Is it time for a new Embedded Linux build system?
yoebuild.org
yoebuild.org
Like the article points out, the software stacks people use for embedded devices are the same as people use non-embedded use; Rust, Go, NodeJS, and sure still C++. The only real difference with embedded devices is non-OS components like the bootloader (u-boot, EDK2) and customizing the device tree. (And of course firmware flashing). Fundamentally those are all just packages that I can describe in Nix. I don't need a separate tool just because the board is small.
IMO the embedded space, especially in the US, is already pretty Niche. Most companies either just ship the vendors BSP example (Ubuntu/Debian/Yocto) and pay very little attention to the detail or re-useablity. Once you vendor declares the BSP EOL you are stuck unless you re-implement it yourself.
Using Nix (or Guix) has the massive advantage of a large and active community that isn't fractured like the Yoctoo/Buildroot community. (By fractured I mean there may by many, many people using those tools at $DAYJOB but due to vendor BSP customization they likely share much less with the upstream community maintained sources).
Because historically, and I would argue should still, refer to very limited hardware, in terms of memory, processing power, and energy.
[1] https://search.nixos.org/options?channel=26.05&query=boot.bi...
Because embedded usually means dealing with complete functionality provided by the soc, bare minimum for the init system to get quick boot times.
If you just want to boot the CPU and run some IoT app, device tree + kernel + uboot should be enough.
But for anything with a GUI, You'd probably need the vendor provided binary blobs for GPU accelerated UI, proper gstreamer packages with all the patches to make it work with the binary blobs... properly packaged Qt etc ... Not to mention read only rootfs and A/B partitions for upgrades to deal with power loss headaches.
I haven't used Nix for any of these but most of these things are available ready to use with the vendor provided buildroot or yocto setups. And it would've been just more work to get this all working with yet another build system/package manager
However, reading through the docs, the author clearly knows about Nix/Nixpks and has factored some of its properties into the design: https://docs.yoebuild.org/nix.html
I’d still probably just implement this in Nix/Nixpkgs so I could leverage the rest of the ecosystem but that’s not trivial at all.
I wouldn’t even say Nix as a language is better than Starlark. The author is certainly well informed in the choices made.
Adding new targets is deceptively easy, just copy an existing template and substitute your values.
https://codeberg.org/guix/guix/src/branch/master/gnu/system/...
https://codeberg.org/guix/guix/src/branch/master/gnu/bootloa...
$ my-arm-system = nixpkgs.lib.nixosSystem { system = "aarch64-linux"; modules = [ ./configuration.nix ]; };
$ :b arm-system.config.system.build.images.iso
nixos-rebuild build-vm (--flake DIR#HOSTNAME / configuration.nix)
./result/bin/SCRIPT
to build and launch a nixos vm with qemu.
build-vm is the same as building the configs “config.system.build.vm” attribute.
nixos-rebuild is fine for getting started but you end up needing to use the attribute path to do anything more complicated like CI caching or inspecting the drvPath.
For me the main downside of Guix is that the maintenance team is very small, which does not seem sustainable.
My impression in recent years is that these image cross build environments are just not as frequently needed as they were back in the day of their invention.
My most recent embedded linux environments were just embedded archlinux.
No need to cross build an image, just install and run the minimized linux environment right on the target.
Of course, a big part of the need for these cross-tools is that it seems most modern embedded linux developers are running windoze on their development workstations 8-/
At the end of the day you need some cross compilation just for board bring up.
If you're playing with some platform for which this has already been done, then sure, but that's not really the "normal" way of doing embedded.
The main reason to not run Debian is that Debian usually makes lots of compile-time decisions for you, and chooses to maximize the enabled options of the software. Alpine makes the opposite choices for you, creating very minimal feature sets.
It's been working well for me, but my use cases is a lot more like a custom in-house DDWRT firmware kind of a thing. If you've experienced additional constraints that make this pattern unworkable, I'd be interested to hear about them.
My preference is to cross build packages for each application on my workstation, and maintain a local package repository from which individual packages are installed on the target.
I find archlinux extremely useful in this configuration, because even if I'm not building on the target, the same package maintenance and other system commands are identical to my development workstation.
Additionally, archlinux provides a well documented, and fairly simple process for hosting your own local repositories.
1) Cross compile. Yoe build still does this where it makes sense. (Go apps, likely kernel builds in the future) 2) QEMU user mode - 5-20x slow, but fine for some things. 3) Run yoe-build on any ARM machine (AWS, rPI5, Jetson) 4) Farm unit builds out to runners on cloud ARM machines. (future)
The yoe-build architecture allows for this. Choose the container, host architecture, and location that makes the most sense on a unit-by-unit basis.
Caching is also part of the vision, so we never rebuild something twice, unless it changes.
