Build #28301886085 failed with 'curl: command not found' on the URL
resolution step. Upstream Dockerfile only installs wget + python3,
not curl. Use urllib.request.urlopen().geturl() to follow Valve's
redirect and pull the versioned IMG name (e.g. steamdeck-recovery-
4-20250603.5-3.7.10.img.bz2) for the cache key.
Adds `--distro steamos` as a first-class build target. Pattern is the
same as bazzite-deck/batocera: distros/steamos/build-rootfs.sh fetches
the upstream rootfs (Valve's Steam Deck recovery .img.bz2 from
steamdeck-images.steamos.cloud), extracts the rootfs-A + var-A btrfs
slots into $CHROOT, swaps in our linux-ps5 kernel pkg.tar.zst, and
rebuilds the initramfs with the modules amdgpu needs for the PS5
Oberon GPU.
PS5-specific fixes applied to the SteamOS rootfs:
- Replace /usr/bin/steamos-session-select with a shim that writes
/etc/sddm.conf.d/zzz-session-override.conf directly + restarts sddm.
The vendored helper goes through steamosctl / steamos-manager whose
Deck-hardware code paths silently no-op on PS5, leaving the UI stuck
at "Switching to Desktop..." with no actual session swap.
- Install a small Python dbus monitor (steam-session-switch-listener)
that catches org.freedesktop.DisplayManager.Seat.SwitchToUser from
Big Picture's "Switch to Desktop" button and writes the same
override file. No -steamos3 / -steampal needed (those trigger Steam's
A/B update reboot loop on a non-Deck box).
- Drop a Return-to-Gaming-Mode.desktop into /home/deck/Desktop pointing
at steamos-session-select gamescope.
- chown /home/deck to deck:deck (the recovery image leaves it root:root,
which breaks ~/.config writability on first Plasma launch).
- Enable sshd + set deck:deck so the box is reachable for further
iteration without dragging a USB keyboard around.
- Patch /etc/fstab: comment out the by-partsets atomic mount entries
(those paths don't exist on a flat-rootfs flash), append a LABEL-based
/efi mount so the FAT boot partition comes up cleanly. Required for
ps5-iw620-firmware.service to find the WLAN blob.
- Append amdgpu.dpm=0 + amdgpu.gpu_recovery=0 to the kernel cmdline.
modprobe.d alone isn't enough — amdgpu loads before the rootfs conf
is read, so the cmdline is the only reliable path. Without these
Switch-to-Desktop produces a black screen on the PS5 HDMI output.
- Polkit-deny the deck user reboot/poweroff to break Steam's update
reboot loop on first boot.
- Patch steam-jupiter to skip its OOBE rm -rf path. The OOBE assumes
Deck factory layout and wipes /home/deck/.local if it runs.
Verified end-to-end on real PS5 (HW-rev EA-1102): autologin lands in
gamescope, Switch-to-Desktop and Return-to-Gaming both work, sshd is
reachable, ps5-iw620 brings up internal WiFi after firmware staging.
Files:
- distros/steamos/build-rootfs.sh (646 lines)
- distros/steamos/README.md
- distros/steamos/return-to-gaming-mode.desktop
- distros/steamos/steam-session-switch-listener.{py,service}
- build_image.sh: steamos → FORMAT=arch (uses linux-ps5 pkg.tar.zst)
- .github/workflows/build-image.yml: steamos in matrix options + distros list
Cachyos build on minipc-gh-2 failed with "Directory not empty"
because a previous run left a loop device backed by a (deleted) file
in image/work/, and rm -rf can't cross the live mountpoint. Lazy-
umount everything under the workspace + detach matching loops before
the rm so stale mounts don't fail subsequent runs.
5h cold-cache download from o2switch (only mirror, rate-limited to
~250 KB/s), and no persistent cache plumbing yet means every run pays
the full cost. Ship bazzite/bazzite-deck for emulation use cases.
distros/batocera/build-rootfs.sh:211 calls python3 inline to patch
libretroControllers.py with the PS5 HOTKEY-on-gamepad fix. Without it
the batocera build exits 127 at the very end of an otherwise-clean
run.
The mirror rotates older builds out — the pinned 43-20260430 went 404
on its own within ~4 weeks. Scrape the `last/` index for whatever
batocera-x86_64-*.img.gz is there now (currently 43.1-20260529). VER
and BUILD are then derived from the discovered filename so the cache
key + log lines remain informative.
BATOCERA_URL still wins if set (override path unchanged).
CACHEBUST only invalidates the umoci layer; the broken cached layer on
the Oracle runner is upstream (apt-install), so step 10 still tried to
exec /bin/sh from a corrupt parent FS. --no-cache rebuilds every layer
fresh — slower (~5 min apt re-run per build) but reliable on every
runner regardless of buildkit storage state.
One org runner has a broken cached layer for the umoci install (runc:
"/bin/sh: no such file"). Pass a unique build-arg each invocation so
that layer + downstream get rebuilt fresh, while the heavy apt install
above stays cached.
