runs first and refuses to continue on a hard mismatch.
None of it has been run on a clean install. It is derived from a machine where all of it already works, and each step detects what is present and skips it — which makes re-running safe, but means the from-scratch path has been reasoned about rather than executed end to end.
The two patched kernel module sources are bundled rather than re-patched. The camera's Bayer order was found empirically, by testing all four orders and matching channel means against a known-good capture; its frame timing is derived from this board's clock. Asking someone to re-derive either from prose would be far worse than shipping the file.
These fixes are per-board, not per-platform. The clock, the Bayer order and the amplifier addresses were all determined on a 965XHD. On a different Galaxy Book they may differ, and applying them blindly gives you a camera with wrong colours rather than one that works. The precheck warns rather than blocks in that case — the warning is the point.
Secure Boot is deliberately excluded from a full run. It needs BIOS steps that cannot be performed from Linux, and it is the only part that can leave a machine unbootable, so it is a guided checklist instead of an automated one.
Everything in it except the notes is a copy of something live, and copies drift — a kernel update rebuilds the DKMS source, a package update rewrites a hook. A record that quietly disagrees with the machine is worse than no record, because it gets trusted.
sudo ./sync.sh # report drift, change nothing
sudo ./sync.sh --pull # refresh the copies, then review with git diff
35 files compared against their live originals, non-zero exit on drift. It has a real limit worth knowing: it tells you when a tracked file has drifted, but it cannot tell you that something new was never tracked at all. That gap is only found by reading the notes against the directory — which turned up four things described but not carried, including a fix that existed in exactly one local clone with no remote of its own.
Secure Boot private keys and NetworkManager profiles both contain secrets — the first can sign anything your firmware will trust, the second holds Wi-Fi passwords in plain text. Neither belongs in a record meant to be shared, and their absence is written down so it reads as a decision rather than an oversight.
tpm-fido (Chromium only, unpackaged). Native support is on the roadmap, not shipping.8086:774c) has no kernel driver. Intel deprecated it; the NPU supersedes it and is fully working. Nothing to do.SEN3.CTYP lookup fails; a firmware bug with no consequence. The _FST fan-status method fails too and was on this list, but it is no longer left alone: a desktop widget was polling it twice a second for a value that never existed, and the resulting 8,000 errors per boot were destroying the evidence any freeze investigation depends on.Enabled, with custom keys, booting a signed unified kernel image. Getting here took two false starts, both worth knowing about if you are following the same path.
There is no dbx variable at all — only dbxDefault, the firmware’s default list, which is not enforced until enrolled. So Secure Boot here enforces zero revocations: a Microsoft-signed but known-vulnerable bootloader would still validate, and Microsoft’s certificate is in db because Windows needs it.
fwupdmgr reports “UEFI dbx: latest available firmware version”. Ignore that. The device’s version reads None because the variable does not exist, and it carries only-version-upgrade — with nothing to compare against it offers nothing. It is not saying dbx is current; it is saying it cannot tell.
Before enrolling, check that nothing which boots would be revoked. The hashes must be computed the way UEFI computes them — Authenticode, not a plain sha256sum of the file — or the comparison is meaningless:
443 revoked image hashes, 0 revoked certificates
ok bootmgfw.efi 96da0da4…
ok systemd-bootx64.efi e424095f…
ok manjaro-7.1-x86_64.efi 384b8c16…
ok grubx64.efi 96241d5d…
SAFE — nothing that boots here is on the revocation list
Zero certificate revocations in the current list means the hash check is complete: nothing can be revoked through its signer instead. Not enrolled yet, because it is close to irreversible — undoing it means restoring firmware default keys, which also wipes the custom PK, KEK and db and means redoing all of this.
Secure Boot: Enabled (user), booted via the UKI entry, Measured UKI: yes. All three out-of-tree modules still load.
Windows, untouched on its own disk. Every hardware fix. 125 checks passing.
