Skip to content

Node power state without /dev/mem, and honest reboot text for v2.4 boards

A small, careful release. The kernel now tells bmcd which nodes are powered through sysfs, so the daemon no longer needs /dev/mem for it. On v2.5 boards a BMC reboot still leaves the nodes running, and the API and tpi reboot help now say plainly that a v2.4 board power-cycles its nodes. Tested on real hardware, including a power pull. Thank you, Daniel Steinhauf (@danste22) and Sven Rademakers, for the shadow-word contract this builds on.

Firmware v2.53.0, 21 November 2026 — the first image to carry bmcd 2.47.0, tpi 1.11.1. Every item below is the repository's own changelog entry, taken apart; the whole entry is a click away under each one.

Thank you, @danste22 (Daniel Steinhauf) and Sven Rademakers. The shadow-word contract that this release exports through sysfs is the one Daniel carried to the 6.18 kernel (tp2/power/0002), from the original idea in Sven's 2024 patch in turing-machines/BMC-Firmware; thanks to both.

Section titled “The latch shadow word and the boot cookie in sysfs”

A new kernel patch, tp2/power/0003, makes the gpio-latch driver export two read-only attributes on its platform device: /sys/bus/platform/devices/gpio-latch/shadow (the shadow word, 0x%08x) and .../shadow_valid (1 when the bootloader's boot cookie is present).

The whole entry

A daemon no longer needs /dev/mem to learn which nodes are powered. The words stay in SoC SRAM and /dev/mem reads them as before, so bmcd 2.46.0 keeps working. The shadow-word contract is the one @danste22 carried to 6.18 (tp2/power/0002, from Sven Rademakers' 2024 patch); that patch is untouched. The attributes are documented in Documentation/ABI/testing/sysfs-platform-gpio-latch (#104).

  • bmcd 2.47.0 reads which nodes are powered from shadow and shadow_valid in sysfs, and no longer needs /dev/mem for it (bmcd#61). On firmware without tp2/power/0003 it falls back to /dev/mem, and on a v2.4 board, which has no latch, to the node enable lines. What counts as a powered node does not change.
  • A v2.4 board power-cycles its modules on a BMC reboot, and bmcd says so (bmcd#62). The /reboot API text and the README now say that a BMC reboot keeps the modules running on a v2.5 board and power-cycles them on a v2.4 board.
  • tpi 1.11.1: tpi reboot --help is correct for v2.4 and v2.5 (tpi#19). It says the same as the API, for both board revisions. Help text only.
  • bmcd 2.47.0, BMC-UI 4.7.0 and tpi 1.11.1. Are the pinned versions (the latest of each).
  • bmcd has new tests that pin how a switch change handles failure at the confirm, apply and revert steps (bmcd#60).
  • On two v2.5 boards with v2.53.0-rc.1: a BMC reboot and a restart of bmcd both adopt the live power state, with no node reset.
  • Earlier the same day, on the same two boards with v2.52.0, which reads the same shadow contract from /dev/mem: a BMC reboot with one node off kept it off, and a real power pull restored the saved state (a node that was off stayed off).
  • One over-the-air write was seen to fail, on one board (v2.52.0). Writing the image with tpi-selfupdate failed with a UBI write timeout on the final block (ubiupdatevol: error 110). The firmware the board was running is untouched and still boots.

    The whole entry

    If an over-the-air write fails, power-cycle the board, then retry or flash from the SD card. The cause is not found yet.

  • The reboot endpoint no longer says the compute modules keep running on every board. The OpenAPI summary of POST /api/bmc/reboot and the README now say what is true of both revisions: on a v2.5 board the modules keep running (the daemon adopts the live power state), on a v2.4 board a BMC reboot power-cycles them. bmcd 2.47.0

    The whole entry

    Text only; no behaviour changes.

  • The switch rollback paths are now covered by tests. A confirm whose switch.json write fails keeps the change pending (and the same token confirms once the disk allows it, or the window puts the old configuration back); an apply that fails in the middle of a plan runs the way back and reports the first error; the timer's revert goes on past a failing command. bmcd 2.47.0

    The whole entry

    The behaviour itself shipped in 2.45.0; this pins it.

  • The latch shadow is read from the kernel's sysfs attributes first. Where the firmware's gpio-latch driver exports shadow and shadow_valid under /sys/bus/platform/devices/gpio-latch/, the daemon reads the boot cookie state and the shadow word from there and no longer maps SoC SRAM through /dev/mem. bmcd 2.47.0

    The whole entry

    On firmware without the attributes (every firmware up to and including v2.52.0), or when they cannot be read or make no sense, it reads /dev/mem as before, and on a v2.4 board or where neither can be read it still reads the node enable lines. What counts as a powered node does not change. The shadow-word contract with the bootloader is the one @danste22 carried to 6.18, from Sven Rademakers' 2024 patch.

  • tpi reboot --help now says what happens on both board revisions. It claimed the compute modules keep running throughout. That is true on a v2.5 board, where the daemon adopts the live power state; on a v2.4 board a BMC reboot power-cycles all four modules. Help text only. tpi 1.11.1

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

Last updated: