Experimenting with Rust in Chromium(chromium.googlesource.com) |
Experimenting with Rust in Chromium(chromium.googlesource.com) |
I worked at Google years ago and there was a presentation once done on some optimizations done on Chrome performance. This is probably 10 years ago now.
So C++ has std::string of course. C libraries however use "const char ", which has lots of problems. The C++ designers allowed you to avoid friction here by allowing you to pass a std::string to a function expending a const char . Technically, this is an operator method.
It was discovered that the Omnibar in Chrome went through many layers of translations between std::string and const char *, back and forth, such that there were approximately 25,000 string copies per keypress in the Omnibar.
So my point is that even with a ton of resources writing good, efficient and performant C++ is still nontrivial. And that's really the point of Rust (well, one of them).
It's the other way round. You can pass a const char* to a function expecting a std::string. Passing a std::string to a function expecting const char* will generate a compile error.
You need to call c_str() on the std::string if you want to pass it as a parameter to a function expecting a const char*.
Those were not difficult to understand C++ features either, but basic ones which were very well known by then. I remember reading Bulka & Mayhew’s Efficient C++ (published in 2000) which mentioned the importance of avoiding copies, calling reserve and many other techniques.
So your point is wrong. Not copying strings all over the place, calling reserve, not creating temporary containers are junior-level C++ skills.
Isn't this a bit oversimplified? IIRC the Rust support was only started as an option for writing device drivers in Rust, not actual kernel code, and even this relatively simple use case looks like a herculean effort which required changes to feed back into Rust - which is a good thing of course, but currently it looks like the Rust ecosystem benefits more from that project than the Linux ecosystem ;)
The beginning of this sentence didn’t surprise me, but the fact that it’s just Linux and Android did. Rust supports macOS and Windows really well, so I wonder what the gap is here?
> Facilities and tooling in Rust are not as rich as other languages yet.
Is this meant in the context of Chromium?
Picking just one piece: Consider that Chromium builds happen in a distributed build farm. There's multiple variants of this (goma, rbe). I'd imagine those systems would have to be modified to support the Rust toolchain for that target.
And it looks like this work is built around making GN/ninja support Rust, just using cargo directly. So that's what they likely mean by "not as rich as other languages yet."
In fact there are articles of people burning out with the release process of Chrome, left behind by the sheer speed at which the project is moving forward.
[1] https://chromium.googlesource.com/chromium/src/+/HEAD/styleg...
What are the alternatives in the context of Chromium development as a replacement for C++?
See "Using safer languages anywhere applicable".
[1] https://en.m.wikipedia.org/wiki/Rust_(programming_language)
All the chaos of the so-called web standards, decades of accumulated C++ complexity and the eccentricity and burgeoning complexity of Rust on top. Getting assigned to such a project must be akin to punishment.
You have to think very very carefully about every line and character in C++ to figure out what it’s doing. Sometimes the easiest way to review it is to compile it and read the assembly.
The point is, with rust you have more options to enforce no-copy through the type system.
Rust for Linux was driven by existing Linux developers. So they obviously think that's great news for Linux not just for Rust. One reason is: Nobody else was delivering a viable way forward. If there were ten of these safe low-level languages and by doing nothing Linux could be sure one would pick "Make ourselves suitable for Linux" as its goal then they could just sit back. Instead Rust is the only game in town, so it's either adjust Rust [ e.g. Rust for Linux alloc crate doesn't have the concatenating + operator on String like your userspace Rust does because that's the sort of thing which makes Linus angry ] or risk not having any safe path forward for the foreseeable future.
And yes, ultimately whether Rust for Linux goes into the Linus tree at some point is always ultimately a decision for Linus Torvalds. It's just that he made friendly noises about it, and there are people putting the work in, so it's like for dozens of other prospective Linux features, we can assume it will land but we shouldn't assume when.
In two words, Google programmers are, as a rule, vastly overrated. They can maybe rope in 100,000 cores on one query, but nothing in their recruiting selects for good coding habits.
