Day 1

The morning got off to a strong start with Sahaj Sarup's Memory-Mapped FPGA Peripherals on Zephyr. He described his quest to memory-map an FPGA into a microcontroller running Zephyr via the FMC and QSPI controllers. I love talks like this.

Marco and I had our talks later that day. Marco's talk made the point that running OP-TEE in the Secure World is not the same as running it securely: loading it takes careful coordination between several firmware stages as there is no shortage of creative alternative ways to try to circumvent memory protections. His work on closing these gaps resulted in patches across TF-A, OP-TEE and barebox.

For my first talk that week, I nerded out about one particular kind of fix: the ones whose commit message opens with a compiler version, followed by a disassembly with a fat arrow pointing at where programmer intent and code generation diverge. The slides collect the caveats we ran into writing bare metal C for barebox.

The talk that drew the most questions that day was the 17:20 slot, Agentic AI Software Debugging in OpenBMC by Jia Chunhui. I have been pondering LLM use for anomaly detection for a while: I have been logging all my interactive Labgrid use for the last three years. With that history at hand, it's possible to see how things like the dmesg output changes across kernel versions, provided you know what to look for. A local LLM with the right harness might just be able to spot such discrepancies on its own.

Day 2

The second day moved a bit away from BMCs and closer to our home turf of SoC bring-up and boot chains.

Bringing up Linux on a memory-safe CHERI RISC-V system by Alice Ziuziakowska was a gentle introduction to CHERI. Jorge Ramirez-Ortiz brought up a fully open TrustZone stack on Qualcomm's QCM6490, from TF-A at EL3 through OP-TEE to U-Boot and Linux at EL2, all built from source. For customers who don't use the evaluation kits, however, documentation and custom board memory designs are still out of reach. Hopefully not for long as these are some quite cool SoCs.

There was an impromptu session on GNU poke by Mohammad-Reza Nabipoor. We look at a lot of hexdumps at Pengutronix and flip our fair share of bits, so this one piqued our interest right away. I am still thinking about how to best use it live on the hardware instead of only on files. Apparently one can instantiate custom I/O spaces for that, maybe something to hook up to barebox RATP.

Simon Glass looked beyond EFI, which was designed in the 1990s for a closed-source world and brings a lot of complexity that, he argues, open source systems don't need. I wouldn't want to force shim on anybody not needing it either, but I could imagine our future projects booting Linux through its EFI stub even if nothing else changes. Keeping the MMU enabled across the handover and physical ASLR? Yes, please.

Jiji Freya Daniel Maslowski made a good case in Boot Chains and Build Systems for why build systems should describe how the components on a SoC interact.

Mate Kukri closed the day with native raminit for Bay Trail and how he reverse engineered the vendor blob by tracing its execution.

After that, the hallway track moved onto two boats for a cruise through Amsterdam's canals.

Day 3

The last day was a mix of lightning talks and open mic slots as well as a parallel workshop track. One of the lightning talks was mine, on barebox security policies (slides). These let a single bootloader image adapt to the development, factory and field stages of a device instead of shipping a zoo of nearly identical images. Security Policies were still in the making when I spoke about bootloader security at last year's ELC-E and Linux Security Summit, both also in Amsterdam, and are now upstream.

We also took part in the Arm SystemReady Workshop. The topic is gaining relevance for us: since v2026.01.0, barebox can act as EFI firmware instead of only running without EFI or under EFI. A future goal would be to designate an upstream platform in barebox as EFI reference platform and get it to pass ARM's Architecture Compliance Suite for the Embedded Base Boot Requirements with Device Tree.

Wrap-up

The conference was well organized and the venue was a nice place to spend three days in. We met nice people, ate well and had interesting conversations with attendees throughout.

OSFC has been alternating between Europe and the US, with last year's edition held in Sunnyvale, CA. We will probably not be there for the next one, even though we hear the US editions are more coreboot-y and therefore likely closer to our day-to-day. We are, however, looking forward to it presumably returning to Europe in 2028.