Chapter 10 Memory model: The hard bits

This chapter describes the details of OCaml relaxed memory model. The relaxed memory model describes what values an OCaml program is allowed to witness when reading a memory location. If you are interested in high-level parallel programming in OCaml, please have a look at the parallel programming chapter 9.

This chapter is aimed at experts who would like to understand the details of the OCaml memory model from a practitioner’s perspective. For a formal definition of the OCaml memory model, its guarantees and the compilation to hardware memory models, please have a look at the PLDI 2018 paper on Bounding Data Races in Space and Time. The memory model presented in this chapter is an extension of the one presented in the PLDI 2018 paper. This chapter also covers some pragmatic aspects of the memory model that are not covered in the paper.

1 Why weakly consistent memory?

The simplest memory model that we could give to our programs is sequential consistency. Under sequential consistency, the values observed by the program can be explained through some interleaving of the operations from different domains in the program. For example, consider the following program with two domains d1 and d2 executing in parallel:

let d1 a b = let r1 = !a * 2 in let r2 = !b in let r3 = !a * 2 in (r1, r2, r3) let d2 b = b := 0 let main () = let a = ref 1 in let b = ref 1 in let h = Domain.spawn (fun _ -> let r1, r2, r3 = d1 a b in Printf.printf "r1 = %d, r2 = %d, r3 = %d\n" r1 r2 r3) in d2 b; Domain.join h

The reference cells a and b are initially 1. The user may observe r1 = 2, r2 = 0, r3 = 2 if the write to b in d2 occurred before the read of b in d1. Here, the observed behaviour can be explained in terms of interleaving of the operations from different domains.

Let us now assume that a and b are aliases of each other.

let d1 a b = let r1 = !a * 2 in let r2 = !b in let r3 = !a * 2 in (r1, r2, r3) let d2 b = b := 0 let main () = let ab = ref 1 in let h = Domain.spawn (fun _ -> let r1, r2, r3 = d1 ab ab in assert (not (r1 = 2 && r2 = 0 && r3 = 2))) in d2 ab; Domain.join h

In the above program, the variables ab, a and b refer to the same reference cell. One would expect that the assertion in the main function will never fail. The reasoning is that if r2 is 0, then the write in d2 occurred before the read of b in d1. Given that a and b are aliases, the second read of a in d1 should also return 0.

1.1 Compiler optimisations

Surprisingly, this assertion may fail in OCaml due to compiler optimisations. The OCaml compiler observes the common sub-expression !a * 2 in d1 and optimises the program to:

let d1 a b = let r1 = !a * 2 in let r2 = !b in let r3 = r1 in (* CSE: !a * 2 ==> r1 *) (r1, r2, r3) let d2 b = b := 0 let main () = let ab = ref 1 in let h = Domain.spawn (fun _ -> let r1, r2, r3 = d1 ab ab in assert (not (r1 = 2 && r2 = 0 && r3 = 2))) in d2 ab; Domain.join h

This optimisation is known as the common sub-expression elimination (CSE). Such optimisations are valid and necessary for good performance, and do not change the sequential meaning of the program. However, CSE breaks sequential reasoning.

In the optimized program above, even if the write to b in d2 occurs between the first and the second reads in d1, the program will observe the value 2 for r3, causing the assertion to fail. The observed behaviour cannot be explained by interleaving of operations from different domains in the source program. Thus, CSE optimization is said to be invalid under sequential consistency.

One way to explain the observed behaviour is as if the operations performed on a domain were reordered. For example, if the second and the third reads from d1 were reordered,

let d1 a b = let r1 = !a * 2 in let r3 = !a * 2 in let r2 = !b in (r1, r2, r3)

then we can explain the observed behaviour (2,0,2) returned by d1.

1.2 Hardware optimisations

The other source of reordering is by the hardware. Modern hardware architectures have complex cache hierarchies with multiple levels of cache. While cache coherence ensures that reads and writes to a single memory location respect sequential consistency, the guarantees on programs that operate on different memory locations are much weaker. Consider the following program:

let a = ref 0 and b = ref 0 let d1 () = a := 1; !b let d2 () = b := 1; !a let main () = let h = Domain.spawn d2 in let r1 = d1 () in let r2 = Domain.join h in assert (not (r1 = 0 && r2 = 0))

Under sequential consistency, we would never expect the assertion to fail. However, even on x86, which offers much stronger guarantees than ARM, the writes performed at a CPU core are not immediately published to all of the other cores. Since a and b are different memory locations, the reads of a and b may both witness the initial values, leading to the assertion failure.

