reterminal-dm4/emmc-provisioning/docs/NETWORK-BOOT-TROUBLESHOOTING.md

Network boot troubleshooting: no DHCP/TFTP during boot, only after OS is up

If you run tcpdump during power-on but see no DHCP/TFTP traffic during boot, and only see traffic after the device has booted to the OS, the reTerminal is almost certainly not on the same L2 segment as the LXC's eth1.

What’s going on

• The Pi’s bootloader (EEPROM) sends DHCP Discover on the Ethernet port when it tries network boot.
• That request only reaches interfaces on the same VLAN / same bridge (same cable/switch segment).
• dnsmasq in the LXC listens only on eth1 (provisioning LAN).
• If the reTerminal is plugged into the main office LAN (or the same segment as the LXC’s eth0), the netboot DHCP never reaches eth1 — so you see no DHCP/TFTP on eth1 during boot.
• After the OS boots, it uses the same Ethernet port and gets an IP from the main LAN; you then see traffic (e.g. on eth0 or from the device’s new IP). That’s why you only see traffic “after the device boots to OS”.

What to do

1. Confirm which interface sees the boot-time DHCP

On the LXC, run tcpdump on both interfaces in two terminals (or run one in background):

# Terminal 1: provisioning LAN (where netboot should happen)
tcpdump -i eth1 -n -e port 67 or port 68 or port 69

# Terminal 2: WAN / main LAN
tcpdump -i eth0 -n -e port 67 or port 68 or port 69

Then power off the reTerminal and power it on. Watch where DHCP (and TFTP) appear:

• If you see DHCP only on eth0 during boot → the reTerminal is on the same segment as eth0, not eth1. So netboot is not using your LXC’s dnsmasq; the device may get an IP from another DHCP server and fall back to eMMC boot.
• If you see DHCP on eth1 during boot → the reTerminal is on the provisioning segment; you should then see TFTP (port 69) as well.

2. Fix: put the reTerminal on the same segment as eth1

• The reTerminal’s Ethernet cable must be connected to the provisioning segment: the same VLAN or bridge as the LXC’s eth1 (e.g. 10.20.50.0/24).
• On Proxmox, eth1 is often on a dedicated bridge (e.g. vmbr1). The reTerminal must be plugged into a switch port that belongs to that same bridge/VLAN.
• If you have one physical switch: either put the LXC’s eth1 and the reTerminal in the same VLAN, or use a dedicated “provisioning” port group / switch.

3. Sanity check: same port as reTerminal

• Plug a laptop (or another device) into the same port (or same VLAN) as the reTerminal.
• Run: sudo dhclient -v <interface> (or let it get DHCP automatically).
• If you get an IP in 10.20.50.x → that segment is your provisioning LAN (eth1); the reTerminal should netboot from there.
• If you get a different range (e.g. 192.168.x.x) → that segment is not the provisioning LAN; move the reTerminal’s cable or VLAN to the segment where 10.20.50.x is served.

Summary table

Symptom	Likely cause	Action
No DHCP/TFTP on eth1 during boot; traffic only after OS	reTerminal on different segment than eth1	Plug reTerminal into same VLAN/bridge as LXC eth1 (provisioning LAN)
DHCP on eth0 during boot, none on eth1	reTerminal on same segment as eth0	Move reTerminal to provisioning segment (same as eth1)
No DHCP on any interface during boot	Cable unplugged, BOOT_ORDER not 0xf21, or device not attempting netboot	Check cable, confirm BOOT_ORDER=0xf21, power cycle with cable in before power

---

I only see DHCP Request/Reply, and the client already has 10.20.50.x

If your tcpdump on eth1 shows something like:

0.0.0.0.68 > 255.255.255.255.67: BOOTP/DHCP, Request from 88:a2:9e:xx:xx:xx
10.20.50.1.67 > 10.20.50.147.68: BOOTP/DHCP, Reply

that is not the bootloader — it is the OS DHCP client (renewal or re-request). The client already has 10.20.50.147, so this happens after the device has booted to the OS.

• Bootloader (network boot): sends DHCP Discover (client 0.0.0.0, no IP yet), then you see Offer, Request, Ack, then TFTP (port 69) for start4cd.elf, kernel, etc.
• OS: sends DHCP Request (renew/rebind, often already with an IP or requesting a known one), then Reply — no Discover, no TFTP.

So the device is on the right segment (eth1, 10.20.50.x). The problem is that you are not seeing the bootloader’s DHCP/TFTP during the first seconds after power-on.