One of the runners has a corrupted buildkit cache for the umoci layer
("/bin/sh: no such file or directory" mid-RUN). --pull refreshes the
ubuntu:24.04 base, which invalidates all downstream cached layers and
forces the umoci step to actually re-execute.
The kernel rpm built by kernel-builder/build.sh now ships mlan.ko / moal.ko
plus modprobe options for the PS5 IW620, so the image side no longer needs
its own out-of-tree build (which was pinned to an old ps5-linux-mwifiex
ref and only applied 1 of 3 patches, breaking against kernel 7.1.x).
Suppress the rpm's modules-load.d autoload so firmware-copy still runs
first via ps5-iw620.service.
git clone https://github.com/... fails with 'server certificate verification
failed. CAfile: none' because ubuntu:24.04 with just 'git' doesn't pull in
ca-certificates. Hit in 'Building NXP IW620 mwifiex driver' stage.
#20 added the distro directories but missed updating the dispatch choice
enum and the 'all' matrix expansion, so the three new distros couldn't
be triggered through CI.
linux-ps5 currently ships no wifi driver — the PS5 onboard wifi is NXP
IW620 (not the upstream mwifiex), which lives at ps5-linux/ps5-linux-
mwifiex as an out-of-tree patch against nxp-imx/mwifiex. Users have to
clone that repo + run install.sh on the target, which needs gcc + kernel
headers and is a manual step every kernel bump.
Build it in the kernel-builder container right after the kernel: clone
both repos, apply ps5-iw620.patch, make against the kernel src tree,
install mlan.ko+moal.ko into /out/staging/lib/modules/$KVER/extra/
ps5-iw620/ with the NXP modprobe options + autoload conf. depmod -b
refreshes the index. All package formats (.deb/.rpm/.pkg.tar.zst) then
include the driver automatically.
MWIFIEX_REPO / MWIFIEX_REF overridable via env in case downstream wants
to test a fork or specific commit.
Steam in -steamos3 mode polls for SteamOS atomic OS updates via
steamos-update, steamos-update-rauc, steamos-atomupd-client. CachyOS
ships none of these (pacman handles rolling updates), so every poll
fires "YieldingApplyUpdateOS: OS update result: 2" and the GamepadUI
shows "Updater apply error: 2: null".
That alone is just a banner, but the first-run OOBE blocks on the
"Software updates" step waiting for a successful update result, so
users can't get past initial setup. Field-reported.
Dropping -steamos3 keeps -gamepadui (controller-friendly UI) without
the SteamOS-mode update polling. OS updates: sudo pacman -Syu.
Polls ps5-linux/ps5-linux-patches for new kernel-X.Y.Z-<sha> tags at
04:00 UTC. If newer than build_image.sh's PATCHES_REF, bumps + pushes
to main; the push fires trigger-builds.yml which rebuilds all images
against the new kernel. Safe-skips if PATCHES_REPO has been forked.
The dispatch in #20 ran build-rootfs.sh without exporting $CHROOT, so
inside the script it was empty — meaning '/' expansions like
`mv "$UNPACK/rootfs"/* "$CHROOT/"` resolved to '/' and the bazzite
rootfs went into the build container's root. Container-contained, but
the build failed with a confusing 'Device or resource busy' on
/etc/resolv.conf.
Two fixes:
- entrypoint exports DISTRO/CHROOT/KVER/ROOT_LABEL/EFI_LABEL when
calling the per-distro script, and asserts $CHROOT non-empty.
- bazzite + batocera scripts have a top-of-file guard that bails if
$CHROOT is unset, missing, or '/'.
Co-authored-by: mia26MAjFm <mia26MAjFm@users.noreply.github.com>
Each distro ships with a self-contained build-rootfs.sh that the entrypoint
calls in place of distrobuilder. This keeps upstream's distrobuilder flow
untouched for the existing distros (ubuntu/arch/cachyos/fedora/etc.) and
isolates the new distros' build logic in their own dirs.
distros/bazzite/
build-rootfs.sh — skopeo+umoci on ghcr.io/ublue-os/bazzite:stable,
promote ostree /usr/etc -> /etc, install linux-ps5 RPM,
mask rpm-ostree services, set up grow-rootfs + DTM-TA
amdgpu reprobe.
image.yaml, grow-rootfs, grow-rootfs.service
distros/bazzite-deck/
Symlinks into distros/bazzite/ — bazzite-deck shares all build logic;
the case branch in build-rootfs.sh dispatches via $DISTRO.
distros/batocera/
build-rootfs.sh — download upstream batocera-x86_64-*.img.gz, mount,
extract /boot/batocera squashfs, swap in linux-ps5
kernel, patch libretroControllers.py, install first-boot
SHARE-partition creator (ps5-share-init).