Clearing the keys in BIOS resets the fingerprint sensor. Enrolments do not survive it, whatever the SDCP persistence fix does — that patch keeps them across reboots, not across a firmware key reset. Do the Secure Boot work first and enrol afterwards, or expect to re-enrol. Losing them at this step is not evidence that the patch failed.
Turning Secure Boot on while the firmware still held only the factory keys meant nothing signed with our own key validated, and both Linux entries failed. Nothing was damaged, but it looks alarming. Enrolment has to happen first, and it can only happen with the firmware in Setup Mode — which means clearing the existing keys in BIOS, not adding anything there.
With Secure Boot on, GRUB 2.14 never reaches a menu. It dies loading its own module:
error: kern/verifiers.c:grub_verifiers_open:119:
verification requested but nobody cares:
(hd1,gpt2)/@/boot/grub/x86_64-efi/normal.mod
The standard remedy is to rebuild with --disable-shim-lock. That was done, verified present, and it did not help. Two diagnostics that look authoritative and prove nothing:
| Test | Verdict | Why |
|---|---|---|
strings | grep shim_lock | Useless | Those strings are present either way. The flag writes a module-table entry, not a string. |
| Reading kern/efi/sb.c | Useless | The copy fetched was from rhboot/grub2, Red Hat's downstream fork — not the code Manjaro ships. |
| Parsing the PE mods section | Valid | Module type 7 appears if and only if the flag was used. |
The marker was present and GRUB still refused. An observed boot failure outranks any amount of source reading. Treat GRUB under Secure Boot on 2.14 as a dead end and move on.
Kernel, initramfs and command line fused into one signed EFI binary the firmware launches directly. No bootloader in the trust chain, so nothing needs to verify loadable modules.
/boot/efi/EFI/Linux/manjaro-7.1-x86_64.efi, 44 MB, signed by the enrolled db keydefault_uki=, and signed by sbctl's hookIf you dual-boot Windows, do this first or you will be locked out of it. Changing Secure Boot state alters PCR 7, which is one of the measurements BitLocker seals its key against. Windows will demand the recovery key at the next boot, and if you do not have that key your data is gone. From an elevated Windows prompt:
manage-bde -protectors -disable C: -RebootCount 0
That suspends protection until you re-enable it, rather than decrypting anything. Re-enable with manage-bde -protectors -enable C: once Secure Boot is settled. Have the recovery key to hand regardless — it is in your Microsoft account under Devices → BitLocker keys.
Related, for afterwards: boot Windows from the firmware boot menu rather than chainloading it, at least until protection is re-enabled. Chainloading changes PCR 7 too.
sbctl enroll-keys -m -f. -m keeps Microsoft's certificates so Windows still validates; -f keeps the vendor's own, which Samsung's firmware components need.grub-install without --no-nvram moves GRUB to the head of the boot order on every GRUB update, sending you back to rescue.bootctl install overwrites that path — which is the better outcome, since the reset fallback then yields a menu rather than one fixed kernel. The frozen-known-good-UKI idea was tried first and displaced; the old image is kept at /root/bootx64.efi.uki-fallback.bak.systemd-boot now provides the OS chooser. It is signed like everything else, discovers the UKI automatically, and lists Windows alongside it. Because Windows lives on its own disk and systemd-boot only auto-discovers on the partition it booted from, \EFI\Microsoft is mirrored onto the Linux ESP — a copy, so it goes stale after a Windows feature update. The symptom is that entry failing with 0xc000000f while Windows still boots perfectly from the firmware menu — which points you at Windows rather than at a stale copy. gb5-refresh-windows-boot.sh re-mirrors it; run it after any Windows feature update, or whenever that entry misbehaves. The firmware’s own Windows entry is never touched and stays as a fallback.
sbctl verify after every update, before rebooting. sbctl's mkinitcpio hook exits successfully when it cannot sign, so pacman can report success over an unsigned image.Docker on a machine that snapshots itself and routes through a mesh VPN is not the stock install. Two independent reviews ran over this before anything was enabled; between them they found a security hole, two settings that were doing nothing, and one place the reviewers disagreed with each other.
