Eliminating Data Races in Firefox
hacks.mozilla.org
hacks.mozilla.org
TSan, ASan, UBSan - if you are writing code and your toolchain supports these, you should use them. Over the past 6-7+ years I've used them, I have never seen a false positive (I'm sure some have occasionally existed but they seem rare in practice).
> However, developers claimed on various occasions that a particular report must be a false positive. In all of these cases, it turned out that TSan was indeed right and the problem was just very subtle and hard to understand.
Yes, I have seen this phenomenon!
UBSan -- again never had a false positive.
Thread Sanitizer does have false positives in some circumstances. They're hard to verify to be falsy and in many ways it's better to refactor to eliminate the complaint because you'll end up with better cleaner code. For example, Qt uses those atomic fences that the article describes [0] [1].
[0]: TSan does not produce false positive data race reports when properly deployed, which includes instrumenting all code that is loaded into the process and avoiding primitives that TSan doesn’t understand (such as atomic fences).
[1]: https://lists.qt-project.org/pipermail/interest/2014-March/0...
In Firefox, media threads use a reduced stack size (128 kB), due to, historically, 32-bit address-space concerns with lots of media elements on a page. Opus, in its default configuration, requires a fair amount of working stack memory (a few dozen kB in some cases). "A few dozen" is much smaller than 128 kB, but apparently enabling ASAN increases the stack usage, quite a bit. So we got crashes under ASAN when it overflowed the stack that would not have been possible without ASAN enabled.
Fortunately, the allocation sizes of these arrays are tightly constrained by the limits of the Opus format (this is not an accident: I helped design the format). I performed a careful analysis (by hand) to ensure the maximum possible stack usage (without ASAN) was not anywhere close to the size allocated for the stack. It is of course possible that I made mistakes, but I tried very hard not to.
The broader point is that this is an entire class of memory access error that ASAN cannot properly validate due to its design. Because it changes the system under test in ways that differ from a real deployment, I cannot use it to mechanically validate my hand analysis.
It is worth noting that valgrind does not have this problem (because it instruments the original binary, instead of making compile-time changes), and is also used by Firefox in testing.
The sanitizers are useful enough that I’ve found adding small targeted changes to support those modes to be valuable vs the increased risk of not testing what you run.
I'm assuming you're still talking about Firefox running Opus [0]? I'm not familiar with it. But I do think that any platform that can run Firefox and wants to listen to audio should have a lot more than 128k lying around. Otherwise your user isn't using the right tool for the job.
I understand memory constraints can be tight and it feels like there's 128k of space right there that you can use. And to achieve that you've written it in C and manually calculated the stack space needed. It feels awesome to have that kind of knowledge and experience with opportunity to leverage it! It's great that you've helped design a media format and implement something to process it!
But Firefox wastes tons of megabytes just to load, store, and transmit telemetry, preferences, search providers... the list can go on. Do you really want to try to convince me that a 128k slab for audio that the user specifically loaded is too expensive in the face of all of the telemetry loaded by Firefox these days?
Why add all the maintenance headache when you can instead allocate a single slab for your own use? If I were to write it and manually optimize then I'd use an allocated slab. And, as a bonus, the slab's ownership can move to any other thread if needed. And it would make it clean in Address Sanitizer. And, even more important, I wouldn't have to worry about what happens if the frames before my entrypoint changes. For example, Firefox's software design could change and adds a couple of wrappers around the manually-sized stack. What happens if I want to compile Firefox on a different platform which doesn't permit a manually-sized thread stack?
So it sounds like you've written something that's going to be very difficult to maintain if you were to vanish. And you've written it for a software product where that level of optimization is, perhaps, better done automatically by tools available to you.
If you build a standalone media player specifically for Opus then that'd be pretty cool. But Firefox shouldn't have pieces that are difficult to maintain.
[0]: https://hacks.mozilla.org/2018/10/introducing-opus-1-3/
As a result, it's not clear that it's a false positive in the technical sense here (as one would need to enable checking for unsigned overflow explicitly).
As a dynamic tool, if a race is observed during execution by TSan, the race is very likely to be real. But it is impossible to "observe" execution for every part of a large code base, and so real races are likely still hiding in some edge case that TSan did not execute.
