Asynchronous IO in Rust
medium.com
medium.com
There are some programming languages that have sufficiently impoverished semantics that this is their best effort that they can make towards concurrency.
But this is Rust. It's the language that fixes all the reasons to be afraid of threading in the first place. What's actually wrong with threads here? This isn't Java. And having written network servers with pretty much every abstraction so much as mentioned in the article, green threads are a dream for writing network servers. You can hardly believe how much accidental complexity you're fighting with every day in non-threaded solutions until you try something like Erlang or Go. Rust could be something that I mention in the same breath for network servers. But not with this approach.
There's plenty to debate in this post and I don't expect to go unchallenged. But I would remind repliers that we are explicitly in a Rust context. We must talk about Rust here, not 1998-C++. What's so wrong with Rust threads, excepting perhaps them being heavyweight? (Far better to solve that problem directly.)
Nothing. If threads work fine, use them! That's what most Rust network apps do, and they work fine and run fast. A modern Linux kernel is very good at making 1:1 threading fast these days.
> And having written network servers with pretty much every abstraction so much as mentioned in the article, green threads are a dream for writing network servers.
Green threads didn't provide performance benefits over native threads in Rust. That's why they were removed.
Based on my experience, goroutine-style green threads don't really provide an performance benefit over native threads. The overhead that really matters for threads is in stack management, not the syscalls, and M:N threading has a ton of practical disadvantages (which is why Java removed it, for example). It's worth going back to the debate around the time of NPTL and look at what the conclusions of the Linux community were around this time (which were, essentially, that 1:1 is the way to go and M:N is not worth it).
There are benefits to be had with goroutine-style threads in stack management. For example, if you have a GC that can relocate pointers, then you can start with small stacks. But note that this is a property of the stack, not the scheduler, and you could theoretically do the same with 1:1 threading. It also doesn't get you to the same level of performance as something like nginx, which forgoes the stack entirely.
If you really want to eliminate the overhead and go as fast as possible, you need to get rid of not only the syscall overhead but also the stack overhead. Eliminating the stack means that there is no "simple" solution at the runtime level: your compiler has to be heavily involved. (Look at async/await in C# for an example of one way to do this.) This is the approach that I'm most keen on for Rust, given Rust's focus on achieving maximum performance. To that end, I really like the "bottom-up" approach that the Rust community is taking: let's get the infrastructure working first as libraries, and then we'll figure out how to make it as ergonomic as possible, possibly (probably?) with language extensions.
My overarching point is this: it's very tempting to just say "I/O is a solved problem, just use green threads". But it's not that simple. M:N is obviously a viable approach, but it leaves a lot of performance on the table by keeping around stacks and comes with a lot of disadvantages (slow FFI, for example, and fairness).
This seems wrong. This shouldn't be the point of green threads -- green threads aren't a "faster" alternative to native threads. How would that work? How could it be the case that virtual threads implemented in user space are faster than native threads provided by the OS? I don't think anyone expects that to be true of green threads in general.
We don't have green threads as a useful abstraction because they're fast, we have them because they're flexible, and because they allow us to easily capture a lot of patterns. They let you easily and efficiently distribute work amongst multiple processors without having to think about the subtleties of the underlying threading model.
Of course, given Rust's other goals, shying away from green threads makes sense.
https://github.com/rust-lang/rfcs/blob/0806be4f282144cfcd55b...
Sadly, that proposal also removed concurrent IO from the standard library, and it hasn't been replaced.
2. Documentation, stability, portability, quality. I'll keep banging that drum until there's a standard. :)
3. The Mio API is subjectively weird, but I might be missing something.
OK, cool! Can you point us to the existing functionality, or any example of it being used for stackless coroutines?
Stackless coroutines are — the compiler (or a clever library) transforms your ordinary function into a state machine.
Could you elaborate on what you consider the interface difference to be?
A thread per connection isn't wrong, though. You're begging the question by assuming that green threads are faster than native threads. I'm specifically arguing against that.
The next point might be if it's harder to get the same performance characteristics with native threads on different platforms.
As I view them, green threads are a useful abstraction just like objects or functions. Do you need something computed, not necessarily now, but at some point? Just make a green thread to compute it and check on the result later. That's the platonic ideal of green threads, anyway: very few languages have gotten to that point (Erlang, maybe Go, on a good day; C# has this with its async functionality but none of it happens in parallel).
But you can't. Green threads have scalability limitations too, and their performance is not predictable (except possibly by the creators of the runtime) in any language or on any processor that does automatic preemption. That's like saying you use i64 instead of i32 because now you can do calculations without worrying about overflow. It may be true for some classes of problem, but it's hardly a different programming paradigm like you're making it out to be. They're definitely not a different abstraction from threads (which are themselves an abstraction!).
while (!ready)
assert(!syscall(SYS_futex, &ready, FUTEX_WAIT, 0, NULL, NULL, 0));
For completeness, I also tested a similar Go program on Go 1.4 (the version available from Debian on the VM); it actually got up to 3,150,000 before OOM, and took 9 seconds to do 2 million - 4.5 microseconds and 2.7KB per thread.
