Repository navigation
Conversation
…rences Abstract derived from the YouTube subtitles (transcribed locally) and cross-checked against the 37-page slides deck and the CFP session page. Render-verified with a Jekyll + just-the-docs build. Signed-off-by: Thierry Laurion <insurgo@riseup.net>
tlaurion
force-pushed
the
add_bsmconf_flashkeeper_conference
branch
from
October 13, 2026 19:53
cb75be1 to
2f17007
Compare
Collaborator
Author
|
@DodoidOfficial do you want to review this before merge? Rendered at https://tlaurion.github.io/heads-wiki/#conferences for now |
There was a problem hiding this comment.
🟡 Changes recommended
The abstract overstates the security guarantees provided by the FPGA’s NVCM configuration.
2 open findings
What changed in this PR
Adds Dodoid’s BSMConf 2026 Flashkeeper presentation to the conference archive.
Changes:
- Adds the talk’s video, detailed abstract, and slide deck.
- Summarizes Flashkeeper’s hardware, firmware, security model, and demo.
| File | Description |
|---|---|
index.md |
Adds the BSMConf 2026 Flashkeeper conference entry. |
🧠 Review effort: Balanced
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
| <details><summary>Abstract</summary> | ||
| Jan Suhr (Nitrokey) leads a design discussion on Heads' measured-boot architecture: a coreboot + Linux-payload firmware using a USB security dongle as root of trust to mitigate evil-maid attacks, paired with Qubes OS on Nitropads. The central UX pain point: users must re-sign after every kernel update because Heads verifies exact hashes of /boot files. "If you compare it to ordinary PC users, this is in fact complicated — this is a burden," Suhr states, noting this blocks mainstream adoption. Alternative architectures debated: UKI (Unified Kernel Image) could solve the resigning problem by reducing changed files from five to one and enabling signature-based rather than hash-based verification. Porting Heads' logic to UEFI/TPM was discussed, but would require reimplementing all the bash-based measurement logic — "if funding is unlimited, everything's possible," Laurion quips. DRTM was considered but dismissed as server-only and requiring PCI-device reset. OpenPGP card support in UEFI was explored but deemed "quite expensive" (requiring Rust reimplementation). The group converged that making attested boot mainstream requires: kernel updates without manual resigning, richer user feedback on what changed in grub configuration, OS-side signing integration, and potentially UKI adoption. | ||
| </details> | ||
| * [2026 - BSMCONF - Dodoid - Flashkeeper - Manage and Protect your SPI Flash](https://www.youtube.com/watch?v=DVMI0dS5t8c) |
| </details> | ||
| * [2026 - BSMCONF - Dodoid - Flashkeeper - Manage and Protect your SPI Flash](https://www.youtube.com/watch?v=DVMI0dS5t8c) | ||
| <details><summary>Abstract</summary> | ||
| Dodoid, embedded developer, PCB designer, free-software sysadmin and system integrator, and administrator of the 15h.org community (100 or more KGPE-D16 and other fam15h users, maintainers of coreboot-15h), and an occasional upstream coreboot contributor, presents Flashkeeper at the Boot Security Mastery Conference (BSMConf) 2026, a project he has worked on for a while with Thierry Laurion of Insurgo, a platform security researcher and long-time maintainer of LinuxBoot and Heads, who offered the Insurgo PrivacyBeast X230, one of the few Qubes OS certified laptops. He opens with a what if: a standardized, open, well-documented expansion port on every motherboard, providing power and SPI on the same bus as the firmware flash, in a standard form factor, adopted by nearly every vendor. That port does not exist. Close things do: SPI TPM headers exist but are vendor dependent, and some lack the BIOS chip select (he showed a TPM module he designed for MSI boards that also exposes the BIOS flash). Older boards with DIP flash take interposer boards, but they are rare, 3mdeb used to sell one for the KGPE-D16 with a write-protect header, and DIP flash chips are hard to find. The one near-universal interface is the flash chip itself: most modern boards have an 8-pin SPI flash (WSON-8 8x6 or SOIC-8 208-mil dominate, WSON 6x5 and SOIC 150-mil are uncommon), the pinout is essentially universal, pitch is usually 1.27 mm, width usually about 8 mm, and manufacturers often draw flexible footprints with exposed ground pads or long pads to allow part substitution. Chip clips exploit this, but WSON and SOIC clips do not interoperate, WSON clips are finicky, clips are unreliable even short term (one flew off the desk at 80% of a flash run), too big to leave in a running system (laptops, chips under heatsinks), and sometimes do not fit at all when passives are close. The Flashkeeper answer is a chip adapter: a 2-layer flex PCB soldered permanently on top of the host flash, with eight flexible legs, four per side, reaching down and soldering to the same pads the chip uses, with variants optimized per package. A breakout (SOIC or WSON) brings the bus out on a ZIF connector. His example is an AAEON UP Squared, an industrial SBC with an Apollo Lake SoC and a coreboot port, whose flash sits under the CPU heatsink: with the breakout installed the heatsink goes back on, the cable runs out of the case, and the board flashes from the desk (his photo shows it going through a small adapter board into a CH341A). He has sent breakouts to laptop owners whose flash is only reachable with the machine fully apart, with the cable run out through a USB port or any chassis hole. A write-protect adapter puts a switch on the chip that holds WP# low. Most SPI flash chips support partial write protection, set in the status registers with flashrom, for example to protect the coreboot bootblock region, but it only enforces if WP# is held low, and many motherboards offer no user access to it, though some have jumpers and some Chromebooks a screw. So the chip just gets a switch. This was the first Flashkeeper item: Laurion asked Dodoid to build it while working on Heads bootblock write protection, replacing the solder blobs and jumper wires people improvised. Then the talk goes bigger: a microcontroller on the chip could monitor reads, debug SPI transactions, read and write the flash remotely, authenticate