Static Analysis tools are interesting in that they have the opposite problem. Static tools can detect races across the entire code base, but without runtime information it is difficult to know if the races are real. This leads to more real races being detected, but also comes with more false positives.
There is a lot of interesting working being done on data race detection and how we can bring these two sides closer together to find more real races, without overwhelming developers with false positives.
On the static vs. dynamic side, I would always opt for the dynamic when it can guarantee me no false positives, even if the results are incomplete. It is pretty much impossible to deploy a tool that produces lots of false positives because developers usually will reject it at some point and question every result.
In my experience TSan finds all possible races if it's built with unit tests and unit tests provide enough coverage. If unit tests don't provide enough coverage then there are likely other bugs lurking beyond just data races.
The code coverage analysis I've seen in the past has only evaluates what code is run at least once, but how do you get coverage analysis to evaluate the coverage of what code runs concurrently with what other code effectively enough to get confidence that you'll detect data races?
For example, if you've got a cache that safely handles concurrent use for most operations, but has a data race when two threads both cache miss on the same key within a short window of each other, are there kinds of code-coverage analysis that could theoretically point out that none of your tests involve a cache miss of the same key in two different threads?
I've seen concurrency permutation testing tools, that can reorder the sequence of operations between different threads searching for a permutation that causes a data race, but that relies on actually executing the potentially-racy behaviour at all during testing, right?
I guess this is also what fuzzing is about, but are there tools or techniques you can use to increase your confidence in how well your tests cover the space of concurrent use of your data structures?
I disagree. Maintenance cost doesn't need to be minimized. It needs to be reasonable.
The goal should be to produce high quality software. High quality software is easy to understand and easy to instrument with tools.
(I'd also argue that unit testing can reduce the quality of the software as it makes the internals of the architecture more costly to change).
I think it's more irresponsible to produce something that the customer isn't going to be able to use for many years or that needs a patch before it's even released.
Coverage can be misleading. Maybe there is some function being executed in parallel:
int global;
void foo() {
int local;
if (complex_condition)
bar(&local);
else
bar(&global);
}
So without true 100% test coverage, it is possible races are being missed by dynamic tools, and true 100% test coverage is probably not possible or feasible in large complex projects.
Yes.
> true 100% test coverage is probably not possible or feasible in large complex projects.
I have yet to encounter a single large complex project where 100% test coverage is not possible. I have encountered such projects where 100% test coverage required massive refactoring of the code to facilitate testing and doing so resulting in fixing most customer complaints received.
For instance, pretty much every piece of code (whether native or JS) I've seen is programmed defensively, assuming that for some reason invariants can be broken - something that typically cannot happen unless there is a memory corruption involved. Testing all these defensive programming branches is not possible without introducing memory corruptions.
Removing those "impossible" code branches is a bad idea. "Impossible" things happen millions of times a day to Firefox.
I used to work for LimeWire, which had tens of millions of users. We were sometimes seeing bug reports of "impossible" stack traces. They went down dramatically when we added a splash screen that calculated SHA-1 hashes of all of the DLLs and JARs before using the reflection API to load and start the actual application. Given that the "corrupted" JARs were still presumably passing the JARs' CRC-32 checksums and bytecode verification at class load time, I think we were most often just detecting dodgy RAM that would later have caused "impossible" errors by corrupting bytecode or native code after bytecode verification.
If you're writing code to defend against "dodgy RAM" whose symptom is a corrupt SHA then it's super trivial to test that code: have a valid JAR file and adjust its content so that it will still "load" but isn't the file you expect. Then verify that the program refuses to load the JAR.
Sorry for being a bit slow today. but what's the difference between "load"ing and loading? You clearly understand that you're using two different definitions of load, otherwise you wouldn't have put one in quotes. (Also, it would be self-contradictory if the two were identical.) However, I can't for the life of me figure out two different definitions for "load" which both involve CRC-32 and bytecode verification checks, and yet aren't identical.
If I'm correct, we saw a dramatic reduction in the "impossible" errors not because we were detecting the actual problem that caused the errors (corruption of internal JVM buffers after bytecode validation), but instead because our SHA-1 failures (loading chunks of files into our own buffers and checksumming those) were highly correlated with the "impossible" failures because they had the same underlying cause.