In other words, Linux is about an order of magnitude slower at managing a lrge number of threads. That looks pretty bad, but on the other hand, it's not that much in absolute terms! I'm pretty sure most server programs don't need more than 250,000 simultaneous connections (or can afford to spend more than 8GB of RAM on them) and don't mind spending an extra 20 microseconds to initiate a connection, so if operating systems other than Linux aren't a concern, they could be written to create a thread per connection without too much trouble. It's not going to give you the absolute maximum performance (meaning it's not appropriate for a decent class of program - then again, I suspect Go isn't either), but it's not terrible either.
I'd like to see it improve. I wouldn't be surprised if there is (still) some low hanging fruit; do kernel developers actually care about this use case?
(And yes, I know this doesn't really test performance of the scheduler during sustained operation. That's its own can of worms.)
Yeah, this matches our results when we did similar tests. It's definitely faster to use green threads if you're just spawning and shutting down the threads, but if you're actually doing I/O work on those threads, the overhead quickly drops down.
It's not the fastest way to do I/O, but Go's approach isn't either. The fastest way to do I/O is to forego the thread management syscalls and the stack, like nginx does.
Though Go's is faster.
>forego the thread management syscalls and the stack, like nginx does.
The problem is nginx uses state machines, which aren't a general solution. Stackless coroutines allow you to write functions that the compiler (or a clever library) transforms into state machines.
Okay, so the test you did doesn't actually reflect the use case in practice. Can I expect to reach 200,000 threads if the threads are not all created at exactly the same moment? What if (God forbid) they're doing memory allocation? And if it does work out, will everything be handled efficiently?
...in fact, neither did removing the setstacksize call and having default 8MB stacks. I guess this makes sense: of course the extra VM space reserved for the stacks doesn't require actual RAM to back it; there is some page table overhead, but I guess it's not enough to make a significant difference at this number of allocations. Of course, on 32-bit architectures this would quickly exhaust the address space.
If increasing max_map_count hadn't worked, it would still be possible to allocate stacks on the fly - but you would get a bunch of them in one mmap() call, and therefore in one VM mapping, and dole them out in userland. However, in this case guard pages wouldn't separate different threads' stacks, you would have to generate code that manually checks the stack pointer to avoid security issues from stack overflows, rather than relying on crashing by hitting the guard page. Rust actually already does this, mostly unnecessarily; I'm not sure what Go is doing these days but I think it does too. Anyway, given that the above result I suspect this won't be an actual issue, at least until the number of threads goes up by an order of magnitude or something.
Coroutines have three awesome advantages here:
Really fast switches (it's basically a function call — no need to trap into the kernel)
Scheduling in response to IO events (a coroutine can yield when making a "blocking" IO call and resume when it completes). You can write asynchronous code that looks synchronous!
Better cache behavior (hugely important on modern processors). You can allocate all your coroutine structures together and don't have to suffer all the cache evictions of a context switch every time a coroutine yields.
It would be interesting to see this benchmark repeated with a latency requirement.
You can also write "actual" synchronous code which does the same thing in the kernel. :P And there are several CPU schedulers to choose from to fine-tune when things get woken up.
I agree it would be interesting to test latency; actually, the numbers would be much more useful than my four seconds. After all, if thread creation throughput or latency becomes an issue, you can keep a thread pool around, while still retaining the advantages of one connection per thread.
You can. But no matter how good your kernel scheduler is, that trip through it is going to cost you. Why? Because it's going to wreck your cache and TLB entries. I think the TLB is a big part of the story, given it's small, specific to the process, and particularly expensive to miss.
http://www.cs.cmu.edu/~chensm/Big_Data_reading_group/papers/...
Anyway, I'm sure you know this already. :) Out of curiosity, did you get a chance to take CS169 before you left Brown?
And I think the kernel already does pool/recycle stacks, but I guess not 200,000 of them given your results. An explicit thread pool would certainly work too.
But yeah, certainly userland context switches are faster; it's just that your comment seemed to postulate "scheduling in response to IO events" as a benefit of coroutines over threads, as if the kernel had no way to distinguish between threads waiting for IO and threads needing to be woken up. I presumably misinterpreted it.
Interesting paper. It seems like it could be a useful step toward my ideal imagined environment where the distinction between coroutines and threads would be meaningless - because scheduling would be a job shared between userland and the kernel, so userland could do fast thread switches, batch system calls, etc., while the kernel would still give them PIDs and let them take signals or be ptraced, and native tools (debuggers) would treat them as threads.