external reprogramming, and verify flash contents at boot, and an FPGA can do all of that. Space is about 25 mm squared, and the module must be self-contained: 3.3 or 1.8 volts from the host with no choice, its own clocks, its own voltage regulation, its own bitstream and firmware storage, and its own development port. Flex PCBs have large minimum feature sizes, and rigid flex is a quote-only item from a few vendors, so the answer is a rigid system-on-module mounted as a component on the flex adapter. The SoM carries a Lattice iCE40UP5K FPGA, chosen for a fully free yosys and nextpnr toolchain, on-chip SPRAM, a built-in high frequency oscillator, about three voltage rails, and one-time-programmable non-volatile configuration memory that after programming refuses other bitstreams and does not let the bitstream be read out, a useful security property; a TI TPS7A20 regulator; a Micro-Crystal RV-3028-C8 real-time clock, because there is no good way to access the host RTC and no clear way to trust it; and a Renesas AT25DF041B SPI flash for the FPGA's own bitstream and firmware, mounted less than a millimeter above the host flash. The SoM is 7.15 by 3.30 mm, six layers, custom LGA-25 footprint, designed in KiCAD, 2 mil trace spacing, 0.4 mm BGA pitch, 0201 passives, no blind or buried vias, on the standard JLCPCB process. It was too small for the fab minimum, so he designed a custom panel with ten SoMs, fiducials, and handling rails to reach 70 by 70 mm, with snap-off development breakout headers. The FPGA chip adapter puts the SoM on the bottom against the host flash, with a 10-pin ZIF debug connector on top. The flex part can fold away or be cut off with scissors for height-constrained builds. A tail adds an onboard speaker for beeps and alerts, possibly coreboot's SPKCON, plus a TDK CeraCharge solid-state battery and charge regulator that keeps the RTC running with the host off, recharged whenever the host runs, expected to last a bit over a month. The FPGA runs a modified PicoSoC: a PicoRV32 RISC-V core in RV32Izmmul (hardware multiply, software divide), 128 KiB of the iCE40 SPRAM, an 8 KiB boot ROM in the bitstream itself, which matters because SPRAM cannot be initialized from the bitstream, so one-time-programmed code can load and verify a firmware image from the external flash instead of trusting it, a ring-oscillator entropy source, and memory-mapped access to both the onboard and host SPI flash. The Flashkeeper firmware reads, writes, and debugs the host flash, includes TweetNaCl lightweight cryptography, and has an extensible Forth-like stack shell (EDE) over serial, not a compliant Forth. It is an FPGA, so whatever you want. Applications: authenticated external flashing; flash image verification; coreboot bootblock verification, coming very soon, which at runtime sniffs the SPI bus, builds a copy of the bootblock in memory, and verifies it against the actual bytes the host reads, and at rest masters the SPI bus with the host powered off and reads the flash directly; logging and debugging SPI flash and SPI TPM transactions, the TPM usually shares the bus; and a coreboot spiconsole decoder, since coreboot can output its console over the SPI bus. The point against vendor verification like Intel Boot Guard: those rely on code the user does not control, in ROM or signed with keys the user does not have. The Flashkeeper FPGA is a general-purpose computer under user control, fully free and open hardware, software, and gateware, built with entirely open toolchains (KiCAD, FreeCAD, yosys, nextpnr, GCC). And fused and read-only do not have to mean proprietary: you develop with the bitstream in flash as long as you like, then burn it to NVCM at home with your own gateware, boot ROM, firmware, and secrets, and nobody can extract it or make the FPGA run anything else. The live demo resets the FPGA over serial with tio, shows the password prompt (no points for guessing what the password is), runs help, pushes 2 and 2 onto the Forth stack and gets 4, reads the host flash ID with HFID (a Windbond chip in his setup), benchmarks the PicoRV32 core, and pulls bytes from the onboard entropy source; the shell can read and write the host flash directly. For authenticated external reprogramming he exits the shell, the FPGA still answers on the serial port with my name is Flashkeeper FPGA, identifies the host flash, and he skips the full read/verify cycle so the audience does not have to sit through it; the password is what stops anyone else from reprogramming the flash, so a laptop you carry can be reprogrammed from outside the case without a cable left in the chassis. Status: the hardware is fully open source, the firmware is open source too, not every application described is done, but everything in the demo works. Repositories are at git.15h.org/Flashkeeper/flashkeeper and github.com/linuxboot/flashkeeper, the site is flashkeeper.org, there is a Matrix room, and contributions, questions, and testing are welcome. Q&A: install guides exist for the chip adapters, the FPGA module guide is planned for the website. On sales and distribution, the iCE40 FPGAs are almost impossible to buy right now (he cannot imagine how they are used for AI, but they are being bought up for it), so he is hoarding his stock and plans a small production run, maybe ten units, as a beta program initially through Matrix, and is open to partners. A question asks whether the write-protect switch still works when the platform uses the flash in quad I/O mode: it is chip dependent, he doubts it works because the switch just pulls that IO low, the only idea is a series resistor if the flash chip overcomes it, he is not sure how flash chips handle the WP pin in quad mode, whether they ignore it or sample the last WP state from single mode, and status-register write protection in quad mode is less standard than the rest of the 25xx SPI flash protocol, so if you rely on write protection on a flash that runs in quad mode, verify that it actually holds. |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.

Abstract derived from the YouTube subtitles (transcribed locally) and cross-checked against the 37-page slides deck and the CFP session page. Render-verified with a Jekyll + just-the-docs build.