ZLT X28 5G CPE (MT6890): custom OpenWrt 23.05 running as my daily router, heading for a stable release, need advice on publishing and the wiki page

Hi all,

I'm building OpenWrt for the ZLT X28, a 5G CPE with a MediaTek MT6890 (4x Cortex-A55). It's at the point where I use it as my actual router, so I want to introduce it here and ask a few things before I publish it the wrong way.

What it is

OpenWrt 23.05.5 userspace (armsr/armv8, fw3 with iptables-legacy) on the vendor 4.19.205 kernel. There's no mainline support for this SoC, so this isn't a clean port. The idea is a normal OpenWrt box that you manage in LuCI and UCI, not a copy of the vendor web UI.

What's done (all on my own unit)

  • Boots from slot A of the A/B layout. Slot B (stock) is untouched for now, so there's still a way back.
  • 5G data. The modem is a normal netifd interface (proto x28_mipc) with its own LuCI protocol page. IPv4, IPv6, IPv4v6, and the APN comes from the SIM if you leave it empty. The IPv6 prefix goes to the LAN.
  • LAN switch (MT7531), DHCP and DNS, firewall and NAT. A Linksys router now sits behind it as a client.
  • Wi-Fi on the MT7915 dual band through the vendor driver, with a netifd handler so it shows up under Network > Wireless. Both radios come up with WPA2. I haven't done real throughput or client testing yet.
  • Front panel LEDs follow radio, signal and WAN state.
  • Settings survive reboots. Root starts with an empty password like stock OpenWrt, and SSH is key-only on port 22, LAN only.
  • The modem stack starts by itself at boot.
  • It runs cool: around 40 C idle on the SoC, and the CPU is almost fully idle. Stock sat around 65 C idle without a fan, and about 46 C with a 12 V fan.
  • Flashing tools that check every write (hashes, neighbouring partitions, the other slot) and a boot counter that behaves like stock's. I lost a lot of time on that one: the preloader adds a try on every boot, stock resets it after a good one, and my first image didn't.

Where I'm taking it: a stable release with a recovery slot

Right now both slots hold a full firmware. The plan is to stop doing that:

  • Slot A: stable OpenWrt, the thing you actually run.
  • Slot B: becomes the recovery slot, a small kernel and recovery path whose job is to rescue the box and install or reinstall OpenWrt on slot A.
  • A way back to the stock firmware from that recovery path, so nobody is stuck with mine.

That way I don't have to keep two full firmwares in flash, which frees up storage, and recovery no longer depends on the download agent and a UART. That's the path to a stable release for me.

Other things on the list:

  • A stable tagged release (every merge to main already gets a 0.x pre-release tag).
  • Proper Wi-Fi and throughput testing, and checking more LED states.
  • Fan control from LuCI. Right now the fan just runs flat out.
  • Local SMS and voice, which is what I originally wanted this box for.
  • Own kernel. I have the SDK of a sibling device with the same SoC (Fibocom FG360, 4.19). I haven't tried building a kernel from it, and I won't until the router itself is finished. If it works, I could publish a matching kmod feed, turn on nftables and namespaces, and move to a newer OpenWrt userland.
  • sysupgrade for the new layout.

What's not great

  • The modem side needs the vendor userspace daemons and libraries, which are closed binaries. I can't put them in a public image. My build script copies them from your own stock firmware dump. So what I can publish is the build tooling and the packages, not a ready-to-flash image.
  • Kernel limits come from the vendor 4.19.205 build: the official feeds' kernel modules don't load (they're built for 5.15), there's no nf_tables so I'm on fw3 and iptables-legacy, and no namespaces so no ujail. Userspace packages from the 23.05.5 feeds install fine.
  • Slot A isn't marked as confirmed in the boot control record. I only reset the tries counter after each good boot, which stops it running out and keeps the fallback to stock B for now.
  • One unit, one SIM, one carrier.

What I need help with

  1. What does a wiki device page or a Table of Hardware entry need for a device that isn't in a release? Is a page for a custom build on a vendor kernel acceptable, and what should it say?
  2. How do other projects publish firmware that needs blobs the user has to extract themselves? Is there a convention?
  3. Has anyone done a recovery slot like this on an A/B NAND device with UBI and a preloader that picks the slot? What did you use for the recovery side, and how did you handle sysupgrade?
  4. Does anyone else have this box, or a rebrand of it (MT6890 + MT7915 + MT7531)? I'd like to compare notes.

