It's like letting a wet dog (who'd just been swimming in a nearby swamp) run loose inside your hermetically sealed cleanroom.
"You can take five actions in unsafe Rust that you can’t in safe Rust, which we call unsafe superpowers. Those superpowers include the ability to:
Dereference a raw pointer
Call an unsafe function or method
Access or modify a mutable static variable
Implement an unsafe trait
Access fields of a union
In addition, unsafe does not mean the code inside the block is necessarily dangerous or that it will definitely have memory safety problems: the intent is that as the programmer, you’ll ensure the code inside an unsafe block will access memory in a valid way.
People are fallible, and mistakes will happen, but by requiring these five unsafe operations to be inside blocks annotated with unsafe you’ll know that any errors related to memory safety must be within an unsafe block. Keep unsafe blocks small; you’ll be thankful later when you investigate memory bugs."
So, in theory, unsafe rust opens the floodgates. In practice, though, you can use small fragments of unsafe code that programmers can fairly easily check to be safe.
Then, once you’ve convinced yourself that those fragments are safe, you can be assured that your whole program is safe (using ‘safe’ in the rust sense, of course)
So, there may be some small islands of unsafe code that require extra attention from the programmer, but that should be just a tiny fraction of all lines, and you should be able to verify those islands in isolation.
This is where the rubber hits the road. Rust does not allow you to do this, in the sense that this is possibly undefined behavior. That "possibly" is why the compiler allows you to write this code, because by saying "unsafe", you are promising that this specific arbitrary address is legal for you to write to. But that doesn't mean that it's always legal to do so.
In practice (in both languages) you check what the actual unsafe code does (or "all" code in C's case), note code that depends on external actors for safety (it's not all C code, nor is it all unsafe Rust blocks), and check their callers (and callers callers, etc).
That's only true at the same level of scrutiny as "all C operations can cause undefined behaviour, regardless of what they are", which I find similarly shallow.
And there is not many things we have statistics on in CS, but memory vulnerabilities being absolutely everywhere in unsafe languages, and Rust cleaning up the absolute majority of them even when only the new parts are written in Rust are some of the few we do know, based on actual, real life projects at Google/Microsoft among others.
A memory safe low-level language is as novel as it gets. Rust is absolutely not just hype, it actually delivers and you might want to get on with the times.
> absolutely laughable c-strings that perform terribly
From my past experiences with Rust, the team never had to think about data race once, or mutable volatile globals. And we all there suffered from those decades ago with C and sometimes C++ as well.
You like those and don't want to migrate? More power to ya! But badmouthing Rust with what seem fairly uninformed comments is just low. Inform yourself first.
Rust's type system (including ownership and borrowing, Sync/Send, etc), along with it's privacy features (allowing types to have private fields that can only be accessed by code in the module that defined them) allows you to create fully safe interfaces around code that uses unsafe; there is provably no combination of uses of the interface which lead to undefined behavior.
Now, yeah, it's possible to also use unsafe in Rust just for applying a local optimisation. And that has fewer benefits than a fully encapsulated safe interface, though is still easier to audit for potential UB than C.
So you're right that it's on a continuum, but the distinction between safe and unsafe code means you can more easily find the specific places where UB could occur, and the encapsulation and type system makes it possible to create safe abstractions over unsafe code.
But using ordinary module encapsulation and private fields, you can scope the code that needs to uphold those preconditions to a particular module.
So the "trusted computing base" for the unsafe code can still be scoped and limited, allowing you to reduce the amount of code you need to audit and be particularly careful about for upholding safety guarantees.
Basically, when writing unsafe code, the actual unsafe operations are scoped to only the unsafe blocks, and they have preconditions that you need to scope to a particular module boundary to ensure that there's a limited amount of code that needs to be audited to ensure it upholds all of the safety invariants.
Ralf Jung has written a number of good papers and blog posts on this topic.
int average(int x, int y) {
return (x+y)/2;
}I might be old, but more than 10 years ago, hardly anyone talked about UB in C and C++ programming. In the last 10 years, it is all the rage, but seems to add very little to the conversation. For example, if you program C or C++ with the Win32 API, there are loads of weird UB-ish things that seem to work fine.
