X-Git-Url: https://git.ralfj.de/rust-101.git/blobdiff_plain/0223210576f27d0743c2d12b890d30f5c2ef6b2d..126f7bfe68883e6465d42c529b00d797b2f8478a:/src/part15.rs diff --git a/src/part15.rs b/src/part15.rs index 7365421..99eb3be 100644 --- a/src/part15.rs +++ b/src/part15.rs @@ -1,25 +1,147 @@ -// Rust-101, Part 15: Interior Mutability (cont.), RefCell, Cell, Drop +// Rust-101, Part 15: Mutex, Interior Mutability (cont.), RwLock, Sync // =================================================================== -//@ [`RefCell`](http://doc.rust-lang.org/beta/std/cell/struct.RefCell.html) -//@ [`is very much like `RwLock`, but it's not thread-safe: "Locking" is done without atomic operations. -//@ One can also see it as a dynamically checked version of Rust's usual borrowing rules. You have to explicitly say -//@ when you want to borrow the data in there shared, or mutably, and Rust will complain at run-time if you have -//@ a mutable borrow while any other borrow is active. You can then write programs that Rust may otherwise not -//@ accept. Sending a shared borrow to this to another thread is dangerous, as the checks are not performed in -//@ a thread-safe manner. However, sending the *entire* `RefCell` is okay, because there's only ever one owner, and all -//@ we need to ensure is that everybody attempting to borrow is in the same thread as the owner.
-//@ [`Cell`](http://doc.rust-lang.org/beta/std/cell/struct.Cell.html) is like a stripped-down version of `RefCell`: It doesn't allow -//@ you to borrow its content. Instead, it has a methods `get` and `set` to change the value stored in the cell, and to copy it out. -//@ For obvious reasons, this requires `T` to be `Copy`. -//@ -//@ You can also think about all these types coming from the other end: Starting with `Cell`, we have a primitive for -//@ interior mutability that provides `get` and `set`, both just requiring a shared borrow. Think of these functions as -//@ mutating the *content* of the cell, but not the cell itself, the container. (Just like in ML, where assignment to a -//@ `ref` changes the content, not the location.) However, due to the ownership discipline, `Cell` only works for types -//@ that are `Copy`. Hence we also have `RefCell`, which allows working with the data right in the cell, rather than -//@ having to copy it out. `RefCell` uses non-atomic operations for this purpose, so for the multi-threaded setting, there's -//@ the thread-safe `RwLock`. And finally, in case a distinction between readers and writers is not helpful, one can use the -//@ more efficient `Mutex`. - -//@ [index](main.html) | [previous](part14.html) | [next](main.html) +use std::sync::{Arc, Mutex}; +use std::thread; + +//@ We already saw that we can use `Arc` to share memory between threads. However, `Arc` can only provide *read-only* +//@ access to memory: Since there is aliasing, Rust cannot, in general, permit mutation. To implement shared-memory +//@ concurrency, we need to have aliasing and permutation - following, of course, some strict rules to make sure +//@ there are no data races. In Rust, shared-memory concurrency is obtained through *interior mutability*, +//@ which we already discussed in a single-threaded context in part 12. +//@ +//@ ## `Mutex` +//@ The most basic type for interior mutability that supports concurrency is [`Mutex`](http://doc.rust-lang.org/stable/std/sync/struct.Mutex.html). +//@ This type implements *critical sections* (or *locks*), but in a data-driven way: One has to specify +//@ the type of the data that's protected by the mutex, and Rust ensures that the data is *only* accessed +//@ through the mutex. In other words, "lock data, not code" is actually enforced by the type system, which +//@ becomes possible because of the discipline of ownership and borrowing. +//@ +//@ As an example, let us write a concurrent counter. As usual in Rust, we first have to think about our data layout: +//@ That will be `Mutex`. Of course, we want multiple threads to have access to this `Mutex`, so we wrap it in an `Arc`. +//@ +//@ Rather than giving every field a name, a struct can also be defined by just giving a sequence of types (similar +//@ to how a variant of an `enum` is defined). This is called a *tuple struct*. It is often used when constructing +//@ a *newtype*, as we do here: `ConcurrentCounter` is essentially just a new name for `Arc>`. However, +//@ is is a locally declared types, so we can give it an inherent implementation and implement traits for it. Since the +//@ field is private, nobody outside this module can even know the type we are wrapping. + +// The derived `Clone` implementation will clone the `Arc`, so all clones will actually talk about the same counter. +#[derive(Clone)] +struct ConcurrentCounter(Arc>); + +impl ConcurrentCounter { + // The constructor just wraps the constructors of `Arc` and `Mutex`. + pub fn new(val: usize) -> Self { + ConcurrentCounter(Arc::new(Mutex::new(val))) /*@*/ + } + + // The core operation is, of course, `increment`. + pub fn increment(&self, by: usize) { + // `lock` on a mutex returns a guard, very much like `RefCell`. The guard gives access to the data contained in the mutex. + //@ (We will discuss the `unwrap` soon.) `.0` is how we access the first component of a tuple or a struct. + let mut counter = self.0.lock().unwrap(); + //@ The guard is a smart pointer to the content. + *counter = *counter + by; + //@ At the end of the function, `counter` is dropped and the mutex is available again. + //@ This can only happen when full ownership of the guard is given up. In particular, it is impossible for us + //@ to borrow some of its content, release the lock of the mutex, and subsequently access the protected data without holding + //@ the lock. Enforcing the locking discipline is expressible in the Rust