In practice, the Linux community is the wild wild west, and sweeping changes are infamously difficult to achieve consensus on, and this is by far the broadest sweeping change ever proposed for the project. Every subsystem is a private fiefdom, subject to the whims of each one of Linux’s 1,700+ maintainers, almost all of whom have a dog in this race. It’s herding cats: introducing Rust effectively is one part coding work and ninety-nine parts political work – and it’s a lot of coding work. Every subsystem has its own unique culture and its own strongly held beliefs and values.
The consequences of these factors is that Rust-for-Linux has become a burnout machine. My heart goes out to the developers who have been burned in this project. It’s not fair. Free software is about putting in the work, it’s a classical do-ocracy… until it isn’t, and people get hurt. In spite of my critiques of the project, I recognize the talent and humanity of everyone involved, and wouldn’t have wished these outcomes on them. I also have sympathy for many of the established Linux developers who didn’t exactly want this on their plate… but that’s neither here nor there for the purpose of this post, and any of those developers and their fiefdoms who went out of their way to make life difficult for the Rust developers above and beyond what was needed to ensure technical excellence are accountable for these shitty outcomes.
…
Here’s the pitch: a motivated group of talented Rust OS developers could build a Linux-compatible kernel, from scratch, very quickly, with no need to engage in LKML politics. You would be astonished by how quickly you can make meaningful gains in this kind of environment; I think if the amount of effort being put into Rust-for-Linux were applied to a new Linux-compatible OS we could have something production ready for some use-cases within a few years.
…
Having a clear, well-proven goal in mind can also help to attract the same people who want to make an impact in a way that a speculative research project might not. Freeing yourselves of the LKML political battles would probably be a big win for the ambitions of bringing Rust into kernel space. Such an effort would also be a great way to mentor a new generation of kernel hackers who are comfortable with Rust in kernel space and ready to deploy their skillset to the research projects that will build a next-generation OS like Redox. The labor pool of serious OS developers badly needs a project like this to make that happen.
Follow up to: One Of The Rust Linux Kernel Maintainers Steps Down - Cites “Nontechnical Nonsense”, On Rust, Linux, developers, maintainers, and Asahi Lina’s experience about working on Rust code in the kernel
Drew DeVault recently wrote a simple but functional UNIX kernel in a new systems programming language named Hare in about a month, which suggests that doing something similar in Rust would be equally feasible. One or two motivated individuals could get something up which is semi-useful (runs on a common x86 PC, has a console, a filesystem, functional if not necessarily high-performance scheduling and enough of the POSIX API to compile userspace programs for), upon which, what remained would be a lot of finishing work (device drivers, networking, and such), though not all of it necessary for all users. Doing this and keeping the goal of making it a drop-in replacement for the Linux kernel (as in, you can have both and select the one you boot into in your GRUB menu; eventually the new one will do enough well enough to replace Linux) sounds entirely feasible, and a new kernel codebase, implemented in a more structured, safer language sounds like it could deliver a good value proposition over the incumbent.
I’m not sure I’d call device drivers ‘finishing work’.
The MAJORITY of the kernel is that pesky bit of ‘finishing work’; there’s a fuckton of effort in writing drivers and support for damn near ever architecture and piece of hardware made in the last 30 years.
You could argue you don’t NEED to support that much stuff, and I’d probably agree, but let’s at least be honest that the device driver work is likely to be 90% of the work and 90% of the user complaints if something doesn’t work.
It depends on if you can feasibly implement compatibility layers for large parts of the “required” but very work-intensive drivers. FreeBSD has the same driver struggles and ended up with LinuxKPI to support AMD/Intel GPUs. I know there’s a whole bunch of toy kernels that implemented compatibility layers for parts of Linux in some fashion too.
It’s a ton of work overall but there’s room to lift enough already existing stuff from Linux to get the ball rolling.
Sure, but isn’t the mixture of rust and c based device drivers and kernel ABIs kinda the whole kerfuffle that’s going on right now? Seems like the rust faction would probably NOT want to grab some duct tape and stuff all this back together, but slightly differently this time.
It’s possible though less than ideal. Drivers that connect to devices are part of the attack surface, and probably the part you’d least want implemented in C when the rest of the kernel is in Rust.
Sure, but even if they started tomorrow it would probably be years before it even could be considered experimental outside of the most daring early adaptors.
Having a combability layer is not ideal but it would mean they could have something worker for more users faster and at the same time see which modules/drivers they should focus on.
This is incredibly true. The hardware manufacture process is a slow turning and cost centric wheel, but it’s always forward looking. If it doesn’t exist today you are building around compromises made outside the scope of your concerns.
Anyone whose had to work on DEC or Sun hardware can describe in excruciating detail about how minor implementation differences in hardware cascade down the chain. (Missing) Rubber washers determined a SAN max writes once, lest the platters vibrating cause the chassis to walk across the floor.
‘Universal’ support is always a myth, and carving up what segment to target is shooting one moving target while standing on another one unless you have exclusive control of implementation of the whole chain (apple).
There’s a Pareto effect when it comes to them, in that you can cover a large proportion of use cases with a small amount of work, but the more special cases consume proportionately more effort. For a MVP, you could restrict support to standard USB and SATA devices, and get a device you can run headless, tethered to the network through a USB Ethernet adapter. For desktop support, you’d need to add video display support, and support for the wired/wireless networking capabilities of common chipsets would be useful. And assuming that you’re aiming only for current hardware (i.e. Intel/AMD boards and ARM/RISC-V SOCs), there are a lot of legacy drivers in Linux that you don’t need to bring along, from floppy drives to the framebuffers of old UNIX workstations. (I mean, if a hobbyist wants to get the kernel running on their vintage Sun SPARCstation, they can do so, but it won’t be a mainstream feature. A new Linux-compatible kernel can leave a lot of legacy devices behind and still be useful.)
I think we’re kind of saying the same thing: making something that boots as an MVP isn’t the most difficult thing but is still a much different and simpler project than making a replacement kernel for Linux.
If you really wanted to be a legitimate actual replaces-Linux-for-all-the-things-you-use-Linux-for kernel, you’d be biting off years and years and years of work on drivers. I mean, just look how long it took Noveau to go from ‘kinda works’ to ‘actually viable’ and that’s just a subset of GPUs from one vendor.
I’d also add that if you cared about server hardware, you’ve got a much larger driver footprint with a lot more weird behavior shit than on desktops which are, honestly, just a couple dozen combinations of chipset, sound, ethernet, and usb controller in any given generation.
And sure, you could lean on the work that’s already been done at this point and probably do it faster, but it’s still a massive undertaking.
ReactOS is probably a good indicator how far you can get with some limited generic drivers.
There’s Redox OS already headed in that general direction.
There’s also this tutorial: https://os.phil-opp.com
Drew mentions this and points out that it’s a new OS design and will therefore take a long time. He argues that an OS based on the linux design would be much easier.
Yeah, I did read that, admittedly after making my comment, but thanks for pointing it out anyways. 🙂
I’d recommend spending some time reading about it. It’s not as hard as he thinks. Applications developed for Linux are quite easy to port to Redox. It supports many of the same system calls and has a compatible libc implementation. The kernel does have abstractions to ease the porting process. And if you’re going to make a new kernel today, you should do it right and make a microkernel like Redox. One of the benefits of having a microkernel is that it doesn’t matter what language you write drivers in. They’re isolated to their own processes. Rust, C, C++, whatever.
I kinda hope this happens, and I even think it’s likely. GNU/Linux came about because of the corporate direction of Unix. If Linux kernel devs are similarly going to shut out Rust devs, it seems like a reasonable path forward is to diverge and start something new.