Skip to content

Notes from armada-os/armada

What a read of the armada repository (a Fedora bootc image with ROCKNIX-derived device support, main at 934b614, 27 September 2026) turned up that bears on the Nova. Kept here so the ideas are not lost when the survey scrolls away. Each item says where it lives there, why it matters here, and what we did with it. Their packages/kernel/PATCHES.md records the provenance of every patch, which makes cherry-picking easy.

Nothing in armada covers RetroArch, a launcher, ScummVM, mpv or panel refresh modes; its display work is gamescope and HDR.

Taken

Item armada Ours
8bpc output dither: the DPU disables dithering at 8bpc, so a 10-bit result is truncated onto the panel; a panel opts in per DT property patches/0049-drm-msm-dpu-panel-opt-in-8bpc-dither.patch (armada-authored) kernel patch 0049 verbatim, armada,dpu-8bpc-dither on the Nova panel (#370)
Gamepad MCU on vdd_mcu_3v3, rail not always-on, rsinput releases it across suspend 1007-input-rsinput-drop-the-mcu-supply-across-system-sleep.patch, dts/qcs8550-ayn-common.dtsi.patch patch 1014, DT change, HTR3212 driver with PM ops from ROCKNIX SM8750 (#370)
Stick calibration through rsinput module parameters apply-input-calibration + input-calibration.json GPcal shipped with its pad name and wrapper fixed (#370), then dropped with pyxel; rsinput's parameters remain
GMU HFI interrupt clearing the reply and OOB ack bits it does not own, unbounded devcoredump wait, HFI interrupt unmasked before GMU boot 0620, 0621 (from Kettle), 0622 patches 1099-1101; 096-cpuidle and its hooks removed (#602)
UFS lane clock helpers enabling every clock, geni UART kept active across sleep (upstream backports) 0533, 0534 patches 1102, 1103 (#603)
ath12k scan priority never sent, scan timeout deadlock 1020, 1021 patches 1104, 1105; 1106 and the timezone's country are ours
GPU submitqueue priorities mapped onto HIGH through LOW, keeping KERNEL for VM_BIND kernel 0619 kernel patch 1113; imported, device verification pending
libliftoff stacking across multiple primary planes gamescope 0020 gamescope patch 0009, including its nine regression cases; imported, device verification pending

Steam: checked again on 18 October 2026

Checked armada at 4c945ec5 against our Gamescope 2b5cf191, its libliftoff 8b08dc1c, kernel 7.2.11, Steam launcher and installer. Kernel 0619 and Gamescope 0020 are now imported as recorded above; neither has been measured on the Nova. The remaining changes below are an import assessment.

The kernel fix applies to both affected 7.2.11 files and corrects an inversion that can give an ordinary Turnip queue kernel priority above Xwayland's GL queue. Armada reports compositor starvation under GPU load. It changes scheduling for every GPU client, so compare Steam and Vulkan emulators before calling the device behavior verified. Both affected kernel objects, msm_submitqueue.o and msm_gpu.o, compile in Linux with the Nova config on 7.2.11. A check using the actual conversion helper and entity-index expression passes 42 priority and range cases; the original mapping fails, and omitting the entity-index correction fails separately.

The libliftoff fix keeps Steam and external overlays above an opaque game when the DPU scans layers out on separate primary planes. It applies to our exact submodule: five of its nine new cases fail before the fix, and a clean Linux build with the fix passes all 67 tests, including the existing suite. Overlay order and direct scanout still need a Nova test.