Though the paper is from 2010; do you know if anyone has tried to implement anything along similar lines in production?
And no, I didn't take that course.
From what I can tell, M:N threading has been thoroughly abandoned by Linux, but it might be worth revisiting given the horrendous I/O scaling of Linux kernel threads. I would imagine scheduling user threads cooperatively simplifies things a lot.
Google has a very interesting talk about cooperative native thread scheduling, but they haven't upstreamed their code:
https://www.youtube.com/watch?v=KXuZi9aeGTw
And here's what linux devs are actually doing (it's underwhelming, something like 50% of what the hardware is capable of):
https://lwn.net/Articles/629155/
In the meantime, people who actually need the performance (HPC and even some enterprise servers) have to bypass the kernel entirely. It looks like that's going to be possible even in virtualized environments, thanks to hardware support:
http://people.inf.ethz.ch/troscoe/pubs/peter-arrakis-osdi14....
I think that mentality, which I supported, may be what ruined green threads for Rust. Perhaps the enemy was the perfect of the good. On the other hand, maybe not having to worry about differences is the better benefit. I would rather be slower-but-acceptably-fast than have to think about 1:1 vs M:N threading in the case that they are not interchangeable.
On the third hand, it could be that Rust is too close to the metal for green threads to be of any significant benefit. If that's the case, Rust ought to trounce all the M:N implementations, subject to reasonable limits on spawn rate, memory pressure and inter-thread communication. Does that sound right?
Incidentally, I've never seen a cooperative threading implementation that didn't have a yield call of some kind available, and you can usually jam any of them up by writing an infinite computation-only loop, so even in the langauges where the scheduler is automatically invoked, correct 1:1 code can deadlock in the M:N scenario. State machines can too, but a trivial example, translated, might read something like the following:
for(;;);
exit(EXIT_SUCCESS);
It's not meant to be a performance boost, but a way of writing code where "might this crash?" is a question you're never bothered to ask. It's wonderfully freeing.
Not true of native threads. If a goroutine were replaced with a native thread, it would swiftly crash on GC.
Is that actually true? Yes, you won't get a partially written int, but you could still run into consistency issues if you don't lock. For example, if you need thread A to update both x and y to have a consistent state, and B is reading both x and y, A might update x, then the scheduler would switch to B, and be might read the new x and the old y. If the update A was making needed to update both x and y for the system to be in a consistent state, you have a problem.
If you use the int as an offset into raw memory, then sure, that will. But if you use it as an offset into a list, then it won't. At most it will cause an exception to be generated. That will cause the reader thread to die, because it was the code that caused the exception, but it still won't crash.
So I don't think it frees you from thinking about concurrency at all.
Basically if you have green threads with only one core, you don't need to do as much work with locking.
This is half the basis for the original design of the Erlang BEAM VM. SMP was a grudging afterthought; the original concept was that each VM "node" was a single scheduler process bound to a single core, and then you tied those nodes together into a (machine-internal) multi-core distributed system, effectively creating a set of "threads" communicating over socket-based message-passing IPC, rather than shared memory.
M:N scheduling—where, critically, a task can be re-scheduled from one core to another—is a much more complex and less-predictable design, that's not always worth the trouble. For Erlang's particular use-cases, it turned out to be worth the trouble enough to justify programming it, eventually. But the SMP VM is still only built via a configure flag, rather than displacing the M:1 VM as "the" VM.
Plus, if you write threaded code, you can change the runtime out. Rust guarantees all the hard parts, anyhow. I wouldn't even be surprised once Rust settles down further it turns out there's some sort of hybrid solution that is better than either 1:1 or M:N on its own because it has superior insight into what its functions are doing.
However, if you are sitting down in front of an editor, faced with the task of writing a network server, and you immediately reach for the overcomplicated solutions like this rather than starting with threads, you are paying an awfully stiff development price for a performance improvement that probably won't even manifest as any useful effect, because the window where this will actually save you is rather small. If you're sitting at 90% utilization, and this sort of tweak can get you to 80% utilization, that's essentially a no-op in the network server space, because it means you still better deploy a second server either way. If you're close to a capacity problem (and the code has been decently optimized to the point where that's not an option anymore), the solution is not to rewrite your code to squeeze out the 10% inefficiency in stack handling, the solution is to deploy another server, and if a rewrite is needed, rewrite to make that possible/effective/practical.
In the Servo design space, it makes perfect sense to be upset that you lost 10% of your performance to green threads, because that's time a real human user is directly waiting, and what does a web browser need with 10,000 threads anyhow? (I mean, sure, set one the deliberate task of using that many threads and they can, but in practice you'd rather be doing real work than any sort of scheduling of that mess.) In the network space it's way less clear... if your infrastructure is vulnerable to 10% variances in performance, your infrastructure is vulnerable to 10% variances in incoming request rate, too.
