Smart pointers for the kernel

https://lwn.net/Articles/992055/

by signa11 • 2 years ago

169 114 2 years ago

smodo• 2 years ago

I’m not very well versed in kernel development. But I am a Rust dev and have observed the discussion about Rust in Linux with interest… Having said that, this part of the article has me baffled:

> implementing these features for a smart-pointer type with a malicious or broken Deref (the trait that lets a programmer dereference a value) implementation could break the guarantees Rust relies on to determine when objects can be moved in memory. (…) [In] keeping with Rust's commitment to ensuring safe code cannot cause memory-safety problems, the RFC also requires programmers to use unsafe (specifically, implementing an unsafe marker trait) as a promise that they've read the relevant documentation and are not going to break Pin.

To the uninformed this seems like crossing the very boundary that you wanted Rust to uphold? Yes it’s only an impl Trait but still… I can hear the C devs now. ‘We pinky promise to clean up after our mallocs too!’

foundry272• 2 years ago

Rust’s whole premise of guaranteed memory safety through compiletime checks has always been undermined when confronted with the reality that certain foundational operations must still be implemented using unsafe. Inevitably folks concede that lower level libraries will have these unsafe blocks and still expect higher level code to trust them, and at that point we’ve essentially recreated the core paradigm of C: trust in the programmer’s diligence. Yeah Rust makes this trust visible, but it doesn’t actually eliminate it in “hard” code.

The punchline here, so to speak, is that for all Rust’s claims to revolutionize safety, it simply(!) formalizes the same unwritten social contract C developers have been meandering along with for decades. The uniqueness boils down to “we still trust the devs, but at least now we’ve made them swear on it in writing”.

wbl2• 2 years ago

The difference is every line of C can do something wrong while very few lines of Rust can. It's much easier to scrutinize a small well contained class with tools like formal methods than a sprawling codebase.

kelnos• 2 years ago

I don't think you're giving Rust enough credit here.

For those projects that don't use any unsafe, we can say -- absent compiler bugs or type system unsoundness -- that there will be no memory leaks or data races or undefined behavior. That's useful! Very useful!

For projects that do need unsafe, that unsafe code can be cordoned off into a corner where it can be made as small as possible, and can be audited. The rest of the code base is just as safe as one with no unsafe at all. This is also very useful!

Now, sure, if most projects needed to use unsafe, and/or if most projects had to write a significant amount of unsafe, then sure, I'd agree with you. But that's just not the reality for nearly all projects.

With C, everything is unsafe. Everything can have memory leaks or data races or undefined behavior. Audits for these issues need to examine every single line of code. Compilers and linters and sanitizers can help you here, but they can never be comprehensive or guarantee the absence of problems.

I've been writing C for more than 20 years now. I still write memory leaks. I still write NULL pointer dereferences. I still struggle sometimes to get my data ownership (and/or locking) right when I have to write multithreaded code. When I get to write Rust, I'm so happy that I don't have to worry about those things, or spend time with valgrind or ASAN or clang's scan-build to figure out what I've done wrong. Rust lets me focus more on what I actually care about, the actual code and algorithms and structure of my program.

jchw5• 2 years ago

I think when people come to these conclusions it's largely due to a misunderstanding of what exactly the point of most programming language safety measures are and why they make sense.

Something that people often ponder is why you can't just solve the null safety problem by forcing every pointer dereference to be checked, with no other changes. Well of course, you can do that. But actually, simply checking to make sure the pointer is non-null at the point of dereference gets you surprisingly little. When you do this, what you're (ostencibly) trying to do is reduce the number of null pointer dereferences, but in practice what happens now is that you just have to explicitly handle them. But, in a lot of cases, there's really nothing particularly sensible to do: the pointer not being null is an invariant that was supposed to be upheld and it wasn't, and now at the point of dereference, at runtime, there's nothing to do except crash. Which is what would've happened anyways, so what's the point? What you really want to do isn't actually prevent null pointer dereferences, it's to uphold the invariants that the pointer is non-null in the first place, ideally before you leave compile time.

Disallowing "unsafe" operations without marking them explicitly unsafe doesn't give you a whole lot, but what you can do is expand the number of explicitly safe operations to cover more of what you want to do. How Rust, and many other programming languages, have been accomplishing this is by expanding the type system, and combining this with control flow analysis. Lifetimes in Rust are a prime example, but there are many more such examples. Nullability, for example, in languages like TypeScript. When you do it this way, the safety of such "safe" operations can be guaranteed, and while these guarantees do have some caveats, they are very strong to a lot of different situations that human code reviews are not, such as an unsafe combination of two otherwise-safe changesets.

It's actually totally fine that some code will probably remain unable to be easily statically verified, the point is that we want to reduce the amount of code that can't be easily statically verified to be as small as possible. In the future we can use much less easy approaches to statically verify unsafe blocks, such as using theorem provers to try to prove the correctness of "unsafe" code. But even just reducing the amount of not-necessarily-memory-safe code is an enormous win, for obvious reasons: it dramatically reduces the surface area for vulnerabilities. Moreover, time and time again, it is validated that most new vulnerabilities come from relatively recent changes in code, which is another huge win: a lot of the unsafe foundations actually don't need to be changed very often.

