Debian sid on a Lenovo Yoga C630: Held GRUB, One-Time Boots and the IPA Reset

This article provides a step by step guide to running Debian sid on a Lenovo Yoga C630, the Snapdragon 850 Windows-on-ARM laptop from 2018. A small set of POSIX shell scripts is built to keep the boot recoverable while the kernel, firmware and desktop change underneath it.

https://github.com/xbill9/linux-arm-adventures

Kernel 7.2.9, GNOME on Wayland, GPU acceleration, audio, Wi-Fi, Bluetooth, video decode, battery and both USB-C ports all work. LTE data is the one thing blocked, because loading the modem's IPA driver resets the whole SoC.

What Is the Problem?

The C630 boots Linux through a firmware built for Windows. It hands the kernel ACPI tables that Linux cannot use, it cannot store EFI variables, and its real-time clock is never set. Each of these breaks a different step of a normal Debian install.

The kernel side is in good shape. Debian's 7.2 kernel ships the C630 device tree, the embedded controller driver (lenovo_yoga_c630) and the USB-C driver (ucsi_yoga_c630). The hard parts are the boot loader, the proprietary firmware and one driver.

At This Point You Should Have…

• A Lenovo Yoga C630 (model 81JL) with Debian installed from a C630 installer stick that carries a shim, GRUB 2.12 and the device tree.
• A second machine on the same network, with an SSH key already installed on the C630.
• A USB Ethernet adapter. Wi-Fi needs firmware that Debian does not ship.

Step 1 — Find the Boot Loader the Firmware Runs

This laptop has two EFI system partitions on its UFS disk:

The firmware has no NVRAM boot entries and efibootmgr reports EFI variables are not supported. It boots only the removable path, EFI/BOOT/BOOTAA64.EFI, on partition 2.

The way to know which loader runs is to make every loader say its own name. Each grub.cfg stub sets a variable before handing over to the shared menu:

search.fs_uuid a81deabf-0654-4f68-b795-284305d26f13 root
set prefix=($root)'/boot/grub'
set c630_from="sda2 EFI/debian"
export c630_from
configfile $prefix/c630.cfg

Every menu entry title starts with [via $c630_from]. The menu on screen read [via sda2 EFI/debian], which settles it.

🔎 Tip: The Installer Stick Has the Same Trap

From the stick, the firmware boots the ISO's internal efi.img. Editing the stick's partition 2 changes nothing on the next boot.

Step 2 — Hold GRUB 2.12

A trixie to sid upgrade installed GRUB 2.14 and a new shim. After it, every kernel went black the instant GRUB handed over, including a 6.12 kernel, with no output even with ignore_loglevel earlycon=efifb.

GRUB is the cause, and the evidence is that nothing else changed:

• vmlinuz-6.12.107 on disk is byte-identical to the stick's kernel, which boots.
• The device tree is byte-identical to the stick's.
• The 6.12.107 initramfs was built the day before the upgrade and was untouched.

The same three files fail under GRUB 2.14 and boot under GRUB 2.12. Why 2.14 fails on this firmware is not known. Its newer Linux EFI handoff is a suspect, not investigated.

The fix is to copy the stick's shim and GRUB 2.12 back onto /boot/efi (scripts/apply.sh) and hold every package that could replace them:

$ apt-mark showhold
grub-common
grub-efi-arm64
grub-efi-arm64-bin
grub-efi-arm64-signed
grub-efi-arm64-unsigned
grub2-common
shim-helpers-arm64-signed
shim-signed
shim-signed-common
shim-unsigned

The held packages are at 2.14-4. They stay installed, and the hold stops them reinstalling themselves over the 2.12 binary on the ESP.

Step 3 — Generate a Boot Menu That update-grub Does Not Touch

update-grub still writes /boot/grub/grub.cfg, and nothing boots from it. What boots is /boot/grub/c630.cfg, written by c630-menu (c630/scripts/c630-menu). For every installed kernel it writes four entries, each with that kernel's own device tree:

A CONTROL entry boots the installer stick's own kernel and installer from the disk. If it reaches the language screen, the boot loader and DTB are fine and the fault is in the kernel or initramfs of the entry that failed.

The menu regenerates itself from hooks in /etc/kernel/postinst.d, /etc/kernel/postrm.d and /etc/initramfs/post-update.d, and it refuses to write a menu whose default kernel has no entry.