(10% is a bit of a made up number. I believe, if anything, it is an overestimate. Point holds even if it's a larger number anyhow.)
If the various threads in your application don't share data, then you don't need to worry about all the pitfalls of threading. But if they don't share data and you don't care about small (~10%) performance differences, why not use processes? Then you can use a really straightforward blocking I/O model, and you get full memory isolation & security provided by the OS.
The interesting questions happen when you a.) want to squeeze as much performance out of the machine as possible or b.) need to share data between concurrent activities. Then all of the different programming models have pros and cons, and I'm not sure you can define a "best" approach without knowing your particular problem. (Which, IMHO, validates Rust's "provide the barest primitives you can for the problem, and let libraries provide the abstractions until it becomes clear that one library is a clear winner" approach.)
FWIW, my experience in distributed systems is that threads+locks is a terrible model for writing robust systems, and that once you're operating at scale, you really want some sort of dependency-graph dataflow system where you specify what inputs are required for each bit of computation and then the system walks the graph as RPCs come back and input becomes available. This lets you attach all sorts of other information to nodes - timeouts, latency statistics, error statistics, whether or not this node is required and what defaults to substitute if it fails, tracing, logging, load-balancing, etc. It also adds a huge amount of cognitive overhead for someone who just wants to make a couple database queries. I wouldn't use this for prototyping a webapp, but it's also invaluable when you have a production system and ops people who need to be able to shut down a misbehaving service at a moment's notice while still keeping your overall product up.
But it wasn't necessarily the threads, so much as the fine-grained locks.
I'm talking about threads-the-programming-model, where we don't throw away the gains we made with structured programming. I suppose I ought to get off my duff and finish the blog post that explains exactly what that means, but if you know what structured programming is, that's actually enough to explain what I mean. (It's just that nowadays most people don't actually know what it is, not because they have no familiarity with it, but because they have too much, and don't realize that it was not actually the "ground state" of programming but actually a hard-fought advance, which we are all collectively on the verge of spending the next, oh, 5-10 years painfully rediscovering the value of.)
If parallelism is off the table I would be surprised by any disagreement.I doubt any one finds control inverted style of programming more pleasurable (I.e. Callbacks) than code that looks synchronous but executed async. One thing that does require watching out for is that there are no blocking calls. Green threads, coroutines, fibres seem so much nicer as a programming abstraction. Is this view even contentious?
This does not address parallelism though. These are about exchanging control between stacks of execution, only one such stack is active in an unit. There could be multiple such units though and rub lies in should they synchronize by messages (less buggy) or shared mutable state (at times faster if serialization is expensive. Think Python).
A model that I like is different units of execution ( threads) in a common address space, each hosting coroutines or their ilk with the units sync'ing exclusively by message passing but without the need of serialization because they live in a common address space.
https://github.com/google/guice
https://github.com/square/dagger
Imagine injecting Futures into your code. Now imagine injecting them, but having the Provider record a lot of metadata about who you were calling, how you were calling it, how long it took, whether an error occurred, letting an SRE turn off the call, etc.
Oh, from an ergonomic point of view I'm in total agreement that imperative-looking code is the best. I've never liked explicit node.js-like continuation passing style or what have you.
https://github.com/chriskohlhoff/asio/blob/master/asio/inclu...
* There is little practical difference between regular threads and green threads in an OS kernel. Most OSes can't handle it if you manually mess with the stack in a kernel context anyway. (thread records are typically accessed by rounding the current stack pointer) So if you want to do synchronous-style I/O with real stacks, you'll need an OS thread for each task.
* Kernel thread stacks are fairly small; 8-16KiB are typical. The reason for this is of course that kernel stacks must use wired (non-pageable) memory, or interrupts will irrecoverably page fault. 8-16KiB is much too small for buffers of course, but also enormous compared to the actual amount of non-buffer state for most I/O tasks. In any case, you never ever want to get too close to utilising the theoretical maximum as you risk crashing the system. So for every kernel thread you create, you know you're wasting precious wired memory. Obviously, this is true for threads created from userspace too, but in the kernel, people usually don't have a choice about running your code, so you try to be as good a citizen as possible, and not fire off hundreds of threads.
* In many contexts, dynamic memory allocations are not reliable, and allocation failure must not impede progress. (I.e. I can't wait for the system to page some memory out to disk if my code is on the critical path for disk I/O.) So typically, it is desirable to pre-allocate enough memory to keep all the state required for a sequence of operations from start to finish. It's even less likely you can just spin up a new thread; so the threaded approach implies keeping a pool of threads around which is guaranteed to be big enough. A.k.a. a waste of resources.
