Repository navigation
Add the NVIDIA N1x kernel, driver and platform packages - #930
Conversation
Pinned to NVIDIA's signed Ubuntu 7.0 kernel tag with two SAUCE commits from the 26.04 branch (MediaTek I2C ACPI support for the internal keyboard and touchpad, gpiolib debounce warn-only), installing the raw arm64 Image for Limine. The Canonical signing keys are added to the build keyring list so makepkg can verify the tag.
Ubuntu-nvidia-7.0-7.0.0-1021.21_24.04.1 carries NVIDIA's MediaTek MT8901 I2C and gpiolib debounce SAUCE, so the two cherry-picks are dropped, and adds the ARML0002 FF-A EC driver (battery, AC, lid, thermal, UCSI on the ASUS ProArt P14) and the MT8901 SoundWire audio stack. Carry three hid-asus patches for the ProArt P14's 0B05:4B42 keyboard: the upstream Zenbook A16 support (Fn keys), a standalone keyboard backlight LED for systems without asus-wmi, and host-controlled Fn-lock.
Ship default/alsa/ucm2 from the Omarchy source under /usr/share/alsa/ucm2 when it is there: profiles for sound cards alsa-ucm-conf has none for, starting with the NVIDIA N1x ASUS ProArt P14, kept under ucm2/Omarchy and picked through card-specific conf.d links so they cannot collide with alsa-ucm-conf.
The ProArt P14's keyboard starts in the same OOBE mode as the ProArt P16's, fading its backlight in and out and ignoring the level the host sets. hid-asus only sends the OOBE disable sequence on the P16, so also send it on the P14.
The README still described the 1018.18 tag, its two cherry-picks and the 610 driver pairing.
linux-omarchy 7.2.5-6 built for aarch64, with NVIDIA's N1x SAUCE from Ubuntu-nvidia-7.0-7.0.0-1021.21 rebased onto it as topic patches (MediaTek MT8901 I/O, the FF-A embedded controller, power management, SoundWire audio and DMA-mode IOMMU for the integrated GPU) and the ASUS ProArt P14 keyboard patches. The config starts from NVIDIA's arm64 config and takes linux-omarchy's choices on top. Validated on the ASUS ProArt P14 H7407BA with nvidia-open-dkms 615.71.09. Like linux-n1x, unscoped repository builds skip it.
Add efi_reclaim_reserved=, which hands named ranges of EFI reserved memory to the kernel. The N1x firmware reserves tens of GiB as the GPU's dedicated memory under Windows, half the RAM of the 128 GB ProArt P14, and Linux's driver never touches it. Omarchy sets the ranges per model. Turn the ProArt P14's keyboard backlight off for sleep; without asus-wmi nothing else does, so it stayed lit through suspend.
Arch's cuda recipe already builds for aarch64 from NVIDIA's SBSA installer, but Arch Linux ARM does not ship it, so ARM machines such as the NVIDIA N1x have no nvcc to build ollama-cuda or llama.cpp with. Carry Arch's 13.4.2-1 for aarch64 only; x86_64 keeps Arch's package. It is a 3 GB package, so it stays out of unscoped builds and is built explicitly with bin/repo deploy.
Arch Linux ARM does not build ollama, so Omarchy's Ollama install on ARM machines such as the NVIDIA N1x had nothing to install. Carry Arch's 0.35.0-1 for aarch64 only, with the same split packages so the install menu's ollama-cuda choice works unchanged. ROCm does not exist on aarch64, so ollama-rocm and its backend are dropped; the CUDA and Vulkan backends and GPU targets are Arch's. x86_64 keeps Arch's package. Compiling the CUDA backend for every target takes about 40 minutes on the N1x, so it stays out of unscoped builds and is built explicitly with bin/repo deploy.
The MT8901 sound card had no subsystem ID, which Cirrus CS35L56 amplifiers name their firmware and tuning after, so they ran on ROM defaults. Pass the SoundWire controller's ACPI _SUB to the card, as AMD's SoundWire driver passes its PCI ID.
…Art P14 The ProArt P14's Cirrus CS35L56 amplifiers ran their ROM defaults, quiet and unvoiced, because linux-firmware has no tuning for this laptop (1043:33A1). ASUS ships it in its Windows audio driver for the H7407BA. Unpack that installer without running it and install Cirrus's DSP firmware and the two speaker tunings unmodified, under the names the cs35l56 driver requests, with the Cirrus licence linux-firmware distributes Cirrus's amplifier firmware under. Drop the package once Cirrus adds this laptop to linux-firmware.
