I bought an enterprise network switch on eBay for $25.

Forty-eight gigabit ports, four 10-gig uplinks — the kind of switch that used to run a corporate network. It shipped with the vendor’s own software, Broadcom’s FASTPATH/ICOS: closed, locked down, and frozen on Linux 4.4 — a kernel that stopped getting security fixes in 2022. That’s why gear like this sells for $25. Not because it stopped working, but because the software stopped.

So I did what I’d done with its bigger sibling: replaced the vendor software with my own — one that’s actually free and open. That distinction matters. The stock NOS here was closed, and the usual alternatives for boxes like this are commercial; even the open-source-branded ones, like Cumulus, sit behind a license. EdgeNOS-4610 doesn’t. It’s a from-scratch, free and open, systemd-based network OS, now running on a current, mainline-derived Linux 6.1 LTS — security-maintained through 2027. No vendor firmware, no closed switch driver, no license key: my own kernel and userland, and a data plane built to drive the Broadcom ASIC directly.

One chip, not two

This box was a completely different beast from its sibling. The AS5610 is a PowerPC machine where the switch ASIC sits on the PCIe bus — a separate CPU and chip, like a giant network card in a computer. The 4610 is the inverse: one chip. The ARM CPU cores are baked into the switch ASIC itself.

That cut both ways. Because the board was well supported, getting onto its original Linux 4.x kernel was the easy part. Upgrading was the hard part — the entire support package for a CPU that lives inside a switch chip sits outside mainline Linux, so every kernel jump meant re-floating that whole out-of-tree BSP by hand.

Climbing the LTS ladder

So I climbed one LTS at a time — 4.14, 4.19, 5.10, 5.15, 6.1 — re-floating the board support package and proving the data plane on real hardware at every rung. A few greatest hits along the way:

  • a dead-battery clock that wedged the boot,
  • a CPU floating-point unit that’s technically optional — soft-float made init crash on an illegal instruction,
  • a 6.x change that NULL-pointered the chip driver at load, until I swapped one API call.

The result: a $25 switch, abandoned on a 2019 kernel, now passing traffic on both copper and fiber — on a kernel newer than a lot of production gear is running today. And it can keep climbing.

“End of life” was a business decision, not a verdict on the silicon.