What to do:

1. Start tcpdump before power-on
   Run tcpdump -i eth1 -n -e port 67 or port 68 or port 69 on the LXC, then power off the reTerminal, wait a few seconds, and power it on. Capture from the first second. Look for:
   • Discover (client 0.0.0.0 → broadcast) at the very start → that’s the bootloader.
   • TFTP (port 69) right after DHCP Ack → bootloader loading files.
2. If you never see Discover or TFTP, only Request/Reply after the OS is up, then the bootloader is either not attempting network boot or is giving up (e.g. link not ready, timeout) and booting from eMMC. Try a full power-off (mains or PSU), wait 10 s, then power on with tcpdump already running.
3. Confirm BOOT_ORDER=0xf21 on the device (eMMC first, then network) and that Ethernet is connected before power-on.

---

reTerminal DM: serial console vs USB boot (rpiboot)

The serial console is not on the same USB as rpiboot.

Port / interface	Purpose
USB Type-C (next to boot-mode switch)	Power, and rpiboot when eMMC is disabled (USB device mode). No serial console here.
40-pin GPIO header (UART)	Serial console. Use a USB‑to‑serial adapter; connect its RX to GPIO 14 (Pin 8), GND to GPIO 15 (Pin 10) or any GND.

Baud rate:

• Bootloader (BOOT_UART=1): use 115200 8N1. This is the Pi EEPROM/bootloader debug output (network boot attempts, DHCP, TFTP, errors).
• OS serial login: some Seeed docs use 9600 for getty; many Pi images use 115200. If you only care about bootloader messages, use 115200.

So: use the same USB‑C cable only for power and rpiboot. For serial console, use a USB‑to‑serial adapter on the GPIO header at 115200 to see bootloader output.

---

Serial shows "Boot mode: SD (01)" and no network attempt

If the bootloader serial output shows something like:

Boot mode: SD (01) order 2

and you never see a line about network (e.g. "Trying DHCP", "TFTP", or "Boot mode: NET (02)"), then the bootloader is not attempting network boot for this boot. It goes straight to SD/eMMC (01). That matches “no DHCP during boot, only after OS”.

Possible causes:

1. BOOT_ORDER not applied or not read
   From the running OS, confirm:
   sudo vcgencmd bootloader_config
   and check that BOOT_ORDER=0xf21 (and optionally NET_BOOT_MAX_RETRIES, DHCP_TIMEOUT, TFTP_IP). If you see different or missing values, the EEPROM config in use at boot may be different (e.g. old EEPROM, or update not applied on cold boot).

2. Network tried but failed before any DHCP
   The bootloader may try network, fail very early (e.g. no link, or timeout before sending DHCP), then fall back to SD without printing a “Trying network” line. Slower link-up (switch, cable) can cause this. Increasing DHCP_TIMEOUT and NET_BOOT_MAX_RETRIES (and setting TFTP_IP) gives the best chance.

3. CM4 / carrier quirk
   On some CM4 carriers the bootloader may skip or shorten the network attempt. Serial is the only way to see what it actually does; if you never see any network-related line, treat it as “network not attempted” for that boot.

What to try:

• Re-apply EEPROM config with network first and timeouts (as in NETWORK-BOOT-TROUBLESHOOTING), then full power cycle (unplug power 10+ s, then power on) with serial connected. Watch from the first character for any “NET”, “DHCP”, “TFTP” or “order” line.
• For a one-off test you can set BOOT_ORDER=0x2 (network only). If network fails, the device won’t boot (no fallback to SD). Use only to confirm whether the bootloader tries network and what it prints; then set back to 0xf21. If the full serial log never shows "NET", "DHCP", or "TFTP" and goes straight to "Boot mode: SD (01) order 2", trying BOOT_ORDER=0x2 (network only) once will force a network attempt and should produce DHCP/TFTP messages on serial.

---

Boot stops after start4.elf ("PCI0 reset" then nothing)

What’s actually going on

The EEPROM bootloader only does TFTP for config.txt, start4.elf, and fixup4.dat. It then starts the GPU firmware (start4.elf) and stops the network. The kernel and initrd are loaded by the GPU firmware, not by the EEPROM: after “Starting start4.elf”, the GPU is supposed to bring the network back up and TFTP kernel8.img, cmdline.txt, and initrd.img. If you never see TFTP for kernel8.img/initrd.img and the log stops at “PCI0 reset”, the GPU stage is not doing that. Common causes:

1. Config not seen by the GPU — The config the EEPROM fetched (e.g. from 0d1ddbda/config.txt) must contain kernel=kernel8.img and initramfs initrd.img followkernel. If that file was a symlink or truncated, the GPU may not see those lines. Use a real copy of the full config in the serial dir (see ensure script below).
2. No visibility into the GPU — The EEPROM logs stop at “PCI0 reset”; the next step is inside the GPU firmware. To see GPU messages (e.g. network bring-up, TFTP, or errors), add uart_2ndstage=1 to config.txt so the GPU logs to the UART. Then power-cycle and watch for lines like MESS:... genet: LINK STATUS or TFTP activity.
3. Firmware/board quirk — On some boards or firmware versions the GPU netboot path can fail silently. Ensuring the latest Pi 4/CM4 boot files in the TFTP root and trying start4cd.elf + fixup4cd.dat (or leaving defaults) is worth a try.

