Embedded Rust or C firmware? Lessons from an industrial microcontroller use case
arxiv.org
arxiv.org
> Two teams concurrently developing the same functionality (one in C, one in Rust) are analyzed over a period of several months. A comparative analysis of their approaches, results, and iterative efforts is provided. The analysis and measurements on hardware indicate no strong reason to prefer C over Rust for microcontroller firmware on the basis of memory footprint or execution speed. Furthermore, Ariel OS is shown to provide an efficient and portable system runtime in Rust whose footprint is smaller than that of the state-of-the-art bare-metal C stack traditionally used in this context. It is concluded that Rust is a sound choice today for firmware development in this domain.
This is even less serious than the typical pattern of grabbing random students for experiments and then drawing conclusions about the general population.
My understand is that both these things are in work, and that neither of these things exist yet.
I’m sure at some point there will be an accepted formal verification toolchain for rust, I hope to never use it.
Edit: With c, you can do memory safety analysis on all system libraries and entire Linux kernel. Some OS kernels, libs and languages do not have dynamic memory allocation at all!
Some languages are memory safe! Learn more about embedded programming!
But I think the lack of a formal specification is really not as big a deal as it's made out to be. It's one of those "think of a technical reason to justify a decision I've already made" excuses.
Obviously it would be great if Rust does get a full formal specification but I think avoiding it because it doesn't is just silly. C++ has a formal specification... which frequently has bugs and ambiguities. They aren't magically right and either way you're going to need to do a lot of non-formal testing as well as formal verification if you want confidence in a design.
This is true even for domains where formal verification is routine like SystemVerilog. I've seen designs pass formal but fail in simulation or vice versa due to subtle differences in the semantics. (Hopefully that can't happen for Rust but you get the point.)
Rust is fantastic for embedded. There are no hard obstacles. The reason to do it IMO is not memory safety, but because holistically the language and tools are (to me) nicer. Enums, namespacing, no headers, `cargo run --release` "just works". (I have found, at least in OSS, compiling C embedded project is a mess. Linker errors, you need to have a certain OS with certain dependencies, there are many scripts etc). Good error messages, easy to structure your programs in a reliable way etc. Overall, I just find it to be a better designed language.
I have found the most fundamental tooling for rust on Espressif's RiscV, and Cortex-M ARM for various STM-32 variants to be great. The cortex-m crate, defmt, and probe-rs, and the PAC project is fantastic.
On the down side, I have have to build my own tooling. I wrote and maintain my own HAL for STM32, and have had to write my own libraries for every piece of hardware. This comes with the territory of a new language, and suspect this will gradually improve this time - especially with vendor support. Because the fundamental libraries are around, this is just reading datasheets and making rust functions/structs etc that do the MMIO as described in the datasheets. Can be tedious (Especially if building a complete library instead of implementing what you need for a given project), but is not an obstacle.
My most complicated rust embedded firmware was a FPV-style UAS. I did it without an RTOS, using interrupt-based control flow.
Furthermore, the code I’ve produced in Rust is generally as fast (or faster) than the code I’ve written in C for the same task, and it’s easier and faster to write.
And I also maintain an open source HAL for an STM32 family! Previously those have just been in house HALs in C because there were no such community efforts.
I think in the future if I do an embedded rust project on a new MCU where there isn't an existing HAL, or one that is more work to repair than start over, I would just implement the subset needed for a project's reqs. Easier to keep track of scope that way. Currently the challenge is "X periph on Y variant of Z STM-32 family doesn't work under A condition" or "Doesn't work after this PAC update changed the syntax".
I would say however that there's still toolchain issues here. There all kinds of MCUs that simply don't/won't have a viable compiler toolchain that would support Rust.
e.g. I recently came from a job where they built their own camera board around an older platform because it offered a compelling bundle of features (USB peripheral support and MIPI interface mainly). We were stuck with C/C++ as the toolchain there, as there was no reasonable way to make this work with Rust as it was a much older ARM ISA
This conclusion was reached with a single experiment.
> Two teams concurrently developing the same functionality — one in C, one in Rust — are analyzed over a period of several months.