Every entry carries the same kernel options. This is the default entry as booted:

$ cat /proc/cmdline
BOOT_IMAGE=/boot/vmlinuz-7.2.9+deb14-arm64 root=UUID=a81deabf-0654-4f68-b795-284305d26f13 ro clk_ignore_unused pd_ignore_unused regulator_ignore_unused usbcore.autosuspend=-1 modprobe.blacklist=ipa quiet

clk_ignore_unused pd_ignore_unused regulator_ignore_unused stop Linux switching off clocks, power domains and regulators the device tree does not claim. The keyboards locked up during the install before regulator_ignore_unused usbcore.autosuspend=-1 were added, and none of the 22 boots since has shown a lockup. That cause is likely, not reproduced.

Use the root UUID, never a device name. The UFS has shown up as sda, sdb and sdd on different boots.

Step 4 — Test Every Risky Change With a One-Time Boot

A test boot that hangs must come back to a working default on the next power cycle. GRUB 2.12 can write grubenv at boot, so the menu clears the one-time entry before booting it:

if test -s $prefix/grubenv; then
  load_env
  if test -n "$next_entry"; then
    set default="$next_entry"
    set next_entry=
    set once="$c630_once_args"
    export once
    set c630_once_args=
    save_env next_entry c630_once_args
  fi
fi

$once is appended to every linux line, so a one-time boot can also carry extra kernel options:

sudo c630-menu --once c630-7.2.9+deb14-arm64 "modprobe.blacklist=qcom_q6v5_mss"

A SoC reset during that boot recovers by itself, because the watchdog resets the chip and GRUB boots the default.

Step 5 — Move to Kernel 7.2

Debian's 6.12 kernel boots, but it has no battery or USB-C drivers and it logs four clk_core_disable warnings from the display driver on every boot. The 7.2.9 kernel from sid has neither problem:

7.2.9 is the default in /etc/default/c630-menu. 6.12 stays in the menu as the fallback.

Step 6 — Install GNOME on sid

gnome-core and Google Chrome's arm64 package install with scripts/gnome-install.sh. Two sid problems need fixing alongside:

• PAM: sid's libpam 1.7.0 does not read /usr/lib/pam.d, so systemd-user and polkit-1 fall back to the generic rules. Copy both files to /etc/pam.d/, and check them again after a systemd or polkitd upgrade.
• Plymouth: gnome-core pulls it in, and it changes the initramfs. Pin it out with Pin-Priority: -1 in /etc/apt/preferences.d/no-plymouth.

$ gnome-shell --version
GNOME Shell 50.5

Step 7 — Fix the Bluetooth Address

The WCN3990 Bluetooth controller comes up with no public address, because none is provisioned. c630-bt-addr derives a stable, locally administered one from /etc/machine-id. hci0 appears before the chip finishes loading its firmware, so the script retries every 0.5 s for up to 30 s:

for i in $(seq 1 60); do
  if btmgmt -i hci0 public-addr "02:$h" 2>&1 | grep -q "complete"; then
    echo "hci0 public address 02:$h set after $i tries"; exit 0
  fi
  sleep 0.5
done

Once it has run, the controller reports the address as public:

$ bluetoothctl show
Controller 02:BF:6F:61:DD:AB (public)
    Powered: yes

A systemd unit, c630-bt-addr.service, runs it at boot.

Step 8 — Install the Lenovo Firmware

The device tree asks for eight signed files under qcom/sdm850/LENOVO/81JL/. Debian does not ship them. They come from Lenovo's Windows driver release 200.0.19.0, mirrored by WOA-Project, in three cabinet files that 7z x unpacks with no Windows needed:

The files are proprietary, so the repository records only their sha256 sums and source URL.

The GPU firmware must be in the initramfs. msm loads before the root filesystem mounts, and sid's initramfs-tools copies no qcom/ firmware. An initramfs hook (c630/initramfs/c630-firmware) adds qcom/a630_sqe.fw, qcom/a630_gmu.bin and the zap shader.

Mask the userspace pd-mapper.service. Debian has no .jsn maps for this SoC, so it fails. The in-kernel qcom_pd_mapper in 7.2 does the job.

Step 9 — Bisect the Reset Down to One Driver

The first boot with all the firmware reset the SoC about 6 s in, right after systemd-udevd started. Each step below is one c630-menu --once boot that blocks a different set of drivers:

IPA is the hardware data path between the LTE modem and Linux. Every normal entry now carries modprobe.blacklist=ipa, and the modem itself still runs, because Wi-Fi depends on it.