In our case, the JVM loaded the code from the JAR, which, unless I misunderstand Sun/Oracle's classloader, means that both CRC-32 checksums and bytecode verification pass, even for unsigned JARs. The chances of random bit flips passing CRC-32 checks are 4 billion to one. The odds of random bit flips passing both a CRC-32 check and bytecode verification are probably at least quadrillions to one. I'm pretty sure the corruption was happening minutes to hours after the classes actually got loaded by the JVM.
In other words, I think we were getting saved by SHA-1 summing a bunch of large files being a poor-man's memtest86: writing semi-random-ism patterns to big chunks of memory and verifying you can read back what you wrote. Failing this correlates heavily with seeing "impossible" errors at runtime, but we weren't directly examining the classloader's internal buffers that were the actual cause of the errors.
On a side note, modern OSes should really have a low rate memory checker built into the kernel's memory manager. It's a crutch, but cheap consumer IoT devices and WiFi APs aren't getting ECC any time soon.
That is:
1) the JAR file successfully loads by the JVM -- as in it will "load"
2) the JAR file doesn't pass the SHA checksum requirement you've added -- as in it isn't the file you expect.
I've had this happen to me personally. Looking again and again at the tsan error and the code it pointed to, I was absolutely sure it was a false positive. Thankfully I got a friend to look at it and after a good night's sleep he was able to explain to me how exactly the code was wrong.
Tl;dr: It is extremely hard to write correct threaded code in C/C++, and it is quite hard to recognize errors even when they are pointed out to you.
* Rust had poor support for sanitizers ("Rust, which has much less mature support for sanitizers)". Work was done to improve this, but it's not clear how well it works now - it would be worth a blog post in itself IMO. This was surprising, because at least I assumed that it works flawlessly thanks to clang.
* Rust / C++ bridges are predictably problematic and weaken Rust's guarantees ("races in C++ code being obfuscated by passing through Rust"). This is the problem to solve for most companies working in a traditionally C and C++ oriented domain, as the C* code will dwarf the Rust code. Because of the shared memory space any problems in that code will be just as critical as if the entire code was unsafe. It would be quite awesome if Rust were able to isolate unsafe code not just at compile time, but also at runtime, by e.g. offering the option of executing the unsafe code in a separate address space. This would require some manual work for moving the data back and forth, but it could offer the tools to support such a work mode e.g. unsafe blocks could take parameters.
I wish Mozilla would be more transparent about the absolutely normal and typical difficulties they must be encountering in a mixed code base. I have the feeling that there's plenty of them, but they don't want to emphasise them in order not to discourage Rust adoption.
On the other hand, integrating Rust and C* isn't terribly hard, there is automated tools that generate bindings for C/Rust so you don't really have to think about it much other than unpacking the unsafe data from C*.
Also, because the .NET GC churns through a lot of memory, statistically almost every problem on the C++ turned up on the C# side first, sometimes with very confusing effects. For example, .NET has a compacting GC, so an object can get corrupted and then randomly moved somewhere else. We didn't have that kind of problem often, but when we did, it was pure hell to debug.
> This is in contrast to many of our C++ races, which often involved things being randomly accessed on different threads with unclear semantics, necessitating non-trivial refactorings of the code.
Here's a video that briefly talks about what Loom can do https://youtu.be/rMGWeSjctlY?t=7519 (it's part of a larger video about memory orderings, and is a great introduction to the topic).
There is a new project out of AWS called Shuttle [1] which is like Loom, but it does random exploration instead of exhaustive exploration, which enables massively distributed testing of really complicated stuff.
Or are these only implemented in rust and they'd still work with C/C++?
http://demsky.eecs.uci.edu/publications/c11modelcheck.pdf
This page describes how to get the code (linked from the paper):
http://plrg.ics.uci.edu/software_page/42-2/
I found some data races in an ODBC driver using TSan. Running it through CDSChecker is one of my round tuit projects.
I think the underlying assumption is more along the lines of: the easy cases are made easier, and the hard cases are made impossible — or rather the concern being that rust hasn’t made the problem easier, it just moved it around.