> Furthermore, Ariel OS is shown to provide an efficient and portable system runtime in Rust whose footprint is smaller than that of the state-of-the-art bare-metal C stack traditionally used in this context.
> The authors thank Davide Aliprandi and Davide Sergi of the STAIoTCraft team, and the wider Ariel OS team.
So one team had Ariel OS developer support, and it's unclear what support the other team had. Seems fair.
In Figure 12, they simply stop optimizing the code once desired rate is reached. Just at the end of the project the Rust firmware gets over a third performance boost, most likely from their OS developers.
Additionally, there is a claim that "Ariel OS is shown to provide an efficient and portable system runtime" - but there are no real tests for portability are conducted. Worst still:
> Where C-based projects require a separate project setup and manual code copying per target, Rust on Ariel OS consolidates everything within a single project [..]
This claim is just not true. This sounds like somebody that is not as familiar with C.
No shit. This is the conclusion reached at the conclusion of this experiment. This part of your comment can be removed with no loss of clarity, I think.
If I ran an experiment where I gave a cancer patient bread, and then they recovered from cancer, I couldn't then say: "It is concluded that <bread> is a sound choice today for <cancer treatment> in this domain.". You would rightfully jump up and down and demand further experiments to increase the confidence of the result before drawing the conclusion.
It could have been concluded instead that there is a case for further experiments to be conducted, or that Rust could be approaching a maturity where it could be considered for some firmware projects. But as it stands, the conclusion is far too strong given the experiments performed.
Yes. The goal was to handle the maximum data rate of the used sensor, and stop there. Time was limited on both ends.
> Just at the end of the project the Rust firmware gets over a third performance boost, most likely from their OS developers.
The ST intern found those boosts all by himself. They compared the exact MCU & peripheral initialization of the C and Rust firmwares, tightened I2C timings (where STM Cube has vendor tuned & qualified values), and enabled the MCU's instruction cache, which somehow is not default in Embassy's HAL. We were quite impressed actually, the last days before the deadline were quite productive, optimization wise.
I understand, and I understand that there were limits to what could be done with the resources there were. What irks me is the strength of the claim made without enough evidence to make it.
> The ST intern found those boosts all by himself. They compared the exact MCU & peripheral initialization of the C and Rust firmwares, tightened I2C timings (where STM Cube has vendor tuned & qualified values), and enabled the MCU's instruction cache, which somehow is not default in Embassy's HAL. We were quite impressed actually, the last days before the deadline were quite productive, optimization wise.
Fair enough, hats off to the intern. This kind of thing is common in MCUs, even on low-end CPUs weird defaults can be selected. But the involvement and influence of the OS developers remains unclear.
Again, there's just not enough data to make such strong claims. I think the paper could easily make recommendations, it could say that at least in some cases (as evidenced) Rust could be a reasonable choice, and it could make an argument for further work.
Megatools is an example of such a code https://xff.cz/megatools/ / https://xff.cz/git/megatools/tree/lib/sjson.c
We're mentioning this in the paper: "The heap is entirely attributable to Parson's dynamic allocation of JSON tree nodes; as memory usage minimization was not a key goal, we kept Parson (the JSON parser used by the PNPL code generator by default), noting that there are less memory heavy options that do not require a heap at all."
> The analysis and measurements on hardware indicate no strong reason to prefer C over Rust for microcontroller firmware on the basis of memory footprint or execution speed.
I admit I have not carefully read the paper, and am collating info from comments here, so I may be fully mistaken. The word "strong" also allows for much interpretation, that I'm not a priori critical of, but am skeptical of.
Yeah of course. Then again - they were one person teams, where the C "team" had years of experience in stm32 / embedded C / stm32 cube development and churned out that handwritten state machine in just days. The Rust "team" was a pre-masters intern with only minimal embedded Rust experience. They ran into all the pitfalls with (async) embedded Rust, but corrected towards the end.
I do applaud you for having the same work done twice but it would have been far more meaningful to have two actual teams of seasoned developers do this sort of thing side-by-side. The biggest item on the checklist would be the number of undiscovered UB or UB related bugs in the C codebase and to compare that with the Rust codebase on 'defect escape rate' or some other meaningful metric.