This behaviour can be explained if a load is allowed to be reordered before a preceding store to a different memory location. This reordering can happen due to the presence of in-core store-buffers on modern processors. Each core effectively has a FIFO buffer of pending writes to avoid the need to block while a write completes. The writes to a and b may be in the store-buffers of cores c1 and c2 running the domains d1 and d2, respectively. The reads of b and a running on the cores c1 and c2, respectively, will not see the writes if the writes have not propagated from the buffers to the main memory.

2 Data race freedom implies sequential consistency

The aim of the OCaml relaxed memory model is to precisely describe which orders are preserved by the OCaml program. The compiler and the hardware are free to optimize the program as long as they respect the ordering guarantees of the memory model. While programming directly under the relaxed memory model is difficult, the memory model also describes the conditions under which a program will only exhibit sequentially consistent behaviours. This guarantee is known as data race freedom implies sequential consistency (DRF-SC). In this section, we shall describe this guarantee. In order to do this, we first need a number of definitions.

2.1 Memory locations

OCaml classifies memory locations into atomic and non-atomic locations. Reference cells, array fields and mutable record fields are non-atomic memory locations. Immutable objects are non-atomic locations with an initialising write but no further updates. Atomic memory locations are those that are created using the Atomic module.

2.2 Happens-before relation

Let us imagine that the OCaml programs are executed by an abstract machine that executes one action at a time, arbitrarily picking one of the available domains at each step. We classify actions into two: inter-domain and intra-domain. An inter-domain action is one which can be observed and be influenced by actions on other domains. There are several inter-domain actions:

On the other hand, intra-domain actions can neither be observed nor influence the execution of other domains. Examples include evaluating an arithmetic expression, calling a function, etc. The memory model specification ignores such intra-domain actions. In the sequel, we use the term action to indicate inter-domain actions.

A totally ordered list of actions executed by the abstract machine is called an execution trace. There might be several possible execution traces for a given program due to non-determinism.

For a given execution trace, we define an irreflexive, transitive happens-before relation that captures the causality between actions in the OCaml program. The happens-before relation is defined as the smallest transitive relation satisfying the following properties:

2.3 Data race

In a given trace, two actions are said to be conflicting if they access the same non-atomic location, at least one is a write and neither is an initialising write to that location.

We say that a program has a data race if there exists some execution trace of the program with two conflicting actions and there does not exist a happens-before relationship between the conflicting accesses. A program without data races is said to be correctly synchronised.

2.4 DRF-SC

DRF-SC guarantee: A program without data races will only exhibit sequentially consistent behaviours.

DRF-SC is a strong guarantee for the programmers. Programmers can use sequential reasoning i.e., reasoning by executing one inter-domain action after the other, to identify whether their program has a data race. In particular, they do not need to reason about reorderings described in section ‍10.1 in order to determine whether their program has a data race. Once the determination that a particular program is data race free is made, they do not need to worry about reorderings in their code.

3 Reasoning with DRF-SC

In this section, we will look at examples of using DRF-SC for program reasoning. In this section, we will use the functions with names dN to represent domains executing in parallel with other domains. That is, we assume that there is a main function that runs the dN functions in parallel as follows:

let main () =
  let h1 = Domain.spawn d1 in
  let h2 = Domain.spawn d2 in
  ...
  ignore @@ Domain.join h1;
  ignore @@ Domain.join h2

Here is a simple example with a data race:

(* Has data race *) let r = ref 0 let d1 () = r := 1 let d2 () = !r

r is a non-atomic reference. The two domains race to access the reference, and d1 is a write. Since there is no happens-before relationship between the conflicting accesses, there is a data race.

Both of the programs that we had seen in the section ‍10.1 have data races. It is no surprise that they exhibit non sequentially consistent behaviours.