Anybody coding C++ in this day'n'age and getting use-after faults needs to go to the back of the line. They will certainly succeed in writing new Rust code that is as bad as their old C++ code. (Note: deadlocks are officially "safe".)
Recently Google made a big push to change the std::string constructor from a null char* to yield the empty string, instead of honestly segfaulting. That failed. They had a half-baked (and hellish, for users) async/await design they tried to put up as worth delaying the whole feature into 2023. That failed.
Rust succeeds, because it does not rely on programmers writing bug-free code. Bad Rust code is not as dangerous as bad C++ code.
BTW: deadlocks are not exploitable for RCE, and are quite easy to debug compared to data races and heap sprays.
70% of all vulnerabilities reported to them are memory safety issues. To my knowledge, 100% of exploits against Chrome in the wild leverage memory unsafety.
However, it gets a lot harder to argue that it's just Chromium when Microsoft found the same thing: https://www.zdnet.com/article/microsoft-70-percent-of-all-se... At that point, the strongest argument shifts from "Chromium is the outlier" to "my code is the outlier". And that's... possible to defend (ex. the OpenBSD folks have a track record that says they can write safe C), but certainly harder.
Doing that seems a bit easier than integrating a whole new language into your project (although I know Rust would have other benefits).
Specifically the "ancient past" here is prior to C++ 11 when C++ decided now it wanted to actually define how its string type works because C++ 98 and C++ 03 strings are both even more dangerous than most things are in C++ and had to be put out of their misery.
... how so ? they were just CoW which is actually I think the better choice most of the time... now there are copies all over the place
10 years ago (when the parent mentioned they were still at Google) c++11 was already out.
Originally C++ doesn't have a string type, the C++ 98 standard does standardize a string type but it's only loosely specified. Most implementations do something "clever" which it turns out is a bad idea (this is a recurring theme in C++. Only in C++ 11 does the standard say OK, we'll prescribe how the string class actually works, making it more complicated but hopefully avoiding the worst problems.
Chrome was launched in 2008, and much of its internal structure was far older having incorporated work by Mozilla and Apple.
I don't know... I see plenty of vulnerabilities in the Java world, no memory unsafety needed.
A browser written in a memory safe language won't be free from vulns but it will be free from a huge class of recurring and very serious vulns.
Security is not a binary yes/no. There are many many ways in which programs can be insecure. Eliminating one class of bugs helps reduce total amount of issues and their severity.
It's like car crashes: cars with seatbelts, airbags, and auto-braking systems kill way fewer people than they used to, but are still deadly.
Yes, the above is a bit oversimplified: most memory-safe languages have an "unsafe" escape hatch, so technically these vulnerabilities are possible; however, these escape hatches are rarely used, explicitly opted-into, and clearly demarcated in the source code such that the number of vulnerabilities in "memory safe" languages is far smaller than "memory unsafe" languages.
I think wider memory safety, SQL injection and things like log4j are related to the former. Some aggregation of data that should be dumb and restricted is given too much trust, so data can be lifted to code and code is too powerful. In essence they are all similar, even though we don't use the same technical terms for each of them.
And yes, if a given programming environment restricts a class of operations, it is given less power so the attack surface is qualitatively smaller. Languages that restrict memory management are an example. Another one would be file/disk access, network access and so on.
As for Java, quite a lot of the exploits against it (when it was a browser plugin) were in fact memory safety issues in the VM. But more recently what we see are serialization issues, which Rust also does not have.
Don’t forget the origin of WebKit, KHTML from the KDE folks.
Long running systems were still using pre-standardisation libraries for string up to 2012 however, so you may well have come across such projects.
Bzzt. C++ 11 completely overhauls how std::string is defined.
C++0x was a thing, with varying levels of support from all major compilers, for years before C++11 was finally ratified.
In the context of my original comment though, no matter how dated the code base, I think it unlikely that Chrome was using any variant of std::string that had an implicit conversion operator for const char* such that string could be passed as a parameter to a function taking const char* without needing to call c_str().
