Metrics

Upstream's firmware exposes no metrics. This fork serves 40 families at /metrics, in Prometheus text format.

This page is generated from the daemon's source, so it cannot describe a metric the daemon does not emit. The authority is always the board:

$ curl -s http://<board>:9110/metrics

Its own port, and no credential

/metrics is served on port 9110, plain HTTP, and takes no credential.

That is a change from earlier releases, which put it on :443 beside the API behind a token of its own. The token existed for one reason: so that a credential sitting in a scrape config could not also reach /api/bmc and power four compute modules off. On a listener that serves nothing but /metrics there is nothing else to reach, so the property survives and the mechanism is a port instead of a secret — one fewer thing to mint, store, rotate and leak.

The TLS went with it. It was always scraped with verification disabled, because the board's certificate is expired and carries no subject-alternative name, so nothing could verify it. Unverified TLS is a handshake per scrape on a Cortex-A7 in exchange for nothing.

What protects the endpoint is the network. It binds the same address as the API, so restricting the daemon to a management network restricts both. And there is nothing secret in it: temperatures, fan steps, port counters, power state, NAND wear and slot versions.

On a board older than v2.15.0

/metrics is on :443 and wants a token, which tpi metrics show printed. That command is gone, along with the token; if you are scraping an older board, keep the old config until you update it.

A Grafana dashboard over every family below ships with each release; see monitor it.

Scraping it

scrape_configs:
  - job_name: turingpi-bmc
    static_configs:
      - targets: ["<board>:9110"]
        labels:
          instance: <the board's hostname>

The instance label is yours, not the board's

The board's exposition carries no instance label and no hostname — 190 lines, and its own name appears in none of them. The labels it emits are device, kind, name, node, port, result, sensor, slot, state, version and volume.

So instance is whatever your scraper assigns, and the example above assigns it deliberately. Renaming a board moves no history; editing that label moves all of it. Both boards here were renamed on 2026-09-11 and not one series moved.

This page said the opposite until then, as did tpi hostname and the API's own description of the endpoint (SQU-191). The claim was written from a worry rather than from the exposition, and a confident wrong warning is worse than none: it makes a safe rename look dangerous, and implies that leaving the name alone protects a history that the scrape config alone decides.

The families

The daemon

metric	type	what it is
bmcd_build_info	gauge	Version of the daemon that produced these metrics. Always 1; the answer is in the version label.

Read this one first when a number looks wrong. Every other family below is produced by the daemon this names, and a metric that does not exist in an older build is simply absent rather than zero.

Temperature and cooling

metric	type	what it is
bmcd_cooling_state	gauge	Step a cooling device is currently at.
bmcd_cooling_state_max	gauge	Highest step a cooling device accepts.
bmcd_cooling_overridden	gauge	1 while a fan is held at a step and its zone's governor is paused.
bmcd_temperature_celsius	gauge	Temperature reported by a kernel thermal zone.

The switch

metric	type	what it is
bmcd_switch_port_link	gauge	Whether a switch port has carrier.
bmcd_switch_port_present	gauge	Whether the kernel has a netdev for this switch port.
bmcd_switch_port_rx_bytes_total	counter	Bytes received on a switch port.
bmcd_switch_port_rx_errors_total	counter	Receive errors on a switch port.
bmcd_switch_port_speed_bits_per_second	gauge	Negotiated line rate of a switch port.
bmcd_switch_port_tx_bytes_total	counter	Bytes transmitted on a switch port.
bmcd_switch_port_tx_errors_total	counter	Transmit errors on a switch port.

Compute modules

metric	type	what it is
bmcd_node_power_on_seconds	gauge	Seconds since a compute module was powered on.
bmcd_node_power_state	gauge	Whether a compute module is powered on.
bmcd_node_usb_boot_armed	gauge	1 when a module's USB-boot pin is asserted, which stops it booting from its own eMMC.
bmcd_usb_config	gauge	The board's persisted USB multiplexer configuration, labelled with its node, mode, route and bus type.

The one to put a rule on

bmcd_node_usb_boot_armed is the only reading here that describes a trap rather than a state. Flashing a module leaves the board's persisted configuration at Flashing(nodeN, …), and the daemon re-applies it on every start — which holds that module's USB-boot pin high and stops it booting from its own eMMC.