* The chain-of-callbacks approach to I/O is even more awful in languages and environments with manually managed memory and resources.
* If I'm going to pick a fancy new language to write my drivers in, I'm still going to have to use the custom alloc/free functions for each type of kernel object (network packet, etc.) that the OS I'm writing against happens to use.
So typically, you end up either splitting your code into a bunch of callbacks, or you create a complicated explicit state machine with a giant dispatch switch() statement. In both cases all state tracked in a giant struct. Keeping track of control flow is tricky. Maintaining invariants is tricky. Making sure you don't leak (or over-free, or use-after-free) any of the resources you touch is tricky.
It'd be really, really nice if the language could help you out with this. Write it as synchronous code, and the compiler turns each location where execution can be suspended into a callback function. The locals that are used across suspension points are stored in an automatically generated struct (bonus points: unions for state which is guaranteed to not have overlapping lifetimes) which can be preallocated before firing off the "task" in question. As far as I'm aware, such a code transformation would effectively be a CPS-transform, (continuation passing style) which has at least been researched quite a bit in theory, if not so much in non-GC-language practice.
I haven't been able to invest large amounts of time into really learning Rust and applying it to my use case in earnest. It's tough without a remotely compliant standard C library around, and I've struggled a bit with trying to pick and choose bits out of Rust's 'core' library without bringing on an avalanche of dependencies. (any kernel module code is kept in wired memory, so unused code is a waste of resources in the kernel) I'm determined to overcome that though and use Rust for something other than a toy kernel module, and see if and how it improves things over C. If Rust does start supporting some kind of advanced I/O pattern that works in that sort of constrained environment, that seems like a significant competitive advantage.
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n445...
http://blogs.msdn.com/b/vcblog/archive/2014/11/12/resumable-...
The first's library implementation is based on boost.coroutine which as far as I'm aware uses stack-switching, so this is more of a syntax demo than a demo of the final implementation details. The fact that you can apply sizeof() to the resumable type is important for the kernel-typical pool allocators.
The second seems to have preview support in VS2015's compiler. Too bad I'm not much of a Windows programmer these days… It's not quite clear what ther memory-allocation story is from that article.
Of course the downside is it's yet another graft onto the C++ language. After ~15 years of using it, my point of view is mainly that C++'s type system and unsafe-by-default behaviour inherited from C are a mistake…
I am also from a low-level C background and I think this is the approach I would like to see in Rust.
It can be implemented in a way that would be safe to run in kernel space and overall provides a pretty nice programming model. It can also be very performant if it's integrated into the compiler.
So, if threads are the way to go today, what has changed in the last 10 years to turn the conventional wisdom on it's head? 64-bit processors and servers with more memory? Hypervisors/containers? Are threads in Rust more practical than they are in C?
(BTW, these aren't rhetorical questions, genuinely interested.)
So, AFAIU the C10K approach is still the correct approach. Although perhaps with current hw and Linux being able to decently handle quite a lot of threads since NPTL, one should nowadays call it the C100K problem?
What has changed is a change in focus. Few people do entirely static sites anymore, and for static assets (images etc.) everybody uses nginx or other nonblocking approaches anyway. Perhaps also nginx/etc. as a reverse proxy. So there is no argument that nonblocking approaches are better for handling a lot of potentially slow connections.
But for the "core" functionality of a network application, that people are actually spending times programming rather than using an off the shelf solution like nginx, nonblocking vs. threads doesn't matter that much, there's all kinds of DB calls, CPU intensive work to do, etc. So people are (correctly) asking whether we can create nonblocking code but with an easier to use programming model (to the extent it does matter for performance) or whether to just use threads.
And: I have to counterbalance the assertion that Node.js is in any way good for anything besides "I need to code in JS, but not in the browser". I'd rather use VAX assembler and the $QIO syscall, typing on a VT100.
But let's talk about what's wrong with threads:
1. Timeout handling is ugly: you need to account a timeout in each read and write operation. At least timeout handling makes coroutine/threaded code no better than state machine code. But actually in state machine approach I can set a deadline once and update it only when changed (note to myself: should add such an example to documentation). In Python, it's usually fixed by making another coroutine with sleep and throw an exception to the one doing I/O. It works well, but Rust will never get exceptions (I hope)
2. When you make a server that receive a request, looks in DB then responds, there is an incentive to own (create or acquire from the pool) a DB connection by each thread. This is a sad trick. The better thing when the DB connection is handled by a coroutine on its own. Because in the latter case you may pipeline multiple requests over a connection, monitor if the connection is still alive, reconnect to DB while no requests are active, or the contrary, shut down idle connections. By pipelining you may keep less number of connections to DB so make the load to the database a little bit lower. When I'm talking about DB in this paragraph, of course, I mean everything for which this application is a client. Sure you can do that in threaded code too, but it's much harder to get right. You need two threads per connection (because one reads network and the other looks at the queue and does write), you need to synchronize both sides, connection cleanup code is complex, there is more than one level of timeouts now, so on.