This is the one that mattered. -p 8080:80 binds 0.0.0.0, and Docker inserts its DOCKER-FORWARD jump at the top of the FORWARD chain. NetBird’s only inbound guard is the last rule:
-A FORWARD -i wt0 -j NETBIRD-RT-FWD-IN
-A FORWARD -m mark --mark 0x1bd20 -j ACCEPT
-A FORWARD -o wt0 -j NETBIRD-RT-FWD-OUT
-A FORWARD -i wt0 -j DROP <-- never reached for a published port
So every published port would have been reachable from all of 100.76.0.0/16 and from the LAN. Checked against the live ruleset rather than reasoned about. Fixed with "ip": "127.0.0.1" in daemon.json, which makes exposure opt-in per port.
/var/lib/docker is its own btrfs subvolume, @docker, a sibling of @. Left inside @ it would put tens of GB into every Timeshift snapshot and make a restore roll container state back with the system. containerd’s root moved alongside it, because Docker 29 makes the containerd image store the default — one flag flip and the whole store relocates to /var/lib/containerd, back inside @, where the guard cannot see it.
Three guards hold that in place, and each catches what the others cannot:
Proven in both directions: with the subvolume unmounted, Docker refuses to start and zero bytes reach @.
The reviewers disagreed about DNS, and the disagreement was real. NetBird owns /etc/resolv.conf. One reviewer wanted a public fallback so containers survive the VPN dropping; the other pointed out that leaks internal names on timeout. Container DNS is pinned to the NetBird resolver with no fallback — the non-leaking side — and the cost is stated rather than hidden: no container DNS while NetBird is down. Neither reviewer noticed the pinned address is this host’s dynamic wt0 address, so gb5-docker-verify now fails if it drifts.
A kernel upgrade deletes /usr/lib/modules/$(uname -r) while that kernel is still running. NetBird’s nftables firewall manager then cannot build its tables, and every restart fails with an error that names netlink and never mentions modules. Established by probing each expression separately instead of arguing from the message:
table / chain / set / counter OK
ct state FAIL nft_ct is =m, .ko deleted
fib FAIL nft_fib_ipv4
masquerade FAIL nft_masq
Not version-specific — 0.77.0 and 0.77.1 fail identically — and no flag avoids it, because the firewall manager is constructed before any flag is read. The running daemon survives only because it predates the upgrade.
The fix did not need a reboot. Every pre-upgrade Timeshift snapshot still holds the module tree, byte-identical; modinfo -F vermagic checked against /proc/version first, then cp -a and depmod. Those are the originals for the running kernel, not a forced version mismatch.
And that repair silently disabled four safety checks. All four detected a pending reboot by testing whether the module directory exists. Putting the tree back made the pending reboot invisible: the unattended updater would have stacked a second kernel upgrade onto an unrebooted one, and hibernation would have been unmasked while a stale-kernel image could still be written. They now compare kernel versions, which cannot be faked by restoring files. The general lesson: after any repair, ask what now looks healthy that is not.
NetBird is installed by upstream’s install.sh straight into /usr/bin, so pacman does not own it and the guarded updater will never touch it. The installer also leaves the binaries owned by the invoking user — while netbird.service runs them as root. Anything running as that user could overwrite the binary and own root at the next restart. Corrected to root:root; re-running the installer undoes it.
Its tray UI had never appeared for a mundane reason: the default build links against WebKitGTK 6.0, which is not installed here. The GTK3 release asset links against the webkit2gtk-4.1 already present, and upstream ships no desktop entry or icon — which is why nothing was ever in the launcher.
taskset on the client does nothing anyway — builds run in the daemon’s cgroup.A full freeze — no cursor, no magic SysRq, held the power button. The journal turned out to hold seven abrupt terminations going back three weeks, four of them while the machine was in use. They are two unrelated faults, and the reason neither left evidence was a deliberate setting in this project’s own tuning file.