This is not how compilers work. Optimization happens based on language semantics, not on what platforms do.
UB in C is often found where different real hardware architectures had incompatible behavior. Rather than biasing the language for or against different architectures they left it to the compiler to figure out how to optimize for the cases where instruction behavior diverge. This is still true on current architectures e.g. shift overflow behavior which is why shift overflow is UB.
int average(int x, int y) {
long sum = (long)x + y;
if(sum > INT_MAX || sum < INT_MIN)
return -1; // or any value that indicates an error/overflow
return (int)(sum / 2);
}There is no guarantee that sizeof(long) > sizeof(int), in fact the GNU libc documentation states that int and long have the same size on the majority of supported platforms.
https://www.gnu.org/software/libc/manual/html_node/Range-of-...
> return -1; // or any value that indicates an error/overflow
-1 is a perfectly valid average for various inputs. You could return the larger type to encode an error value that is not a valid output or just output the error and average in two distinct variables.
AI and C seem like a match made in hell.
That used to be the case for 32-bit platforms, but most 64-bit platforms in which GNU libc runs use the LP64 model, which has 32-bit int and 64-bit long. That documentation seems to be a bit outdated.
(One notable 64-bit platform which uses 32-bit for both int and long is Microsoft Windows, but that's not one of the target platforms for GNU libc.)
average(INT_MAX,INTMAX) should return INT_MAX, but it will get that wrong and return -1.
average(0,-2) should not return a special error-code value, but this code will do just that, making -1 an ambiguous output value.
Even its comment is wrong. We can see from the signature of the function that there can be no value that indicates an error, as every possible value of int may be a legitimate output value.
It's possible to implement this function in a portable and standard way though, along the lines of [0].
[0] https://stackoverflow.com/a/61711253/ (Disclosure: this is my code.)
That won't occur on an 'LP64' platform, [0] but we should aim for proper portability and conformance to the C language standard.
[0] https://en.wikipedia.org/wiki/64-bit_computing#64-bit_data_m...
C libraries tend to support a lot of exotic platforms. zlib for example supports Unicos, where int, long int and long long int are all 64 bits large.
- Exposing an unsafe API and relying on the caller to manually uphold invariants
- Doing things like defensive copying at a performance cost
In many cases Rust gives you the best of both worlds: sharing memory liberally while still having the compiler enforce correctness.
And even if you try to provide some kind of safer abstraction, you're limited by the much more primitive type system, that can't distinguish between owned types, unique borrows, and shared borrows, nor can it distinguish thread safety properties.
So you're left to convention and documentation for that kind of information, but nothing checking that you're getting it right, making it easy to make mistakes. And even if you get it right at first, a refactor could change your invariants, and without a type system enforcing them, you never know until someone comes along with a fuzzer and figures out that they can pwn you
Sure, you can technically just write your own vulnerability for your own program and inject it at an unsafe and see the whole world crumble... but the exact same is true for any form of FFI calls in any language. Is Java memory safe? Yeah, just because I can grab a random pointer and technically break anything I want won't change that.
The fact that a memory vulnerability error may either appear at no place at all OR at the couple hundred lines of code thorough the whole project is a night and day difference.
By the way, the rust compiler does generate such code because under the hood LLVM runs an autovectorizer when you turn on optimizations. However, for the autovectorizer to do a good job you have to write code in a very special way and you have no way of controlling whether or not it kicked in and once it did that it did a good job.
There’s work on creating safe abstractions (that also transparently scale to the appropriate vector instruction), but progress on that has felt slow to me personally and it’s not available outside nightly currently.