The N1x USB4 host routers are ACPI platform devices that nothing bound, and the firmware leaves the connection manager to the OS, so docks and PCIe-tunnelled devices never worked. Add a platform glue driver for the Thunderbolt core, power the routers through power_wrap, fix UCSI's PPM init and drop NVIDIA's USB4-disable _DSM, and keep the PCI bus numbers below a hot-added dock shared between its downstream ports.
With CL states on the N1x host router's links, a monitor behind a DisplayPort tunnel never synced to its first link training and kept dropping out, so a dock's display usually stayed dark at boot and often on hotplug. Turn CL states off for that router.
The connection manager asked the routers to sleep on suspend, but a router only leaves sleep through a reset the N1x never gives it, so a port with a device attached stopped answering after resume. Keep the routers awake in system sleep, and stop powering them down on unbind, which the SSPM cannot undo. Number the package's own releases from here instead of tracking linux-omarchy's release with a decimal: this is the ninth, so 9.
Ship default/alsa/ucm2 like omarchy-settings does, so an ISO built from a local source checkout gives the NVIDIA N1x ASUS ProArt P14 its speakers, headphones and microphones.
…rnel Omarchy installs
…ysctl drop-in omarchy ships etc/sysctl.d/90-omarchy-arp.conf so a dock's Ethernet and Wi-Fi on one network each answer ARP only for their own address. It reaches the packages with the rest of etc/, so back it up like the other sysctl drop-ins, wherever the source ships it, the way 00-omarchy-hooks.conf is handled. An older source without it gets no backup entry for a file it does not ship.
Three patches for the Dell XPS 16, the second N1x laptop, each limited to that machine: - 1011: the EC's battery service answers GetBst with zeros on Dell's firmware, so the battery read empty at 0 V. Fill the status from EC RAM when the reply is empty. - 1012: PSYS and PSOC report the system and SoC power in milliwatts as thermal zones, which showed up at 600-1245 C. Skip them on the XPS 16. - 1013: dell-arm64-hotkeys reports the mic mute key, which Dell's firmware sends as a WMI event, without the x86-only ACPI-WMI core. The XPS 16's EC never raises its FF-A notifications on BIOS 1.2.0, so on that machine the driver polls the EC's event queue every 250 ms. Its mic-mute LED can only be toggled and stays lit across reboots, so the driver tracks it as platform::micmute, which Omarchy's mute script sets, and turns it off at shutdown. pkgrel 10 (6.13 on n1x-xps16). Verified on an XPS 16: battery 100% at 13.1 V, 23 sane thermal zones, the mic mute key mutes and unmutes with its LED in step. [Ryan: cherry-picked onto n1x, where pkgrel is an integer.]
…A's EGL crashes on
…setup from failing units
…ackend is torn down
…inance On the Dell XPS 16 (N1x) the brightness keys moved the OSD and the backlight device, and the panel stayed as it was. RM leaves brightness to the GPU firmware, which switches the panel to its VESA eDP level mode and writes a level; the eDP 1.5 OLED panel has no PWM input and ignores that level. It takes a target luminance in nits instead, which the firmware never sets. 0003 has nvkms set the target luminance after RM, as the kernel's drm_edp_backlight helpers do for i915 and amdgpu, only on internal eDP 1.5 panels without a PWM input that support luminance control. It reads the level back for actual_brightness and sets it again after a modeset powers the panel up. Every other panel stays with the firmware. pkgrel 1.4. The overlay is aarch64-only; GB10 desktops have no internal panel, so nothing changes there.
… has them The recipe installed them unconditionally, so it no longer built from a source without them: today's quattro, an older tag, or the test fixtures. Guard each like omarchy-settings does, and have the runtime-profile test check that a source without them ships none and one with them ships all three.
The N1x GPU firmware's head routing map lets each DP pad-link drive one display, and a USB4 port's two DisplayPort IN adapters share one, so a dock's second display never lit. When the firmware rejects a set of displays, nvkms now asks again with one USB-C DP connector per shared pad-link. Verified on the ASUS ProArt P14 with a CalDigit TS4: an LG at 4K 240 Hz, a Corsair Xeneon Edge and the internal panel all driven from boot.