3. You need to avoid deadlocks. Rust takes care to avoid data races, but deadlocks are possible. And they are not always simple or reproducible, so you will have a hard time debugging them. In the single-threaded async code, you are the only user. Even if you have an async thread per processor, you are more likely to own resources instead of locking on them. You may duplicate many things for every thread. You can have more coarse-grained locks, so never hold two of them. But it's almost impossible to write lock-free threaded server.
All of the issues above are neither fixed with async/await nor with any M:N or 1:1 threading approaches.
1: In either approach, somewhere in your event loop you're setting yourself a timeout to fire. Haskell & Erlang do use exceptions, but Go does not, it simply makes this a first-class concern of the core event loop. This is only a problem in languages where the threading was bolted on after-the-fact. Which is a lot of languages, which matter because they have a lot of code. I don't mean to dismiss those real problems. But it's not a fundamental problem, only accidental.
2. In practice, this is not a problem I ever worry about. You get a DB library, it provides pools, unless you're talking to a very, very fast DB (like, memcached on localhost fast) this is one of those cases where IO really does dominate any minor price of thread scheduling.
3. This has been solved for a long time. Go has the nicest little catch phrase with "share memory by communicating instead of communicating by sharing memory", but each of Haskell, Erlang, and Go have their own quite distinct solutions to their problems, and in practice, all of them work. There's other solutions I merely haven't used, but I hear Clojure works, too. (Perhaps arguably a subset of the several approaches Haskell can use. Haskell kind of supports darned near everything, and you can use it all at once.)
This is part of why I write this sort of thing... at its usual glacial pace (despite how much we like to flatter ourselves that we move quickly), the programming community is finally getting around to being really seriously pissed off about how bad threading was in the 1990s. Good. We should be. It sucked. Let us never forget that. But what has not been so well noticed is that the problems with threading have basically been fixed, and in production for a long time now (i.e., not just in theory, but shipping systems; go ask Erlang how long it's been around). You just have to go use the solutions. Don't mistake debates about the minutia of 1:1 OS threading vs. M:N threading and which is single-digit percent points faster than the other for thinking that threading doesn't work.
Lest I sound too pollyannaish about what is still a hard domain, the way I like to put this is that threading has moved from an exponentially complex problem to a polynomially complex problem. (And Rust is leading the way on making even the polynomial have a small number in the exponent.) There's still a certain amount of complexity in making a threaded program go zoom, and it does require some adjustments to how you program, it's not "free", but rather than requiring wizards, it merely requires competent programmers who take a bit of care and use good tools and best practices now.
Thread 1 sends message A to thread 2 and waits for a response.
As part of processing message A, thread 2 sends message B to thread 1 and waits for a response... forever, since thread 1 is blocked waiting for thread 2...
And to be fair, it does significantly reduce the risk of deadlock by avoiding the complexity of explicit mutexes and condition variables.
Certainly a bit more contrived then deadlock with locks, though (which you can easily do with even a single thread, even in Rust).
Aborting once the allocator has actually been invoked is important for perf and exception safety, though we have briefly mused allowing it to panic: https://github.com/rust-lang/rust/issues/26951
> you're free to provide an alternative stdlib
But like... should you have to?
> for those rare occasions
In kernel programming, allocation failures are common and handling them is essential. To quote my (very talented) classmate, who actually wrote part of a kernel in Rust:
"The only really major issue is with how allocation failure works, which makes rust a (very) poor choice for real kernel development"
https://www.reddit.com/r/rust/comments/341v3n/cs_honors_thes...
This is the status quo (in all languages) if you're doing embedded work: you lose most of the standard library. Rust goes slightly beyond that default by bundling `core`, which is designed for these environments and serves as a building block on which embedded/bare-metal development can flourish. (Cargo works fine with such things, so one can even distribute alternate standard libraries on crates.io.)
> "The only really major issue is with how allocation failure works, which makes rust a (very) poor choice for real kernel development"
I think this was the result of bravely trying very hard to shoe-horn `std` and the inflexible `box` syntax into a kernel environment, somewhere both are 100% not designed for. The recommended approach is to just use `core` and layer non-`std` libs on top of that.
2. The libcore allocator API is marked unstable, so even if you go through the trouble of implementing it, how long will it last?
That's what libcore is. A lot of the stdlib is reexports over libcore.
new and delete are C++ isms, their Rust counterparts are `Box::new` (and destructors are automatic). `Box::new`, like `new`, has the OOM issue. If you're okay with that, you are free to link to the `alloc` crate, which gives you `Box` without pulling in additional deps.