With C++, if you have char*'s (because you don't need to own the memory) and you pass it to a function that takes a const std::string& (because it also doesn't want to own the memory), then there will still be an implicit conversion to a temporary std::string (involving an allocation) despite neither the caller or the callee needing to own any memory.
With Rust, if you have a &str (because you don't need to own the memory) and you pass it to any function that takes a String (or even the unidiomatic &String), then you will get a compile error. There won't be any implicit conversion of types and therefore no implicit allocation. If you really want to pass it, you need to explicitly convert it, making the cost of the allocation explicit.
Rust's "too many strings" model says "there are many different ways in which you can use string-like objects, each with their own performance tradeoffs. Know which one you want to use in your code or I won't compile".
You should try to avoid calling OS APIs in general and cache the results as much as possible. Who knows what the performance characteristics are of an API that has to serve 7 layers of historical OSes simultaneously. Unless you're directly interfacing with the kernel you shouldn't expect much. Omnibar-like layered calls between your app and the OS are a worst-case scenario regardless of conversions.
At some point there will need to be an allocation when crossing Rust -> C++ boundary because Rust strings are not null-terminated.
The difference Rust makes is that unlike C++, it is always explicit when the allocation occurs.
If you are using C strings in C++ you are either doing something incredibly low level or don't care about performance at all. C strings require strlen calls or something equivalent for basic operations and you can easily run into code with exploding runtime if you aren't extremely careful.
…or interoperating with C code?
Note: chrome only supported c++17 features in Dec 2021 [0], and whether std::string_view is allowed to be used is still 'to be determined'.
0: https://chromium.googlesource.com/chromium/src/+/HEAD/styleg...
Technologically, Rust's only built-in string type, &str, is a reference to a string slice - that is, you can't change it (the reference isn't mutable) and it is both a pointer to the start of some UTF-8 and the length of the UTF-8.
What encoding? Always UTF-8. Only UTF-8. Not "Well, it's kinda UTF-8 but..." it's just always UTF-8. This moves the burden to a single place, your text decoding code, to do things correctly, and great news - the entire world is moving to UTF-8, so you're on a downhill gradient where every week this works better without you lifting a finger.
That reference knowing the length is brilliant. Trimming whitespace off a string? You can just make another immutable reference to the smaller trimmed string. Zero copies. Slicing a URL up into components? You can do that too, zero copies. And yet it's all memory safe.
Now, Chromium is not some raw firmware for a $1 micro-controller, so it has library types like Rust's alloc::string::String (you can just name it "String" in normal Rust code but that is its full name) which, as its presence in alloc suggests, is an allocating String type, you can concatenate them, you can make them by formatting a bunch of other variables, the default ones are empty, the data goes on your heap and so on. But, String is AsRef<str> which means if what you've got is a String, and what you're doing is calling a function that wants &str Rust is OK with that and it costs nothing at runtime. Why? Because that &str is just two of the elements of the String type you had, the pointer into the heap and the length, it's easy.
Rust has lots of other types for stuff like Foreign Interfaces, like CStr and CString (for the C-style NUL-terminated array of bytes which might be text) but your pure Rust code shouldn't care about those, often it can say (unsafely) "Look, the C++ promises this is UTF-8, we'll take their word for it" or "I only need it to have bytes in it, let's make [u8] and we're done".
Culturally, Rust programmers write &str when that would do. There's a strong cultural pressure not to write String when you really mean &str, and the compiler won't let you write &str if you needed String. So this results in less thunking of the sort complained about in C++
Rust makes all of that easier to see during code review. It is very explicit about these things.
I'm a Google employee working on chromium and chromeOS and have been asking internally about rust support for over a year now, so it's exciting that it's making progress.
The string slice for example is an Unicode-capable view into bytes of the string (immutably pre-compiled static bytes in the binary, bytes of fixed length on the stack or heap-allocated). The aliasing rules are enforced by the compiler, so it is safe to throw around pointers and sizes and not to worry about buffer overflows, as long as it compiles.