Two of the four freezes happened while aircrack-ng was running against the internal BE201. The firmware asserts, the driver tries to recover, and the recovery path is where it dies:
iwlwifi: Microcode SW error detected. Restarting 0x0.
iwlwifi: 0x00000070 | ADVANCED_SYSASSERT
ieee80211 phy0: Hardware restart was requested
Oops: general protection fault, probably for non-canonical
address 0xdead000000000122
RIP: 0010:ieee80211_free_chanctx+0x25/0x80 [mac80211]
Comm: airodump-ng Tainted: G OE
0xdead000000000122 is LIST_POISON2. The chanctx was freed and then walked again — a use-after-free in the new iwlmld op-mode’s teardown, reached only because the firmware restart and airodump’s next channel hop race each other. It never appears in a boot without monitor mode: zero microcode errors across every ordinary boot in the journal.
Upgrading the kernel does not fix it, and the usual check would have said otherwise. 7.1.9 to 7.1.13 looked like a chance to inherit a fix. Comparing srcversion is the easy test and it is weak, so the modules were decompressed and their compiled code hashed directly. iwlmld, iwlwifi, mac80211 and cfg80211 are SHA256-identical across the two kernels, with ieee80211_free_chanctx at the same offset 0x614f0. The whole-module files differ by two bytes: the vermagic string. Updating linux-firmware is equally pointless — this driver caps at ucode c102 and will not load the c103 and c106 already sitting on disk.
It is also only sometimes fatal, which is worth stating because it changes what a fix has to do. The same oops fired in two other boots that survived — one rebooted cleanly, one shut down degraded — and a third boot logged twenty-four firmware asserts with no oops at all and shut down normally. The freeze happens when the dying worker thread happens to hold a contended net lock. In the fatal boots the machine limped on for about seven and a half minutes, the power-profile timer still ticking once a minute, before it stopped answering entirely.
Only one thing removes this: not using this adapter for monitor mode. An external USB adapter on ath9k_htc or mt76 bypasses iwlmld altogether. Short of that, gb5-monitor-mode on reloads iwlwifi with fw_restart=0 for the length of a session, so an assert leaves the adapter down instead of entering the teardown path. That is a session toggle and deliberately not a modprobe.d default: permanently disabling firmware restart means any blip kills Wi-Fi until a reload.
The other two freezes have no crash signature of any kind. Machine checks, EDAC, thermal, GPU resets, PCIe AER, RCU stalls, hung tasks, OOM — all clean, and the log simply stops mid-sentence. The tempting conclusion is a platform-level hang below the kernel’s ability to log anything. That conclusion is wrong, and correcting it mattered more than it first appeared.
This machine had kernel.nmi_watchdog = 0, committed months earlier to save a PMU counter and one timer interrupt per CPU. An ordinary software hard lockup — a CPU spinning with interrupts disabled, in any driver — produces exactly this signature when the hard-lockup detector is switched off and there is no crash-dump backend. “No evidence” was fully explained by “nothing was looking”. These freezes are undetected, not proven to be firmware.
Three settings between them guaranteed that anything going wrong would end as a dead machine rather than a reboot:
All three are now inverted, with hardlockup_panic and softlockup_panic alongside them and a twenty-second reboot. The PMU counter is spent deliberately: recoverability is worth more here than a few milliwatts. Because a perf-based detector cannot fire if the local APIC itself stops, the PCH’s own iTCO watchdog is armed too, at RuntimeWatchdogSec=30s — it sits outside the CPU cores and recovers the machine when nothing on them is still running.
A panic that reboots still teaches nothing if it takes its evidence with it, so ramoops now persists the dump across the warm reboot into an 8 MiB region reserved with reserve_mem=8M:4096:oops. The name-based form avoids hardcoding a physical address. EFI pstore was considered and rejected: it writes crash logs into UEFI NVRAM, and this firmware has not earned that. A signed 6.12 LTS unified image was added as a second boot entry at the same time, because arming automatic reboot without an escape hatch is how a boot-time panic becomes a loop.
