The Tale of a Bug in Arc: Synchronization and Data Races

While I was busy doing Rust-unrelated research, RustBelt continues to move and recently found another bug (after a missing impl !Sync that we found previously): It turns out that Arc::get_mut did not perform sufficient synchronization, leading to a data race.

Notice that I am just the messenger here, the bug was actually found by Hai and Jacques-Henri. Still, I’d like to use this opportunity to talk a bit about weak memory, synchronization and data races. This is just a primer, there are tons of resources on the web that go into more detail (for example here).

Synchronization and Data Races

Consider the following example (full of unsafe code because I am demonstrating behavior that Rust does its best to rule out):

extern crate rayon;

static mut DATA : Option<String> = None;
static mut FLAG : usize = 0;

fn main() { unsafe {
  rayon::join(
    || {
      // Initialize data
      DATA = Some("Hello World!".to_string());
      // Signal that we are done
      FLAG = 1;
    },
    || {
      // Wait until the other thread is done
      while FLAG == 0 {}
      // Print data
      println!("{}", DATA.as_ref().unwrap());
    }
  );
} }

The question is: What will this program print? If you are thinking “Hello World!”, you are not wrong, but you are not fully right either. “Hello World!” is one possibility, but this program can also panic at the unwrap and in fact it can do anything it wants, because it has undefined behavior.

Weak Memory

But let us first talk about the possibility of a panic. How can that happen? The reason is that accessing memory is slow, at least compared to other operations like arithmetic or accessing a register. Both compilers and CPUs do everything they can to make sure the program does not wait for memory. For example, your compiler might decide that it is advantageous to reorder the two instructions in the first thread, putting the FLAG = 1 before the DATA = Some(...). These are writes to two distinct memory locations, so reordering is fine as the overall result is the same – right?

Well, that was true when all programs were single-threaded, but with a concurrent program, another thread can actually tell that the writes are now happening in a different order! That’s how the second thread might see FLAG = 1 but still read DATA = NONE. Moreover, even if the compiler didn’t do anything like that, we could still see a result of “0” if you are running this program on an ARM chip. These chips are very aggressive in how they cache memory accesses, with the result that even if you write a program like the above in assembly, you may still be up for a surprise.

We call memory that can exhibit such strange reordering effects “weak memory” because it is weaker in the sense that it provides fewer guarantees than one might naively expect.

Data Races

But why does the program have undefined behavior? The reason is that Rust has tons of primitives whose behavior is essentially impossible to specify once concurrency gets involves. Consider the assignment DATA = Some(...) above, where an entire Option<String> is being copied. What would happen if some other thread reads DATA while the assignment is still in progress? The other thread would see some kind of mixture between the old and the new DATA! That’s just complete garbage. So systems programming languages essentially give up at this point and declare data races to be undefined behavior.

A data race is defined as follows:

Two accesses to the same location form a data race if:

1. at least one of them is a write, and
2. at least one of them is not a special atomic access, and
3. the accesses are not ordered.

We will talk about the “order” mentioned in condition (3) later. For now, let us look at “atomic accesses”.

Atomic Memory Accesses

In our program above, the reads and writes to FLAG satisfy all three conditions. To fix that, we will negate condition (2): We will use atomic accesses.

All usual accesses (like all the ones in our example code above) are so called non-atomic accesses. They can be compiled efficiently and are heavily optimized, but as we have seen they quickly cause trouble when writing low-level concurrent code. To make it possible to write such code, but still keep most of the existing optimizations, the C++11 standard introduced special functions that can be used to perform atomic read and write accesses to a location. We also have these atomic accesses in Rust, and we can use them in our program as follows:

extern crate rayon;

use std::sync::atomic::{AtomicUsize, Ordering};

static mut DATA : Option<String> = None;
static FLAG : AtomicUsize = AtomicUsize::new(0);

fn main() {
  rayon::join(
    || {
      // Initialize data
      unsafe { DATA = Some("Hello World!".to_string()); }
      // Signal that we are done
      FLAG.store(1, Ordering::Relaxed); },
    || {
      // Wait until the other thread is done
      while FLAG.load(Ordering::Relaxed) == 0 {}
      // Print data
      println!("{}", unsafe { DATA.as_ref().unwrap() });
    }
  );
}

Notice how we are using AtomicUsize as type for FLAG, and AtomicUsize::load and AtomicUsize::store instead of non-atomic memory accesses. These are atomic accesses, and hence we no longer have a data race on FLAG. (I will come back to this Ordering::Relaxed parameter soon.) Notice also that we significantly reduced the amount of unsafe operations, because AtomucUsize is actually completely safe to use from multiple threads.