When a DisplayPort tunnel through a USB4 dock did not finish DPRX negotiation, the connection manager took the N1x host router's DP IN adapter out of use until it reported a hotplug, which these adapters never do. Every display behind the dock on that port then stayed black until reboot. Offer them again when a monitor is plugged in.
When a DisplayPort monitor came back after an unplug while its head was still attached, nvkms revived that head and skipped link training because the link status still read as trained. Behind a USB4 dock on the N1x the DisplayPort tunnel is new by then, and the monitor stayed black while the desktop kept it: whenever the dock was replugged within a few seconds, or a monitor dropped out as it woke. Train the link again after an unplug, and on tunnelled links report a monitor that comes back as unplugged until its old head is shut down, so that the desktop sets a mode on it afresh.
…reload When a monitor was added and removed within moments, Hyprland destroyed its wl_output global at once, and every client that had not bound it yet was disconnected with "global wl_output (N) is unavailable": Xwayland, the shell, the portals, the input method and every app. Plugging a dock into an N1x did this as the dock's monitors came up. Keep a removed output for five seconds before destroying it, as wlroots does. A config reload also ignored a monitor rule's vrr unless something else about the rule changed, so turning VRR on for the internal panel did nothing until the next session. Apply it on reload.
…ecode it Behind a USB4 dock on the N1x, a monitor could come out of standby black while its link read as fully trained: clock recovery, equalisation, symbol lock and alignment on every lane, yet the monitor's symbol error counters were pinned on all of them. Any retrain brought the picture back. After each modeset on a tunnelled link, read the monitor's symbol error counters a second later and, if every lane keeps counting errors, train the link again, up to three times.
… set Take the kernel.org source and the signed shared patches from linux-omarchy-eevdf 7.2.8-5, the EEVDF kernel linux-omarchy split into when it became a meta package for linux-omarchy-bore, and carry the N1x patches forward as 7.2.8-1: - drop 1033 (rt712-sdca PLL2): Omarchy's ASoC fixes now include it - 1000: rebase NVIDIA's EINT wake patch onto 7.2.6's EINT teardown - 1075: keep Omarchy's PCI shutdown host reset next to nhi_remove() - 1032, 1083, 1084: now apply as posted or as in linux-next - 1101, 1102, 1104: rebase onto 7.2.6's hid-asus worker rework None of the N1x fixes is in 7.2.6-7.2.8 stable. The bpftool build now drops the makepkg flags, as linux-omarchy-eevdf does, instead of carrying 9999-bpftool-strip-wformat-bootstrap.patch. config.aarch64 went through olddefconfig on 7.2.8 and takes linux-omarchy's architecture-independent changes since 7.2.5-6: RESET_ATTACK_MITIGATION and SWIOTLB_DYNAMIC off, IMA_WRITE_POLICY and IMA_READ_POLICY on. Built natively, not booted yet.
UPGRADING-7.3.md goes through every N1x patch against 7.3-rc6 and next-20261008: what upstream already has, what needs rework (1094 on the SPI PM macros, 1100 dropped, the MediaTek pinctrl module build, the CS35L56 feedback port change), how NVIDIA's hierarchical _LPI series and the sbsa_gwdt rework would go in, the arm64 s2idle and NVMe changes, the open sleep power items, and a checklist with the verification tests.
…and ADIOS schedulers and its scheduler patch rebases
… the N1x at boot in an interrupt storm
…model on the N1x, so EAS no longer runs the desktop on the A725 cores at their lowest clocks
The Snapdragon X laptops' drivers were already on in the config, but the laptops get no device tree from their firmware and the package built none. It now builds the Windows-on-ARM families' trees and installs them under /boot/dtbs/linux-omarchy-n1x/qcom, under the package's name so that no file is shared with Arch Linux ARM's linux-aarch64. Omarchy's Snapdragon setup lists them for the UKI. Booted on a Lenovo Yoga Slim 7x with the X1 platform drivers named for the initramfs and iommu.passthrough=0, both of which Omarchy's Snapdragon setup adds. Nothing changes for the N1x apart from 6 MB of device tree files.
The config builds the Iris video decoder, which the Snapdragon laptops' device trees enable. Without this fix its firmware can assert when a video starts after the decoder has been idle. It is cc330f558996 from the media tree, queued for 7.4, and was reported and tested on the Yoga Slim 7x. On this kernel it is build-tested.
