The command line the bootloader threw away

the symptom

Three lines, and two of them are good news

This is everything the console said. It looks like a catastrophe and it is very nearly a success — which is the useful thing about reading it carefully rather than reaching for the card reader.

[  OK  ] Finished Create Volatile Files and Directories in the Real Root.
You are in emergency mode. After logging in, type "journalctl -xb" to view
system logs, "systemctl reboot" to reboot, or "exit" to continue bootup.

Cannot open access to console, the root account is locked.
See sulogin(8) man page for more details.

the unit

initrd-find-nixos-closure.service, and its one demand

From nixos/modules/system/boot/systemd/initrd.nix. It reads /proc/cmdline, looks for init=, and if it does not find one it refuses to continue — because without it, stage 1 does not know which of the closures on the disk it is supposed to be booting.

# initrd-find-nixos-closure.service
# RequiresMountsFor = "/sysroot/nix/store"
# requiredBy        = [ "initrd.target" ]

closure=
for o in $(< /proc/cmdline); do
    case $o in
        init=*) … closure="${initParam[1]}" ;;
    esac
done

if [ -z "${closure:-}" ]; then
  echo 'No init= parameter on the kernel command line' >&2
  exit 1
fi

RequiresMountsFor = "/sysroot/nix/store" is what orders it immediately after the sysroot tmpfiles unit, and requiredBy = [ "initrd.target" ] is what turns its exit 1 into emergency mode. The entire failure therefore reduces to one sentence: there was no init= on the kernel command line.

the cause

U-Boot honours /chosen/bootargs only when it feels like it

The SD image derivation writes the command line into the device tree with fdtput -t s … /chosen bootargs, and that step is genuinely required — an earlier boot with no /chosen/bootargs at all was silent on both UARTs. It is also, on its own, useless. Vendored U-Boot rewrites the property on its way past.

// u-boot/board/canaan/common/k230_img.c:110
char *board_fdt_chosen_bootargs(void){
    char *bootargs = env_get("bootargs");
    if(NULL == bootargs) {
        …
        bootargs = "root=/dev/mmcblk1p2 loglevel=8 rw rootdelay=4 \
                    rootfstype=ext4 console=ttyS0,115200  earlycon=sbi";
    }
    printf("g_bootmod = %d, bootargs=%s\n", g_bootmod, bootargs);
    return bootargs;
}

The vendor environment never defines bootargs, so every boot took the second branch. The g_bootmod = 3 line printed on every one of those boots is line 126 of this same function — it had been sitting in the log the whole time, naming the function that was doing it.

the fix

Define the variable, and win the test outright

env_get("bootargs") is checked first, so anything that defines it wins. The interesting part is where the value is allowed to live: the command line names the NixOS closure and therefore changes on every rebuild, while the stage 1 environment image must not, because it is vendored firmware data with a recorded hash.

blinux = k230_set_dtb &&
         ext4load mmc ${mmc_boot_dev_num}:1 0x7000000 /bootargs.txt &&
         env import -t 0x7000000 ${filesize} &&
         ext4load mmc ${mmc_boot_dev_num}:1 0x8000000 /fw_jump_add_uboot_head.bin &&
         ext4load mmc ${mmc_boot_dev_num}:1 0x200000  /${k} &&
         ext4load mmc ${mmc_boot_dev_num}:1 0x8400000 /${dtb} &&
         ext4load mmc ${mmc_boot_dev_num}:1 0x9000000 /initrd.uimg &&
         bootm 0x8000000 0x9000000 0x8400000

mmc_boot_dev_num = 1     k = Image     dtb = force.dtb

corrections

Three things this project believed that the board disproved

Written from the QEMU side before the board had ever run our image, repeated in this repository, and wrong. Each one cost a boot cycle, which is why they are named here rather than quietly edited out.