Repo: https://github.com/mralinp/zlt-x28-openwrt

I can post logs, the partition layout and the boot flow if that helps.

As long as there's no mainline support, there's not going to be one ... ?

Skip the wiki, add all info to this tread, it is a community build, after all.

Ah, got it, thanks for clarifying!

Actually, we've made quite good progress so far. The custom firmware is already booting on the X28, and everything seems to be working fine, including cellular, WAN, LAN, Wi-Fi, DHCP, firewall, SMS, and VoIP.

I'm updating the docs and will share the findings here as well.

My goal is to eventually get proper OpenWrt support for this device, and I'm happy to maintain it myself and welcome anyone who wants to contribute.

I'd appreciate any guidance on how we can get this upstream.

Only happening if you I can get the hw supported upstream, or are willing to keep maintaining it yourself, forever.

Quick recap with details, since the first post is a few weeks old. Everything below is on one unit, one SIM, one carrier. Current release is v0.38, and the evidence for each item is in docs/TESTING.md.

Repo and docs: https://github.com/mralinp/zlt-x28-openwrt
Status table: https://github.com/mralinp/zlt-x28-openwrt/blob/main/README.md

Working

  • Native OpenWrt 23.05.5 userspace on the vendor 4.19.205 kernel (fw3 + iptables-legacy), running from slot A. Stock slot B is untouched.
  • 5G data as a normal netifd interface (proto x28_mipc). It does IPv4, IPv6 and IPv4v6, takes the APN from the SIM, and has its own LuCI page.
  • LuCI over HTTPS, a four-port LAN bridge (MT7531), dnsmasq DHCP/DNS, IPv6 on the LAN, firewall and NAT.
  • Wi-Fi on the MT7915 dual band through the vendor mt_wifi driver and a netifd handler. Both radios come up with WPA2 and a phone works through them. Throughput tests are still pending.
  • Front panel LEDs follow radio, signal and WAN state.
  • Settings survive reboots. SSH is key-only on port 22, LAN only.
  • A/B boot counter handled like stock: the preloader adds a try on every boot and x28-bootok resets it after a good one.
  • About 40 °C idle on the SoC. Stock idles at 65 °C with no fan, or 46 °C with a 12 V fan.

New since the first post: VPNs and Passwall2
The vendor kernel has no WireGuard and no TPROXY, and the official kmods don't load on it. So I build them for this kernel (wireguard-linux-compat, TPROXY and socket modules) and every image ships them preinstalled.

  • WireGuard: up to 715 Mbit/s sending, with the CPU not saturated.
  • OpenVPN client: about 100 Mbit/s, limited by one core.
  • Passwall2 (xray): the whole LAN goes through the node (TCP, UDP and IPv6). It's switched from LuCI and survives a reboot.
    • From a LAN client I measured 206 Mbit/s down and 40 Mbit/s up through the tunnel. The link ran out before the CPU did: the CPU peaked at 56% of 4 cores.
    • My estimate is about 0.45 cores per 100 Mbit/s of tunnelled traffic, so a ceiling of roughly 300–450 Mbit/s.
  • L2TP and PPTP are installed but not yet tested against a server.

Not done or not proven

  • SMS and voice: they worked on the stock-firmware overlay (voice was blocked by the SIM) and are not ported to the native image yet.
  • Slot A is deliberately not marked successful in the boot record. That keeps the fall-back to stock B.
  • The download-agent write to misc was never acknowledged, so misc is only changed live, using stock's dd-nand method.
  • Memory slowly falls over a long soak (#213).
  • Recovery slot on B, sysupgrade, fan control from LuCI and an own kernel built from the Fibocom FG360 SDK are all still planned.

Publishing
Thanks frollic, that makes sense: no wiki page without mainline support, and I'll keep everything in this thread and the repo. The vendor modem userspace is closed, so no ready-to-flash image can be published. The repo has the build tooling and packages, and the image is assembled from your own unit's backup.

If anyone has an X28 or a rebrand (MT6890 + MT7915 + MT7531), I'd like to compare notes. I can post logs, the partition layout and the boot flow if useful.

How to Get Started

All instructions for dumping your stock firmware, pulling the blobs, and compiling/flashing the image are detailed in the repository's README.

Please check out the code, test it on your device, and drop your logs or feedback here. Let's get this tricky hardware closer to a rock-solid, stable community release!