> However, for the autovectorizer to do a good job you have to write code in a very special way
Rust encourages using unsafe to "teach" the language new design patterns and data structures; and uses this heavily in its standard library. For example, the Vec type is a wrapper around a raw pointer, length, and capacity; and exposes a safe interface allowing you to create, manipulate, and access vectors with no risk of pointer math going wrong -- assuming the people who implemented the unsafe code inside of Vec didn't make a mistake, the external, safe interface is guaranteed to be sound no matter what external code does.
Think of unsafe not as "this code is unsafe", but as "I've proven this code to be safe, and the borrow checker can rely on it to prove the safety of the rest of my program."
> For example, the Vec type is a wrapper around a raw pointer, length, and capacity; and exposes a safe interface allowing you to create, manipulate, and access vectors with no risk of pointer math going wrong -- assuming the people who implemented the unsafe code inside of Vec didn't make a mistake, the external, safe interface is guaranteed to be sound no matter what external code does.
Instead, all that functionality is written as Rust code in the standard library, such as Vec. This is what I mean by using unsafe code to "teach" the borrow checker: the language itself doesn't have any notion of growable arrays, so you use unsafe to define its semantics and interface, and now the borrow checker understands growable arrays. The alternative would be to make growable arrays some kind of compiler magic, but that's both harder to implement correctly and not generalizable.
> you can do exactly the same in C by using an opaque pointer to protect the data structure. Then you write a bunch of functions that operate on the opaque pointer. You can use assert() to protect against unreasonable inputs.
That's true and that's a great design pattern in C as well. But there are some crucial differences:
- Rust has no undefined behavior outside of unsafe blocks. This means you only need to audit unsafe blocks (and any invariants they assume) to be sure your program is UB-free. C does not have this property even if you code defensively at interface boundaries.
- In Rust, most of the invariants can be checked at compile time; the need for runtime asserts is less than in C.
- C provides no way to defend against dangling pointers without additional tooling & runtime overhead. For instance, if I write a dynamic vector and get a pointer to the element, there's no way to prevent me from using that pointer after I've freed the vector, or appended an element causing the container to get reallocated elsewhere.
Rust isn't some kind of silver bullet where you feed it C-like code and out comes memory safety. It's also not some kind of high-overhead garbage collected language where you have to write unsafe whenever you care about performance. Rather, Rust's philosophy is to allow you to define fundamental operations out of small encapsulated unsafe building blocks, and its magic is in being able to prove that the composition of these operations is safe, given the soundness of the individual components.
The stdlib provides enough of these building blocks for almost everything you need to do. Unsafe code in library/systems code is rare and used to teach the language of new patterns or data structures that can't be expressed solely in terms of the types exposed by the stdlib. Unsafe in application-level code is virtually never necessary.
A simple example might be modifying a const value deep down in some class, where it only becomes apparent later in the program’s execution. Hence their analogy of the wet dog in a clean room - whatever beliefs you have about the structure of memory in your entire program, and guaranteed by the compiler, could have been undone by a rogue unsafe.
There's even unsafe usage in the standard library and it's used a lot in embedded libraries.
Sorry but horrible comparison ;)
If you need to rely on unsafe in a memory-safe language for performance reasons, then there is a issue with the language compiler at that point, that needs to be fixed. Simple as that.
The whole memory-safety is the bread and butter of the language, the moment you start to bypass it for faster memory operations, you can start doing the same in any other language. I mean, your literally bypassing the main selling point of the language. \_00_/
It actually means "Rust needs to interface with many other systems that are not as stringent as it". Your interpretation has nothing to do with what's actually going on and I am surprised you misinterpreted the situation as hugely as you did.
...And even if everything was written in Rust, `unsafe` would still be needed because the lower you get [to the kernel] you get more and more non-determinism at places.
This "all or nothing" attitude is boring and tiring. We all wish things were super simple, black and white, and all-or-nothing. They are not.
Once you can internalize that you could unlock the power of encapsulation.
All safe code in existence running on von Neumann architectures is built on a foundation of unsafe code. The goal of all memory-safe languages is to provide safe abstractions on top of an unsafe core.
> What language is the JVM written in?
I like your second paragraph. It is well written.