type system, so we don't have to worry + //@ about data races *even though* we are mutating shared memory! + //@ + //@ One of the subtle aspects of locking is *poisoning*. If a thread panics while it holds a lock, it could leave the + //@ data-structure in a bad state. The lock is hence considered *poisoned*. Future attempts to `lock` it will fail. + //@ Above, we simply assert via `unwrap` that this will never happen. Alternatively, we could have a look at the poisoned + //@ state and attempt to recover from it. + } + + // The function `get` returns the current value of the counter. + pub fn get(&self) -> usize { + let counter = self.0.lock().unwrap(); /*@*/ + *counter /*@*/ + } +} + +// Now our counter is ready for action. +pub fn main() { + let counter = ConcurrentCounter::new(0); + + // We clone the counter for the first thread, which increments it by 2 every 15ms. + let counter1 = counter.clone(); + let handle1 = thread::spawn(move || { + for _ in 0..10 { + thread::sleep_ms(15); + counter1.increment(2); + } + }); + + // The second thread increments the counter by 3 every 20ms. + let counter2 = counter.clone(); + let handle2 = thread::spawn(move || { + for _ in 0..10 { + thread::sleep_ms(20); + counter2.increment(3); + } + }); + + // Now we watch the threads working on the counter. + for _ in 0..50 { + thread::sleep_ms(5); + println!("Current value: {}", counter.get()); + } + + // Finally, we wait for all the threads to finish to be sure we can catch the counter's final value. + handle1.join().unwrap(); + handle2.join().unwrap(); + println!("Final value: {}", counter.get()); +} + +// **Exercise 15.1**: Add an operation `compare_and_inc(&self, test: usize, by: usize)` that increments the counter by +// `by` *only if* the current value is `test`. +// +// **Exercise 15.2**: Rather than panicking in case the lock is poisoned, we can use `into_innter` on the error to recover +// the data inside the lock. Change the code above to do that. Try using `unwrap_or_else` for this job. + +//@ ## `RwLock` +//@ Besides `Mutex`, there's also [`RwLock`](http://doc.rust-lang.org/stable/std/sync/struct.RwLock.html), which +//@ provides two ways of locking: One that grants only read-only access, to any number of concurrent readers, and another one +//@ for exclusive write access. Notice that this is the same pattern we already saw with shared vs. mutable borrows. Hence +//@ another way of explaining `RwLock` is to say that it is like `RefCell`, but works even for concurrent access. Rather than +//@ panicking when the data is already borrowed, `RwLock` will of course block the current thread until the lock is available. +//@ In this view, `Mutex` is a stripped-down version of `RwLock` that does not distinguish readers and writers. + +// **Exercise 15.3**: Change the code above to use `RwLock`, such that multiple calls to `get` can be executed at the same time. + +//@ ## `Sync` +//@ Clearly, if we had used `RefCell` rather than `Mutex`, the code above could not work: `RefCell` is not prepared for +//@ multiple threads trying to access the data at the same time. How does Rust make sure that we don't accidentally use +//@ `RefCell` across multiple threads? +//@ +//@ In part 13, we talked about types that are marked `Send` and thus can be moved to another thread. However, we did *not* +//@ talk about the question whether a borrow is `Send`. For `&mut T`, the answer is: It is `Send` whenever `T` is send. +//@ `&mut` allows moving values back and forth, it is even possible to [`swap`](http://doc.rust-lang.org/stable/std/mem/fn.swap.html) +//@ the contents of two mutably borrowed values. So in terms of concurrency, sending a mutable borrow is very much like +//@ sending full ownership, in the sense that it can be used to move the object to another thread. +//@ +//@ But what about `&T`, a shared borrow? Without interior mutability, it would always be all-right to send such values. +//@ After all, no mutation can be performed, so there can be as many threads accessing the data as we like. In the +//@ presence of interior mutability though, the story gets more complicated. Rust introduces another marker trait for +//@ this purpose: `Sync`. A type `T` is `Sync` if and only if `&T` is `Send`. Just like `Send`, `Sync` has a default implementation +//@ and is thus automatically implemented for a data-structure *if* all its members implement it. +//@ +//@ Since `Arc` provides multiple threads with a shared borrow of its content, `Arc` is only `Send` if `T` is `Sync`. +//@ So if we had used `RefCell` above, which is *not* `Sync`, Rust would have caught that mistake. Notice however that +//@ `RefCell` *is* `Send`: If ownership of the entire cell is moved to another thread, it is still not possible for several +//@ threads to try to access the data at the same time. +//@ +//@ Almost all the types we saw so far are `Sync`, with the exception of `Rc`. Remember that a shared borrow is good enough +//@ for cloning, and we don't want other threads to clone our local `Rc`, so it must not be `Sync`. The rule of `Mutex` +//@ is to enforce synchronization, so it should not be entirely surprising that `Mutex` is `Send` *and* `Sync` provided that +//@ `T` is `Send`. +//@ +//@ You may be curious whether there is a type that's `Sync`, but not `Send`. There are indeed rather esoteric examples +//@ of such types, but that's not a topic I want to go into. In case you are curious, there's a +//@ [Rust RFC](https://github.com/rust-lang/rfcs/blob/master/text/0458-send-improvements.md), which contains a type `RcMut` that would be `Sync` and not `Send`. +//@ You may also be interested in [this blog post](https://huonw.github.io/blog/2015/02/some-notes-on-send-and-sync/) on the topic. + +//@ [index](main.html) | [previous](part14.html) | [next](part16.html)