There is absolutely nothing special about code written in Rust, it's doing the same shit that C code has been doing for decades (well, on the abstract anyway; I'm not trying to downplay how much more expressive it is by any means). What Rust mainly offers is a significantly more advanced type system that allows validating many more invariants at compile-time. God knows C developers on large projects like the Linux kernel care about validating invariants: large amounts of effort have been poured into static checking tools for C that do exactly this. Rust is a step further though, as the safe subset of Rust provides guarantees that you basically can't just tack onto C with only more static checking tools.

stackghost• 2 years ago

I'm not a rust fanboy but isn't the point of rust to dramatically decrease the area in which null pointer dereferences and friends can occur, and thus make them more likely to be spotted?

Rusky• 2 years ago

It's not the same unwritten social contract: in Rust even the unsafe code has the same stricter type signatures as the safe code, so there is a formal way to judge which part of the program is at fault when the contract is broken. You might say the contract is now written. :-)

In C, the type system does not express things like pointer validity, so you have to consider the system as a whole every time something goes wrong. In Rust, because the type system is sound, you can consider each part of the program in isolation, and know that the type system will prevent their composition from introducing any memory safety problems.

This has major implications in the other direction as well: soundness means that unsafe code can be given a type signature that prevents its clients from using it incorrectly. This means the set of things the compiler can verify can be extended by libraries.

The actual practice of writing memory-safe C vs memory-safe Rust is qualitatively different.

thfuran• 2 years ago

Would you similarly say that Russian Roulette is the same game whether the revolver has two chambers or ten thousand?

sqeaky• 2 years ago

Constrained and minimized trust in programmer diligence is better than unconstrained and omnipresent trust in the same.

Ar-Curunir3• 2 years ago

This is nonsense. Just because some small parts of the code are must be annotated as unsafe doesn’t mean that we’re suddenly back to C land. in comparison, with C the entire codebase is basically wrapped in a big unsafe. That difference is important, because in Rust you can focus your auditing and formal verification efforts on just those small unsafe blocks, whereas with C everything requires that same attention.

Furthermore, Rust doesn’t turn off all checks in unsafe, only certain ones.

mort96• 2 years ago

It is literally impossible to build systems where you never at any point trust that underlying systems work correctly. This is a boring and uninformative criticism. It is the case for every language ever invented.

pornel2• 2 years ago

Building safe abstraction around unsafe code works, because it reduces the scope of the code that has to be reviewed for memory safety issues.

Instead of the whole codebase being suspect, and hunting for unsafety being like a million-line "Where's Waldo?", it reduces the problem to just verifying the `unsafe` blocks against safety of their public interface, "is this a Waldo?". This can still be tricky, but it has proven to be a more tractable problem.

gpm• 2 years ago

This kind of use of unsafe has been in rust forever, for example with `Sync`, `Send`. Implementing an unsafe marker trait to promise to the compiler that other methods on the structure act a certain way.

The scope of an unsafe block has always been interpreted to include guaranteeing things about it's surroundings up to other things in the same module. E.g. if I'm implementing `Vector::push` I'm going to rely on the fact that `self.capacity` really is the size of the allocation behind `self.ptr` without verifying it, and I'm going to feel good about that because those fields aren't public and everything within the module doesn't let you violate that constraint, so it's not possible for external safe code to violate it.

The same applies to marker traits. If I'm writing `unsafe impl Send for MyStruct {}` I'm promising that the module exposes an interface where `MyStruct` with will always comply with the requirements of `Send` no matter what safe external code does (i.e. sending MyStructs across threads is safe given the exposed API). With this proposal if I write `unsafe impl PinCoerceUnsized for MyStruct {}` I'm promising the same with respect to that (whatever the documentation for that trait ends up saying, which should essentially be that I've implemented `Deref for MyStruct` in the same module and I don't expose any way for safe external code to change what reference `Deref` returns).

mise_en_place• 2 years ago

I'm not convinced you can have your cake and eat it too. Not a dig at Rust specifically; it's more like, you can have managed memory, or manually manage it yourself. There is no in-between. The implementation ends up being a kludge. Terry had it right, just put everything in the lower 2 GB and DMA it.

remram• 2 years ago

I think the part that uses `unsafe` and can break Pin if done wrong is the implementation of the smart-pointer type, not its use.

GolDDranks2• 2 years ago

Try imagining trait AlwaysIndexableUnder100. There's a generic codebase/library that takes in types that implement that trait, and do indexing with indexes that are always below 100. Like `usersCustomSliceA[4] = usersCustomSliceB[5];`

You'd be tempted, for performance, to use `get_unchecked` methods that skip the boundary checks. After all, the trait says that this should always succeed.

However, if the user passes in a type that is not indexable with integers smaller than 100, whose fault it is if the program segfaults? The users? But they managed to get the program to segfault _without_ writing unsafe code. The provider of the library? They are using `unsafe` to call `get_unchecked` after all.

