Skip to content

Settings survive a power cut, and the switch no longer cuts the board off

A settings file is no longer left empty by a power cut right after you confirm a change, and going from Flat to Split or Trunk no longer strands the board's own address on the wrong VLAN. Both were found and first fixed by @danste22, who reported them as #81 and #82, and @cloudlabworks fixed a forgotten session token counting toward a ban (bmcd#39). Thank you both.

Firmware v2.50.0, 20 November 2026 — the first image to carry bmcd 2.44.0, BMC-UI 4.6.1. Every item below is the repository's own changelog entry, taken apart; the whole entry is a click away under each one.

Fixes https://github.com/excavador-turing/BMC-Firmware/issues/81 (a power cut could leave a settings file empty) and https://github.com/excavador-turing/BMC-Firmware/issues/82 (Split and Trunk cut the BMC off), and lets a node's console speed go back to its default.

Thanks to the people outside the project whose work is in this release: @danste22 found both problems by reading the code, reported them as #81 and #82, and wrote the first fixes (bmcd#48 and bmcd#49, carried on as bmcd#51 and bmcd#52); @cloudlabworks wrote the fix for a forgotten session token counting as a wrong password (bmcd#39) and corrected the toolchain note in the daemon's Justfile (bmcd#40).

  • BMC-UI 4.6.1. Takes its API types from bmcd 2.44.0. Nothing in the interface changes.

  • The toolchain named in the Justfile and the README is 1.98.1. Both still said 1.85, which CI left behind long ago; rust-toolchain.toml has pinned 1.98.1, the version the firmware builds with. Documentation only. bmcd 2.44.0

    The whole entry

    (thanks @cloudlabworks, #40; the README in #53)

  • The API types are generated from bmcd v2.44.0. bmcd-release.txt and src/lib/api/schema.d.ts, with npm run api:refresh. BMC-UI 4.6.1

  • A power cut just after confirming a change could leave a settings file empty (bmcd 2.44.0, #81). The switch table, /etc/network/interfaces, the access overrides, the managed client CA and the TLS key and certificate were written to a temporary file and renamed, but never flushed to flash.

    The whole entry

    On the board's flash a power cut right after confirm could leave a file with the right name and nothing in it: the switch fell back to Flat with no message, or the board came up with no network. Each of them is now flushed, and so is its directory, before the change counts as saved. Thanks @danste22 (bmcd#48 → #51).

  • Going from Flat to Split or Trunk cut the BMC off and the change was reverted (bmcd 2.44.0, #82). The switch plan took the front ports out of VLAN 1 but left the BMC's own bridge address in it, so the board became unreachable and the confirm window undid the change.

    The whole entry

    The plan now puts the BMC in the VLAN the table gives it, joining the new one before leaving the old, so it is never in none. Thanks @danste22 (bmcd#49 → #52).

  • A session token the board had forgotten counted as a wrong password (bmcd 2.44.0). A browser tab left open across a reboot sent its old token on every request, and each refusal counted toward banning that address, so the right password typed from the same machine could then be refused.

    The whole entry

    An unknown token is still refused, but no longer counts. Thanks @cloudlabworks (bmcd#39).

  • Setting a node's console speed to null did nothing (bmcd 2.44.0). A node's speed starts as null, meaning 115200, but sending null again answered 200 and kept the speed that was stored.

    The whole entry

    null now clears it and the open port goes back to 115200 at once, without disconnecting the web console. Leaving the speed out of a request still keeps it.

  • A session token the daemon had forgotten counted as a wrong password. Tokens live in memory, so a restart forgets all of them, and an expired token is dropped from the store. bmcd 2.44.0

    The whole entry

    A browser tab or script still sending one got a 401 for each request, and each 401 counted toward the peer's ban exactly like a wrong password. A page that fires several requests at once could therefore ban its own address after a BMC reboot, and the correct password typed from that machine was then refused with a 429. An unknown token is still refused with the same 401 and Bearer challenge, but no longer counts. A 64-character random token cannot be guessed, so counting it gave no protection. A ban that already stands still refuses tokens. (thanks @cloudlabworks, #39)

  • Confirmed config files are flushed to flash before rename. switch.json, /etc/network/interfaces, access overrides, the managed client CA, and TLS key/cert pairs were written with temp+rename and no fsync. bmcd 2.44.0

    The whole entry

    On the UBIFS overlay a power cut just after confirm could leave a correctly named empty or truncated file — Flat with no UI error for the switch document, or no network at boot for interfaces. One helper, write_durably, now syncs the data and the directory (one <name>.tmp per target, opened exclusively, a stale one from an interrupted write removed first; a directory sync failure after rename is logged, not treated as a failed write). (thanks @danste22, #48 → #51; fixes BMC-Firmware#81)

  • Split and Trunk left the BMC on VLAN 1. plan() skipped PortId::Bmc because it has no netdev. The BMC's membership is the bridge self entry (bridge vlan … dev br0 … self). bmcd 2.44.0

    The whole entry

    Without that, Flat→Split/Trunk cleared VID 1 from the front ports while br0 stayed on the kernel default, the board went unreachable, and the confirm window reverted the change. The planner now emits those self lines, the BMC's new VLAN added BEFORE the old one is deleted so the BMC is never in none, and a failed delete of a self entry (a bridge that never had it) no longer stops the plan. (thanks @danste22, #49 → #52; fixes BMC-Firmware#82)

  • {"NodeN":{"uart_baud":null}} answered 200 and changed nothing. The serial speed of a node starts as null (the default, 115200), but there was no way back to it: a request could not tell null from a field left out, and both kept the stored speed. bmcd 2.44.0

    The whole entry

    Now null clears the stored speed and the open port goes back to 115200 at once, without disconnecting the web console; a field left out of a request still keeps what is stored. A configuration import behaves as before: a file with no speed for a node leaves it alone.


Every release of the firmware · the roadmap · follow by feed

Last updated: