#async #operating-system #os #resources

no-std lilos

A tiny embedded OS based around Futures and async

19 releases (4 stable)

1.2.0 May 5, 2024
1.1.0 Apr 29, 2024
1.0.0-pre.2 Mar 6, 2024
0.3.6 Jul 11, 2023
0.1.2 Apr 29, 2021

#81 in Embedded development

Download history 121/week @ 2024-02-26 133/week @ 2024-03-04 93/week @ 2024-03-11 64/week @ 2024-04-01 527/week @ 2024-04-22 394/week @ 2024-04-29 24/week @ 2024-05-06 9/week @ 2024-05-13 18/week @ 2024-05-20 13/week @ 2024-05-27 14/week @ 2024-06-03

54 downloads per month
Used in 4 crates

MPL-2.0 license



This is a wee operating system written to support the async style of programming in Rust on microcontrollers. It fits in about 2 kiB of Flash and uses about 40 bytes of RAM (before your tasks are added). In that space, you get a full async runtime with multiple tasks, support for complex concurrency via join and select, and a lot of convenient but simple APIs. (If you want to see what a lilos program looks like, look in the examples directory, or read the intro guide.)

lilos has been deployed in real embedded systems since 2019, running continuously. I've built about a dozen systems around it of varying complexity, on half a dozen varieties of microcontroller. It works pretty okay! Perhaps you will find it useful too.

See the repository for example code and getting started instructions, or the Rustdoc on the top level lilos module for API details and descriptions of modules.


A simple but powerful async RTOS based around Rust Futures.

This provides a lightweight operating environment for running async Rust code on ARM Cortex-M microprocessors, plus some useful doodads and gizmos.

lilos is deliberately designed to be compact, to avoid the use of proc macros, to be highly portable to different microcontrollers, and to be as statically predictable as possible with no dynamic resource allocation.

These are the API docs for the OS. If you'd like a higher-level introduction with worked examples, please have a look at the intro guide!

About the OS

lilos is designed around the notion of a fixed set of concurrent tasks that run forever. To use the OS, your application startup routine calls exec::run_tasks, giving it an array of tasks you've defined; run_tasks never returns.

The OS provides cooperative multitasking: while tasks are concurrent, they are not preemptive, and are not "threads" in the traditional sense. Tasks don't even have their own stacks -- they return completely whenever they yield the CPU.

This would be incredibly frustrating to program, were it not for Future and async.

Each task co-routine must be a Future that can be polled but will never complete (because, remember, tasks run forever). The OS provides an executor that manages polling of a set of Futures.

Rust's async keyword provides a convenient way to have the compiler rewrite a normal function into a co-routine-style Future. This means that writing co-routines to run on this OS looks very much like programming with threads.

For detailed discussion and some cookbook examples, see either the intro guide or the examples directory in the repo.

Feature flags

lilos currently exposes the following Cargo features for enabling/disabling portions of the system:

  • systick (on by default). Enables reliance on the ARM M-profile SysTick timer for portable timekeeping. Disabling makes the executor smaller at the cost of losing all time API. On platforms where the SysTick timer stops during sleep, such as Nordic nRF52, you may want to disable this feature and use a different timekeeping mechanism.

  • mutex (on by default). Enables access to the mutex module for blocking access to shared data. Leaving this feature enabled has no cost if you're not actually using mutex.

  • spsc (on by default). Enables access to the spsc module for single-producer single-consumer inter-task queues. Leaving this feature enabled has no cost if you're not actually using spsc.

Composition and dynamic behavior

The notion of a fixed set of tasks might seem limiting, but it's more flexible than you might expect. Because Futures can be composed, the fixed set of OS tasks can drive a dynamic set of program Futures.

For instance, a task can fork into several concurrent routines using macros like select_biased! or join! from the futures crate.

Concurrency and interrupts

The OS supports the use of interrupt handlers to wake tasks through the Notify mechanism, but most OS facilities are not available in interrupt context.

By default, interrupts are masked when task code is running, so tasks can be confident that they will preempted if, and only if, they await.

Each time through the task polling loop, the OS unmasks interrupts to let any pending interrupts run. Because the Cortex-M collects pending interrupts while interrupts are masked, we don't run the risk of missing events.

Interrupts are also unmasked whenever the idle processor is woken from sleep, in order to handle the event that woke it up.

If your application requires tighter interrupt response time, you can configure the OS at startup to permit preemption of various kinds -- including allowing preemption by only a subset of your interrupts. See the exec module for more details and some customization options.


Co-routine tasks in this OS are just Futures, which means they can be dropped. Futures are typically dropped just after they resolve (often just after an await keyword in the calling code), but it's also possible to drop a Future while it is pending. This can happen explicitly (by calling drop), or as a side effect of other operations; for example, the macro select_biased! waits for one future to resolve, and then drops the others, whether they're done or not.

This means it's useful to consider what cancellation means for any particular task, and to ensure that its results are what you intend. There's more about this in The Intro Guide and the technical note on Cancellation.

lilos itself tries to make it easier for you to handle cancellation in your programs, by providing APIs that have reasonable behavior on cancel. All of the core APIs aim for strict cancel-safety, where dropping a future and retrying the operation that produced it is equivalent to not dropping the future, in terms of visible side effects. (Obviously doing more work will take more CPU cycles; that's not what we mean by side effects.)

If some code is useful, but can't achieve strict cancel safety, it should go in a separate crate. For example, the lilos-handoff crate used to be part of the core OS, but was evicted because it can only achieve a weaker notion of cancel safety.

Any deviations from this principle are considered bugs, and should be reported if you notice them!


~24K SLoC