View Source Document


Matchbox is a tool which finds race conditions. It is currently a work in progress.

It is not a production-quality tool; rather, it is more of a demonstration. For that reason, it is implemented in Javascript and runs in a web browser. It may be educational for understanding what a race condition is, and the sort of reasoning that is needed to find and prevent them.

Matchbox takes, as input, two programs written in a toy assembly-like language. These two program each have their own set of registers, but the main memory which they write to and read from is shared between them.

It then computes all possible interleavings of these two programs, and executes them all, each on freshly-zeroed registers and shared memory. Unless all the possible interleavings leave the memory in exactly the same state after execution, the programs have a race condition.

Basic Examples

These two programs do not have a race condition. The reason is that Matchbox's INC instruction atomically updates memory; two INCs cannot run at the same time.





But the following two programs do have a race condition:


MOV M0, R0
MOV R0, M0


MOV M0, R0
MOV R0, M0

The reason is that the second program's MOV M0, R0 might happen before the first program's MOV R0, M0 — in which case M0 = 1 at the end — or it might happen after — in which case M0 = 2.

Syntax and Semantics

Each line may contain a pragma, an instruction, a comment, or be blank. Comments and blank lines are ignored. Comments begin with a ; in the first column.

There are two pragmas. The first, DESC, indicates a section of the program which contains a human-readable (and computer-ignored) description of the following program(s). This description is not free-form; it should be given in terms of comments.

The other pragma, PROG, is followed by an integer which specifies which program is being given next in the file.

These pragmas are only used when the source is a single text file; if the two programs are being given separately, for example in individual textboxes in a GUI, they are not used (and in fact their presence is an error.)

Instructions consist of an opcode which is followed by zero, one, or two data references.

A data reference may be an immediate value (denoted by an integer literal,) a register reference (denoted by an R immediately followed by an integer literal,) or a memory reference (denoted by an M immediately followed by an integer literal.)

When an instruction has two data references, often the first data reference is called the source and the second is called the destination. In this case, the destination may not be an immediate value.

All instructions are atomic in their execution.


MOV takes 2 data references. It reads a value from the source, and writes that value into the destination. Thus it can write a constant value into a register or memory, copy a register to another register or to a memory location, or copy a memory location to another location in memory or to a register.

INC takes one data reference, which may not be an immediate value. It increments the value stored at the given register or memory location.

WAIT takes a memory reference and an immediate value, and simulates a busy wait for that memory location to contain that value. Note that this is merely a simulation: if the memory location does not contain that value when WAIT is executed, the program run is merely discarded.

This is justified on the grounds that, if it were a busy wait for that value, the present interleaving would simply never have occurred. (This doesn't entirely make sense if WAIT is the final instruction in a program, though.)

Advanced Examples

There are some classic synchronization algorithms in computer science. We can try some of them here, and see if Matchbox can tell us if they work.

Peterson's algorithm is implemented in eg/petersons-no-race.mbox. Inside the critical sections, the programs shown above, the ones which have the race condition, are embedded. Matchbox takes a while to find all the interleavings, but once it does, it confirms that there is no race condition.

Dekker's algorithm will not be possible to implement in Matchbox, because it contains a nested while loop (see "Limitations", below.)

Note that these algorithms are generally not needed on modern computer architectures which provide operations like atomic test-and-set. They are demonstrated here for their beauty as algorithms and their historical importance only.


Why should I care? I don't program in assembly language!

Doesn't matter; you still face race conditions. Race conditions are everywhere, and every software developer should be familiar with them (I like to use the questions "Can you tell me what a race condition is? Can you give me an example? Can you tell me about one you had to find and fix?" during phone screens.)

A glaring example that is relevant to virtually every programmer is the filesystem. It's a big, shared, mutable store. There are often some operations, like renaming a file, that are guaranteed to be atomic; but most filesystem operations do not have this guarantee.


There are of course two significant limitations to this toy machine language: it has no loops, and no indirect references.

This is because both of those features would add significantly more dynamic factors to the run. Taking account of them would add a lot of overhead, while not adding much to the basic presentation.

However, to consider how they might work:

To handle loops, we would probably want to split each program up into basic blocks, and run every basic block alongside every other basic block, to see if there are any race conditions there.

Even that is not quite enough in practice, for there is nothing stopping there from being, for example, a loop in-between two mutexes. In this case, we would probably report some false positives — race conditions which look like they can happen, in theory, but would never happen in practice.

To handle indirect references (computed offsets, like arrays with indexes,) would not be terribly difficult — unless they were combined with loops. Then we would need to determine the entire range of possible memory locations that could be affected, and check all of them.

These features would be better handled with abstract interpretation, than with direct simulation like we're doing here.