Comment by tialaramex | Hacker News Reader

Comment by tialaramex | Hacker News Reader

Parent Full thread

So let's take a closer look at the fairly recent CVE-2022-21658.

Rust's standard library provides std::fs::remove_dir_all("/some/dir"), in C++ you will find the very similar std::filesystem::remove_all("/some/dir")

As you might anticipate, these functions get rid of the directory /some/dir if necessary deleting sub-directories and files recursively in order to achieve that.

They both explain that they won't delete other things, even if there's a symlink in /some/dir or some child that points to those outside things, the symlink gets deleted but not the thing it pointed to.

What kind of terrible unsafety was discovered in CVE-2022-21658?

Until Rust 1.58.1 you could exploit a TOCTOU race condition to blow away other people's stuff using symlinks when they made this call.

And where is the CVE for the same bug in C++? There isn't one. C++ standards people argued that the C++ standard already actually says that none of their filesystem API is safe to use if you have a "concurrent environment" in which such races are possible. Standard C++ programs are only intended to run on a machine where they have exclusive control over the entire system for the entire length of their runtime and so when your C++ program deletes files you thought it couldn't delete that was actually your fault for not being familiar with every caveat and ruling in the entire C++ standard.

What should you do about that? I believe the answer is obvious: Use Rust

It is a choice between (1) acknowledge it is an OS problem the library can't fix, or (2) pretend the library can fix it. C++ chose one; Rust, you say, chose the other.

The Posix file API is shot through with races. Patching around them is a game of whack-a-mole. C++ cannot fix Posix. Rust cannot fix Posix.

It is dishonest to assert that "C++ programs are only intended to run on a machine where they have exclusive control over the entire system". The system a program is a part of is expected to have enough control over what it operates on that its semantics are well defined. No Standard can do more. Rust can do no more. I doubt Rust pretends to more.

The C++ standard does not, in fact "acknowledge it is an OS problem the library can't fix" in respect of remove_all it plainly says this function doesn't remove files resolved to by symbolic links even though sometimes it does.

The extent of its "acknowledgement" is that it says elsewhere in the standard that if there's any filesystem race you get Undefined Behaviour. C++ programmers can't do anything with this except, as I explained, insist on exclusive control over the system where their program runs during the entirety of its runtime to avoid this Undefined Behaviour.

C++ programmers rely on definitions from a wide variety of places, not limited to the ISO C++ Standard, and not even limited to Standards. Just writing "#include <unistd.h>" is Undefined Behavior as far as the ISO C++ and C Standards are concerned, but ISO Posix defines it. Linux, the BSDs, MSVC, iOS, Intel, AMD, ARM, and NVidia provide numerous additional definitions programmers rely on. A C++ program's behavior is defined by the aggregate of all of them as they apply to the runtime environment of the program.

ISO Standards do not editorialize about things they do not define; they simply do not define them. For what the ISO C++ Standard does not define, you must look elsewhere if you want a definition. Often there is one, elsewhere.

Every last jot and tittle of what a Rust program does is Undefined Behavior as far as every ISO Standard is concerned, but Rust coders rely on other definitions, too, including a specified binding to ISO Posix where relevant, and the various CPU and OS specs.

It is generally advisable not to pontificate on things you do not understand to people who do.

Your comment would be better without the last sentence.

And the handling of that CVE does illustrate a key difference between Rust and C++. Rust says "it's our fault if we get the intuitive semantics wrong". C++ says "it's your fault for relying on your intuition". It does no good to pretend there's no difference between the two philosophies.

You seem to be saying Rust patches up the most obvious Posix holes but leaves the more subtle ones, to make it seem more secure without, in fact, being more secure.

If you fix some issues but not others (the fact that we aren't being specific about which POSIX issues we're referring to is driving me crazy, but anyway) you have in fact improved security. The "well, if it can't be perfect, there's no point in doing anything at all" attitude is deeply pernicious, and it would be unfortunate if that is what C++ has adopted.

If you fix a filed complaint, but leave in place the same means to break in as before, that is what we call "security theater", which has a well-deserved poor reputation.

Can you expand how this is the case here? Where does the Rust standard library leave a “subtle Posix bug” unpatched?

There are so many, if Rust wants to do it right across the board, then for every POSIX API where the behavior isn't fully specified, they need to provide a correct way.

https://pubs.opengroup.org/onlinepubs/9699919799/xrat/V4_xbd...

And if you want a specific example, thread cancelation cannot be wrapped in safe Rust, because many details depend on the actual UNIX is being used.

https://pubs.opengroup.org/onlinepubs/9699919799/xrat/V4_xsh...

> cannot be wrapped in safe Rust

There are a lot of things that cannot be exposed in safe Rust (and that's OK, that's what `unsafe` is made for).

There's a big difference between “using this is unsafe, and you must use the `unsafe` keyword and be sure to manually check your invariants” and “oops you used a safe function from the standard library and it was in fact unsafe too bad”.

The safety problem that gets lost, is that the only guarantees safe Rust can offer are related to memory corruption.

We're literally in a thread about a guarantee that Rust kept (but C++ didn't) for filesystem APIs.

So "related to memory corruption" is either a deliberately vague statement such that you can count "deleted files you didn't intend" as "memory corruption" or you're missing the true extent of what's offered here.

