Validation record
Tato stránka zatím není přeložená, zobrazuje se anglická verze.
This page is the honest answer to “how well does this work?”. Everything below is a local x86_64 development result unless stated otherwise. Nothing here is a claim about your hardware.
What is covered
Section titled “What is covered”| Area | Status |
|---|---|
| x86_64 ISO, BIOS and UEFI | Booted, installed to XFS, rebooted — on every commit, PR, tag and full-matrix run |
| Other five architectures | Built and published; no VM is started |
| Device images | Built on tags; no boot test configured |
| Physical GPU, Wi-Fi, suspend, multi-display | Not certified; needs separate hardware tests |
| Bluetooth and WireGuard VPN | Verified against BlueZ and ConnMan VPN fixtures, not physical hardware |
| Byte-identical whole-ISO rebuild | Not established |
Publication requires green build reports for all six architectures and green boot reports for both x86_64 firmware modes. A configured test is not evidence of a successful run — check the workflow results and attached build records.
Measured on the installed reference target
Section titled “Measured on the installed reference target”The reference target is an installed x86_64 system with 512 MiB of RAM, one display and no swap. Both firmware paths passed ISO boot, XFS installation, installed-system reboot, partition-table preservation, and a complete hash check of an unrelated partition. Root remained locked and the installed account used its new password. Neither run reported OOM.
| Firmware | Live serial login | Background components PSS | Flyout PSS | Full UI PSS |
|---|---|---|---|---|
| BIOS | 44.13 s | 6.29 MiB | 31.22 MiB | 34.77 MiB |
| UEFI | 54.29 s | 6.05 MiB | 31.34 MiB | 34.78 MiB |
These are QEMU software-rendering measurements. Background figures include the session service and the temporarily active privileged helper, not the entire desktop. Live boot timing uses a separate 2 GiB live-image test, not the installed 512 MiB test.
Native headless measurements: session service 5.72 MiB PSS, Settings UI 36.77 MiB, flyout alone 31.24 MiB. A two-second idle sample recorded 0% CPU for the service and GUI. First usable frames were 36 ms for the full UI and 36 ms for a new flyout process; a warm flyout activation in the existing GUI took 6 ms. These are local process-launch measurements with build caches retained, not a disk-cache-cold hardware benchmark.
Budgets are ≤16 MiB PSS for background Settings components, ≤80 MiB for the full UI and ≤64 MiB for the flyout, with a cold usable frame target of 1.5 s and warm 500 ms. Those are native reference targets, not emulated timing guarantees.
Stripped executables: uflinux-settings 989,176 B, uflinux-settingsctl 735,328 B,
uflinux-settingsd 1,370,184 B, uflinux-settings-system 775,912 B, Ironbar 6,703,128 B.
Those artifacts shipped Ironbar as their taskbar; the current tree replaces it with the
in-repo uflinux-desktop shell, and the numbers are kept as recorded.
What the tests actually exercise
Section titled “What the tests actually exercise”- Real headless labwc output management: test and apply, exact modes, last-display rejection, the 15-second deadline, caller-loss rollback, fractional scaling, rotation, hotplug and saved-layout fallback.
- PipeWire subscriptions over outputs, defaults, volume, mute, recording and playback streams,
and reconnection after a
pipewire-pulserestart. - A controlled ConnMan D-Bus fixture for credentials, cancellation, network power, addressing, DNS, proxies and forgetting — this does not certify wireless association.
- Real
mac80211_hwsimand hostapd tests for WPA2 personal, WPA3 SAE, hidden networks, PEAP/MSCHAPv2, TTLS/PAP and EAP-TLS, with DHCP, reconnect and forgetting, plus route restoration, unrelated-route preservation and server-domain mismatch rejection. - ConnMan’s SNTP client against a controlled server, as sole synchronizer, reporting the actual kernel state.
- The real BusyBox
acpidinput parser in a private chroot: only power-key presses invoke the hook; release, repeat and unrelated events are ignored. The installed VM verifies real TLP transitions and physical-key Do nothing and Shut down behaviour. - Signing: resigning preserved payloads; damaged archives and mismatched keys were rejected; temporary private keys were removed on success and failure.
- Repeat build: all custom APK control and data payloads matched after a fresh compiler rebuild, signatures excluded.
- Configuration preservation: modified system defaults survived package reinstallation, and a simulated prior defaults generation upgraded without replacing local edits. That is a simulated previous generation, not a previously published UFLinux release.
Test-only hostapd and dnsmasq packages are not desktop dependencies, and the installed test’s Python interpreter is a harness dependency, not an installed-desktop dependency. Missing test facilities fail the gate rather than silently claiming coverage.
Live image memory
Section titled “Live image memory”The live root unpacks the desktop and the full Papirus icon theme into tmpfs, whose default capacity is 75% of RAM. The current Papirus, artwork and theme package set uses about 993 MiB in the live root. A 2 GiB live guest exhausted its root filesystem during boot, so live tests now allocate 4 GiB — 3 GiB for armv7’s QEMU highmem restriction. The runner rejects boot-time APK database and kernel memory failures explicitly and logs byte and inode capacity.