I've got my own set of restrictions for when I'm coding in C based on many nights spent poring over various pieces of code and trying to find a way to do it better and safer without outright switching languages. I do believe it is possible. But at the end of all that you have essentially redefined the language in a way that probably no other C programmer would like or agree with, and it would still require very good discipline.
So having languages with fewer footguns is good, as long as the lack of one kind of footgun isn't replaced by a other kinds of footguns. It is one of the reasons I'm interested in the FIL-C project.
My personal view is that good C code looks a lot like Rust where ownership is clear and a borrow checker would approve. The mindset that Rust forces you into is the same one you should be using when writing C.
The longer-term concern is that, if you’re spending late nights learning Rust, it’s probably with the borrow checker. Late nights with C, it’s probably with memory management. One of those two is a bit more applicable to understanding computing at a deeper level.
2. Well, ST has released official Rust drivers for a bunch of their sensors. They're built on embedded-hal(-async), so can directly be used with Ariel OS. There is probably more.
I mean sometimes efficiency matters a lot, but a lot of other times, interoperability is more important.
Text based IO with microcontrollers over tty has been quite a standard thing even decades ago.
> The command-response protocol and binary data format are described in device models generated using DTDLv2 [10], a JSON-based language for describing digital twins. These models are used within the Azure IoT Plug and Play (PnP) framework [11], which STAIoTCraft adopts for datalogging
I’d guess that’s an area where C tooling is pretty far ahead of Rust tooling at present?
"Author's" is possessive; "authors" is plural.
Memory layouts are often pre planned, with hand-written linker files, which are sometimes even necessary, as there are quirks like DMA only being able to access certain addresses etc.
In embedded, engineers often want hard real time guarantees, very high (essentially unfailing) reliablity, at the lowest possible price points.
Dynamic memory allocation is often no good - embedded allocators either waste RAM, have large code sizes, have worse runtimes than their desktop-grade cousins - something like jemalloc generates 10x as much code as the rest of your app, and assumes your heap is at least megabytes in size.
On the other hand, embedded-grade allocators often use algorithms that are prone to fragmentation, unpredictable runtimes, and tend to be less tested in general, while still wasting RAM and Flash.
Having an allocator that has a bad runtime behavior can basically ruin your hard runtime guarantees, and if you prealloc all you memory you will never run out - on the other hand, even the paper states they measured not calculated the max memory usage, meaning we have no knowledge of what the actual max is.
In contrast, when I go back and read Go or C code I wrote years ago, I have no trouble at all quickly figuring out what I was doing in the small programs I write.
The way these issues manifest themselves, as it recently did, was I went back to add a simple addition to a CLI tool I wrote for myself a year ago, and I was having trouble doing it because I couldn't quickly understand what I had been doing a year ago ... so I just had an AI agent do it for me. This was the kind of change that if it was a Go program, I would've done manually myself in about 5 or 10 mins.
- Sensor agent is such a rancid name for a remote sensor that I feel a need to public say so. Please don't use marketing names for things that already have more descriptive names.
- Rust uses a full RTOS and C uses the mediocre ST HAL (vendor specific). Immediately apples to oranges. Also I've never heard of the C JSON library and it looks sketchy at a glance so that will also hurt the comparison.
- Streaming slow sensor data with a 160MHz 786KB/2MB MCU is not a good test in the slightest. You could probably use something like micro python here and be done. No one is reaching for bare metal C here. Also no one serious about performance is using JSON serdes. If you're using bare metal C, you're likely trying to push the limits of your hardware or doing something so simple that you won't be tempted to reach for terrible third party libraries.
- Does the Rust code base use the 'unsafe' keyword anywhere, including the RTOS? If so, it's not memory safe without additional formal verification.
Overall I'd say this paper has approximately zero value wrt its stated goal of comparison.
-> paper is not final. And IIUC ST will be releasing the code at some point.
For me it comes down to the old standby: just vendor the .cargo build chain into your repo and be done with it. There you go. Lock in the year edition, the version, the features, the quirks, and the bugs. Just like game devs did with game engines or OpenGL or Direct X versions.