Type safety could also be argued to just be about "memory corruption" but because of the newtype idiom Rust's type safety ensures all sorts of critical things that you wouldn't associate with "memory corruption" as it is ordinarily understood, starting with obvious built-ins like str really being UTF-8 and Saturating<i16> really having saturating arithmetic operations despite only being an ordinary 16-bit integer - but building up to Mutex<Thing> genuinely not letting you access the Thing without taking the Mutex, and GPIO types that reflect the reality that this GPIO is either a PCM device, or it is a serial port, but it by definition is only one thing at a time and must be explicitly switched between them if that's what you want to do.

A fake guarantee, because it is impossible to actually enforce it for every filesystem in use.

If Rust safety goals is to be dependent typed safety language instead of memory, then quite a few knobs are missing.

> A fake guarantee, because it is impossible to actually enforce it for every filesystem in use.

For some time now the situation is that a POC exists which works just fine against C++ programs, but no longer works against Rust programs with current Rust. That seems much more like a real guarantee to me than a fake one.

> If Rust safety goals is to be dependent typed safety language

Rust does not offer full-blown dependent types today, and so far as I know has no plans to do so, although it would be convenient if it was easy (it isn't easy).

But (safe) Rust does have working type safety and this already meaningfully makes us safer as I explained already.

Until, of course, a very slightly different POC is coded that works against Rust programs built with current Rust.

Erecting a fencepost and expecting attackers to run headlong into the fencepost and knock themselves out is not security. Actual attackers just go around.

Again, it is unwise to explain things you do not understand to people who do understand them.

Of course, just "a very slightly different POC". And let me guess, because you are "people who do understand them" you could certainly conjure up such a thing, but alas, you just don't have the time right? So, Rust is vulnerable to an imaginary attack that Nathan Myers claims to be able to write, while C++ is vulnerable to an actual attack that we know works.

As to the fencepost analogy I will note that Rust fixed the problem a little over a month ago and ever since C++ proponents have been running into that particular fencepost quite certain that the resulting brain damage is proof of just how smart they are for not fixing bugs.

You embarrass yourself. It brings me no pleasure to watch it.

But it reflects also, perhaps unfairly, on Rust and its creators, who did not ask for foolish promotion.

Almost this, but it's slightly broader: Rust guarantees means you won't hit an undefined behavior in safe Rust (as long as the underlying unsafe Rust is correct, this is an important caveat). This isn't equal to memory corruption, especially because optimizing compilers exist and UB can lead your code to be completely miscompiled (which is both dangerous and really annoying to debug).

Also, while this isn't a guarantee, Rust's type system and control flow regarding error handling can help avoid mistakes you sometime see in C (the infamous “goto fail” for instance). Obviously, since C++ is already much more expressive than C, Rust is less of an improvement from C++ in that regard.

But pthread_cancel isn't exposed in the Rust standard library! It's marked unsafe in libc, as it should be.

That is not the reason why it is marked unsafe, rather how the C FFI works on Rust.

So now is safety memory safety, or POSIX cross platform semantics safety?

Better make your mind what safe actually means in Rust.

the libc crate exposes functions in Rust, some are safe, others aren't, you don't call the libc crate function through FFI. (The libc crate indeed calls the underlying libc through FFI but this isn't the same thing!)

> Patching around them is a game of whack-a-mole. C++ cannot fix Posix. Rust cannot fix Posix.

Rust cannot “fix Posix”, but the stdlib (and others) can definitely work around its quirks. See how this[1] library sidesteps a thread-safety issue when using localtime_r(3).

https://github.com/x-hgg-x/tz-rs

It is sometimes possible to work around "quirks", particularly when no malice is intended. When it is a question of security, and fending off attackers, the most obvious fix is rarely any kind of fix at all. Attackers can see what you did, and simply step around it.

Bragging about how you put in your broken fix and somebody else didn't tends to attract what may be unwelcome attention.

> Bragging about how you put in your broken fix and somebody else didn't tends to attract what may be unwelcome attention.

The specific thing being fixed here was a TOCTOU race and thus the fix is just to hold on to the thing being checked so you can use the same thing and eliminate the race. Well, of course I say "just" but it was a lot of work.

You obviously have no idea what you're talking about.

There is no attacker involved in this discussion: the entire purpose of Rust is to prevent the user to use a dangerous method without opting into unsafe.

If The user opt into unsafe, there maybe a way for an attacker to exploit a user's mistake, but as long as the user doesn't they'll be safe, because what Rust does is forbidding the user to shoot themselves in the foot.

You can downplay those CVEs, but the fact is they exist.

It is a big difference to sell Rust stating that the number of exploits cases will be lower, which they are, or asserting they don't happen at all.

When the seller tells a message the audience doesn't believe into, the audience loses interest.

> You can downplay those CVEs, but the fact is they exist.

Indeed. Rust decided by policy that it needs to not only fix mistakes, but also tell people about them.

There's a big difference though between the fact that such a large piece of software isn't perfect (to be expected) and the choice to silently ignore whole categories of problem and pretend they just don't exist or blame the programmer for holding it wrong.

This "..., making it difficult to shoot _anything_ besides the correct target with rust" doesn't seem to follow that policy, rather sends the message that everything just works.

C++ isn't perfect and there are many issues to improve, but I also don't agree that Rust is such a perfection.

Looking forward to a perfect abstraction of POSIX in stable Rust, taking into consideration all the caveats on the standard.

After all, I want my Rust code to be perfectly safe.

Reply on news.ycombinator.com