About me

I do embedded Linux and networking engineering across the full stack - from the circuit to the kernel. That means schematic-level analysis, microsoldering and component-level rework, firmware, kernel drivers, and board integration in OpenWrt. Most engineers work at one of those layers. I work across all of them, which matters when a bug refuses to stay on one side of the hardware/software boundary.

Upstream Linux kernel work

• net: stmmac: dwmac-motorcomm - fix eFUSE MAC address read failure
  • on the YT6801 the eFUSE controller needs a settling window after reset; without it the read returns zeros and the driver falls back to a random MAC. Found by comparing against a custom U-Boot port where the same read succeeded because the driver came up later in boot. Fixes a user-reported regression on TUXEDO hardware. Merged.
• net: dsa: realtek: rtl8365mb - SGMII/HSGMII support for the RTL8367S
  • boards wiring the switch to the CPU over the SerDes had no working CPU port. The register sequence was recovered from the vendor GPL drop and cross-checked against live hardware by chainloading a custom chained U-Boot ahead of the stock firmware and logging the real SerDes accesses. Merged.
• net: stmmac - raise the TX completion interrupt at the end of an xmit burst
  • TX skbs are only freed on completion, and completion was gated on 25 frames or a 5 ms timer. Paced flows never queue 25 frames, so every burst waited out the timer: BBR over a 23 ms path was pinned at 5.24 Mbit/s where CUBIC reached 207. Setting the interrupt bit on the last descriptor of each burst takes it to 447 Mbit/s with interrupt load unchanged. A 2016 report of the same starvation was closed asking for a real fix; this is it. In review.

Other work

• Armbian - board support for the Orange Pi 5 Pro, and a follow-up rework extending it.
• games-on-whales/inputtino
  • fixed multi-touch slot allocation in the evdev backend. Contacts released out of order collided on the same MT slot, making fingers teleport between positions over Sunshine/Moonlight touch passthrough. Now allocates the lowest free slot with monotonic tracking IDs.
• OpenWrt - board support for a Mercusys MR80X v2.20 router. Still in process to publish the testing image.

Current focus

Networking: MACs, PHYs, Ethernet switches - the data path between the silicon and the Linux network stack, with most of my recent kernel work in and around stmmac. I got here through electronics and circuit design, and that hardware background is what lets me debug the problems that cross the boundary: a driver that misbehaves because of a strapping pin, a link that flaps because of a clocking issue the datasheet buries in a footnote, an interrupt that never fires because the errata says it won't.

I work upstream because that's where the code outlives the project. A driver merged into mainline keeps running on devices for decades, maintained by the community, long after any single vendor has moved on.