I am still skeptical. In the late 90s or early 2000s Linus was interviewed on TV and what he said stuck with me to this day. When asked about competitors he roughly said:
No one likes writing device drivers and as long no one young and hungry comes along who is good at writing device drivers I am save.
I think he was already well aware at that time that keeping the driver interface unstable is his moat. A quarter of a century later kernels that run on virtualized hardware are a dime a dozen but practically useable operating systems in the traditional sense of abstracting away real hardware can still be counted on one hand.
Maybe we will have young and hungry AI-for-system researchers who would like to take on the job of developing AI agents that translate Linux drivers in C to Asterinas ones in (safe) Rust.
Another feasible approach is to reuse Linux drivers by running a Linux kernel inside some kind of isolated environments. For example, the HongMeng kernel leverages User-Mode Linux to reuse Linux drivers on HongMeng [1]. Asterinas can take a similar approach.
[1] https://www.usenix.org/conference/osdi24/presentation/chen-h...
(It's also why regressions are pretty common: it's completely infeasible to test all of linux on each release, some people test some parts of it, but it's all very ad-hoc, very little is automated, and it's not at all unified)
There was also an academic project that combined virtualization with Windows drivers.
Does Linus have/want a moat? He's not a tech startup founder. He's a kernel hacker who has had success beyond his wildest dreams, and whose needs will be met for the rest of his working life no matter what happens.
It seems like projection to talk about this aspect of the kernel as if it's some intentional strategy for preventing certain kinds of competition.
Ignoring performance or adoption, the main weaknesses are: abstraction, gap attacks; unsafe code bypassing the whole thing; compiler or JIT-induced breakage of safety model; common, hardware failures like cosmic rays. This is still far safer than kernels and user apps in unsafe languages.
One can further improve on it by using static analysis of unsafe code, ensuring all unsafe functions respect type-safe/memory-safe interfaces, compilers that preserve abstraction safety during integration, and certified compilers for individual components. We have production tools for all except secure, abstract compilation which (a) is being researched and (b) can be checked manually for now.
I think this is telling. There are plenty of 'standards' for interfaces in the hardware world. Some (mostly USB) are even 'followed', but the realities of hardware are that it never behaves nominally. So without someone willing to spend the time writing code to handle the quirks that can't be patched out and the errata it's very hard to run on physical hardware with any performance or support.
98% of Linux I interact with is running virtualized: on my desktop/laptop systems it’s either Virtualbox full-screened so I can use Windows for drivers, or a headless VM managed by Docker.app on my Mac. All my employer’s production workloads are AWS virtual machines.
My only Linux bare metal hardware is a home server, which I’m planning to retire soon-ish, replaced by a VM on an ebay Mac mini to reduce power bill & fan noise.
If someone can make a Linux compatible kernel that’s more secure and just as performant, it’s much easier these days to imagine a large new user base adopting it despite a dearth of drivers.
My point is that every virtualized environment needs a layer that talks to real hardware down below. We have enough diversity in the upper layers but not enough in the lowest layer.
[1] I heard it expressed like this from an Azul Systems employee first, but unfortunately don't remember who it was.
I don’t see any support for an os that doesn’t have good driver support for accelerators, whether GPU/TPU or otherwise. And if your look into some of the accelerators built into modern amd and intel chips that becomes a nightmare just supporting the CPU, never mind USB host controllers and network interfaces etc
Its the old guys who write the best drivers, naturally.
For me, Asterinas represents a refreshing way to approach some thorny problems in the embedded space, in which embedded-Linux on ARM, RISC-V and MIPS is a viable, economically-speaking, platform for a great deal of industry.
While Asterinas is really sexy, if this same approach were taken for, say, FreeRTOS as well along the way .. then there could at least, also, be "on one hands" worth of operating systems, abstracted, in the "lets just use rust' camp ..
It's basically like npm update, at the kernel level.
I thought newer microkernels... Reduced that? Fixed it? I forget, I just had the impression it wasn't actually that bad except that the industry is still traumatized by Mach.
From the project website:
> Only the privileged Framework is allowed to use unsafe features of Rust, while the unprivileged Services must be written exclusively in safe Rust.
That feels backwards to me. If an unprivileged task is unsafe, it's still unprivileged. Meanwhile the unsafe code that requires extra verification... Is only allowed in the part where nothing can safeguard it?
And from https://asterinas.github.io/book/index.html (because it was one of my first questions on seeing 'Linux replacement in rust'):
> Licensing
> Asterinas's source code and documentation primarily use the Mozilla Public License (MPL), Version 2.0. Select components are under more permissive licenses, detailed here.
Not GPL, but not BSD either.
[1]: https://www.destroyallsoftware.com/talks/the-birth-and-death...
We're actively implementing key features like Linux namespaces and cgroups, and we're also working on the first Asterinas-based distribution. Our initial focus is to use Asterinas as the guest OS inside Confidential VMs. This use case prioritizes security, where Asterinas has clear advantages over Linux thanks to its memory safety guarantee and small TCB.
Somewhat comforting to see deeply technical people still misconstruing why approaches/projects don't get adopted.
I am sorry that the doc is a kind of misleading. I wrote that... The statement need to be interpreted in the context of framekernel. An entire Rust-based framekernel runs in the kernel space but is logically partitioned into the two halves: the privileged OS framework and the de-privileged OS services. Here, "privileged" means safe + unsafe Rust kernel code, whereas "de-privileged" means all safe Rust kernel code. And this is all about the kernel code. Framekernels do not put restrictions on the languages of user-space programs.
The kernel is logically divided into two parts: the privileged OS framework (akin to a microkernel) and the de-privileged OS services. Only the privileged framework is allowed to use unsafe, while the de-privileged services must be written in safe Rust completely. As the TCB, the privileged framework encapsulates all low-level, hardware-oriented unsafe operations behind safe APIs. Using these safe APIs, the de-privileged OS services can implement all kinds of OS functionalities, including device drivers.