If the serial log shows TFTP for config.txt, start4.elf, fixup4.dat, then "Starting start4.elf", "Stopping network", "PCI0 reset", and no TFTP requests for kernel8.img or initrd.img, use the checks below.

Fix on the LXC: ensure /srv/tftpboot/config.txt contains (and that 0d1ddbda/config.txt is a real copy with the same content):

enable_uart=1
kernel=kernel8.img
initramfs initrd.img followkernel
uart_2ndstage=1

enable_uart=1 is required for the kernel serial console when netbooting (otherwise the firmware can set 8250.nr_uarts=0). uart_2ndstage=1 makes the GPU firmware log to the UART so you see MESS: lines after "PCI0 reset" (e.g. network bring-up, TFTP, or errors).

You can run:

# On the LXC (or from your machine)
ssh root@<LXC-IP> 'bash -s' < emmc-provisioning/scripts/ensure-tftpboot-config-kernel-initrd.sh

Also ensure the TFTP root has kernel8.img and initrd.img (and the serial subdir has symlinks or copies). Then power-cycle the device; you should see TFTP_GET for kernel8.img and initrd.img, then the kernel and initramfs (e.g. rescue shell or provisioning client) run.

If it still stops after “PCI0 reset”:

• Add uart_2ndstage=1 to the TFTP config.txt (root and serial copy). Re-run the ensure script so the serial dir gets the updated config, then power-cycle. Watch the serial log for MESS: lines from the GPU (e.g. genet: LINK STATUS, TFTP, or errors). That shows whether the GPU is bringing the network up and trying to load the kernel.
• On the LXC, confirm the config the device gets has the right size and content:
  ssh root@<LXC-IP> 'wc -c /srv/tftpboot/0d1ddbda/config.txt && grep -E "kernel|initramfs|uart_2ndstage" /srv/tftpboot/0d1ddbda/config.txt'

---

Kernel loads but serial stops at "Baud rate change done" (no rescue shell)

If you see the GPU load kernel8.img and initrd.img, then "Baud rate change done..." and nothing else (no rescue shell, no kernel messages), the kernel is likely hanging very early because of a missing or invalid Device Tree. The GPU log may show dterror: Failed to load Device Tree file '?'.

The GPU loads files from the serial-prefix dir (e.g. 0d1ddbda/). If the .dtb files (e.g. bcm2711-rpi-cm4.dtb, bcm2711-rpi-cm4-io.dtb) are only in the TFTP root and not in that dir, the firmware can fail to load the right DTB and the kernel gets no valid device tree.

Fix: Ensure the TFTP root has the Pi 4/CM4 DTB files (from the Raspberry Pi firmware boot/ folder) and that each serial-prefix dir has symlinks to them. Re-run the ensure script (it now links *.dtb into each serial dir):

ssh root@<LXC-IP> 'bash -s' < emmc-provisioning/scripts/ensure-tftpboot-config-kernel-initrd.sh

If the TFTP root has no *.dtb files, populate it from the Pi firmware (e.g. run populate-tftpboot-from-git.sh or copy bcm2711-rpi-cm4.dtb, bcm2711-rpi-cm4-io.dtb, and other bcm2711*.dtb from the firmware repo into /srv/tftpboot), then run the ensure script again and power-cycle the device.

Serial stops at "Baud rate change done" (no kernel/initramfs output): On Pi 4/CM4 netboot, the firmware can force 8250.nr_uarts=0, which disables the kernel serial driver so you get no console after the GPU handoff (raspberrypi/firmware#1575). The workaround is enable_uart=1 in config.txt (within the first 4KB). The ensure script adds it; re-run the script so the root and serial-prefix configs have it, then power-cycle. Keep serial at 115200 baud.

---

TFTP "file .../SERIAL/start4.elf not found" — serial-number prefix

The Pi bootloader may request files under a path named after the board serial number (e.g. 0d1ddbda/start4.elf). If the TFTP root has no such subdirectory, those requests fail and the bootloader falls back to the root (e.g. start4.elf). To avoid "not found" for the first requests, on the LXC create the serial directory and symlink the boot files:

# On the LXC (replace 0d1ddbda with your Pi's serial from vcgencmd or serial output)
mkdir -p /srv/tftpboot/0d1ddbda
cd /srv/tftpboot/0d1ddbda
for f in start4.elf start4cd.elf start.elf fixup4.dat fixup4cd.dat config.txt cmdline.txt kernel8.img initrd.img; do
  [ -f ../$f ] && ln -sf ../$f $f
done

After that, the bootloader’s first TFTP requests succeed. The device already had this directory created for serial 0d1ddbda.

