Skip to content

ImageBuilder limits for RAM-bootable LibreMesh

tools/ci/build_image.sh produces RAM-bootable LibreMesh images by repacking three artifacts that come out of make image PROFILE=...:

  1. The pre-built kernel binary <profile>-kernel.bin.
  2. The pre-built device tree blob image-<dts>.dtb (FIT only).
  3. A freshly-built rootfs CPIO of the LibreMesh root.

These are wrapped in a FIT (mediatek/filogic, mt7622) or in a legacy IH_TYPE_MULTI uImage (ath79) that the testbed TFTP-boots from RAM. The strategy works only when ImageBuilder ships a KERNEL_INITRAMFS recipe for the profile - otherwise it cannot emit a kernel binary that is RAM-bootable on its own.

This page documents two devices that fall outside that path and the prototype work that was rejected.

Devices not in the matrix

librerouter_librerouter-v1 (ath79/generic, MIPS)

Removed from targets.yml entirely. ath79/generic boots the legacy uImage format and OpenWrt builds a single kernel image with the initramfs CPIO statically linked at compile time via CONFIG_INITRAMFS_SOURCE. The shipped <profile>-kernel.bin already has the upstream OpenWrt rootfs hardcoded into the kernel, and ImageBuilder cannot substitute our LibreMesh CPIO without recompiling the kernel - which it cannot do, because it ships no kernel sources, no compiler and no build infrastructure.

The labgrid YAML libremesh-tests/targets/librerouter_librerouter-v1.yaml is preserved for manual local/remote test runs against a pre-staged *-initramfs-kernel.bin.

linksys_e8450 legacy (mediatek/mt7622, NAND non-UBI)

Replaced in targets.yml with linksys_e8450-ubi, which targets the same Belkin RT3200 hardware via the UBI boot path. The legacy profile in target/linux/mediatek/image/mt7622.mk does not define a KERNEL_INITRAMFS recipe, so linksys_e8450-kernel.bin is never produced under build_dir/.../linux-mediatek_mt7622/. The -ubi profile defines KERNEL_INITRAMFS and is the upstream-recommended boot path.

Prototypes rejected for ath79

Three build paths for librerouter_librerouter-v1 were tried before removing the device from CI:

  1. image_format: multi-uimage (legacy IH_TYPE_MULTI uImage with [kernel.lzma, rootfs.cpio]). The kernel boots, but the LibreRouter U-Boot 1.1.x fork (LibreRouterOrg/u-boot, lib_mips/mips_linux.c) does not propagate sub-image-1 to the MIPS kernel as initrd_start / initrd_size. The kernel cmdline ends up console=ttyS0,115200n8 rootfstype=squashfs,jffs2, the initramfs unpacker never runs, and the device falls through to the on-flash squashfs. CI symptom: root@margarita:/# instead of root@LiMe-XXXXXX:/#.

  2. OpenWrt SDK (ghcr.io/openwrt/sdk:ath79-generic-*). make image inside the SDK fails immediately:

    make[1]: *** No rule to make target 'image'. Stop.

The SDK ships only the package/ subtree and the host toolchain; its target/ and include/image.mk are deliberately stripped. By design the SDK compiles .ipk files against a pre-built kernel; it cannot rebuild the kernel.

  1. ImageBuilder with CONFIG_TARGET_ROOTFS_INITRAMFS=y. Even with the kconfig flag forced, ImageBuilder only emits *-squashfs-sysupgrade.bin. include/image.mk skips the initramfs recipe entirely under $(if $(IB),,...), again because there are no kernel sources to recompile.

The only viable path: full source build

The only path that produces a real RAM-bootable LibreMesh image for librerouter_librerouter-v1 is a full OpenWrt source-tree build (make world) with CONFIG_INITRAMFS_SOURCE pointing at the LibreMesh CPIO. A prototype (build_image_source.sh, ~780 lines) wired this up against v24.10.6 with GHA caching: cold runs took ~50-60 min, warm runs ~10-20 min.

Discarded because:

  • LibreMesh CI testing on ath79 is not a release-blocker.
  • The full source build dominates wall-clock time for the entire workflow even when run in parallel with the other matrix entries.
  • The maintenance surface (toolchain bumps, OpenWrt release tag drift, feed src-link plumbing, libremesh.mk symlink) is significant.

The prototype lives in the git history if it ever becomes worth reviving.

Resurrecting source build for ath79

If anyone resurrects the source-build path, the key inputs are:

  1. Check out openwrt/openwrt at the matching tag (v24.10.6).

  2. make defconfig with:

    CONFIG_TARGET_ath79=y
CONFIG_TARGET_ath79_generic=y
CONFIG_TARGET_ath79_generic_DEVICE_librerouter_librerouter-v1=y
CONFIG_TARGET_ROOTFS_INITRAMFS=y
CONFIG_TARGET_INITRAMFS_COMPRESSION_LZMA=y

  3. Wire pi-lime-packages as a src-link feed and make sure libremesh.mk is reachable from package/feeds/lime_packages/ (otherwise the include ../../libremesh.mk lookups in lime-proto-* / lime-hwd-* silently drop those packages).

  4. make -j$(nproc) world produces bin/targets/ath79/generic/openwrt-*-librerouter_librerouter-v1-initramfs-kernel.bin.

Manual lab runs

For local lab runs against a manually-built artifact:

labgrid-client -p labgrid-fcefyn-librerouter_1 acquire
export LG_PLACE=labgrid-fcefyn-librerouter_1
export LG_ENV=targets/librerouter_librerouter-v1.yaml
export LG_IMAGE=/srv/tftp/firmwares/librerouter_librerouter-v1/libremesh/<artifact>.bin
uv run python -m pytest tests/test_libremesh.py tests/test_base.py tests/test_lan.py -v
labgrid-client -p labgrid-fcefyn-librerouter_1 release

See also

  • CI: firmware build pipeline - the high-level pipeline that consumes BUILD_INITRAMFS=1 artifacts.
  • tools/ci/build_image.sh - the manual mkimage repack flow.
  • .github/ci/targets.yml - build_initramfs / test_firmware keys.