…agon laptops get cpufreq policies The X1 and X2 CPUs scale through SCMI's performance protocol. scmi-cpufreq and scmi_perf_domain were modules and neither carries an alias, so nothing loaded them: the laptops booted with no cpufreq policies and stayed at the clock the firmware left. Arch Linux ARM's linux-aarch64 builds both in. The N1x and GB10 scale through ACPI CPPC, and sc8280xp through cpufreq-hw, whose modules autoload, so they are unaffected. Validated on a Surface Pro 11 (X1E80100) on this recipe's 7.2.8-4 patch set plus the Surface Pro 11 patches: with the module, no policies and 9.5 W idle; with modprobe scmi-cpufreq, policies on all three clusters and 4.8 W; built in, the policies come up at boot (scmi, schedutil) and idle is 5.1 W, against 5.2 W on linux-sp11, measured the same way. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… at under 1 W (#991) Suspend-to-idle drew 6-15 W on the N1x because the SoC never reached its deepest idle state, Standby-DRIPS: - 1023: the firmware reports DRIPS with an exit latency one below the package state's until PEPD.ELST(1) sets MPEL, so suspend-to-idle never picked it. Add DRIPS's wake latency to those states' exit latency and keep their 0xFFFFFFFF residency, so only suspend-to-idle enters them. (Calling ELST(1) at boot lets the idle governors enter DRIPS with devices running and hangs the boot.) - 1024: in DRIPS the SoC was woken by PCIe every ~60 ms because the NVMe link sat in plain L1 with CLKREQ# asserted. Allow ASPM L1 and PCI-PM L1.1/L1.2 on the NVMe link on N1x systems. On the ASUS ProArt P14 a 7.3-hour night used 0.77 W in a single DRIPS at PPS4, woken only by the lid. pkgrel 6.
|
This completes Spark hardware validation for replacing the DisplayPort patch carry introduced by #601. NVIDIA's 615.78.08 release contains its own unplugged-detach cleanup fix, superseding Martin Stark's now-closed NVIDIA #1359. Publishing the matching pair to ARM edge and verifying normal updates remain the delivery closeout. Tested this PR's NVIDIA pair at Please publish the matching NVIDIA pair together to ARM edge, which still carries 615.71.09. Keep the ARM userspace library correction; its README still describes 615.71.09-1.1 and needs updating. Fresh installation, published-repository updates and system suspend were not tested here. |
… the GPU reservation The N1x firmware reserved 62.5 GiB as the GPU's Windows memory, and patch 1050 let the command line hand most of it back as RAM. NVIDIA and the OEMs agreed on a UEFI variable, EXPOSED_DISP_FRM, that sizes the reservation itself; Omarchy now sets it to 4 GiB on every N1x and stops passing the option. On the ASUS ProArt P14 the firmware then reserves 4 GiB, and the kernel has 512 MiB more RAM than it reclaimed with the patch. Ship after the Omarchy release with the migration that sets the variable: until it runs, a machine on this kernel boots with half its RAM.
Brings in the build tooling master has gained since the branch was cut, among it the rule that sends aarch64 kernel builds to Namespace (#1028): linux-omarchy-n1x built in 24 minutes there against 105 on GitHub's 4-core runner.
linux-omarchy-n1x began as the N1x laptops' kernel and is now the one for the Snapdragon laptops too. It is the same kernel.org source and Omarchy patch set as linux-omarchy on x86_64, so it takes that name: one kernel name on both architectures, and nothing downstream has to know a second one. The recipe stays apart from the x86_64 one, which would gain nothing from carrying patches and a config it never builds. Its folder is linux-omarchy-aarch64, and that is its pkgbase too, because the build and publish tooling matches a build's files to their folder by pkgbase. The packages it makes are linux-omarchy and linux-omarchy-headers. With the name, these follow: - the kernel reports 7.2.8-8-omarchy, as x86_64's does, where it said -omarchy-n1x; - mkinitcpio and Limine name it linux-omarchy, from the modules' pkgbase file; - the Snapdragon device trees are in /boot/dtbs/linux-omarchy/qcom. Both packages replace and conflict with the ones named linux-omarchy-n1x, so a machine that has those moves over on its next update.