As others have mentioned, various kinds of string types are baked into the language which makes it ergonomic to do “the right thing” from the get go, but hard to say. I would be skeptical of claims that it would make a difference, especially in the interim where you now have an added impedance mismatch with C++, Rust, C.
The expression of what C++ does here in Rust is awkward and, I think, nobody has proposed it because you'd never write that. Basically C++ char* is a raw pointer. Rust does have those, but you'd never use them in this context.
What you would use is either the borrowed slice reference &str or the owning String type, but in both cases we have an owned object and there's our crucial difference. If you've got the owned String, and I needed an owned String, I should ask for your owned String, and we're done.
In C++ "dropping" ownership as you describe is no big deal, the C++ design doesn't care, but in Rust if you actually drop(foo) it's gone. The references to it can't out-live that, if it's gone then they're gone. If you write code that gives away references and then tries to drop the thing they're references to, Rust will object that this is nonsense, because it is nonsense, you need to ensure those references are gone before dropping the thing they refer to.
As a result I feel you're greatly under-estimating the ergonomic difference.
I think you’ve built a straw man of my argument and then argued with that.
Clearly I meant that it seems possible that a sufficient complicated call stack could still be set up to jump between needing the owned String type and the borrowed &str type. That’s what I meant by dropping ownership as that’s what’s happening in the c++ code when you go between char*/string (the API is dropping its need for ownership). The argument of “ If you've got the owned String, and I needed an owned String, I should ask for your owned String, and we're done” is weak because that same argument would apply to C++ code and yet the code still ended up that way when you pasted together components in a very large code base. Now maybe it’s a bit simpler because you have string, string&, const string&, and const char* and doing that antipattern that happened in C++ just wouldn’t be ergonomic in Rust. Maybe. But that feels like a very thin argument and not “this is impossible in Rust”.
So I've got the string "IR Baboon big star of cartoon" and I take references into it, which are cheap and then you use your C++ 98 copy constructor to get another string, which of course also says "IR Baboon big star of cartoon", when you took it -- and then I scrawl "I AM Weasel" on top of my string using my reference and now your string was changed because it was COW.
If you liked COW for this purpose Rust has std::borrow::Cow which is a smart pointer with similar flavour, Cow<T> is a sum type that's either a thing you own T (and thus you could modify it) or it's a reference, perhaps &T (and thus you can't modify it) but which promises you could get an owned thing (e.g. for strings by deep-copying the string) if you need one. Methods that would be OK to call on the immutable reference (e.g. asking how many times an ASCII digit appears in the string) work on Cow<T> and if you find you need to mutate it (maybe in a rare case) you can ask the Cow for the mutable version, if it already had the owned version you get that, if not it will make one for you.
Rust's traits kick in here, Cow<T> requires T: ToOwned, which is a trait saying "I can make an immutable reference to T into a new thing T you own", obviously types you shouldn't do that to simply do not implement ToOwned and so you can't make a Cow of those types. The standard library provides in particular an implementation of ToOwned for &str which makes Strings from it.
> and then I scrawl "I AM Weasel" on top of my string using my reference and now your string was changed because it was COW.
I mean, that's the point of references... no ? If I wanted a different object I'd make a copy.
Like, even with just one string, without any CoW, your post makes it sound like you'd be surprised than if you had:
void set_some_config(const char*);
char* get_some_config();
std::string s = "foo";
set_some_config(&s[1]);
s = "bar";
get_some_config();
In the explanation I posted, you do make a copy to get a different object "you use your C++ 98 copy constructor to get another string".
The problem happens because both strings share the same bytes to represent the text "IR Baboon big star of cartoon" as part of the COW optimisation. But my reference can scribble on this shared text.
I don't see how your get_some_config is similar at all. Notice that with C++ 11 strings, the copy constructor gives you a deep copy of that "IR Baboon" text and so my references can't smash your string.