docker/image-builder/entrypoint.sh
- detect_kver() helper from /kernel-debs/ package filenames
- dispatch to distros/$DISTRO/build-rootfs.sh when present (skips
distrobuilder for that distro)
- EFI assembly detects /boot/efi/ vs /boot/bzImage layouts so batocera
(which mounts FAT at /boot, not /boot/efi, for batocera-part SHARE
detection) works alongside the standard layout
build_image.sh
- IMG_SIZE defaults: bazzite* -> 24 GB, batocera* -> 16 GB
- FORMAT: bazzite* -> rpm (linux-ps5 RPM); batocera* stays deb
CI workflows are NOT modified — the upstream build-image.yml /
trigger-builds.yml continue to apply and will skip these distros unless
explicitly dispatched (they're not in MULTI_DISTROS). Local builds:
./build_image.sh --distro bazzite
./build_image.sh --distro bazzite-deck
./build_image.sh --distro batocera
Co-authored-by: mia26MAjFm <mia26MAjFm@users.noreply.github.com>
Both packagers copied the staged UAPI headers with:
cp -a /out/staging/headers/usr "$DEST/usr"
Because $DEST/usr already exists (modules live under $STAGING/usr/lib/modules,
and other files may be staged under $PKG/usr), cp placed the source directory
inside the destination, producing a nested usr/usr/include/ path. The headers
were therefore installed at the wrong location and not found by builds.
Merge the contents with cp -a .../usr/. "$DEST/usr/" so the headers land at
the expected /usr/include/linux/... path in both .deb and .pkg.tar.zst
packages.
Plain Debian 12 Bookworm with XFCE desktop, LightDM autologin, Firefox
ESR, and standard desktop utilities. Fills the gap between the heavier
Kali (Debian + security tools) and the headless Proxmox (Debian +
hypervisor) — a clean, lightweight Debian desktop for PS5.
Tested on PS5 hardware: XFCE displays over HDMI, boots to desktop.
[skip ci]
The release step now fetches the existing release body and merges the
new distro rows into it, instead of rebuilding the table from scratch.
This means a single-distro build (e.g. --distro fedora) only updates
that distro's row and checksum — all other distros stay in the table.
Previously, every single-distro CI run wiped the other distros from
the release page, requiring a manual API patch each time.
[skip ci]
Debian 12 Bookworm base with Proxmox VE 8.4 (pve-manager, pve-qemu-kvm,
qemu-server). Headless hypervisor managed via web UI at https://<ip>:8006.
Login: root/proxmox.
The PS5 kernel can't satisfy proxmox-ve's hard dependency on a Proxmox
kernel, so the recipe installs an equivs stub (ps5-proxmox-kernel-stub)
that Provides: proxmox-default-kernel before installing proxmox-ve. First
boot auto-bridges the PS5 ethernet into vmbr0 with DHCP. WiFi is not
included (Proxmox VM networking requires wired bridging).
Also wires proxmox into the CI matrix, entrypoint staging, and release
table.
[skip ci]
Ubuntu's apt umoci (0.4.7) crashes on the self-hosted arm64 runner with
"fatal error: lfstack.push invalid packing" while unpacking Fedora's OCI
base image. The image-builder is an amd64 image, so it runs under qemu on
the arm64 runner, and the emulated process inherits the host's 52-bit-VA
high addresses that the old Go in apt's umoci mis-packs in lfstack. Build
umoci v0.5.0 from source with the Go 1.25 toolchain already installed
(Go >=1.21 handles high virtual addresses correctly). Only Fedora uses
umoci (OCI base); the other distros use debootstrap/pacstrap.
[skip ci]
New fedora distro: distrobuilder fedora-http recipe with full GNOME
desktop (GDM autologin), the PS5 IW620 internal WiFi modules built from
nxp-imx/mwifiex + ps5-linux-mwifiex, and an RPM packaging path for the
shared PS5 kernel (docker/kernel-builder-rpm).
Notable PS5-specific fix: the kernel patches write into the
request_firmware() buffer to skip the firmware signature header
(gfx_v10_0_early_init, amdgpu_sdma_init_microcode). Fedora ships
firmware xz-compressed, and the kernel maps xz-decompressed firmware
read-only (fw_decompress_xz_pages -> fw_map_paged_buf -> vmap
PAGE_KERNEL_RO), so the write oopses amdgpu before /dev/dri exists and
the display never comes up. Distros with zstd or uncompressed firmware
decompress into writable buffers, which is why only Fedora was
affected. The recipe therefore ships the cyan_skillfish GPU firmware
uncompressed, materializing linux-firmware's dedup symlinks first
(unxz refuses symlinks).
Also installs umoci/skopeo in the image builder (Fedora 41+ bases ship
as OCI archives) and wires fedora into build_image.sh (FORMAT=rpm) and
the image-builder entrypoint.
Tested on PS5 hardware: GNOME displays over HDMI, internal WiFi scans
and connects.
[skip ci]
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