Accessing disjoint array indices and fields of a record in parallel is not a data race. For example,

(* No data race *) let a = [| 0; 1 |] let d1 () = a.(0) <- 42 let d2 () = a.(1) <- 42
(* No data race *) type t = { mutable a : int; mutable b : int } let r = {a = 0; b = 1} let d1 () = r.a <- 42 let d2 () = r.b <- 42

do not have data races.

Races on atomic locations do not lead to a data race.

(* No data race *) let r = Atomic.make 0 let d1 () = Atomic.set r 1 let d2 () = Atomic.get r

Message-passing

Atomic variables may be used for implementing non-blocking communication between domains.

(* No data race *) let msg = ref 0 let flag = Atomic.make false let d1 () = msg := 42; (* a *) Atomic.set flag true (* b *) let d2 () = if Atomic.get flag (* c *) then !msg (* d *) else 0

Observe that the actions a and d write and read from the same non-atomic location msg, respectively, and hence are conflicting. We need to establish that a and d have a happens-before relationship in order to show that this program does not have a data race.

The action a precedes b in program order, and hence, a happens-before b. Similarly, c happens-before d. If d2 observes the atomic variable flag to be true, then b precedes c in happens-before order. Since happens-before is transitive, the conflicting actions a and d are in happens-before order. If d2 observes the flag to be false, then the read of msg is not done. Hence, there is no conflicting access in this execution trace. Hence, the program does not have a data race.

The following modified version of the message passing program does have a data race.

(* Has data race *) let msg = ref 0 let flag = Atomic.make false let d1 () = msg := 42; (* a *) Atomic.set flag true (* b *) let d2 () = ignore (Atomic.get flag); (* c *) !msg (* d *)

The domain d2 now unconditionally reads the non-atomic reference msg. Consider the execution trace:

Atomic.get flag; (* c *)
!msg; (* d *)
msg := 42; (* a *)
Atomic.set flag true (* b *)

In this trace, d and a are conflicting operations. But there is no happens-before relationship between them. Hence, this program has a data race.

4 Local data race freedom

The OCaml memory model offers strong guarantees even for programs with data races. It offers what is known as local data race freedom sequential consistency (LDRF-SC) guarantee. A formal definition of this property is beyond the scope of this manual chapter. Interested readers are encouraged to read the PLDI 2018 paper on Bounding Data Races in Space and Time.

Informally, LDRF-SC says that the data race free parts of the program remain sequentially consistent. That is, even if the program has data races, those parts of the program that are disjoint from the parts with data races are amenable to sequential reasoning.

Consider the following snippet:

let snippet () = let c = ref 0 in c := 42; let a = !c in (a, c)

Observe that c is a newly allocated reference. Can the read of c return a value which is not 42? That is, can a ever be not 42? Surprisingly, in the C++ and Java memory models, the answer is yes. With the C++ memory model, if the program has a data race, even in unrelated parts, then the semantics is undefined. If this snippet were linked with a library that had a data race, then, under the C++ memory model, the read may return any value. Since data races on unrelated locations can affect program behaviour, we say that C++ memory model is not bounded in space.

Unlike C++, Java memory model is bounded in space. But Java memory model is not bounded in time; data races in the future will affect the past behaviour. For example, consider the translation of this example to Java. We assume a prior definition of Class c {int x;} and a shared non-volatile variable C g. Now the snippet may be part of a larger program with parallel threads:

(* Thread 1 *)
C c = new C();
c.x = 42;
a = c.x;
g = c;

(* Thread 2 *)
g.x = 7;

The read of c.x and the write of g in the first thread are done on separate memory locations. Hence, the Java memory model allows them to be reordered. As a result, the write in the second thread may occur before the read of c.x, and hence, c.x returns 7.

The OCaml equivalent of the Java code above is:

let g = ref None let snippet () = let c = ref 0 in c := 42; let a = !c in (a, c) let d1 () = let (a,c) = snippet () in g := Some c; a let d2 () = match !g with | None -> () | Some c -> c := 7

Observe that there is a data race on both g and c. Consider only the first three instructions in snippet:

let c = ref 0 in
c := 42;
let a = !c in
...

The OCaml memory model is bounded both in space and time. The only memory location here is c. Reasoning only about this snippet, there is neither the data race in space (the race on g) nor in time (the future race on c). Hence, the snippet will have sequentially consistent behaviour, and the value returned by !c will be 42.