Candidates

Change there Value here Assessment
Gamescope 0024, report native Wayland windows to Steam A game with an xdg-shell window can remain behind Steam's launch spinner because only Xwayland windows are reported as focusable. Publishes the Wayland window's app ID and PID too. Applies after our current patches. Useful when adding native Wayland games or Proton's Wayland driver; verify launch, focus, overlay and return to Steam. Our launcher would also need to expose Gamescope's Wayland socket deliberately.
Gamescope 0025, native Wayland touch passthrough Lets an xdg-shell application receive touch rather than Steam's emulated mouse clicks, retaining passthrough for that touch until it ends. Applies with 0024. Conditional on supporting native Wayland applications; test touch across focus changes and keep X11 mouse emulation working.
Gamescope 0018, ARM64 virtual white Recovers the missing white-point component from the ARM Steam client's packed property, making its color-temperature slider effective. Applies directly. Needs the Steam color-management capability flags and on-device verification, including direct scanout. It does not supply the DPU color pipeline by itself.
launch-steam, preserve SYSTEM_PATH and SYSTEM_LD_LIBRARY_PATH Steam restores these when starting compatibility tools. Armada sets them explicitly to keep the tools' command and library search paths available. Small launcher change worth testing. Our direct ARM client launch sets neither; use our own runtime paths rather than copying Armada's paths.
Proton-CachyOS DXVK probe Requires storageBuffer8BitAccess as well as descriptorIndexing before choosing modern DXVK; otherwise allows the bundled fallback. Low priority on the Nova: our Mesa enables 8-bit storage for A7xx, so this should not change its choice. The guarded patcher works on the real 11.0-20261005-slr Proton script; its missing/duplicate-probe tests pass. Our installer downloads changing Proton releases, so any adoption must detect the installed script rather than assume that release's shape forever.

Sources: GPU priorities, Gamescope patch ledger, Steam launcher, DXVK probe patcher.

Useful designs requiring adaptation

  • Per-game FEX profiles. armada-game-launch writes FEX_APP_CONFIG into the user's cache, which pressure-vessel shares with the game. It omits RootFS, ThunkHostLibs and ThunkGuestLibs, allowing each compatibility tool's FEX to use its own paths. Our system FEX settings alone cannot be assumed to configure a bundled FEX. Worth adapting for per-game compatibility, with an actual game launch confirming consumption.
  • Controller recovery. The same wrapper injects missing x86 Wayland and xkbcommon libraries into applicable x86 Proton tools, and repairs stale Valve-controller registry entries left when changing Proton architectures. These are specific workarounds, involving installed tools and prefixes: reproduce the missing-library or stale-registry failure before adopting either. They do not calibrate the Nova's physical sticks.
  • Steam repair. armada_steam.py stages and verifies replacement client files while preserving games, account data and compatibility tools. Its Linux tests pass (13 passed, two integration cases skipped). Our installer instead removes a real Steam directory when replacing it with the library symlink. Adapting preservation and staging is worthwhile; Armada's factory payload/OSTree exporter cannot be copied into PortareOS unchanged.

Leave out for now

  • Gamescope 0028 is the three-commit frame-pacing series from Valve PR 2417, merged on 11 October. Our 14 October Gamescope pin already has the scheduling helpers, independent nested refresh clocks and Wayland wakeup.
  • 0027 changes the fixed-refresh limiter. Our normal Steam launch uses VRR and its separate limiter; our newer fixed-refresh branch also differs from the one Armada patches. Revisit for a measured fixed-refresh problem, including external displays.
  • 0026 makes Steam's FSR selection run SGSR on supported SDR inputs. Our Gamescope already offers SGSR explicitly. It is a selection policy change, with no Nova quality or performance comparison yet.
  • The synthetic EDID, nominal Nova HDR profile and hardware color-management stack (0002 through 0006, 0017, 0019, 0021) are a separate display project. Armada's Nova profile explicitly uses unmeasured nominal P3 primaries; the LUT patches require its kernel's 3D LUT/colorop interfaces, which ours does not expose. Do not treat these as a small Steam fix.
  • DRM leases, companion-screen MangoApp fixes and the generic fallback splash app ID serve Armada's multi-screen/session design. We have one screen and already carry a targeted Steam-bootstrap focus fix.
  • Gamescope already requests realtime Vulkan queues when it has CAP_SYS_NICE. Check that capability on our root-run Steam path before considering Armada's bypass; it is separate from the kernel's submitqueue mapping fix.

The applicability checks used a fresh archive of our exact Gamescope pin with all five of our patches applied first. 0018, 0024 and 0025 were then applied together and reverse-checked in that temporary tree. The libliftoff tests ran in an ARM64 Debian container against the pinned dependency, using a fresh build after applying the allocator fix. These checks establish applicability and regression behavior, not game performance.

Suspend and sleep power

armada's commit 431bf57 reports Nova sleep power on USB input falling from 0.67 W to 0.27 W with the following together. The MCU rail above is one part; the rest is open here and relevant to the overnight drain item (#62).