Is a single atomic increment really that expensive? I mean we are not even talking about a full memory barrier here, just the atomic increment's implied acquire and release on the single variable. Other operations not dependent on a subsequent read could still be re-ordered in both directions.
And also keep in mind that the alternative was copying the whole string instead. Which means both heap memory allocation (which is often pretty expensive, even with per-core heaps), plus the actual copying. Unless a platform has a terrible implementation of atomic increment, or you have a std::string that is frequently getting copied on multiple cores (so as to have meaningful contention), I would have expected the actual copying implementation to be slower. But I'm not super familiar with the timings of these things, so i certainly could be mistaken.
My understanding was that the change was more for about being able to set proper bounds on some operations, ensuring .c_str() is O(1), and not O(n) sometimes, and similarly with string writes, etc.
std::string to const char doesn’t (usually?) require copying.
https://docs.microsoft.com/en-us/windows/apps/design/globali...
[1]: Windows has two variants of many API calls with either -A or -W suffix, where the -A suffix is for strings formatted as 1-byte ASCII (or a specified codepage) and the -W suffix is for strings formatted as 2-byte UTF-16 (kinda). Example: DlgDirListA / DlgDirListW, https://docs.microsoft.com/en-us/windows/win32/api/winuser/n...
Chromium has "string_piece": https://chromium.googlesource.com/chromium/src/base/+/refs/h... which is at least 9 years old (was moved from elsewhere in the repo into base/ then)
Despite having better options available (either in the standard library or custom code), the less optimal ways still get used.
Rust had the benefit of learning from c++'s mistakes and separated the concepts of owned vs unowned strings in to separate types with explicit conversions required whenever an allocation would occur. This was baked in to the language from the beginning and so you don't get a mix of different types in signatures to convey the concept of pointing to a slice of a string owned by someone else, you just have &str.
Even if you get fancy with your interface and do things like AsRef<str>, there's still no concern about implicit or hidden allocations. Any time you need to own the memory (either for yourself or to pass in to another function) you need to do so explicitly and you end up with a different type (much to the chagrin and confusion of newcomers to the language).
C++ is trying to correct its mistakes also, but not everyone can use those latest features and even if they can, the mistakes still have to be left in for compatibility reasons.
The "I should ask for your owned String" argument does not apply equally well in C++ because of a crucial design infelicity in C++. Your caller may well not have an owned std::string.
In C++ raw char pointers are totally a thing. Because std::string is a late addition (if you learned C++ in the early 1990s a "string" class was maybe an interesting exercise, not a library type) the string literals aren't a built-in string type, and much of the API isn't shaped for such a type either.
Now, the effect is those are (sometimes) owning pointers, it is possible I own some C++ string in this sense, and all I have is a pointer into it. If I give you that pointer, it's not because I didn't give you the owned string, that pointer is my owned string. You want a std::string and there's no reason I would have one at all.
You can mutate these strings, but of course you can't extend them because you've got no way to know how to communicate with the allocator, maybe they live on the stack, or in a private heap. At the time this seemed like a good idea, today we don't think so.
// some process managed by team a
fn caller1(...) {
String s;
level1Callee(&s);
}
// some process managed by team b
fn caller2(s: &str) {
level1Callee(s);
}
// library 1 by team c
fn level1Callee(iDontNeedOwnershipOrDoI: &str) {
level2Callee(iDontNeedOwnershipOrDoI.to_owned())
}
// library 1 by team d
fn level2Callee(iNeedStrongOwnershipBecauseIMutateTheStringAnyway: String) {
level3Callee(&iNeedStrongOwnershipBecauseIMutateTheStringAnyway)
}
In a lot of places in Chromium the impedance choices are arbitrary. You have a raw pointer but need a string or vice versa. Ownership isn't the problem, so in Rust you literally just always choose &str for these APIs and pay nothing. A team who design their API taking &String in this situation get the same treatment as a team who name all their types Data1, Data2, Data3 and so on. Somebody senior fetches the water spray, "No. Bad programmer".
You might be outraged, surely C++ programmers also never get this wrong. But nope, happens all the time as Chromium illustrates.