This remote build runner is not implemented yet, just ideas.
- You're building on the same native system as GNU and Linux packages, you install them globally in the same places that servers and desktops use
- Your C, C++ compiler and entire toolchain and other binary utilities with the kernel is a one single unit that you can only change one part at a time
- You use the same up to date headers with glibc, gcc and Linux kernel
- You're building software in the same universe of all the other packages, especially gcc libraries (libgcc_s, libstdc++), glibc (especially bad since ld-linux.so is part of it)
- The build system only uses standard paths
The reason Yocto is so complicated is that developing in a Linux environment actually sucks when you're not writing web-oriented or server / VM software. Yocto fixes it. It introduces a good set of abstractions that work around terrible design decisions that were made in overall Linux ecosystem. There are a lot because the OS design is fundamentally broken, especially with C-based toolchains which is 99.999% of the ecosystem. Current C toolchains including MSVC strongly ties OS with the C's internal types and bad decisions of 70s.
As always all articles whose title asks a question are answered with NO, 99% time. By taking away the cross-compiling abilities and the workarounds doesn't fix the brokenness of Linux and overall FOSS ecosystem.
If you're looking for how a better embedded environment looks like, look at Rust toolchains. For Linux take a look at musl-libc based ones (you 100% need a systemd distro to get away from nss complexities that musl introduces). Or even better take a look at relibc. There are barely any assumptions about the target filesystem and tooling in Rust toolchains, unlike C/C++/Make toolchains. There is redox OS but it is still in slow development and they stuck with Make, which I think was a bad decision. Android uses its own build tooling but cannot run away from C/C++ tooling unless Google revives Fuschia.
Architecture emulation with qemu-userspace is 5-10x slower than running a native build.
The only 'proper' no-friction cross-compilation system I know is Debian/Ubuntu, because it allows to install foreign architecture libraries into your native system, and has all the wrappers for cross-compilation in dpkg-buildpackage.
This way you don't have to maintain cross-compiler or toolchain/sysroot, you just install whatever dependencies you need with the regular `apt install` of another architecture.
I'd like to have more friendly tooling around that instead of a slow compilation of architecture emulation.
Here'a s quick start for a C++ application compilation with libncurses6 and libtinfo6 dependencies as an example. `apt build-dep` will install armhf library dependencies and native other build dependencies.
# Adding armhf architecture support to existing system
dpkg --add-architecture armhf
apt update
# Installing cross-compiler for armhf
apt install crossbuild-essential-armhf
mkdir nload; cd nload
apt source nload
cd nload-0.7.4
# Download build dependencies for nload package, for armhf
apt build-dep -aarmhf nload
apt install libstdc++6:armhf # build-dep misses this one
# Cross-compiling for armhf, get .deb package for armhf
dpkg-buildpackage -a armhf -Pcross
Here's all manual process which involves tinkering with your running system, but there are 'proper' compilation tools like sbuild which bootstrap the (container) OS, configure it for cross-compiling, and do all the required steps from a clear start, with just a `sbuild --host=armhf` call more or less.
Yes, skipping builds from source can be faster, and, I don't know that you need to throw away yocto to do that https://rootcommit.com/pub/conferences/2024/elce/yocto-binar...
From the statement about "Caches builds so no piece of software is built twice." I'm guessing that the author has not enabled sstate caching https://docs.yoctoproject.org/dev/overview-manual/concepts.h...
I'm a little doubtful on the post's assertion that yocto is only needed for "deeply embedded regulated products." For products that have to follow the https://en.wikipedia.org/wiki/Cyber_Resilience_Act coming up, they will almost certainly need:
* SBOM management https://docs.yoctoproject.org/next/dev-manual/sbom.html
* CVE (or its successor) tracking https://docs.yoctoproject.org/dev/security-manual/vulnerabil...
* License management https://docs.yoctoproject.org/next/dev-manual/licenses.html
And I'll echo what some other folks have said here: if you don't need features like that, bundle your custom bootloaders and kernel with a binary distribution like Debian and call it a day.
Reversing and reimplementing the closed bits are what I want my Buildroot pipelines to do.
That’s my use case though; targeting COTS devices with (mainline) Linux support that don’t do all they could.
I have done it the ugly way (in production) and I want elegance in my old age.
yoebuild does seem interesting at least as a proof-of-concept.
When someone starts with "why cross-compiling the system if it's bearable to compile it directly on the device?", it really sounds like they are just not comfortable with cross-compilation. To me it would be similar to say "why would you need a compiled language, with this weird compilation step that nobody understands, when you can just write in an interpreted language?". Except that developers wouldn't generally proudly admit that they don't know how to use a compiler.
