Will there be performance and security improvements?
Memory safety would be the main advantage.
It’s a better, more modern language in general. It has way better tooling (better, more user friendly compiler, better package manager), really good set of modern features (null-safety, good error handling, type-classes, algebraic types), it’s easier to modularize your code (workspaces, modules). Rust does a lot of things right and is fun to work with. That’s why it’s the most liked language overall. It’s not hype, it really is that good. It will just make working on the kernel easier. And on top of that it offers some memory safety and concurrency features.
P.S. I forgot about amazing documentation. Again, way better then what you can find for C.
P.P.S Zero cost abstractions.
The “tooling” argument is kind of backwards when we’re in the kernel. The package manager is not allowed to be used. Even the standard library is not allowed to be used. Writing code free of the standard library is kind of new in the Rust world and getting compiler support for it has been one of the major efforts to get Rust into the kernel. Needless to say tools around no-stdlib isn’t as robust as in the user world.
Not true, you can use cargo to build Linux drivers: https://github.com/not-matthias/kernel-driver-with-rust/blob/master/Cargo.toml
That link appears to be for a Windows driver.
Ups, you’re right. Still, better tooling attracts more devs to a language increasing the number or people that in the future may work on the kernel and in the future it could be possible to build Linux drivers using cargo. So yeah, today it’s not that important but it still a benefit.
I’ve been watching Asahi Lina develop a big GPU driver for Apple silicon and development was so much faster because a whole category of bugs were largely absent once the code compiled, and memory issues are notoriously difficult to fix. Also error handling is easier and much cleaner.
Security? Probably. I wouldn’t expect any measurable improvements to performance but the with compiler being able to do more checks it might enable some clever optimization trickery that would be harder to maintain in C.
Still, Rust on the kernel probably won’t leave the realm of drivers any time soon, so it all depends on if you have the hardware that will use a driver written in Rust.
Also to add: allowing Rust can bring in more developers to the kernel, given the growing popularity.
Memory safety is likely to prevent a lot of bugs. Not necessarily in the kernel proper, I honestly don’t see it being used widely there for a while.
In third party drivers is where I see the largest benefit; there are plenty of manufacturers who will build a shitty driver for their device, say that it targets Linux 4.19, and then never support/update it. I have seen quite a few third party drivers for my work and I am not impressed; security flaws, memory leaks, disabling of sensible warnings. Having future drivers written in rust would force these companies to build a working driver that didn’t require months of trawling through to fix issues.
Now that I think about it, in 10 years I’ll probably be complaining about massive unsafe blocks everywhere…
Haha. At least you immediately caught exactly what they’ll end up doing if they bother to use Rust at all 😂
Performance? Not really no. I believe C is slightly faster with Rust and C++ competing for second place. The benefit is safer code as Rust is built with performance and safety in mind. It highlights what potential errors can be found where making human error way less common. Instead of potential null errors types are wrapped in an option enumerator which ensures you know there can be a lack of a value. Expections are also enumerators done similarly with a result object so you know which functions may fail. Instead of using memory and potentially forgetting to free it we have the ownership system.
How is C faster than C++? Unless you use virtual functions, it’s as performant as C. And you definitely wouldn’t use virtual functions in a kernel.
The use of basic classes can very quickly become a performance issue because of data locality issues, confusing the branch predictor, and generally using instructions where C wouldn’t need to.
They’re honestly insignificant compared to the value you get in return, especially with the better typing and application design, but they’re there.
If you stick to the builtin classes, no capturing lambdas, don’t use too many generics, and use
constoften enough, you should be able to produce code that’s as fast as C. At that point you may as well use C, though, especially because one mistake with semantics and you’re back to hitting C++ related performance issues anyway.C++ is only as fast as C if you use only the parts of C++ that are identical to C. In other words, C is faster than C++
You can use compile time polymorphism in C++ without any runtime performance cost.
Compile time has got to be part of the convo esp when it comes to the kernel. The Linux kernel is one of the few bits where end-users are actively encouraged to compile from source. It is a feature!
Adding C++ compilitis is pain for what gain, from a kernel pov.
I am not a big fan of c++ overall however that is because other languages have emerged that are sweeter than C that gate some of the people issues with C++.
Anyone who has ever had a thing that was like a thing but not exactly the thing, in C, knows C ain’t great at that.
Unfortunately, that’s not true
The description says:
In this video, we’ll do a deep dive on what C++ Polymorphism is, what “virtual” does under the hood, and ultimately why it is SUCH a performance hit compared to languages like C and Rust.
This is not about compile-time polymorphism.
Here is an alternative Piped link(s): https://piped.video/watch?v=aq365yzrTVE
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source, check me out at GitHub.
That’s just plainly false. A
std::sort()for examples beats aqsort()easily. C just doesn’t have the tools to handle that kind of thing without a lot of manual code duplication.The reason for not using C++ is simply that it’s a huge monster of a language, that makes it difficult to find common ground for a programming style. In C that is much easier as you have much less features to worry about. It is more verbose and error prone, but also more predictable and easier to review, as the code you see is what you get in the binary. In C++ can have a mountain of stuff hidden behind operator overloading, exceptions and other stuff, which makes it very difficult to reason about.