Which is often the case with claims of solving long-lasting problems
This is largely the point of Rust, though. Rust takes problems that used to manifest at runtime and causes them to manifest at compile-time, shortening the feedback loop drastically. It also takes problems that used to manifest all over the codebase and roots their causes in the leaf nodes of the program that are making use of `unsafe` blocks. Moving around the problem turns out to be a valuable way of addressing the problem, without ignoring any of the essential complexity or making anything fundamentally impossible.
More importantly, forcing you to catch them. Problems at runtime don't get caught until that condition actually occurs during an actual run of the software.
This is what I really like about programming language theory: problems which are literally undecidable in existing languages, can be made almost trivial by designing a language with that problem in mind. The trick is to have the language keep track of the information needed for the solution, so we're not dealing with "black boxes".
A classic example is annotating every variable with a type, and propagating that information through compound expressions; lets us spot problems like trying to use a number as a string, without worrying about the halting problem (we do this by forbidding some technically-valid programs, like `1 + (true? 10 : "hello")`).
Of course, this leaves the much harder problems of (a) keeping the burden imposed by this extra information down to a reasonable level, and (b) trying to make the language compelling enough for people to actually use. Rust is one of the few examples of going all-in on a new ecosystem, and it actually being embraced. Often it's easier to "embed" a new language inside an existing language, to piggy-back on the existing tooling (e.g. the variety of auto-differentiating languages built in Python)
Rust does not, by any means, solve this problem. What it does do is help you prove that the problem is contained to a smaller subset of your codebase that you can audit/reason about much more easily.
Except in this case the article linked demonstrates that they've made things dramatically easier overall. Data races went from being endemic in a C++ codebase to being in a few low-level and easy-to-locate areas.
The easy stuff stayed easy, the bulk of the rest was folded into the easy case, and even the hard cases appear to have been made easier to find, identify, and fix.
In addition, `shared` types cannot be accessed directly. They can only be accessed via library routines, such as D's atomic library.
Having an `immutable` data type also helps, as immutable data can be accessed by multiple threads without need for synchronization.
This is called Sync in rust. There is also a Send trait for types types which can be moved (but not necessarily shared), as there are data items which can not leave their origin thread e.g. thread locals or non-atomically refcounted instances.
I ended up just casting to and from `shared`, and that seemed to work, but it was pretty verbose for highly parallel code and I was never quite sure under what circumstances doing that might violate the compiler's invariants.
Also such casting appeared to eliminate most of the benefits, since what appeared to be local data might have been cast and had a pointer sent across to another thread elsewhere. In the end `shared`, @nogc antics (the language internals weren't fully annotated), closures requiring the GC (compare to C++), and the friction of interfacing with highly templated C++ code such as the Eigen library caused me to abandon the attempt. I sure learned a lot in the process though!
It's very similar to C++'s memory model, but it isn't necessary to understand it to be effective with D.
> casting
Casting to/from shared is necessary, but is restricted to @system code.
I mean, it's great that you can write threaded code with more confidence, but it seems the conclusion here is that Rust gives you more confidence but not absolute confidence and you still need the normal stable of runtime detection tooling. The second fix[1] in particular resembles the bad pattern of "sprinkle atomics it until it works" you see in non-Rust languages.
(If this comment raises your hackles, please read the "no true Scotsman[2]" fallacy before responding.)
[1]: https://hg.mozilla.org/integration/autoland/rev/044f8b594c71 [2]: https://en.wikipedia.org/wiki/No_true_Scotsman
(emphasis mine)
The important thing is that the races only happened in unsafe { } blocks. It's well-established that these blocks should basically be treated like C++ in terms of scrutiny, but (I believe) you can still have roughly "absolute confidence" in non-unsafe code.
It's true and interesting that unsafe { } blocks deserve sanitization analysis like C++ code does - and maybe this hasn't been discussed enough - but I don't think it's fair to suggest that the enormous language benefits are outweighed by some false sense of confidence. The proportions in that idea just seem way out of whack.
Looking at https://searchfox.org/mozilla-central/rev/ee9dab6aa95f167a34..., SwCompositeGraphNodeRef has "safe" methods (line 286) with unsafe blocks, handing out raw &SwCompositeGraphNode and &mut SwCompositeGraphNode to the interior of an UnsafeCell, without locking or runtime checking. I think SwCompositeGraphNodeRef "caused" the UB because it was invoked by some other caller to create a &mut SwCompositeGraphNode without proper synchronization... And worse yet, SwCompositeGraphNodeRef has Deref and DerefMut implementations that implicitly create & and &mut, unsynchronized, at call sites that merely look like method calls...
