Building Rust code for my OpenWrt Wi-Fi router

I recently got interested in running Rust code on my router (stay tuned for a future blog post?). This is supposed to be easy, but I never tried it, so let's see how it goes.

A test project

Hello, worlds are somewhat boring, so I'd like to build something more realistic — a DNS client. We're not going to implement DNS here, but rather piggy-back on the trust-dns-resolver crate, which looks pretty good.

After skimming the trust-dns and tokio docs, and a cargo new delve (shouts to dig and drill), we have this:

name = "delve"
version = "0.1.0"
edition = "2021"

anyhow = "1.0"
tokio = { version = "1.18", features = ["net", "rt-multi-thread"] }
trust-dns-resolver = { version = "0.21", features = [
] }
use std::env;

use tokio::runtime;
use trust_dns_resolver::{
    config::{ResolverConfig, ResolverOpts},
    proto::{rr::RecordType, xfer::DnsRequestOptions},

async fn run() -> Result<(), anyhow::Error> {
    let resolver =
        TokioAsyncResolver::tokio(ResolverConfig::cloudflare_tls(), ResolverOpts::default())?;
    let query = env::args().nth(1).unwrap();

    let response = resolver
        .lookup(query, RecordType::A, DnsRequestOptions::default())

    for address in response.iter() {
        println!("{}", address);

fn main() -> Result<(), anyhow::Error> {
    let runtime = runtime::Builder::new_multi_thread().enable_all().build()?;
    runtime.block_on(async { tokio::spawn(run()).await })?

The code is pretty simple. It creates a tokio multi-threaded Runtime, sets up a trust-dns TokioAsyncResolver, then queries the Cloudflare DNS server for the hostname given in the command line. RecordType::A means asking for an IPv4 address record.

If you're wondering about the block_on / spawn dance, apparently tokio prefers that you don't do a lot of work in block_on, but rather spawn a root future and do your stuff from there. It doesn't really matter in this case, and trust-dns-resolver helpfully provides a synchronous resolver, but it will matter in the project I have in mind.

And it does appear to work fine:

$ cargo run --release --
    Finished release [optimized] target(s) in 9.66s
     Running `target/release/delve`

Inside my router

My router is an AVM FRITZ!Box 4040 running OpenWrt. As far as embedded systems go, OpenWrt is pretty close to a Linux PC. I've already enabled SSH and added my key, but I don't know what architecture it's running. Fortunately, uname -a works just as expected:

root@OpenWrt:~# uname -a
Linux OpenWrt 4.14.171 #0 SMP Thu Feb 27 21:05:12 2020 armv7l GNU/Linux
root@OpenWrt:~# cat /proc/cpuinfo
processor	: 0
model name	: ARMv7 Processor rev 5 (v7l)
BogoMIPS	: 67.03
Features	: half thumb fastmult vfp edsp neon vfpv3 tls vfpv4 idiva idivt vfpd32 lpae evtstrm
CPU implementer	: 0x41
CPU architecture: 7
CPU variant	: 0x0
CPU part	: 0xc07
CPU revision	: 5

processor	: 1
model name	: ARMv7 Processor rev 5 (v7l)
BogoMIPS	: 67.03
Features	: half thumb fastmult vfp edsp neon vfpv3 tls vfpv4 idiva idivt vfpd32 lpae evtstrm
CPU implementer	: 0x41
CPU architecture: 7
CPU variant	: 0x0
CPU part	: 0xc07
CPU revision	: 5

processor	: 2
model name	: ARMv7 Processor rev 5 (v7l)
BogoMIPS	: 67.03
Features	: half thumb fastmult vfp edsp neon vfpv3 tls vfpv4 idiva idivt vfpd32 lpae evtstrm
CPU implementer	: 0x41
CPU architecture: 7
CPU variant	: 0x0
CPU part	: 0xc07
CPU revision	: 5

processor	: 3
model name	: ARMv7 Processor rev 5 (v7l)
BogoMIPS	: 67.03
Features	: half thumb fastmult vfp edsp neon vfpv3 tls vfpv4 idiva idivt vfpd32 lpae evtstrm
CPU implementer	: 0x41
CPU architecture: 7
CPU variant	: 0x0
CPU part	: 0xc07
CPU revision	: 5

Hardware	: Generic DT based system
Revision	: 0000
Serial		: 0000000000000000
root@OpenWrt:~# ldd --version
musl libc (armhf)
Version 1.1.24
Dynamic Program Loader
Usage: ldd [options] [--] pathname

So it appears a 4-core ARMv7 CPU with hardware floating-point support, running a MUSL-based distro.


Rust supports dozens of targets, but ARMv7 is pretty common, so it's hopefully well-supported. I'm not sure how the target is called, so rustup target list is handy:

$ rustup target list | rg armv7
armv7-unknown-linux-musleabihf # sounds like a winner
$ rustup target add armv7-unknown-linux-musleabihf
info: downloading component 'rust-std' for 'armv7-unknown-linux-musleabihf'
info: installing component 'rust-std' for 'armv7-unknown-linux-musleabihf'

Great, let's try it!

$ cargo build --release --target armv7-unknown-linux-musleabihf
error: linking with `cc` failed: exit status: 1
   [snip half a screenful of errors]

Oh, we also need a linker. This would normally be a version of the BFD linker (ld). Unlike clang and rustc, gcc and ld only support one target at a time, so I'd have to track down or compile a compatible version. My Linux distro actually has one, available in AUR as muslcc-arm-linux-musleabihf-cross-bin.

But since we're compiling pure-Rust code — did you notice the fancy rustls feature of trust-dns-resolver? — the LLVM linker, lld, will do the trick with less work.