Bingo. It's the library dev's fault. The API they provide is not sound. However, they can make it sound by marking the _trait_ unsafe to implement. Then the user needs to type `unsafe` when implementing the trait for their type. "I solemnly swear that this type is actually indexable with all integers smaller than 100." That shifts the blame to the mistaken implementation, and the user is to blame.

It's the same situation here. Deref is not unsafe to implement. That's why if you need to uphold a trust boundary, you need an unsafe trait.

So, the whole thing doesn't account to crossing the boundary willy nilly, a big point of Rust's unsafe is for the compiler to force documenting and checking for accountability: who is required to do what, and who is allowed to rely on that.

rocqua• 2 years ago

Rust is all about ring-fencing the scary parts in unsafe. A rust program that doesn't use unsafe, and only uses dependencies that are sound with respect to unsafe, is guaranteed to be fine. And it is very easy to write code without using unsafe. Unlike C, where code style that is guaranteed to be memory safe is nigh impossible.

The difficult bit with Rust is still the sound use of unsafe, but it is quite feasible to do that by hand. It does, sadly, require looking at the entire module that contains the unsafe code.

DoubleDecoded• 2 years ago

"could break the guarantees" is a weak guarantee then.

josephcsible• 2 years ago

It's still a guarantee. The point of the guarantee isn't that no code can cause certain kinds of problems, but rather that any code that can must be marked unsafe.

dzaima2• 2 years ago

Basically any guarantee can be broken by an evil guaranteer. All this is saying is that this is a case of the user being required to uphold a guarantee, instead of the stdlib/language.

acbits2• 2 years ago

https://github.com/acbits/reftrack-plugin

I wrote this GCC plugin for this exact reason. Let's see whether the kernel team is interested in adopting it.

brenns10• 2 years ago

I searched lore.kernel.org and couldn't find any postings that propose using this in the kernel. I'd encourage you to share a proposal, otherwise the "kernel team" will never be interested, because they'll never hear of it.

acbits2• 2 years ago

I will send it again. It probably got lost in the high traffic volume of LKML last year.

pornel2• 2 years ago

That's cool, but it'd be nice to also have a distinction between a get and a copy (like ObjC) or borrow/move (like Rust) to avoid redundant increments and decrements.

glouwbug• 2 years ago

I love high quality C++ that reads like C but doesn’t throw away the STL or litter everything with const

stonethrowaway2• 2 years ago

Anyone who wants to do kernel-level development should first do Embedded hardware/software interfacing. No RTOS, plain “Embedded C”, with some bit banging and dealing with voltage spikesc transients and people doing stupid shit to hardware (yes, really) and other things. Know the memory map and recite it from memory. Some might think I’m joking or being facetious - no, I’m pretty serious actually. I’d rather have an embedded person writing kernel drivers in slapped-together C than a Rustacian that complains about unsafe code and being an idiot about it. See [0] for detailed explanation.

People need to learn the niceness of safety and perfect execution is a continuum of tolerances and flimsy guarantees from unmarked silicon that could be made in US, but is most likely a knock off made in China that will fail in 1/3rd of the expected time and gives a false reading if you so much as look at it the wrong way.

[0] https://www.usenix.org/system/files/1311_05-08_mickens.pdf

sureglymop• 2 years ago

You can do that in rust. Yes, you probably will have unsafe blocks. You can even write "C like" rust code that has a bunch of unsafe blocks but you'll benefit from better tooling. But maybe I misunderstand the article and there is somehow an implication that unsafe blocks are bad?

When I was doing some embedded development using rust it was actually a great experience, with hal and pac crates available already for a lot of hardware or easy to generate.

jendjdndn• 2 years ago

You sound a lot like "kids these days!"

What applicable skills would someone writing a kernel driver gain from reciting a memory map? Abstractions exist for a reason.

The skill is in creating useful an performant abstractions.

eru6• 2 years ago

You sound a lot like "kids these days!"

Exactly! Kids like stonethrowaway with their C. Real engineers write their compilers in assembly.

(Less snarky, why stop at C, and not complain about even lower level stuff?)

exmadscientist• 2 years ago

See, I read the parent post's point not as "abstractions are bad" but as "it's much better to be someone who doesn't need the abstraction, but chooses to use it". I have worked with a number of crappy embedded developers in my career. Somehow, the ones who are capable of going beneath the many levels of abstraction are always really, really good at their jobs.

So it's not that embedded Rust is bad. It's that developers who can't do their jobs without embedded Rust are usually very bad indeed. It's great when it's a choice. It's terrible when you lack the skills or perspective to work with what's under the hood.

a5c11• 2 years ago

The problem starts when the abstraction fails and you have to dive deeper, but no one taught you how to dive, only swim with a head above the water level. Those who can dive can also swim, it doesn't work the other way round, though.

rcxdude• 2 years ago

Plenty of people writing rust are coming from that background. And a substantial part of the reason for wanting it in linux is that the slapped-together C in linux drivers is often awful, both in terms of understanding of the hardware and in terms of the quality of the software. Rust can at least help the poor quality of the latter not affect the security and stability of the kernel so much.

Smart pointers for the kernel

Related Stories