The OCaml memory model guarantees that even for programs with data races, memory safety is preserved. While programs with data races may observe non-sequentially consistent behaviours, they will not crash.

5 An operational view of the memory model

In this section, we describe the semantics of the OCaml memory model. A formal definition of the operational view of the memory model is presented in section 3 of the PLDI 2018 paper on Bounding Data Races in Space and Time. This section presents an informal description of the memory model with the help of an example.

Given an OCaml program, which may possibly contain data races, the operational semantics tells you the values that may be observed by the read of a memory location. For simplicity, we restrict the intra-thread actions to just the accesses to atomic and non-atomic locations, ignoring domain spawn and join operations, and the operations on mutexes.

We describe the semantics of the OCaml memory model in a straightforward small-step operational manner. That is, the semantics is described by an abstract machine that executes one action at a time, arbitrarily picking one of the available domains at each step. This is similar to the abstract machine that we had used to describe the happens-before relationship in section ‍10.2.2.

5.1 Non-atomic locations

In the semantics, we model non-atomic locations as finite maps from timestamps t to values v. We take timestamps to be rational numbers. The timestamps are totally ordered but dense; there is a timestamp between any two others.

For example,

a: [t1 -> 1; t2 -> 2]
b: [t3 -> 3; t4 -> 4; t5 -> 5]
c: [t6 -> 5; t7 -> 6; t8 -> 7]

represents three non-atomic locations a, b and c and their histories. The location a has two writes at timestamps t1 and t2 with values 1 and 2, respectively. When we write a: [t1 -> 1; t2 -> 2], we assume that t1 < t2. We assume that the locations are initialised with a history that has a single entry at timestamp 0 that maps to the initial value.

5.2 Domains

Each domain is equipped with a frontier, which is a map from non-atomic locations to timestamps. Intuitively, each domain’s frontier records, for each non-atomic location, the latest write known to the thread. More recent writes may have occurred, but are not guaranteed to be visible.

For example,

d1: [a -> t1; b -> t3; c -> t7]
d2: [a -> t1; b -> t4; c -> t7]

represents two domains d1 and d2 and their frontiers.

5.3 Non-atomic accesses

Let us now define the semantics of non-atomic reads and writes. Suppose domain d1 performs the read of b. For non-atomic reads, the domains may read an arbitrary element of the history for that location, as long as it is not older than the timestamp in the domains’s frontier. In this case, since d1 frontier at b is at t3, the read may return the value 3, 4 or 5. A non-atomic read does not change the frontier of the current domain.

Suppose domain d2 writes the value 10 to c (c := 10). We pick a new timestamp t9 for this write such that it is later than d2’s frontier at c. Note a subtlety here: this new timestamp might not be later than everything else in the history, but merely later than any other write known to the writing domain. Hence, t9 may be inserted in c’s history either (a) between t7 and t8 or (b) after t8. Let us pick the former option for our discussion. Since the new write appears after all the writes known by the domain d2 to the location c, d2’s frontier at c is also updated. The new state of the abstract machine is:

(* Non-atomic locations *)
a: [t1 -> 1; t2 -> 2]
b: [t3 -> 3; t4 -> 4; t5 -> 5]
c: [t6 -> 5; t7 -> 6; t9 -> 10; t8 -> 7] (* new write at t9 *)

(* Domains *)
d1: [a -> t1; b -> t3; c -> t7]
d2: [a -> t1; b -> t4; c -> t9] (* frontier updated at c *)

5.4 Atomic accesses

Atomic locations carry not only values but also synchronization information. We model atomic locations as a pair of the value held by that location and a frontier. The frontier models the synchronization information, which is merged with the frontiers of threads that operate on the location. In this way, non-atomic writes made by one thread can become known to another by communicating via an atomic location.

For example,

(* Atomic locations *)
A: 10, [a -> t1; b -> t5; c -> t7]
B: 5,  [a -> t2; b -> t4; c -> t6]

shows two atomic variables A and B with values 10 and 5, respectively, and frontiers of their own. We use upper-case variable names to indicate atomic locations.