- Java 74.1%
- C++ 14.0%
- C 7.9%
- Assembly 2.7%
And those values have been increasing for Java with each OpenJDK release.
No, not even close. You only lose Rust's safety guarantees when your unsafe code causes Undefined Behavior. Unsafe code that can be made to cause UB from Safe Rust is typically called unsound, and unsafe code that cannot be made to cause UB from Safe Rust is called sound. As long as your unsafe code is sound, then it does not break any of Rust's guarantees.
For example, unsafe code can still use slices or references provided by Safe Rust, because those are always guaranteed to be valid, even in an unsafe block. However, if from inside that unsafe block you then go on to manufacture an invalid slice or reference using unsafe functions, that is UB and you lose Rust's safety guarantees because of the UB.
At the same time, unsafe doesn't just turn off all compiler checks, it just gives you tools to go around them, as well as tools that happen to go around them because of the way they work. Rust unsafe is this weird mix of being safer than pure C, but harder to grasp; with lots of nuanced invariants you have to uphold. If you want to ensure your code still has all the nice properties the compiler guarantees (which go way beyond memory safety) you would have to carefully examine every unsafe block. Which few people do, but you generally still end up with a better status quo than C/C++ where any code can in principle break properties other code was trying to uphold.
This isn't a wet dog in a cleanroom. This is cleanroom complex that has a very small outhouse that is labeled as dangerous.
As soon as you start playing with FFI and raw pointers in Python, NodeJS, Julia, R, C#, etc you can easily loose the nice memory-safety properties of those languages - create undefined behavior, segfaults, etc. I'd say Rust is a lot nicer for checking unsafe correctness than other memory-safe languages, and also makes it easier to dip down to systems-level programming, yet it seems to get a lot of hate for these features.
Normally in safe code you can’t violate the language rules because the compiler enforces various rules. In unsafe mode, you can do several things the compiler would normally prevent you from doing (e.g. dereferencing a naked pointer). If you uphold all the preconditions of the language, safety is preserved.
What’s unfortunate is that the rules you are required to uphold can be more complex than you might anticipate if you’re trying to use unsafe to write C-like code. What’s fortunate is that you rarely need to do this in normal code and in SIMD which is what the snippet is representing there’s not much danger of violating the rules.
To continue the analogy of the dog, you let the dog get wet (=you use unsafe), but you put a cleaning room (=the sound and safe API) before your sealed room (=the safe code world)
Inside that block, both yes and no. You have to enforce those nice guarantees yourself. Code that violates it will still crash.
In other words, unsafe works if you use it carefully and keep it contained.
Not sure why would one resulted in all. One of Rust's advantages is the clear boundary between safe/unsafe.
The usual retort to these questions is 'well, the standard library uses unsafe code, so everything would need a disclaimer that it uses unsafe code, so that's a useless remark to make', but the basic issue still remains that the only clear boundary is whether a function 'contains' unsafe code, not whether a function 'calls' unsafe code.
If Rust did not have a mechanism to use external code then it would be fine because the only sources of unsafe code would be either the application itself or the standard library so you could just grep for 'unsafe' to find the boundaries.
The point is that you don't need to. The guarantees compose.
> The usual retort to these questions is 'well, the standard library uses unsafe code
It's not about the standard library, it's much more fundamental than that: hardware is not memory safe to access.
> If Rust did not have a mechanism to use external code then it would be fine
This is what GC'd languages with runtimes do. And even they almost always include FFI, which lets you call into arbitrary code via the C ABI, allowing for unsafe things. Rust is a language intended to be used at the bottom of the stack, and so has more first-class support, calling it "unsafe" instead of FFI.
Yes, there is a boundary, and usually it's either the function itself, or all methods of an object. For instance, a function I wrote recently goes somewhat like this:
fn read_unaligned_u64_from_byte_slice(src: &[u8]) -> u64 {
assert_eq!(src.len(), size_of::<u64>());
unsafe { std::ptr::read_unaligned(src.as_ptr().cast::<u64>()) }
}
O_o