The module carries on running, because it is already booted. The failure arrives at its next reboot, which can be weeks later, and it looks like dead hardware: nothing at all on the serial console, not even a bootloader banner, because the loader does not use the console; nothing on the network; and the board still reporting the rail on. tpi usb status cannot tell you either — it prints the same route for a normal configuration and for this one.

So alert on duration, not on the state. Flashing a module is a legitimate thing to be doing; a module still armed fifteen minutes later is a forgotten tpi usb device -n N, and the page reaches whoever armed it while they still remember.

- alert: ModuleLeftArmedForUsbBoot
  expr: bmcd_node_usb_boot_armed == 1
  for: 15m
  annotations:
    summary: "{{ $labels.node }} will not boot from its own eMMC - run `tpi usb device`."

The family is four samples rather than one because a single "which node" gauge cannot say none, and none is the normal answer.

The BMC itself

metric	type	what it is
bmcd_load1	gauge	1-minute load average of the BMC.
bmcd_load15	gauge	15-minute load average of the BMC.
bmcd_load5	gauge	5-minute load average of the BMC.
bmcd_memory_available_bytes	gauge	Memory available to a new allocation on the BMC, reclaim included.
bmcd_memory_free_bytes	gauge	Free memory of the BMC.
bmcd_memory_total_bytes	gauge	Total memory of the BMC.
bmcd_process_resident_bytes	gauge	This daemon's own resident set. Board memory says the board is being consumed; only this says by whom.
bmcd_process_threads	gauge	Threads this daemon has. Read beside the resident set: a heap leak grows memory with this flat, while a leaked task or an unreaped blocking thread grows both, because every thread carries a stack.
bmcd_uptime_seconds	gauge	Seconds since the BMC booted.

The certificate it is serving

metric	type	what it is
bmcd_tls_certificate_expiry_timestamp_seconds	gauge	When the certificate the listener is serving expires.
bmcd_tls_certificate_info	gauge	The key behind that certificate, in a key label. Always 1; the answer is the label.

Both describe the running listener, not whatever is on disk now. They are read once, when the certificate is read, which is at start — so replacing the file and not restarting is exactly the case where a fresh read would lie, and these keep telling you what the board is actually presenting.

Alert on the expiry with a long fuse. A board whose certificate has expired still serves, but nothing that verifies can reach it, and the recovery needs a hand on the board.

NAND

metric	type	what it is
bmcd_nand_available_bytes	gauge	Unallocated space on the BMC's NAND.
bmcd_nand_eraseblock_size_bytes	gauge	Size of one logical eraseblock of the BMC's NAND.
bmcd_nand_eraseblocks	gauge	Eraseblocks of the BMC's NAND, by what UBI counts them as.

Clock

metric	type	what it is
bmcd_clock_offset_seconds	gauge	The BMC's system clock minus true time; negative when the board is behind.
bmcd_clock_stratum	gauge	Stratum of the time source the BMC is synchronised to.
bmcd_clock_synchronised	gauge	Whether the BMC's system clock is disciplined. Absent when that cannot be determined.
bmcd_rtc_info	gauge	A real-time clock the kernel registered on the BMC.
bmcd_rtc_present	gauge	Whether the BMC has a real-time clock at all.

Firmware

metric	type	what it is
bmcd_firmware_promotion_total	counter	Boots that ran the firmware health gate, by what the gate decided. promoted is derived as attempts minus rollbacks, because the gate has no single line meaning kept; a board cut off mid-gate therefore counts as promoted.
bmcd_firmware_last_promotion_timestamp_seconds	gauge	When the firmware health gate last decided, so a dashboard can say when and not only how often. Absent rather than wrong when the board's clock cannot be trusted: the gate records the board's own time, which on a BMC that has just come up may be well before the real one, and only UTC is accepted.
bmcd_firmware_slot_info	gauge	A firmware slot on the BMC's NAND. The rollback slot is not mounted, so it has no version.
bmcd_firmware_slot_size_bytes	gauge	Size of the firmware in a slot.
bmcd_firmware_update_staged	gauge	Whether a firmware update is staged for the next boot. Absent when the U-Boot environment cannot be read.

A gauge that is absent is not a zero

Several of these are omitted rather than sent as 0 when the board cannot answer: bmcd_clock_synchronised on a board with no chrony, bmcd_firmware_update_staged when the boot environment cannot be read, bmcd_process_resident_bytes on a kernel without /proc/self/statm, bmcd_process_threads on one without /proc/self/status. An absent series means not known; a zero would be a claim.