The correct incantation is then:

$ CARGO_TARGET_ARMV7_UNKNOWN_LINUX_MUSLEABIHF_LINKER=rust-lld cargo build --release --target armv7-unknown-linux-musleabihf
    Finished release [optimized] target(s) in 9.50s

Is that all? That was surpisingly easy!

And by the way, you can also set the linker in .config/cargo.toml, like this:

linker = "rust-lld"


We could test using QEMU, but that seems too much of a hassle, as the real hardware is already available.

$ scp target/armv7-unknown-linux-musleabihf/release/delve [email protected]:~
scp: Connection closed

The router is running OpenSSH 8.0p1, but scp is deprecated, and the version on my PC uses SFTP by default. The -O flag reverts to the deprecated protocol:

$ scp -O target/armv7-unknown-linux-musleabihf/release/delve [email protected]:~
delve                                        100% 8598KB 176.1KB/s   00:48

That's a larger binary than I'd like, but 176 KB/s seems pretty slow. I know my network is faster that this, but it could be the file system:

root@OpenWrt:~# mount
/dev/root on /rom type squashfs (ro,relatime)
proc on /proc type proc (rw,nosuid,nodev,noexec,noatime)
sysfs on /sys type sysfs (rw,nosuid,nodev,noexec,noatime)
tmpfs on /tmp type tmpfs (rw,nosuid,nodev,noatime)
/dev/mtdblock14 on /overlay type jffs2 (rw,noatime)
overlayfs:/overlay on / type overlay (rw,noatime,lowerdir=/,upperdir=/overlay/upper,workdir=/overlay/work)
tmpfs on /dev type tmpfs (rw,nosuid,relatime,size=512k,mode=755)
devpts on /dev/pts type devpts (rw,nosuid,noexec,relatime,mode=600,ptmxmode=000)
debugfs on /sys/kernel/debug type debugfs (rw,noatime)

Oh, okay, let's copy it to /tmp instead:

$ scp -O target/armv7-unknown-linux-musleabihf/release/delve [email protected]:/tmp
delve                                        100% 8598KB  11.6MB/s   00:00

More importantly, does it work?

root@OpenWrt:~# /tmp/delve

Honestly, I was more suprised to see it working than you are.

What about the binary size?

Our executable packs in quite a bit: Tokio, an async DNS client, and a TLS implementation for DNS-over-TLS. But at 8.5 MB, it's pretty large for a Wi-Fi router with 128 MB of flash.

Let's see if we can get it to weigh less, testing with the PC version. I know that a large part of the binary must be the symbols. I'd normally use strip -s, but cargo can do this by itself. This is also a good excuse to try the custom profiles feature.

Let's start by adding a new profile to Cargo.toml:

inherits = "release"

, then build it with cargo build --profile minsize.

There's a lot of resources out there with tips for reducing the binary sizes (it's a common complaint), so I'll just list each thing I've tried, incrementally:

Rebuilding it for the router, we get a 1.4 MB binary, which is still a bit large, but workable.

Bonus: packaging for OpenWrt

I was curious about making a binary package for OpenWrt, which appears to use a format inspired by Debian (ipkg). Reading the docs isn't fun, but we can find where the package manager (okpg) downloads stuff from (/etc/opkg/distfeeds.conf), then get a random package from there.

I picked attr_20170915-1_arm_cortex-a7_neon-vfpv4.ipk, which appears to be a .tar.gz archive, containing three files:

control.tar.gz has one metadata file, control, and some pre- and post-install scripts we can copy over or ignore:

Package: attr
Version: 20170915-1
Depends: libc, libattr
Source: feeds/packages/utils/attr
License: LGPL-2.1 GPL-2.0
LicenseFiles: doc/COPYING doc/COPYING.LGPL
Section: utils
Maintainer: Maxim Storchak [email protected]>
Architecture: arm_cortex-a7_neon-vfpv4
Installed-Size: 11080
Description:  Extended attributes support
 This package provides xattr manipulation utilities
 - attr
 - getfattr
 - setfattr

Interestingly, my router uses vfpv4 packages, even though only vfpv3 appears in /proc/cpuinfo. Well, whatever makes it happy. The Installed-Size field is also strange, as it doesn't seem to match the size of the files. It's optional optional, so it doesn't matter too much.

We can make a similar file:

Package: delve
Version: 0.0.1
Depends: libc
License: MIT
Section: utils
Architecture: arm_cortex-a7_neon-vfpv4
Installed-Size: 1414464
Description: DNS testing tool

Finally, data.tar.gz contains the installed tree. Putting things together, we have:

├── control
├── control.tar.gz
├── data.tar.gz
├── debian-binary
├── delve_0.0.1_arm_cortex-a7_neon-vfpv4.ipk
└── usr
    └── bin
        └── delve

2 directories, 6 files

And on the router:

root@OpenWrt:/tmp# opkg install delve_0.0.1_arm_cortex-a7_neon-vfpv4.ipk
Installing delve (0.0.1) to root...
Configuring delve.
root@OpenWrt:/tmp# delve

If you ever try this, make sure to actually compress the archives using gzip (or tar czf). I forgot that tar only guesses the format when extracting, and opkg install greeted me with a fun Segmentation fault message.

Naming apologies

I just realized that the delve name is already taken by a Go debugger. Not that it matters, as nobody will be using this, but it's a common source of complaints.

Final words

To sum up, we wrote a simple DNS client, tweaked things a little to reduce the binary size, cross-compiled it for an ARM Wi-Fi router, packaged and tested it there. Surprisingly, it actually worked.

Thanks for staying with me, and if you found this interesting please consider buying me a coffee.