There are plans for in-place boxing, and AFAICT that API will be available to embedded users.
Already exists, unstably: http://doc.rust-lang.org/nightly/core/ops/trait.Placer.html
I see Box::from_raw and core::ops::Placer. Is that what you mean by in-place boxing? But now there are two problems:
You have to manually allocate a buffer of the right size, check if the pointer is null, and box it. For every allocation. Forget a null check? Allocate too small a buffer? Guess what, you've got undefined behavior!
It also doesn't generalize to other parts of the standard library. Did you want to use Rust's built in containers in your kernel? Well, sorry, you're going to have to write your own ones that don't panic on allocation failure.
Yeah, I mean that it's used like `new` is. We don't have the notion of constructors.
> You have to manually allocate a buffer of the right size, check if the pointer is null, and box it. For every allocation.
No. With the placement API, you'll be able to use the `box` syntax (which at the moment just supports `Box`, which has OOM issues) for your custom wrapper (i.e., `Box` with OOM support, or whatever).
What `box` gives us is that `let x = box make_inner()` would have the returned value of `make_inner()` written directly to the heap location. (Unlike `let x = Box::new(make_inner())`, which may be a move of `make_inner()`)
Placer is a part of this (https://github.com/rust-lang/rfcs/blob/master/text/0809-box-... is the whole design). from_raw isn't, that's just a useful API for FFI.
> Did you want to use Rust's built in containers in your kernel? Well, sorry, you're going to have to write your own ones that don't panic on allocation failure.
When you're writing a kernel you're not supposed to be using the standard library at all, just libcore (which exists for this purpose). The same issue exists in C++.
>When you're writing a kernel you're not supposed to be using the standard library at all, just libcore (which exists for this purpose). The same issue exists in C++.
Not quite. C++ standard containers are polymorphic on the allocator. That means you're welcome to use std::list or std::string in kernel mode. Just plug in a kernel allocator and you're good to go.
There's no technical reason why every Rust core project must reinvent the linked list and the hash map, like every C project.
It doesn't literally offer working dynamic allocations, but I don't see how a cross-platform and cross-use-case freestanding library can do that, too many different environments/constraints to encode an built-in allocator. In fact, core says nothing about how allocation has to work or even if it needs to exist.
The rustc distribution has the additional `alloc` and `collections` crates layered on top of `core`, which offers some extra functionality when allocation does exist (without assuming other OS support, like IO), and the allocator used is pluggable.
(Do you have an example of a standards complaint C++ freestanding standard library?)
2. There will be some way to do this long term. I can offer no guarantees that it will be what's there right now, but something will exist. This is also something that has been on the radar for a long time, but isn't core to the language. In any case, what is in core is essentially just optimising stack usage, all of the sematics it offers can be done with normal function calls, e.g. `Box::new` is literally just a normal function and perfectly implementable in any environment, if you have some allocation routine (the compiler/language doesn't need to know about either). So the unstable libcore functionality is important and desirable, but not killer.
There's also the meaning of "allocators" as in allowing `Vec` (etc.) to use some other algorithm than what Rust does by default; again something desirable and been on the radar for a while, but not key to the language. I don't think this needs any new language features, just people to explore the library space. (NB. this is different to what libcore offers now, which is basically what is called "placement new" in C++.)
Yes, you've probably heard of libstdc++:
https://gcc.gnu.org/onlinedocs/libstdc++/faq.html#faq.what_i...
In other words, freestanding C++ allows you to to use new and delete normally (after providing malloc and free). Note that this a minimum requirement. There's nothing stopping you from using STL containers in embedded systems, and people do (especially now that STL containers can take user defined allocators).
It seems like Rust is trying to achieve the same result with libcore. That's good! Thanks for clarifying.
But now you lose all the nice parts of the standard library, and probably much of the safety. If you want to use Box, you have to manually do the allocation, check the pointer, and construct the box. Buffer too small? Forget the pointer check? Welcome to undefined behavior.
If you want a container, you have to roll your own. If you want a container that takes user-defined allocators, I'm not even sure what'd you do.
> probably much of the safety
Only if you want to make life hard for yourself. The standard library is not at all special in its ability to create safe abstractions for `unsafe` code, and doing this makes things so much smoother: you don't have to scatter `unsafe` all over your code, and you let the compiler help you as much as possible. Also, I think there's a lot of nice stuff in `core` too; even string formatting works.