  • PCIe D3cold (0522-PCI-host-common-let-only-endpoints-veto-d3cold.patch, armada-authored). The generic PCI host asks every device in the hierarchy whether D3cold is possible; the root port, still in D0, says no, so the PCIe controller, PHY and clocks stay on through s2idle. The patch lets only endpoints (the WCN7850) answer. Removes a spurious veto rather than adding one. Risk: Wi-Fi must return from D3cold on resume, which our PCIe suspend patches already exercise.
  • Codec rails in LPM during s2idle: regulator-state-mem with regulator-on-in-suspend and RPMH_REGULATOR_MODE_LPM on vreg_l15b_1p8 and vreg_bob1. We have the enabling driver patches (our 0218 and 1047) and use them for bob2 and the fan, not for these.
  • ADSP sleep with audio open at suspend (0528, 0529, 0530, ASoC sc8280xp card PM ops and AudioReach graph release across suspend). Measured on an Odin 3 as about 0.09 W when a stream was open at suspend. The launcher, RetroArch or mpv can hold a PipeWire stream at suspend. armada then raised its PipeWire min-quantum to 1024 against post-resume crackle; that part conflicts with our 256-frame quantum.
  • tsens lower-threshold IRQs masked across suspend (0204): a more targeted version of our 0203, which skips arming the whole uplow IRQ.
  • IPCC summary IRQ masked only for suspend-to-RAM (0504): keeps IRQF_NO_SUSPEND and masks unlazily for s2ram only, scoped to SM8550/SM8750, said to stop cable-attach wakes on the AYN Thor. Our 0502 removes IRQF_NO_SUSPEND outright.
  • ICE clocks unwound on resume failure (0510); our 0207 already matches their 0511 (UFS IRQ on host reset failure).
  • iommu-map cell count (0514 upstream by Mani Sadhasivam, plus 0514a): with five-cell iommu-map entries the fixed four-cell stride misprograms BDF to SID and, with fw_devlink.strict=1, links PCIe to random CoreSight nodes. We boot with fw_devlink.strict=1. Check the sm8550 dtsi before the next kernel bump.
  • Wake ledger (system-sleep/40-armada-wake-ledger): 20 lines appending pm_wakeup_irq, the interrupt name and a timestamp to a log on every resume, since the kernel keeps only the last wake. We record nothing about wake reasons. Trivial, and the first thing to add when #62 is worked.
  • Sleep tracer (armada_sleep_debug.py, 1391 lines): a per-cycle ftrace instance over power, rpm, clk, regulator, interconnect icc_set_bw, rpmh_send_msg (filtered to the sleep and wake sets) and UFS events, then a report. Heavy, but the event list is a recipe.
  • Fan supply in suspend: their powerd sets pwm-fan pwm1_enable to 3, cutting the fan supply at duty 0 and in suspend. We rely on regulator-off-in-suspend on the fan regulator instead. Probably equivalent; worth one measurement.
  • Already equivalent: dwc3 USB autosuspend (our 81-usb-runtime-pm.rules), PCIe L23 skip (our 0215), PCIe suspend OPP (our 1048/1049), geni UART IRQ masking (our 1006), rsinput quiesce (our 1004), UART frame reassembly (our 1013).

Charging

  • Their 0903 exposes the charge current limit as our 0510 does. Their note adds: writing 0 stops battery charging while USB keeps powering the system, which is bypass charging. A "bypass while docked" option would be one line in charge-throttle if the Nova firmware behaves the same.
  • armada-charge-debug: a one-shot charging diagnostic (sysfs, journal, the QCOM_PMIC_PDCHARGER_ULOG firmware log). 0900 logs the USB adapter type when it changes.