A valid reason I see to require a new embedded linux build system would be if the existing ones (mostly buildroot and Yocto) were unable to build a modern OS (e.g. by not being able to integrate some security features). But I don't think that's the case. This article sounds more like very verbose marketing for a system claiming that it can do similarly without having to learn as much (which I honestly doubt).
To me it's often the same pattern:
* Meson is nice, but it doesn't remove the need to use CMake, and it's not fundamentally simpler or more powerful (it's more comfortable at most).
* Jujutsu is nice, but it doesn't remove the need to learn git, and it's not fundamentally simpler or more powerful.
* Yoebuild sounds like even if it got traction like Meson and Jujutsu, it wouldn't remove the need to interact with Yocto for embedded linux engineers.
As an embedded linux engineer, if anyway I have to interact with Yocto and therefore learn Yocto, why would I learn another one that isn't fundamentally better? Usually people who go for those "simpler" (at least on the surface) tools do it because they don't want to learn Yocto. And for an embedded linux engineer, isn't it a red flag to not want to learn the most famous embedded linux build system?
It reminds me of Balena/Resin, which seems like it's targetting people who are not embedded linux engineers, don't want to become embedded linux engineers, but still want to distribute an embedded linux. To me that's generally a red flag.
You can make great things with that, but it was easy to be "holding it wrong".
I still recommend CMake at work, it's the least offensive solution in most situations.
Probably CMake is not a good fit for very, very big projects. But the vast majority of projects are not that big.
Usually what I see with CMake is self-inflicted pain, often where the devs start adding custom functions and mix it with python and stuff. If you keep it simple, it's very clear and does the job with surprisingly few keywords, I find. And I have setup tens of projects with CMake.
I agree this is true a lot of the time, speaking as someone who has made many Yocto messes and later regretted. But you also can't discount all of the messes made by silicon OEMs that you are forced to deal with to use their SoCs.
E.g. I recently did a project with Yocto on a Zynq US+. Look at the meta-xilinx layerset (or, God forbid, petalinux) and tell me they haven't made a complete mess of it.
The complexity and lack of upstreaming in kernel and userspace for so many of these SoCs, coupled with non standard boot process and devicetrees means you will be forced to use these terrible OEM layers. I don't think a new build system can fix this at all, it's a cultural/commitment that needs to change from the OEMs.
At the end of the day, of you are doing any complex embedded Linux work as a small team, you must use Yocto because it's what the OEM supports. I don't have time or resource to debug DFX not working because some userspace tool was built with slightly different flags than expected by the flaky ass xilinx upstream.
Yes, totally! It's a bit sad because we don't usually have the luxury to select the hardware based on the quality of the BSP. First because they all seem to be terrible, and second because it is somehow "hidden" cost. One selects the hardware based on its price and capabilities, and the software part is left as an exercise.
But as you say it's not a problem of Yocto, it's a problem of the OEMs being in a position to make a mess because it doesn't have an impact on them. A bit like Apple is terrible for developers (in my experience) because anyway developers don't have a choice.
Note, this is generated by Claude as I've not used Bazel and may not be 100% accurate. PR's welcome.
Instead of cross-compiling with a large meta-layer stack, the tool builds kernel, rootfs, and applications together using one engine, with a CLI, TUI, and AI assistant talking to the same core. All you need is the tool, Docker, and Git — no global SDKs or hidden state.
It’s pre-1.0 and rough around the edges; I’m sharing it early to get feedback from people who live in Yocto/OpenEmbedded, Buildroot, Nix, etc. I’d love to hear where this breaks on your boards, what workflows feel wrong, and whether the “native builds + AI-aware build graph” direction seems promising.
The workflow for user space can definitely improve some of this pain but I feel like a large portion of any embedded Linux development effort still ends up in the weeds for boot related items (secure boot, proper updates, nuanced kernel patches, bootloaders, device trees, and supporting machine variants, etc). Solving those to make them easy is a hard problem for sure.
This is the complete opposite way, actually.
We need cross-compiling that is just as effortless as native compilation.
You should be able to build complex software on a powerful computer and perform costly optimization, then run it on a low-powered device.
the various embedded build systems make cross-compiling easy at scale, which means you often don't need to think about it when adding a new package
Even if you don't have a beefy server of the same architecture, you can probably run it in qemu instead of docker to the same effect. And even if qemu is slow, you can run a build of the kernel and your in-house stuff in parallel on 64 cores and not really be affected by the qemu slowdown.
I'm interested to hear counterexamples, though.
What if this hegemony in ISAs is partially because of the high costs of porting software and toolchains?
Maybe the next generation of embedded devices will have ISAs that are, say, extremely optimized for real time industrial controllers, but terrible at compiling C++.
It seems a little silly to give up on cross-compilation just because _today_ everything runs on ARM.