The unified desktop
Section titled “The unified desktop”The unified desktop is the default session: uflinux-autostart starts
uflinux-desktop --supervise, and the Ironbar panel it replaced is no longer packaged. Do not
run a second desktop beside the session’s own taskbar — both provide panel and tray services.
Its memory work compared the replacement against that legacy panel as the baseline, with three sessions per implementation: median process-set proportional size fell from 86.8 to 73.5 MiB once settled and from 89.2 to 75.9 MiB after one pass through all four menus, both a 15% reduction against a 10% requirement. No replacement session retained more than 0.4 MiB across 100 cycles, none grew its descriptor count, and desktop idle CPU peaked at 0.067%.
The quick-settings tabs and notification surfaces are checked by a protocol test that asserts
the Wi-Fi, Ethernet, Bluetooth and VPN tab labels, the Output and Input audio tabs, and
identical window extents across tab switches. A separate fixture test drives the daemon against
BlueZ and ConnMan VPN fixtures, covering adapter power, scanning, pairing, trust, connect and
disconnect plus a WireGuard import, connect and removal, and checks that an imported private
key is never echoed back. A taskbar phase seeds taskbar.json, restarts the daemon and requires
the file to survive unchanged, then corrupts the file and requires the daemon to start anyway
with no popup.
Those figures come from one headless-pixman environment on one kernel and are not a release approval. The collector does not measure interactive frame latency, widget or D-Bus subscription growth, real audio devices, multi-architecture builds, or an installed low-memory VM.
Still open
Section titled “Still open”Physical Wi-Fi devices, GPU acceleration, HDR, VRR, DDC/CI, laptop backlights, battery transitions and suspend/resume need separate hardware tests — see Physical hardware validation. Generic architecture boot support does not imply those features work on every board. Other CPU architectures and public-repository upgrades remain unverified locally. Clean-build comparisons cover unsigned APK control and data payloads, not byte-identical whole ISOs or signing streams.
Physical hardware validation
Section titled “Physical hardware validation”QEMU and headless tests establish software behaviour. They cannot establish panel HDR or VRR activation, I2C reliability, laptop firmware behaviour, or physical resume. The procedure is to run, from an installed UFLinux desktop on each reference device:
- Graphics — check the renderer in Displays and About, the actual GPU and driver, normal launch and logout, supported Vulkan and OpenGL plus the software fallback. A configured renderer is not evidence of hardware acceleration.
- Displays — apply exact advertised refresh modes, scaling, rotation and arrangement; leave a change unconfirmed to check the 15-second rollback; unplug and reconnect to confirm a usable display recovers and the saved layout returns.
- VRR — verify with the panel’s refresh indicator or a moving test pattern. Protocol support alone is not a physical VRR pass.
- HDR — use labwc’s global HDR option with compatible hardware Vulkan and record an observation. The app does not claim to report per-monitor activation.
- Brightness — test the internal backlight in Settings and the flyout, then with a hardware key; test monitor-specific DDC/CI, disconnect during a request and reconnect. Physical monitor-button changes need a rescan.
- Power — switch all three TLP modes and reboot to check the selected default; test charger transitions and each physical power-button action on AC and battery; check inhibitors, suspend, resume, and networking and audio after resume.
Record results as JSON with optional keys hardware_gpu, vrr, hdr, ddc_ci, backlight,
battery, suspend_resume, each with a status of pass, fail, unavailable or
unverified plus notes identifying device, method and observation. Then run:
python3 scripts/hardware-record.py --results results.jsonThe recorder reads state and metadata; it does not change hardware settings or include Wi-Fi credentials. Its report distinguishes observed configuration from operator-supplied evidence and rejects physical passes on known QEMU or VM hardware.
Device image targets have no hardware evidence yet. When one is tested, record at least that the board boots from the image as written without re-partitioning, that the root partition grew to the media on first boot, the serial console and getty behaviour, the renderer reported by the desktop, and whether the display works at all. The SpacemiT K1 image is expected to have no display output with the current mainline device tree.
Historical artifacts
Section titled “Historical artifacts”The initial development build was checked on 8 November 2026: 24 repository tests, 30 Ironbar
upstream tests with the selected features, ShellCheck and actionlint, a successful abuild of the
runtime, debug, documentation and distro subpackages, a schema-valid SPDX 2.3 inventory, and a
fresh runtime root check. That ISO was 918,552,576 bytes with SHA-256
7b1e22f21357e560c513aa07937662db60de0c9803f9cedc1f6e6369cd382992, and used a development
signing key and an example update URL.
Later development ISOs — with Settings, then artwork, then the splash — are recorded upstream with their own hashes and sizes. Each describes its own artifact; historical ISO checksums in those documents are not measurements of the current build.
Derived from docs/validation.md inUFlinux-project/uflinux3(b4f4817). Report documentation problems inUFlinux-project/docs.