---

Stuck in network-only boot (BOOT_ORDER=0x2): get back to Raspbian and change boot order

If you set BOOT_ORDER=0x2 (network only) for testing, the device will never try eMMC. To fix the EEPROM config, use rescue mode: the network boot chain loads the Alpine-based provisioning initramfs which includes Python and rpi-eeprom-config; with a special kernel cmdline it drops to a shell so you can run rpi-eeprom-config directly from the initramfs (no chroot into eMMC needed).

Prerequisites

• Initramfs with rescue support — Build the Alpine-based initramfs (it includes /rescue-eeprom.sh, rpi-eeprom-config, and EEPROM firmware) and copy it to the LXC TFTP root and into the serial dir:

  cd emmc-provisioning/network-boot-initramfs && ./build.sh
  scp initrd.img root@<LXC-IP>:/srv/tftpboot/
  ssh root@<LXC-IP> 'cp /srv/tftpboot/initrd.img /srv/tftpboot/0d1ddbda/ 2>/dev/null || true'
  

  Building requires Docker or Podman with arm64 emulation (qemu-user-static).
• TFTP config — Ensure /srv/tftpboot/config.txt (and thus 0d1ddbda/config.txt if it’s a symlink) has kernel=kernel8.img and initramfs initrd.img followkernel so the full kernel+initrd chain runs.

Steps

1. On the LXC, enable rescue for this device by serving a cmdline that includes provisioning_rescue=1. The Pi loads 0d1ddbda/cmdline.txt; replace that with a real file (not a symlink) so this device gets the rescue cmdline:

   # On the LXC (replace 0d1ddbda with your Pi serial if different)
   CD="/srv/tftpboot/0d1ddbda"
   rm -f "$CD/cmdline.txt"
   # Same as root cmdline plus rescue flag (one line, space-separated)
   cat /srv/tftpboot/cmdline.txt | tr '\n' ' ' > "$CD/cmdline.txt"
   echo -n ' provisioning_rescue=1' >> "$CD/cmdline.txt"
   echo >> "$CD/cmdline.txt"
   

2. Power on the reTerminal (or reboot). It will network boot, load kernel + initramfs, and rescue mode will start a shell (serial or console). You should see:
   === RESCUE MODE (provisioning_rescue=1) ===

3. In the rescue shell, run the rescue script. It automatically sets BOOT_ORDER=0xf21 and writes the EEPROM update to the eMMC boot partition:

   /rescue-eeprom.sh
   

   The script runs rpi-eeprom-config directly from the initramfs (no chroot, no dependency on the eMMC OS). It creates a pieeprom.upd file on the eMMC boot partition with the updated config. For manual editing instead, use /rescue-eeprom.sh --edit.

4. Disable network boot and reboot — The EEPROM update is only applied when the bootloader boots from the same storage where the update file was written. You wrote it to eMMC, so the bootloader must boot from eMMC once to apply it. With BOOT_ORDER=0x2 (network only) the next reboot netboots again, so the bootloader never reads eMMC and the update is never applied. Do this before rebooting from the rescue shell:

   • On the LXC, disable PXE so the next boot does not advertise TFTP:
     ssh root@<LXC-IP> '/opt/cm4-provisioning/toggle-network-boot-dhcp.sh disable'
   • Then power cycle the reTerminal (or run reboot -f / echo b > /proc/sysrq-trigger in the rescue shell). The bootloader will get DHCP without option 66/67; it may then try eMMC (depending on firmware) and apply the update. If it still netboots (e.g. cached TFTP), unplug the Ethernet cable and power cycle so it has no choice but eMMC.

5. After you are back in Raspbian, restore normal cmdline for the device so the next network boot runs the provisioning client, not rescue:

   ./emmc-provisioning/scripts/disable-rescue-cmdline-on-lxc.sh root@<LXC-IP> 0d1ddbda
   

   Or on the LXC: rm -f /srv/tftpboot/0d1ddbda/cmdline.txt && ln -s ../cmdline.txt /srv/tftpboot/0d1ddbda/cmdline.txt

Why did my boot order not change? The update file was written to the eMMC boot partition. The bootloader applies it only when it boots from that partition. When you rebooted, the device netbooted again (TFTP), so the bootloader read the “boot” files from the network, not from eMMC, and never saw or applied the update. Disable PXE (and optionally unplug Ethernet) before rebooting so the next boot is from eMMC and the update is applied.

See also NETWORK-BOOT-LXC.md for setup and monitoring.