Why Does IPA Reset the SoC?

modprobe ipa on a running system fails the same way, only slower. The cause is in the driver and its firmware, at any point in the boot:

[1361.110] ipa 1e40000.ipa: IPA driver initialized
[1364.556] ipa 1e40000.ipa: GSI command 0 for event ring 2 timed out, state 1
[1364.648] [dpu error]enc35 frame done timeout
[1371.361] Error sending AMC RPMH requests (-110)
[1372.024] watchdog: CPU1: Watchdog detected hard LOCKUP on cpu 2

The driver comes up. About 3 s later the GSI, IPA's DMA engine, stops answering commands. The display stalls, then the RPMh resource manager stops answering, and the whole SoC hard-locks. The watchdog resets it, and the default entry is back up in about 50 s.

Setting qcom,gsi-loader = "modem" in a test device tree stops the lockup. In that mode Linux waits for the modem to load the GSI firmware, and the Lenovo modem firmware never does, so no rmnet_ipa0 interface appears. Safe, but no data.

Two hypotheses are untested. The Lenovo ipa_fws.elf may be a different build from the one the upstream driver was written against, and postmarketOS ships a different IPA blob that is a candidate. Or an interconnect or power vote for IPA may be missing from the device tree.

🔎 Tip: Make a Crash Test Cost Nothing

The watchdog reset takes several seconds, so the kernel log of the failed boot survives in journalctl -b -1 -k. Flush and freeze the filesystem before a test that may lock up:

sudo journalctl --flush; sync; sudo sh -c 'fsfreeze -f / && fsfreeze -u /'

Step 10 — Get Wi-Fi Onto 5 GHz

wlan0 is the WCN3990 on ath10k_snoc. At boot the regulatory domain is the world default, 00, and an access point on channel 112, a DFS channel, refused the association:

wlan0: ae:17:d8:07:9a:8d denied association (code=18)

The same network on 2.4 GHz connected at once. cfg80211 is a module loaded after the initramfs, so one modprobe option sets the country with no initramfs rebuild:

options cfg80211 ieee80211_regdom=US

That file is /etc/modprobe.d/regdom.conf. Not yet verified across a reboot.

The Wi-Fi MAC address is random on every boot (invalid MAC address; choosing random), so NetworkManager uses a stable MAC per network (wifi.cloned-mac-address=stable). A captive portal login should then survive a reboot. Not yet verified.

What Works?

$ glxinfo -B | grep -E 'renderer|Accelerated'
    Accelerated: yes
OpenGL renderer string: FD630

Firmware Quirks Worth Knowing

How Do You Recover When It Does Not Boot?

The runbook (c630/RECOVERY.md) keeps typing on the laptop to two lines:

1. Boot the installer stick, choose Rescue, pick the root filesystem and open a shell in it.
2. Type mkdir /run/sshd and /usr/sbin/sshd.

Everything else happens over SSH from the second machine. Before rebooting out of rescue, check every file the boot depends on as GRUB will see it. GRUB cannot replay the ext4 journal, so debugfs -R 'stat /boot/grub/c630.cfg' reads the raw disk, and grub-script-check parses the menu.

Summary

The goal of this article was to run a current Debian with a full desktop on the Yoga C630 without losing the ability to boot. The key to the solution was a boot path that stays fixed while everything above it changes: a held GRUB 2.12, a generated menu with a SAFE entry per kernel, and one-time boots for every risky change. The results were:

• 🟢 Debian sid, kernel 7.2.9 and GNOME Shell 50.5 on Wayland, with GPU acceleration, audio, Wi-Fi, Bluetooth, video decode, camera, battery and USB-C.
• 🟢 One-time boots recover by themselves from a SoC reset, which turned a firmware install into a bisect.
• ❌ GRUB 2.14 blacks out every kernel on this firmware. The cause inside GRUB is unknown.
• ❌ Loading IPA resets the SoC, so LTE data is blocked.
• ⚠️ Speakers, suspend and the 5 GHz fix across a reboot are not yet tested.

Scope: one C630 (81JL, BIOS 9UCN34WW), Debian sid as of 2026-10-10, Lenovo firmware release 200.0.19.0, tested in one day, with no SIM in the modem.

The strategy for bringing up Debian on the Yoga C630 was validated with an incremental step by step approach.

References

• Repository: https://github.com/xbill9/linux-arm-adventures
• Lenovo C630 firmware mirror: WOA-Project/Qualcomm-Reference-Drivers
• postmarketOS C630 firmware: sdm845-mainline/firmware-lenovo-yoga-c630
• Arch Linux ARM packages for the C630: rigoorozco/yoga-alarm
• Earlier C630 firmware notes: aarch64-laptops/build#62