Table of Contents
I recently got a few Milk-V Duo boards to play with and thought I’d share my experience with them here.
The boards have a similar form factor to the Raspberry Pi Pico, but they run a full Linux distribution rather than custom firmware like the pico.
The first thing I wanted to try was using the GPIO pins just like you would on a pico or any microcontroller. The way Milk-V expects you to do this is using their modified version of the
wiringX library, which directly writes to memory mapped registers.
That’s not the recommended way to access GPIO on Linux because it’s error-prone and insecure. Instead, Linux provides special interfaces to control GPIO.
There’s the deprecated sysfs interface which you use by writing text to files in
/sys, for example,
echo 440 > /sys/class/gpio/export would allow you to access pin 440. This interface has several problems. For one, there’s no way to know if a program is currently using the pin, so you could end up with two programs trying to control the same pin. Also, when the programs finish, they have to manually unexport the pins, and if they don’t, they’ll stay exported even after the program exits.
That problem was solved by the character device interface. Linux provides character device files at
/dev/gpiochipX, on which you can use special
ioctl commands, in order to control the GPIO pins. Once a program closes the file or exits, Linux automatically releases all the GPIO lines it was using.
Later, it was decided that the API was inadequate, so there’s a v2 character device API, which is what I wanted to use. The recommended way to do that is using
libgpiod, which is a C library that invokes the various syscalls for you and provides a clean API.
However, that didn’t sound fun. I’d just be importing a C library, and besides, I was really bored of using C and wanted to learn a new low-level language, like Zig. So, I decided that in order to learn Zig, I’d implement a GPIO library in it as my first project. A bit complicated for a first project, but very fun.
Writing the GPIO library #
I started by searching for examples of using the character device API. Unfortunately, what little information I could find was for the deprecated v1 API which, to be fair, still works, but I wanted to use the v2 API, so I decided to read the kernel’s source code to figure out how these ioctls worked. Specifically, I found the
gpio.h file in Linux kernel 5.10 (the first version that implements the v2 API).
That file has very good descriptions of what the various structs and fields are meant to be used for, but it doesn’t provide any actual examples of how to use the API. I decided to just try it anyway.
I started by rewriting each struct as a Zig
extern struct. In the process, I learned a lot of interesting stuff about Zig, such as
packed structs, for example, which make bit sets really easy to implement because each
bool is stored as one bit, so you can just have a struct of
bool fields that will act as a bit set.
I also learned a lot about Zig’s type system as I had to do things like accept an unknown number of offsets to request from the chip, use enums, unions, etc.
Eventually, I finished that and started implementing the actual syscalls. The kernel represents ioctls as macros, such as
_IOWR(0xB4, 0x05, struct gpio_v2_line_info). This macro contains the ioctl type (
0xB4), the ioctl number (
0x05), and the type which gets sent or received (
struct gpio_v2_line_info). The kernel uses all of these in order to calculate an ioctl ID that you use when running the syscall.
Zig provides functions to do the same thing. For example, the
_IOWR macro above becomes
std.os.linux.IOCTL.IOWR(0xB4, 0x05, LineInfo). I used that to implement each of the ioctls implemented by the gpio interface.
Doing this gave me a low-level base upon which I could build my public API.
However, I later found out that I misinterpreted the API. Requesting GPIO lines requires offsets, while operations on GPIO lines, such as setting the values or config, requires indices. An index corresponds to the index of the offset in the request. For example, if you request the offsets
[22 20 21], then
22 will be
20 will be
21 will be
Because I didn’t realize this, my GPIO library wasn’t working, so I decided to compile a Go program using the gpiod library and then used
strace to analyze the syscalls it made and see how they were different from mine. That’s when I noticed that it was using indices rather than offsets. Once I fixed that, my library finally started working, so I finished writing it, set up zig package manager so it could be used by others, and published it.
Here are the links to my Zig GPIO library:
zig-gpio with the Milk-V Duo #
The first thing I wanted to try was to make the Duo’s LED blink, which is like the “Hello World” of GPIO projects.
To do that, I needed to know which chip controlled the LED and what offset it was at. Unfortunately, there’s no documentation for that. However, there is an example script that uses the deprecated sysfs interface to blink the LED. That script uses pin
440 as the LED, which I correlated with
gpiochip values in sysfs to determine that the LED was offset 22 on
I tried it, and it worked! The LED was blinking using entirely my code, which made me really excited. (I’m probably the only person in the world who gets excited over a blinking LED)
So now I needed the rest of the GPIO offsets so I could do some more testing
Finding the Milk-V Duo GPIO offsets #
Since Milk-V doesn’t provide any documentation about the offsets, I had to look further. I found the GPIO Operation Guide from the manufacturer of the chip that the Duo used, which let me know how to figure out the GPIO offsets and which chip they belong to.
To do that, I had to look at the Duo’s schematic, which had the prefixes
GPIOA would correspond to
gpiochip0, so I tried it with offset 28, which corresponds to
GP0 on the official Duo pinout. Sure enough, my multimeter showed that it was blinking.
Next, I assumed that
GPIOC would correspond to
C is the third letter), and tried that, and it worked again!
Since 0 and 2 are known, I assumed
GPIO with no letter would be
gpiochip1. However, when I tried it, nothing happened, so I decided to check
gpiochip4, and that did work. It turns out
PWR_GPIO is also
gpiochip4, so I’m not sure why there’s a difference in the naming.
Anyway, I used this to calculate all the GPIO offsets, which I’ll provide below for anyone who wants to use them.
GPIO offset table #
The Milk-V Duo is a really fun board to play with, and I encourage anyone who’s interested in Linux and embedded programming to get one and try it out. You’ll probably learn a lot and it’ll be a lot of fun!
They cost only $9, and you can usually find them for as low as $5. All you need is a USB-C cable and a microSD card. No serial adapter or debugger is required (though a serial adapter is nice to have for troubleshooting boot issues).