Ostd provides a small yet expressive set of safe OS development abstractions, covering safe user-kernel interactions, safe kernel logic, and safe kernel-peripheral interactions. Of particular note is the untyped memory abstraction, which addresses the challenge of safely handling externally-modifiable memory (e.g., MMIO or DMA-capable memory) – a longstanding obstacle in safe driver development.
So the privileged part is privileged because it does unsafe stuff. It's also quite minimal, so that the "business logic" of a driver or similar can be implemented using safe code in the de-privileged code, which is de-privileged because it doesn't need privileged unsafe access. At least that's my understanding.
SeL4 is a microkernel like this. They’ve apparently aggressively optimized IPC far more than Linux ever has. Sending a message via sel4 ipc is apparently an order of magnitude or two faster than syscalls under Linux. I wouldn’t be surprised if most programs performed better under sel4 than they do under Linux - but I’d love to know for real.
They have.
Actually the elephant in the room is modern hardware which makes even syscalls into monolithic kernels expensive. That's why io_uring and virtio perform well - they queue requests and replies between the OS and applications (or the hypervisor and the guest for virtio) avoiding transitions between address spaces. Any operating system in the future is going to need some kind of queuing syscall mechanism to perform well, and once you've got it doesn't much matter if you structure the components of your OS as a monolith or microkernel or something else.
It's an unfortunate overloading of terminology that you entirely reasonably interpreted according to the more common usage.
The unprivileged task is running in the same memory space as the core kernel, and thus there are no runtime checks to ensure that it doesn't do something which it is not allowed to do. The only way you could enforce it at runtime would be to adopt a microkernel architecture. The alternative architecture proposed here is to enforce the privileges statically by requiring that the code doesn't use unsafe features.
Designing a modern and secure kernel in 2025 as a monolith is a laughable proposition. Microkernels are the way to go.
Those containers run on a monolithic kernel; what's your point?
In fact I sometimes wonder whether it's feasible to write a new kernel while somehow shimming into Linux's driver model, while still keeping your own kernel unique (ie. not just a straight clone of Linux itself.) Some way of "virtualizing" the driver layer so that a driver can "think" it's in a Linux kernel but with a layer of indirection to everything somehow.
Isn't this basically the situation on Android?
They're not open and maintained as part of the broader kernel release which is why the driver situation degrades so quickly once the hardware is no longer actively supported.
What requires a custom driver is when your device adds its own non standard features.
So two things:
1. Containers don't have a meaningful performance hit. (They are semi-frequently used with things that can have a perf hit, like overlay filesystems, but this is generally easy to skip when it matters.)
2. I don't think containers meaningfully mimic microkernel features. If I run everything on my laptop in a container, and a device driver crashes, then the machine is still hosed.
2. It depends on what the containers are being used for. Microkernels aren't only about using drivers in userspace.
* In modern times, the practical benefit from a microkernel is minimal. Hardware is cheap, disposable, and virtual machines exist. The use case for "tolerate a chunk of the kernel misbehaving" are minimal.
* To properly tolerate a partial crash takes a lot of work. If your desktop crashes, you might as well reboot.
* Modern hardware is complex and there's no guarantee that rebooting a driver will be successful.
* A monolithic kernel can always clone microkernel functionality wherever it wants, without paying the price elsewhere.
* Processes can't trust each other.
The last one is a point I hadn't realized for a while was an issue, but it seems a tricky one. In a monolithic kernel, you can have implicit trust that things will happen. If part A tells part B "drop your caches, I need more memory", it can expect that to actually happen.
In a microkernel, there can't be such trust. A different process can just ignore your message, or arbitrarily get stuck on something and not act in time. You have less ability to make a coherent whole because there's no coherent whole.
> A different process can just ignore your message
> arbitrarily get stuck on something and not act in time
This doesn't make sense. An implementation of a microkernel might suffer from these issues, it's not a problem of the design itself. There are many ways of designing message queues.
Also:
> In a microkernel, there can't be such trust [between processes]
Capabilities have solved this problem in a much better and scalable way than the implicit trust model you have in a monolithic kernel. Using Linux as an example of a monolith is wrong, as it incorporates many ideas (and shortcomings) of a microkernel. For example: how do you deal with implicit trust when you can load third-party modules at run-time? Capabilities offer much greater security guarantees than "oops, now some third-party code is running in kernel mode and can do anything it wants with kernel data". Stuff like the eBPF sandbox is like a poor-man's alternative to the security guarantees of microkernels.
Also, good luck making sure the implicitly trusted perimeter is secure in the first place when the surface area of the kernel is so wide it's practically impossible to verify.
If you allow me an argument from authority, it is no surprise Google's Fuchsia went for a capability-based microkernel design.
It’s design largely failed at being a modern generic operating system and it’s become primarily an os used for embedded devices which is an entirely different set of requirements
It’s also not that widely used. There’s only a handful of devices that ship fuschia today. There’s a reason for that.
I've seen this exact opinion before, only the year in it was "1992". And yet Linux was still made and written regardless of it.
Someone may come along and correct me about BSD. Apologies I'm not super familiar with it's history.
It’s quite disingenuous to use “success” as a metric when discussing the advantages microkernel vs monolithic, as the only kernels you can safely say have succeeded in the past 30+ years are three: Linux, NT and Mach, one of which is a microkernel (of arguably dated design), and the other is considered a “hybrid microkernel.”
Did L4 fail? What about QNX?
This topic was considered a flame war in the 90s and I guess it still isn’t possible to have a level-headed argument over the pros and cons of each design to this day.