Display and colour

  • LUTDMA IGC and 3D gamut LUT (0068 to 0070, 1515 lines): an independent implementation of what our never-run 1070/1071 do, verified on an Odin 2 Portal. Facts their notes record, to check ours against: IGC components ordered G, B, R; 257 samples at i/256, so the 256-entry LUT is resampled; command buffers rewritten in place, so an update waits on the done bit; in-flight marks dropped on runtime resume. They also expose the 17^3 gamut as a CRTC blob, which we do not.
  • DSPP joins the topology only on a modeset (from their gamescope patch 0017): colour management on the CRTC needs a modeset when it is switched on or off. Relevant to how the launcher applies a profile.
  • DSI brightness transfer (0016-rp5-smooth-brightness-adjustment.patch, generic despite the name): msm_dsi_host_xfer_prepare() re-cycles link clocks on every DCS transfer while the display is on, cutting the byte and pixel clocks mid-frame and underflowing the HS FIFO. Our Nova panel driver sends a DCS command per brightness step. Try it if brightness changes ever glitch the picture.
  • DP audio prepare while the display is off (0618): a failing prepare made PipeWire drop the whole card profile, losing speaker and headphone sinks too. Directly relevant to our unverified HDMI sink switch, since Nova HDMI is USB-C DP.
  • Nova firmware: their rp6 DTS points ADSP, adsp_dtb, AW883xx and battmgr.jsn at files extracted from a Pocket 6 vendor image rather than the Odin 2 fallback ours still names. Relevant to the audio topology our link-rate patches run against and to charger behaviour. The blobs are vendor licensed.
  • Stale jack report on removal (0064, wcd-mbhc, shared with our codec): only if the Nova ever shows a stuck headphone state.
  • Mesa 26.2.3 workarounds for A740 translation-fault storms (disable-turnip-sparse-sync, ir3-disable-bindless-ubo-const-lowering): candidates if Vulkan cores ever show SMMU faults.

Input

  • Haptics 1003: the rumble bridge slept under event_lock ("scheduling while atomic" on every rumble stop, intermittent lockups on an Odin 3); mutex, deferred work, workers cancelled on suspend. Compare against our 1005.
  • Haptics 1004: the periodic-sine patch (our 1002 is the same) leaves a 5-second delayed brake tap after rumble ends. We likely share it.
  • axis-deadzone DT property for all four stick axes (1300); the Nova DTS there uses trigger deadzones only, as ours does.
  • Volume-up is not a wake source there (wakeup-source removed from KEY_VOLUMEUP); ours keeps it.

Wi-Fi after resume

  • NetworkManager 0002 and wpa_supplicant 0001/0002 (Collabora): scan only the last-associated channel after resume. We use iwd, so only applicable if the backend changes.
  • ath12k 1020: cmd->scan_priority never assigned, so every host scan goes out as VERY_LOW and is refused while the 11d scan is pending; reassociation took 4.4 to 8 s with iwd. Same WCN7850, same iwd here. Taken as patch 1104, 2026-10-09: the Nova connected 9.4 to 9.7 s after resume, 8 s of it in that wait, and 1.9 to 3.0 s with it, 12 of 13 resumes (one quick scan was aborted, 8.5 s). The wait was what let 11d find a country: ath12k stops 11d on association, so with a fast reconnect the chip stayed on world rules, without 6 GHz. The country now comes from the timezone (setregdomain, with ATH_REG_DYNAMIC_USER_REG_HINTS), and patch 1106 sends it again when the radio starts after the power cycle every suspend is.

Power and performance policy

Informational; opposite choices to ours in places.

  • IRQs pinned to cores 3 to 7 on SM8550 with the note "the A740 GMU wedges if a GPU interrupt wakes CPU 0-2 out of power collapse". We boot with irqaffinity=0-2. Worth knowing when chasing GPU hangs.
  • conservative governor for eco and balanced, per-policy underclock tables, fan curves as temp:pwm pairs with smoothing. We run schedutil, teo and a fancontrol curve.
  • Kernel: SCHED_CLASS_EXT with scx_lavd, UCLAMP_TASK, LRU_GEN, PSI, NTSYNC. A udev rule copies GPU and GMU devcoredumps before the kernel's 5-minute auto-delete; we capture none.
  • bfq plus read_ahead_kb=4096 on mmcblk; we set bfq only.

Build and process

  • CI computes a content hash per package, checks the registry for an image with that tag and builds only the missing ones, one job per package on arm64 runners. Our incremental nightly restores saved state; the hash-as-tag idea is the only new element.
  • tests/ holds about 48 shell and Python tests run by just check, among them a UCM duplicate check and a power-button suspend hook test.

Access

GitHub's API and HTML were not reachable from the session; the survey worked from a shallow clone of main. Older history, issues and pull request discussion were not read.