Thanh Le / bench notes
EMBEDDED · BOARD BPI-M3 · DONE

First contact: what the vendor image says about eight cores and 2 GHz

Plugged in a serial adapter, read the old vendor kernel instead of waiting for mainline, then checked FEL

Result8 cores online, 11 OPPs 480 to 2016 MHz, FEL at 12 Mbit/s
Last change2026-10-03: measured on the vendor Ubuntu 16.04 image, read-only
Boardall stories on this board →

First contact: what the vendor image says about eight cores and 2 GHz

Back to the Banana Pi M3 hub

  1. Situation
  2. Task
  3. Action
  4. Result
  5. What I’d do next

Situation

October 2026. I had a list of open questions from reading the docs two days before: the spec says 1.8 GHz but mainline Linux tops out at 1.2 GHz, the board has two CPU clusters and I didn’t know how the second one gets started, and I wasn’t sure FEL works on this SoC at all. I also had a ready mainline SD image, and no way to write an SD card on this PC.

Spec 1.8 GHz, mainline 1.2 GHz, vendor kernel 480 to 2016 MHz, FEL ROM 408 MHz The clock question I walked in with: 1.8 GHz on the spec sheet, 1.2 GHz in mainline, and the vendor kernel on the board says 11 steps up to 2016 MHz.

So the board came up on a new FTDI adapter on the debug UART (CON2, 115200 8N1), still running whatever was on its eMMC.

Task

Answer those questions on the real board, today, with the software that’s on it. And not break anything: it’s an installed system, not a scratch image.

Action

I listened for 5 seconds before sending anything, so I wouldn’t mess up a boot log in progress. Two NUL bytes, then silence. One empty line got me a login prompt:

Ubuntu 16.04.2 LTS bpi-iot-ros-ai ttyS0

bpi-iot-ros-ai login:

That’s the vendor image, so every number below is for the vendor BSP, not mainline. Fine by me: that’s exactly the comparison my mainline notes were missing.

A small script of mine logs in and runs each command with echo off, so the output comes back clean. Only read commands:

$ uname -a
Linux bpi-iot-ros-ai 3.4.39-BPI-M3-Kernel #1 SMP PREEMPT Thu Jul 13 09:32:18 UTC 2017 armv7l ...
$ cat /proc/device-tree/model
cat: /proc/device-tree/model: No such file or directory

No device tree at all: this kernel takes its board description from script.bin. Then the numbers: RAM 2009 MiB, eMMC 7.3 GiB, the vendor boot0 (eGON.BT0) at sector 16. All eight Cortex-A7 r0p5 cores online, one cpufreq policy with 11 OPPs from 480 to 2016 MHz, governor ondemand, 840 mV on the CPU rails at idle, 54 °C.

The two-cluster question turned out to have a nice answer. CPU4 to CPU7 first fail with -22 and come up a bit later. The address a secondary core lands on sits in a “soft entry” register, and I could read it:

$ busybox devmem 0x01F01DA4 32
0x400201A0
$ grep mcpm_entry_point /proc/kallsyms
c00201a0 T mcpm_entry_point

0xC00201A0 virtual is 0x400201A0 physical. So the vendor kernel parks the wake-up address of every secondary core there, and mainline uses the very same register (mc_smp.c), just with its own entry code.

Soft-entry register 0x01F01DA4 holds the physical address of mcpm_entry_point How a sleeping core finds its way back: 0x01F01DA4 holds 0x400201A0, which is mcpm_entry_point (0xC00201A0 virtual) on the vendor kernel.

Last, FEL. The kernel has no symbol with efex or fel in it, so there’s no “reboot into FEL” here: it’s the U-Boot key or nothing. I held the key and replugged OTG, and the PC saw 1f3a:efe8 ... in FEL/flashing mode. xfel version answered with SoC ID 0x1673 (A83T), full speed.

What I skipped on purpose: no CPU hotplug, no fsck, no writes. The boot log has two EXT4-fs error (device mmcblk0p2) lines at 19.9 s and 23.5 s, so the root file system was already damaged before I touched it.

Result

All 8 cores come up on the vendor kernel, with 11 OPPs from 480 to 2016 MHz. So the “1.8 GHz” on the spec sheet is the boot speed, not the top. At idle it sat at 840 mV and 54 °C.

The second cluster wakes through the soft-entry register 0x01F01DA4, which holds 0x400201A0, the physical address of mcpm_entry_point. FEL works, but only with the key, and only at USB full speed (12 Mbit/s). Small extra: the board clock was 2 h 8 min ahead of the PC (no RTC battery, no NTP).

What I’d do next

Boot the mainline image and read the same register: it should then hold the physical address of sunxi_mc_smp_secondary_startup. If it still says 0x400201A0, mainline’s SMP code didn’t run. Also read the dmesg lines between “Brought up 4 CPUs” and the late CPU4 start, which I didn’t capture.