Synchronization

However, we are not done yet: We still have a race on DATA! We also cannot use atomic accesses for DATA because of its type, Option<String>, which is just too big to be read/written atomically. This brings us to the third condition in our definition of data races: We have to somehow “order” the two accesses to DATA.

So, what is that “order” about? The idea is that while in general, operations on a concurrent program can happen in parallel, with strange effects like our first program panicking, there are still some operations that we can rely on to be properly “ordered”. For example, all operations within a single thread are ordered the way they are written, so e.g. DATA = Some(...) is ordered before FLAG.store. However, what is missing in the program above is some way to obtain an order between operations on different threads. This is where the Ordering::Relaxed parameter comes in: All atomic accesses come with an order mode that indicates under which circumstances this access will induce an order between operations. Our accesses above are Relaxed, which means no order is induced.

So, what we will have to do is to strengthen the mode of our FLAG accesses to induce an order between the write of 1 in the first thread, and the operation that reads 1 in the second thread. To this end, we use the Release and Acquire pair of modes:

extern crate rayon;

use std::sync::atomic::{AtomicUsize, Ordering};

static mut DATA : Option<String> = None;
static FLAG : AtomicUsize = AtomicUsize::new(0);

fn main() {
  rayon::join(
    || {
      // Initialize data
      unsafe { DATA = Some("Hello World!".to_string()); }
      // Signal that we are done
      FLAG.store(1, Ordering::Release); },
    || {
      // Wait until the other thread is done
      while FLAG.load(Ordering::Acquire) == 0 {}
      // Print data
      println!("{}", unsafe { DATA.as_ref().unwrap() });
    }
  );
}

This program, finally, is free of data races and hence has no undefined behavior and is guaranteed to print “Hello World!”.

They key point is that when a load(Acquire) reads from a store(_, Release), an order is induced between these two accesses. We also say that the two accesses synchronize. Together with the per-thread order, this means we have an order between the DATA = Some(...) in the first thread and the load of DATA in the second thread, through the accesses to FLAG. Intuitively, the store(_, Release) in the first thread “releases” everything that has been changed by this thread so far, and the load(Acquire) in the second thread then “acquires” all that data and makes it accessible for access later in this thread.

Now, most of the time a Mutex or RwLock is good enough to protect your data against concurrent accesses, so you do not have to worry about such details. (And, thanks to Rust thread safety guarantees, you never have to worry about such details in safe code!) But based on what you learned so far, it should make perfect sense that when a lock is released by thread A, that will happen through a Release access, and when a lock is acquired by thread B, that happens through an Acquire access. This way, the lock makes sure that there is an order between the accesses thread A performed when it held the lock (before the Release), and the accesses thread B will perform while it has the lock (after the Acquire).

Coming Back to Arc

I said that Mutex/RwLock are good enough most of the time. However, Arc is one of those cases where the overhead induced by an exclusive lock is just way too big, so it is worth using a more optimized, unsafe implementation. As such, you are going to find plenty of atomic accesses in the source code of Arc.

And it turns out, as Hai and Jacques-Henri noticed when attempting to prove correctness of Arc::get_mut, that there is one place where Relaxed was used as an ordering, but it really should have been Acquire. Discussing the exact details of the bug would probably fill another blog post (Arc is really subtle), but the high-level story is exactly like in our example above: Thanks to Acquire, an ordering is induced between the code that follows the get_mut and the code in another thread that dropped the last other Arc, decrementing the reference count to 1. The PR that fixed the problem contains some more details in the comments. With Relaxed, no such ordering is induced, so we have a data race. To be fair, it is very unlikely that this race could lead to real misbehavior, but I am still happy to know that we now have a proof that Arc is mostly¹ correctly synchronized.

One last thing: I have previously claimed that our first RustBelt paper already verified correctness of Arc, how can there still be bugs? In that paper, we did not (yet) have the tools to reason realistically about these ordering issues we have been discussing here, so instead we worked with a “sequentially consistent” logic which essentially assumes the strongest possible mode, SeqCst, for all atomic accesses. (We did not discuss SeqCst above, and in fact there are not many cases where it is really needed. Release and Acquire are enough most of the time.) This is one of the simplifications we made compared to real Rust to make the verification feasible. We were realistic enough to find another bug, but not realistic enough for this one. Hai and Jacques-Henri are currently working on remedying this particular simplification by extending the first RustBelt paper to also cover weak memory, and that’s when they ran into this problem.

Update: Turns out Servo has a copy of Arc that has the same problem. So we got two bugs for the price of one. :) /Update

Footnotes

Posted on Ralf's Ramblings on Jul 13, 2018.
Comments? Drop me a mail or leave a note on reddit!