…x again, so NVIDIA's power profiles can switch EAS (#1048) 1022 kept cppc_cpufreq from registering its artificial energy model on the N1x's Cortex-X925 and Cortex-A725 cores, so EAS never ran. EAS itself had run unnoticed while schedutil picked the A725 cores' frequency; the desktop lagged once omarchy-power-profiles-daemon's first driver ran balanced in CPPC autonomous selection at EPP 64, which left those cores at 338 MHz through Hyprland's short bursts. NVIDIA's RTX Spark driver for power-profiles-daemon, which replaces that driver, uses the governor as the EAS switch: schedutil in power-saver and in balanced on battery, performance otherwise. Without the model that switch does nothing. With it, in balanced on battery on an ASUS ProArt P14, EAS kept Hyprland on the A725 cores with p99 IPC round trips of 1.7-2.3 ms and no stall over 8 ms; idle draw matched EAS off, and a bursty background load drew 17% less but ran 2.9 times slower. Ship after omarchy-power-profiles-daemon 0.30-2, which brings NVIDIA's driver: under 0.30-1's balanced, EAS brings the lag back.
…on the N1x Resume treated the routers that 1080 keeps awake as reset: it tore down their tunnels and built them again. The USB3 tunnel did not get its bandwidth back, so the dock's xHCI stopped responding, and the DP tunnels waited for plug events that a router kept awake never sends, so the dock's displays stayed dark until it was replugged. 1085 keeps all the tunnels of those routers across suspend-to-idle. Validated with a CalDigit TS4 and two monitors on a ProArt P14.
…ock is unplugged Unplugging a dock with two monitors tore down both DP tunnels about 2.5 ms apart. The N1x GPU firmware handled only the first HPD drop and kept the other monitor connected, so it stayed black after every replug until the next suspend. 1086 waits 100 ms between DP tunnel teardowns on the N1x host router (the dp_teardown_delay module parameter). Six unplugs of a CalDigit TS4 with two monitors, from 5 seconds to 6 minutes, all brought both monitors back.
…ounts a dock's displays nvidia-drm reported the N1x's USB-C outputs as Unknown, which systemd-logind does not count as external displays, so closing the lid on a dock suspended instead of staying docked.
… connected A monitor behind a USB4 dock on the N1x can wake slowly and connect with no modes; when its modes arrive later, reconnect the output so it lights up instead of staying dark.
Publish succeeded → livePackages: aquamarine (aarch64, PR artifact); asus-proart-p14-speaker-firmware (aarch64, PR artifact); cuda (aarch64, PR artifact); hyprland (aarch64, rebuilt at merge); linux-n1x (aarch64, PR artifact); linux-omarchy-aarch64 (aarch64, PR artifact); nvidia-open-dkms (aarch64, PR artifact); nvidia-utils (aarch64, PR artifact); ollama (aarch64, PR artifact); omarchy-settings (x86_64, already published, skipped); omarchy-settings (aarch64, already published, skipped); omarchy-settings-dev (x86_64, already published, skipped); omarchy-settings-dev (aarch64, already published, skipped)
Some planned slots did not run because an earlier slot failed. |
|
Hi @ryanrhughes , thanks for all the great work on NVIDIA N1X Linux support! Thanks again for your work! |
Summary
Packages for the NVIDIA N1x laptops (ASUS ProArt P14 H7407BA, Dell XPS 16 DX16263), used by the omacom/omarchy and omacom/omarchy-iso n1x branches.
linux-omarchyandlinux-omarchy-headersfor aarch64, replacing the earlierlinux-omarchy-n1xnames on upgrade. It carrieslinux-omarchy-bore7.2.8-5's patch set (BORE and ADIOS, shared patches byte-identical) plus the N1x platform: NVIDIA's 1021 SAUCE, the Dell XPS 16 battery, thermal and hotkey patches, the USB4 host router driver and dock fixes (tunnels kept through suspend, DisplayPort teardowns spaced so both monitors of a dock are seen leaving), NVIDIA's 1022.23 device-sleep patches, an s2idle tick-freeze fix, s2idle reaching the platform's deepest idle state (about 0.77 W asleep), HDA/CS35L56/UCSI resume fixes, the Snapdragon laptops' device trees and SCMI cpufreq, and kexec handover off by default. Built locally (skip_build);UPGRADING-7.3.mdis the porting guide for the next major.wl_outputglobal alive while clients can still bind it (a dock's flapping output disconnected every client), and apply a monitor rule'svrron reload.linux-omarchyyet.The hyprland and aquamarine patches reach every aarch64 edge machine, not only the N1x.
Test plan
tests/settings-runtime-profile.shandtests/settings-boot-config.shpass.