During atomic reads, the frontier of the location is merged into that of the domain performing the read. For example, suppose d1 reads B. The read returns 5, and d1’s frontier updated by merging it with B’s frontier, choosing the later timestamp for each location. The abstract machine state before the atomic read is:

(* Non-atomic locations *)
a: [t1 -> 1; t2 -> 2]
b: [t3 -> 3; t4 -> 4; t5 -> 5]
c: [t6 -> 5; t7 -> 6; t9 -> 10; t8 -> 7]

(* Domains *)
d1: [a -> t1; b -> t3; c -> t7]
d2: [a -> t1; b -> t4; c -> t9]

(* Atomic locations *)
A: 10, [a -> t1; b -> t5; c -> t7]
B: 5,  [a -> t2; b -> t4; c -> t6]

As a result of the atomic read, the abstract machine state is updated to:

(* Non-atomic locations *)
a: [t1 -> 1; t2 -> 2]
b: [t3 -> 3; t4 -> 4; t5 -> 5]
c: [t6 -> 5; t7 -> 6; t9 -> 10; t8 -> 7]

(* Domains *)
d1: [a -> t2; b -> t4; c -> t7] (* frontier updated at a and b *)
d2: [a -> t1; b -> t4; c -> t9]

(* Atomic locations *)
A: 10, [a -> t1; b -> t5; c -> t7]
B: 5,  [a -> t2; b -> t4; c -> t6]

During atomic writes, the value held by the atomic location is updated. The frontiers of both the writing domain and that of the location being written to are updated to the merge of the two frontiers. For example, if d2 writes 20 to A in the current machine state, the machine state is updated to:

(* Non-atomic locations *)
a: [t1 -> 1; t2 -> 2]
b: [t3 -> 3; t4 -> 4; t5 -> 5]
c: [t6 -> 5; t7 -> 6; t9 -> 10; t8 -> 7]

(* Domains *)
d1: [a -> t2; b -> t4; c -> t7]
d2: [a -> t1; b -> t5; c -> t9] (* frontier updated at b *)

(* Atomic locations *)
A: 20, [a -> t1; b -> t5; c -> t9] (* value updated. frontier updated at c. *)
B: 5,  [a -> t2; b -> t4; c -> t6]

5.5 Reasoning with the semantics

Let us revisit an example from earlier (section 10.1).

let a = ref 0 and b = ref 0 let d1 () = a := 1; !b let d2 () = b := 1; !a let main () = let h = Domain.spawn d2 in let r1 = d1 () in let r2 = Domain.join h in assert (not (r1 = 0 && r2 = 0))

This program has a data race on a and b, and hence, the program may exhibit non sequentially consistent behaviour. Let us use the semantics to show that the program may exhibit r1 = 0 && r2 = 0.

The initial state of the abstract machine is:

(* Non-atomic locations *)
a: [t0 -> 0]
b: [t1 -> 0]

(* Domains *)
d1: [a -> t0; b -> t1]
d2: [a -> t0; b -> t1]

There are several possible schedules for executing this program. Let us consider the following schedule:

1: a := 1 @ d1
2: b := 1 @ d2
3: !b     @ d1
4: !a     @ d2

After the first action a:=1 by d1, the machine state is:

(* Non-atomic locations *)
a: [t0 -> 0; t2 -> 1] (* new write at t2 *)
b: [t1 -> 0]

(* Domains *)
d1: [a -> t2; b -> t1] (* frontier updated at a *)
d2: [a -> t0; b -> t1]

After the second action b:=1 by d2, the machine state is:

(* Non-atomic locations *)
a: [t0 -> 0; t2 -> 1]
b: [t1 -> 0; t3 -> 1] (* new write at t3 *)

(* Domains *)
d1: [a -> t2; b -> t1]
d2: [a -> t0; b -> t3] (* frontier updated at b *)

Now, for the third action !b by d1, observe that d1’s frontier at b is at t1. Hence, the read may return either 0 or 1. Let us assume that it returns 0. The machine state is not updated by the non-atomic read.

Similarly, for the fourth action !a by d2, d2’s frontier at a is at t0. Hence, this read may also return either 0 or 1. Let us assume that it returns 0. Hence, the assertion in the original program, assert (not (r1 = 0 && r2 = 0)), will fail for this particular execution.

6 Non-compliant operations

There are certain operations which are not memory model compliant.