Adding that fallback kernel armed a latent bug that would have booted the wrong one. gb5-ensure-uki.sh picked the default boot entry by mtime, so the most recently built image won — which is the older LTS. Worse, it compared that against Default Boot Loader Entry:, an empty section header in bootctl output; the value lives under Default Entry:. The comparison could never match, so it rewrote the default on every run. Its hook fires on Target = linux*, which matches all seventeen linux-firmware-* packages, so the next weekly unattended update would have quietly pointed the machine at 6.12 — a kernel whose camera and speaker modules do not build. Found by a review, not by the person who introduced it. It now sorts by version, excludes fallback images, and gb5-verify.sh asserts the result.
Every boot logged around 8,000 copies of one ACPI error: this firmware’s fan-status method, \_SB.PC00.LPCB.FAN0._FST, fails on every call, twice a second. It was the last line written before one of the silent freezes, which made it look like a cause. It is not — but it was destroying the ring buffer that any future investigation depends on.
thermald was the obvious suspect and was wrong. Stopping it for fifteen seconds changed the rate by nothing; pausing KDE’s ksystemstats dropped it to zero. A desktop widget was polling the sensor at 2 Hz. The readout it was fetching never worked at all: fan1_input returns -EIO unconditionally, so the widget had been displaying nothing while generating eight thousand kernel errors per boot for it.
The first fix measured clean and was still wrong. Turning the fan off in the widget’s settings gave zero errors over thirty seconds — and the flood returned at exactly 2.00/s after the next reboot. That setting governs whether the applet displays the sensor, not whether ksystemstats subscribes to it. A measurement taken in the same session proves nothing about a poller that re-subscribes at login; the only honest test is a reboot. The fix that holds is a udev rule making fan1_input root-readable, so an unprivileged poll never reaches the failing AML. One error per boot now, from before udev runs.
None of this diagnoses the two silent freezes. It makes them diagnosable: the next one reboots the machine instead of killing it, and leaves the dump behind. The outstanding lead is a BIOS update. That cannot be done from here: Samsung does not publish to LVFS, so fwupdmgr reports no updates and always will — an ESRT entry exists, but with nothing to apply. The route is Samsung Update from the Windows install on the other disk, which is worth knowing before concluding the firmware is current because Linux said so.
After 12 September nobody touched this machine’s setup. The weekly guarded updater kept running: eight times, six updates, one correct refusal on a broken mirror, one “nothing to do”. Every update ended “Update complete and verified.” The first kernel-series bump it ever met alone showed what that sentence did not cover.
On 11 November it applied 602 packages on battery, including Manjaro retiring linux71 (7.1.13) for linux72 (7.2.9) in a single transaction, and GRUB 2.14 → 2.16. What held: the new kernel got a signed unified image, the default boot entry followed it, all three DKMS modules rebuilt, the hibernation resume offset stayed put, and the patched fingerprint library was protected. Four check failures sat unread in the log.
sbctl sign -s on an already-signed file returns success without recording it.mkinitcpio -P fail, breaking the printed rollback recipe. The .pacnew check compared mtimes, which pacman sets to the package’s build date, so it could never fire.A new hook, gb5-boot-guard.sh, runs last in every boot-relevant pacman transaction, however it was started. It checks signatures with sbverify against this machine’s own key and keeps four things true:
It also keeps systemd-boot itself current through a pre-signed copy, and has a read-only --check mode. Manjaro’s GRUB hook is disabled. The updater now detects a kernel change from what changed on disk, never from a name, and tells a running-kernel replacement apart from an LTS-only update. Snapshot pruning and a weekly Flatpak update were switched on, so the machine maintains itself.
The first draft of the fix repeated the bug it was fixing. An adversarial review, run before anything was committed, found four faults:
rm -rf’d on it.All four are fixed and tested by fault injection: a deliberately removed database entry was reported by --check and left alone, then repaired. A real GRUB reinstall then exposed one more pre-existing gap, a GRUB-only update leaving grubx64.efi unsigned, which the guard now closes.
Two corrections to earlier chapters:
Still unproven on 7.2: one hibernate cycle, a fingerprint touch (nothing is enrolled), speaker and camera use, a boot of the LTS image, and the rewritten kernel branch itself. It meets real input at the next point release.