-// Rust-101, Part 15: Mutex, Interior Mutability (cont.), Sync
-// ===========================================================
+// Rust-101, Part 15: Mutex, Interior Mutability (cont.), RwLock, Sync
+// ===================================================================
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. If however,
-//@ some care would be taken at run-time, then mutation would still be all right: We have to ensure that whenever
-//@ someone changes the data, nobody else is looking at it. In other words, we need a *critical section* or (as it
-//@ is called in Rust) a [`Mutex`](http://doc.rust-lang.org/stable/std/sync/struct.Mutex.html). Some other languages also call this a *lock*.
+use std::time::Duration;
+
+//@ 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.
//@
-//@ As an example, let us write a concurrent counter. As usual in Rust, we first have to think about our data layout.
-//@ In case of the mutex, this means we have to declare the type of the data that we want to be protected. In Rust,
-//@ a `Mutex` protects data, not code - and it is impossible to access the data in any other way. This is generally considered
-//@ good style, but other languages typically lack the ability to actually enforce this.
-//@ Of course, we want multiple threads to have access to this `Mutex`, so we wrap it in an `Arc`.
+//@ ## `Mutex`
+//@ The most basic type for interior mutability that supports concurrency is
+//@ [`Mutex<T>`](https://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.
//@
-//@ 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<Mutex<usize>>`. 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.
+//@ 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<usize>`. 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<Mutex<usize>>`. 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<Mutex<usize>>);
ConcurrentCounter(Arc::new(Mutex::new(val))) /*@*/
}
- //@ The core operation is, of course, `increment`. The type may be surprising at first: A shared borrow?
- //@ How can this be, since `increment` definitely modifies the counter? We already discussed above that `Mutex` is
- //@ a way to get around this restriction in Rust. This phenomenon of data that can be mutated through a shared
- //@ borrow is called *interior mutability*: We are changing the inner parts of the object, but seen from the outside,
- //@ this does not count as "mutation". This stands in contrast to *exterior mutability*, which is the kind of
- //@ mutability we saw so far, where one piece of data is replaced by something else of the same type. If you are familiar
- //@ with languages like ML, you can compare this to how something of type `ref` permits mutation, even though it is
- //@ itself a functional value (more precisely, a location) like all the others.
- //@
- //@ Interior mutability breaks the rules of Rust that I outlined earlier: There is aliasing (a shared borrow) and mutation.
- //@ The reason that this still works is careful programming of the primitives for interior mutability - in this case, that's
- //@ `Mutex`. It has to ensure with dynamic checks, at run-time, that things don't fall apart. In particular, it has to ensure
- //@ that the data covered by the mutex can only ever be accessed from inside a critical section. This is where Rust's type
- //@ system comes into play: With its discipline of ownership and borrowing, it can enforce such rules. Let's see how this goes.
+ // The core operation is, of course, `increment`.
pub fn increment(&self, by: usize) {
- // `lock` on a mutex returns a *guard*, giving 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.
+ // `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 another example of a smart pointer, and it can be used as if it were a pointer to the data protected
- //@ by the lock.
+ //@ 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!
+ //@ This can only happen when full ownership of the guard is given up. In particular, it is
+ //@ impossible for us to take a reference to 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.
+ //@ 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.
let counter1 = counter.clone();
let handle1 = thread::spawn(move || {
for _ in 0..10 {
- thread::sleep_ms(15);
+ thread::sleep(Duration::from_millis(15));
counter1.increment(2);
}
});
let counter2 = counter.clone();
let handle2 = thread::spawn(move || {
for _ in 0..10 {
- thread::sleep_ms(20);
+ thread::sleep(Duration::from_millis(20));
counter2.increment(3);
}
});
// Now we watch the threads working on the counter.
for _ in 0..50 {
- thread::sleep_ms(5);
+ thread::sleep(Duration::from_millis(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.
+ // 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 14.1**: 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.) Change
-// the code above to use `RwLock`, such that multiple calls to `get` can be executed at the same time.
+// **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 14.2**: Add an operation `compare_and_inc(&self, test: usize, by: usize)` that increments the counter by
-// `by` *only if* the current value is `test`.
-
-//@ ## Sync
-//@ In part 12, 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/beta/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.
+// **Exercise 15.2**: Rather than panicking in case the lock is poisoned, we can use `into_inner`
+// 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`](https://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 references. 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?
//@
-//@ 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 `&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.
+//@ 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 reference 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`](https://doc.rust-lang.org/stable/std/mem/fn.swap.html) the contents of two mutable
+//@ references. So in terms of concurrency, sending a mutable, unique reference is very much like
+//@ sending full ownership, in the sense that it can be used to move the object to another thread.
//@
-//@ 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<T>` is `Send` *and* `Sync` provided that
-//@ `T` is `Send`.
+//@ But what about `&T`, a shared reference? 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.
//@
-//@ In the next part, we will learn about a type called `RefCell` that is `Send`, but not `Sync`.
+//@ Since `Arc` provides multiple threads with a shared reference to its content, `Arc<T>` 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.
//@
-//@ 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.
-
-// FIXME TODO some old outdated explanation FIXME TODO
-
-//@ [`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. <br/>
-//@ [`Cell<T>`](http://doc.rust-lang.org/beta/std/cell/struct.Cell.html) is like a stripped-down version of `RefCell<T>`: 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`.
+//@ Almost all the types we saw so far are `Sync`, with the exception of `Rc`. Remember that a
+//@ shared reference is good enough for cloning, and we don't want other threads to clone our local
+//@ `Rc` (they would race for updating the reference count), so it must not be `Sync`. The rule of
+//@ `Mutex` is to enforce synchronization, so it should not be entirely surprising that `Mutex<T>`
+//@ is `Send` *and* `Sync` provided that `T` is `Send`.
//@
-//@ 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)
+//@ 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) | [raw source](workspace/src/part15.rs) |
+//@ [next](part16.html)