Especially because cross-compilation is not a particularly hard problem, the problem is with decades of tooling assuming the build and target environments are the same.
I had been maintaining a fork of the U of KY BDR with SBCs and Arm devices in mind, might need to revive it for RISC-V.
"Ampere processors natively support both 32 & 64 bit Android applications and require no binary translation for maximum instance density" ... "Run up to 120+ 3D cloud game instances per socket" ... "DRAM: 384GB - 512GB"
https://amperecomputing.com/solutions/arm-native
I've not personally tested this, but it very much appears that just like x86, you can use a 64-bit ARM kernel that has 32-bit support enabled to run a 32-bit userspace, likely compatible with docker, or even just in a chroot. Then with 120 cores, "make -j 120"
I’m already sitting at an expensive computer, if everything else were equal I’d rather cross compile than buy more hardware.
I’m open to reasons that’s presently too hard of a workflow to enable so I’m (literally) paying for a penalty, but I don’t see how it’s a better idea.
As a long time user of Bitbake and Yocto, and watcher of Yoe Linux... I kind of have to respectfully disagree with the approach.
One look at the Python or Node packaging ecosystem, and you can see how difficult trying to integrate those in a 'sane' way in embedded, with repeatability and security in mind, nevertheless wrangling something like C/C++ dependencies and native packages in a qemu-cross (or binfmt emulation) environment.
I feel like Bitbake and the wider Yocto ecosystem has essentially 'solved' cross compiling. Sure there are the incredibly complex codebases like Chromium that require lots and lots of study and patience, or esoteric compilers, etc. but for most applications I feel that especially AI tooling can write up a good basic recipe for integration with Yocto.
The unfortunate thing about AI tooling is Kernighan's law (where debugging is twice as hard as writing it the first time.) Especially in Embedded Linux, where 99.99% of the product code is written by others, trying to figure out where the AI didn't quite get it right or missed something can break things unexpectedly and impossibly for the unaware developer to fix, even breaking at runtime. So the build environment has to be simple and predictable. I think Yocto/Bitbake already strikes a good balance here, with other nice things like offline builds, repeatable builds, sstate caching across machines, PRbuild servers... things useful for big teams, so consider those as well.
Although, if someone does a 'uv'-like rewrite of Bitbake into Rust with the same feature set but faster, I am all for it...
Maybe, having Yoe or some other centralized distro make specific opinionated choices and then provide an open sstate feed, working inside Bitbake, could get a lot of the speed and customizability gains back without a big rewrite effort.
RAM is expensive now. Reasonably priced seems to be counter to that fact.
The whole article feels like a promo-tour.
> The catch is that nobody handed us the tools to keep the place running.
Software such as homebrew exists. Why is the article ignoring this? Not only does it feel like promo, it now feels like propaganda. Homebrew is also just one example; LFS/BLFS also exists for embedded linux. Perhaps the quality is less than for desktop computer systems but it exists.
> Edge devices have started behaving like cloud systems. They run increasingly complex stacks, are frequently updated, and are managed remotely over their entire life.
Companies want to become dependent on others? Because that is what cloud is all about. Top-down control. A mafia for later blackmail rack-up-the-price strategies.
> That puts the cross-compile burden squarely on embedded developers, and maintaining recipes for thousands of packages has been a steady drain on the Yocto community
Seems untrue. If you have a compiler that works, you can compile. Thousands of packages? I am tracking 3888 right now. That covers really most software by far, including python pypi addons. If a single person can do that, a team has an even easier time. The whole article seems to narrate a story that is just made up.
> Building everything from source in Yocto means long builds, heavy memory use, and powerful workstations.
No it does not. I do so every day. You need a reasonably fast computer and RAM, which is now driven up by the existing hardware mafia that must be put in jail, but it is easily possible for solo persons too. Now imagine a team.
The whole article is really really crap. Absolute garbage. The claims it makes are for the most part simply not correct. It narrates a story as if all linux developers are incompetent. I highly doubt that. It also never mentions any existing software here that would change the rhetoric such as homebrew, LFS/BLFS and so forth. It feels as if that article is just engineered to tell a "we thus need xyz". That's not good writing. I mean just look at this:
"We can leverage these change agents and rethink the world of connected products."
What does that even mean? Did AI write this? Because humans are usually smarter when it comes to opinion pieces. "connected products" - everything is now connected?
I'm glad Yocto is working for you.
I'm by no means a Yocto expert like many of you, but I've used OpenEmbedded/Yocto since the start, so I'm quite familiar with it. I started the following project, which Khem mostly maintains these days:
And "I've been experiment with" seems to just mean "I've been vibe coding"
Yes, much of Yoe build is implemented using Claude Code, driven by specs.