Aside from that I think C is more performant than C++ (indeed when you use the bells and whistles that C++ offers), you are comparing the libraries with each other.
The fact that the implementation of one random
std::Sortis faster than the implementation ofqsort()is comparing libraries, not the languages. You are comparing the algorithm of the Rust Sort with quicksort (which is obviously theqsortyou are referring to.I am certain there are sort implementations in C which outperform Rust.
Having said that, I immensely enjoy Rust because it forces me to think about the error handling and it does not give me the quirks of C/C++ (index out of bounds, memory corruption).
you are comparing the libraries with each other.
It’s not the libraries that make the difference, but that C++ templates allow you to write generic code that is easy to use and easy to optimized for the compiler due to having all the real data types readily available, while C has to fiddle with
void*and function pointer to get similar flexibility, which are much harder to optimize, as none of the type information is there.while C has to fiddle with void* and function pointer to get similar flexibility, which are much harder to optimize, as none of the type information is there.
I thought we were discussing speed, not ease of use?
If you manually hand optimize every line of code, C and C++ are identical. That’s a worthless discussion to have. The point is that you can write in C++ high level code that would require substantial code duplication to match the speed in C, as C just doesn’t offer the tools for that level of abstraction.
Since C++ has Turing complete compiling, I guess technically it can go infinite compilation time
Potentially stability improvements as well (for the same reasons as the security improvements), especially for lesser used drivers and stuff.
Yes, enhanced security is pretty much the entire pitch of Rust. There wouldn’t be any reason for it to result in performance enhancements, though.
There are theoretical performance improvements possible by compiler optimizations if the guarantees Rust provides are met. However, the kernel relies on a lot of
unsafecode to interoperate with C so I don’t expect that to actually happen, because all of the safety guarantees go out of the window the moment you use that keyword.Not all guarantees are gone, even with unsafe
Well, it largely removes an attack surface for memory bugs, which is a huge thing. If we’re writing a big driver (see the Rust driver for the Apple GPU) then suddenly waving hands incoherently 90% or more of the driver (depending) is likely to be much more memory safe and stable. As has been demonstrated with that particular driver already.
I was watching the streams and when it compiled Asahi Lina usually only had to deal with logical type errors, not memory issues, it was basically a great showcase for Rust and memory safely. Unsafe is perfectly fine Rust, but it’s a contract where the developer says to the compiler: “I know you can’t guarantee this block is safe, so I’ll keep a special eye on that, peer review more, test, etc. while you keep an eye on all the other code I can’t fit in my head”. In the case of Linux an Unsafe blocks means “we’ll trust the Linux kernel code we connect to, though review it carefully”.
So saying all safety goes out the window is wrong, see it as a vastly reduced potential for memory problems, better error handling and more stable drivers, as demonstrated by the Apple GPU driver.
Rust code calling Rust code definitely brings safety improvements. The problem is that a lot of Rust code also needs to interact with C code (to work with pointers, for example) and that’s where
unsafebecomes a requirement, and where the compiler’s optimizations don’t get applied automatically anymore.Unsafe Rust code in the kernel is as safe as the existing C code because unsafe code is the norm, and that’s why Rust only makes things safer. However, in terms of performance improvements alone, you need to have in-depth knowledge of what abstractions you can or cannot use, and
unsafecan make a bunch of easy automatic optimisations stop working.
It just depends on how isolated that part of the kernel is. Unsafe code should be done only in interop, and so it still theoretically has a memory safety benefit over C in that sense.
In terms of how much interop code needs to be written for Rust at this point is another discussion though.
I would choose zig over rust; while zig isn’t as safe as rust (nor is that zig’s aim) zig also is also safer than C. Zig still has a bit to go.
Also rust compile times … in the kernel … that might matter more?
Then why would you choose zig?
multi-arch compile that is getting faster by the release, sweeter & safer than C — it isn’t unsafe, mind you — why wouldn’t you pick zig for systems development? And it’s compiled bins sizzle!
Plus, there is no C interop; C just plugs in.
The grammar is still in the oven, but once it bakes, I predict zig takes over C as a perf layer in script pkgs (python, node, etc).
Compile times are crisp but some foundational elements (incremental compile etc) will help it positively fly. Already it’s smoking rust however we expect more.
The grammar is why it is one to watch.
So. Compile times?
I’m willing to have slower compile times for more stable software.
Makes for a nice call out on a slide but in reality? Are we sure? Why not both?
I think at some level you can’t really get both. The rust compiler is constantly being sped up, but the amount of checking it does simply takes time.
It’s the age old adage “The fastest code is the one that doesn’t exist”. The fastest compilation checks are the ones you don’t do.
Situation is everything.
speed matters when it comes to systems-level logic/code. Zig strikes that balance very nicely.
Industry cares at the systems level. If a “safe” kernel is 4x slower (compilation, run-time) and there is an “unsafe” option that is just as secure, outside of a cadre of philosopher kings/queens (more power to 'em), practical teams will hit that
unsafeperformance boosting switch every time.C is as safe as the situation & the code dictates. Zig makes it easier to be safe in most situations and excels at promoting systems-level safety.
App level, or app API? Where there is more of an x-factor? rust has a great use case. In “user-land”, availability matters almost as much as performance.. An app that always crashes is not as valuable as a slower app that never crashes…