Added: One cause that shows up in my review is this:
C++ strings know how long they are. The raw pointer does not. As a result if we have lots of people asking if their thing is "some text" it's tempting to demand they give us a string, since if the string's length isn't 9 we don't need to look at the text itself. It's an optimisation! Rust's &str knows how long it is.
Certainly plenty of widely copied strings are short strings, so a COW implementation that lacks the short-string optimization could very easily be a bad bottleneck for multi-core compute.
My impression through the fog of history is that what happened was a really clever GNU person with little foresight and no access to an SMP system implemented std::string with CoW. Its performance in practice was so poor that the standard committee intentionally changed the standard to make it an illegal implementation, thereby eradicating the GNU CoW string. There was no higher principled logic.
https://blogs.msmvps.com/gdicanio/2016/07/09/is-copy-on-writ...
https://oribenshir.github.io/afternoon_rusting/blog/copy-on-...
Also, the point of that was to improve multithreading of string: I think this very idea is problematic. I've written at this point hundreds of thousands of line of C++, and the number of times where strings are really, by design, supposed to be shared across threads is honestly counted on the fingers of one hand, just like e.g. justification for using Arc over Rc in rust. 99% of string handling is done as some GUI work on the main thread or as part of some task processing done in some network thread, which stays in that thread.
> A team who design their API taking &String in this situation get the same treatment as a team who name all their types Data1, Data2, Data3 and so on. Somebody senior fetches the water spray, "No. Bad programmer". You might be outraged, surely C++ programmers also never get this wrong. But nope, happens all the time as Chromium illustrates.
So the thrust of your claim is that Rust programmers are better. The more generous interpretation I’ll read here is that Rust has stronger conventions here. Even still. I don’t see it. The problems are still the same. I think there’s a hyper focusing on ownership when it has nothing to do here. In c++ Const Char (and now string_view) == &str. Chrome developers were going from &str to String and back many times. If Rust had some way to convert &mut str to String without a copy if none was needed, then I think that might apply.
As a Rust programmer this is what I'd expect, because that's what alloc::string::String is for but this is not how std::string is used in C++ & especially not how it was used at that time
So I spent some more time reading. Beyond signifying ownership std::string has other properties that the raw pointer char * does not have which influence the C++ programmer and especially the enthusiastic but perhaps less experienced C++ programmer
1. It's a real C++ type whereas char * is left over from C
2. Unlike char * the std::string remembers the length of the string which also speeds up equality comparison (we know "classification" != "class" from the length before we even look at the text data)
3. std::string has a "Small String Optimisation" which will be emphasised repeatedly to you by C++ gurus. This means small local strings don't need heap space which is good. So... you should use std::string?
Now. If we compare &str and alloc::string::String:
1. Both Rust types, not left over from some prior language
2. Both know how long they are
3. Neither has "Small string optimisation". You can do this trick (oh boy and how) in Rust, but Rust's standard library intentionally does not provide it and nobody's public APIs expose such a thing.
> &mut str would let you mutate the existing characters (similar to char) but it doesn’t give you permission to resize (since doing so obviously might involve a reallocation and change underlying pointers referenced elsewhere).
Yes it would, and this exists but I've rarely seen it put to any use.
> The more generous interpretation I’ll read here is that Rust has stronger conventions here.
Ultimately yes, the conventions are stronger, a cultural difference. You can go look for yourself, at both the sprawling vastness of public Rust library and Chromium's own APIs in that era which often take string despite having no interest in ownership.
Chromium has a map (of configuration parameters) inside it, in which the keys are std::string. If you understand this as an owning object for mutation that sounds insane but if you just think it's a convenient object that knows how long the text is and keeps shorter text out of the heap it's awesome. Right?
Of course, you can't just compare a char * to a std::string, the map has no idea that would be possible, so you make a std::string from your char * and compare that. Don't worry there are only a few dozen configuration parameters to check, what do you mean this hash lookup now incurs a heap allocation ?