My theory as to how this happened:
- C++ atomics are not mutable through a const *, so the authors used &mut to indicate mutability (even though Rust's &mut has stronger guarantees, according to the Stacked Borrows memory model, "all other pointers to this object are invalidated, except the pointers we reborrowed from"... though Stacked Borrows may change to make async fn and intrusive linked lists sound, I suspect it won't make this code sound).
- Maybe they thought "we want to prevent calling &mut SwCompositeGraphNode methods from a &SwCompositeGraphNodeRef". IDK. Semi-related: SwCompositeGraphNodeRef is a #[derive(Clone)] Arc<UnsafeCell<SwCompositeGraphNode>> but !Sync, so only threads that own a SwCompositeGraphNodeRef can create either &SwCompositeGraphNodeRef or &mut SwCompositeGraphNodeRef. And I don't see &SwCompositeGraphNodeRef floating around, so I don't see why access control would matter.
(EDIT: The commit introducing the unsound abstraction is https://hg.mozilla.org/integration/autoland/rev/8d47e408c578, filed under the bug at https://bugzilla.mozilla.org/show_bug.cgi?id=1670328.)
(That notwithstanding, I agree that Rust is a vast improvement over C/++.)
In C++, data races can cause jumps to invalid addresses and the compiler will assume data races away and delete checks that are redundant in the single-thread case. The equivalent problem happens in most languages where data races are possible (two Python threads writing at once can put a value into an impossible/illegal state)
An ordinary race condition is just an observable effect of the nondeterministic progress of threads, so it's possible to have a truly benign race condition, so long as the correctness of the whole program doesn't depend on the outcome of the race. I/O causes nondeterminism anyway, so it's not really practical to avoid all ordinary races.
One of the reasons that Rust's ownership model is necessary is most data races cannot otherwise be statically detected. The compiler will be unable to apply any of its theoretical UB optimizations. It will emit its native code and you get the defined behavior of the underlying hardware. There will be no undefined behavior. If that ever changes, the bigger news, I imagine, will be that we solved the halting problem.
(Worth noting that data races have a clear, unambiguous definition, and therefore are possible to eliminate through tooling. Race conditions rely on program intent or logic, and are therefore... not really.)
Also, I think it matters, generally, that we are truthful about what Rust prevents and does not prevent.
This is a great point. Usually people only talk about the two categories in terms of vague severity, but this is a hard distinction which has lots of implications for how each can be dealt with.
This is why I think Rust discussions about this subject don't focus enough that the language only validates a very tiny subset of races.
Some data races really are benign while some other race conditions can lead to data corruption or security vulnerabilities. Why then should data races be strictly more severe than data races? What am I missing?
What safe Rust doesn't have is _data races_.
I think when you build Rust projects that require reliability, it's foolish to lean back and believe that marketing slogan and not look for concurrency bugs. However, once you found them, it's way easier to fix them, at least when your unsafety is well encapsulated like how it's recommended.
I think the big challenge for Rust in this context is to improve its tooling so that using it is comparable or even easier than the C++ tooling. This blog post is proof that such tools should be ran otherwise you are missing out on important bugs.
I hate that you're getting so many downvotes, and I wish people wouldn't do that...but I still want to disagree slightly :) I think it's interesting to note that the meaning of "you" changes between the two cases. If I write safe code on top of parking_lot, you could argue that I don't really need to do runtime race detection. Rather, someone in the ecosystem needs to do it, but once it's been done, the shared code is improved for everyone. If my code is all safe, I can benefit from these global improvements over time, and I can be confident that I haven't introduced any new races myself.
So Rust did exactly what we wanted here, the problem code was narrowed down to the code that was explicitly unsafe and needing caution. Humans remained fallible, and the ThreadSanitizer spotted their mistake.
Even worse, there are very few, if any, tools that could help you analyze a multiprocess program for any kind of race conditions.
[1] https://servo.org/ [2] https://mjtsai.com/blog/2020/08/11/mozilla-layoffs/