You can still implement equally safe version of them tuned for your environment or, more likely (as the ecosystem develops), use a crate that someone else has written that does it for you. Here's the start of an implementation of a version of `Box` that returns `Result` instead of crashing on allocation failure:
extern crate core;
use core::{mem, ops, ptr};
extern {
fn malloc(size: usize) -> *const ();
fn free(p: *const ());
}
pub struct MyBox<T> {
p: *const T,
}
impl<T> MyBox<T> {
fn new(x: T) -> Result<MyBox<T>, ()> {
unsafe {
let p = malloc(mem::size_of::<T>()) as *const T;
if p.is_null() {
Err(())
} else {
ptr::write(p as *mut T, x);
Ok(MyBox { p: p })
}
}
}
}
impl<T> Drop for MyBox<T> {
fn drop(&mut self) {
unsafe {
drop(ptr::read(self.p));
free(self.p as *const _);
}
}
}
// allow `*` to work
impl<T> ops::Deref for MyBox<T> {
type Target = T;
fn deref(&self) -> &T { unsafe { &*self.p } }
}
The whole of Rust's standard library (including `Box`, all the containers, all of IO) is built via this process: wrap the raw `unsafe` functionality into safe interfaces that manage things like checking `malloc`'s return value, bounds checking buffer accesses.
> If you want a container that takes user-defined allocators, I'm not even sure what'd you do.
Do exactly what C++ does: have some interface (a trait, in Rust) that an allocator needs to satisfy and have the collections take a type parameter that implements that interface. We "already" know what this looks like in Rust (broadly speaking):
struct Box<T, A: Allocator = DefaultAllocator> {
p: *const T,
allocator: A,
}
An out-of-tree version aimed at a single use-case (e.g. embedded development) is likely even easier, since the problem space is smaller.
But doesn't that seem kind of redundant? Box and std containers are great as is, except for the one line that aborts on OOM. Would it be possible to use traits to choose OOM behavior at compile time?
To be clear, this isn't just important for embedded. A production database or web server should be able to handle an allocation failure without blowing up the whole process.
And I'm glad there's some talk about a standard allocator trait. I'll look for the relevant issue.
https://github.com/rust-lang/rfcs/issues/538 covers the allocator stuff.
But could Box and containers be polymorphic on an OOM handling trait?
Alternatively, one could imagine an OOM-safe container that returns results, and a convenience wrapper class that unwraps them. That might harm efficient code generation though.
> OOM is an abort, not a panic (contrary to the official docs, interestingly)
1. Distinguish between abort and panic.
2. Explain that allocation failure is an abort.
3. Explain that a vec can allocate up to twice its nominal size, which is a common cause of OOMs. To their credit, the docs do explain that vec over-allocates, but not that it allocates double the memory. They could also explain that you don't get that memory back unless you ask for it, even if you pop() or remove() elements.
Scenario:
I'm a new systems programmer excited to try Rust (fast and safe? wow!). I have 8GB of memory in my computer. I push ~4GB of data into a vector so I can process it, something I'm used to doing in Python.
POW! My program exits with no error message. What happened? I look at the error handling docs:
https://doc.rust-lang.org/book/error-handling.html
Well, it sure wasn't a recoverable failure! Must have been a panic. So I look at the docs for vec.push:
https://doc.rust-lang.org/std/vec/struct.Vec.html
"Panics if the number of elements in the vector overflows a usize."
A usize is big, right? I didn't overflow that. I guess I could have run out of memory, but I have 8GB. That should be plenty!
Agreed! http://pyparallel.org
> But this is a kernel interface issue not a language issue,
Absolutely agree! The thing that surprises me most about Rust and Go in particular is that they're trying to solve an OS problem with new language abstractions. You don't need a new language -- what you really want is thread-agnostic I/O primitives, thread-agnostic "threadpool" primitives, and completion-oriented APIs that provide both synchronous and asynchronous primitives. Basically, you want what Windows has.
Those are all entirely reasonable things to block for if you understand the mechanics behind why those calls may block. And none of it should be fast-path stuff; it's easy to architect a solution where you avoid those things except in corner cases.
The other thing you're missing is that when you architect around Windows completion ports and threadpool I/O facilities, it doesn't matter if one thread blocks every so often. Windows will detect this and schedule another one to run, such that there is one active thread scheduled on every core.
I/O completion ports facilitate thread-agnostic I/O completion; the thread that blocked because it was extending an encrypted file won't impede the latency of other clients because there is no thread-specific association.
I exploit all of this with PyParallel. http://pyparallel.org
For more details on asynchronous I/O on Windows and why it's fundamentally better than the UNIX I/O model in every possible way: https://speakerdeck.com/trent/pyparallel-how-we-removed-the-...
AFAIK, the kernel APIs only deal with abstract handles and descriptors - they aren't written specifically for disk IO or network IO. And C#'s async/await and BeginXXX also work fine with all kinds of IO.
(On Windows, as I understand it, you can do completion-based I/O on all handle types.)
And then, even if you manage to use async disk I/O, you can only really use it for reads and writes. Filesystem calls that manipulate the directory tree usually don't have async versions. At some point you have to give up and do things in threads.