When we write programs that use threads operating in multiuser systems, multiprocessing systems, or a combination of the two, we often discover that we have critical sections of code, where we need to ensure that a single process (or a single thread of execution) has exclusive access to a resource. Semaphores have a complex programming interface. Fortunately, we can easily provide ourselves with a much-simplified interface that is sufficient for most semaphore-programming problems. a semaphore is a special variable that takes only whole positive numbers and upon which only two operations are allowed: wait and signal. Since “wait” and “signal” already have special meanings in Linux and UNIX programming, we’ll use the original notation:
❑ P(semaphore variable) for wait
❑ V(semaphore variable) for signal
These letters come from the Dutch words for wait (passeren: to pass, as in a checkpoint before the critical section) and signal (vrijgeven: to give or release, as in giving up control of the critical section). You may also come across the terms “up” and “down” used in relation to semaphores, taken from the use of signaling flags.
Semaphore Definition
The simplest semaphore is a variable that can take only the values 0 and 1, a binary semaphore. This is the most common form. Semaphores that can take many positive values are called general semaphores. For the remainder of this chapter, we’ll concentrate on binary semaphores.
The definitions of P and V are surprisingly simple. Suppose we have a semaphore variable sv. The two operations are then defined as follows:
P(sv) If sv is greater than zero, decrement sv. If sv is zero, suspend execution of this process.
V(sv) If some other process has been suspended waiting for sv, make it resume execution. If no process is suspended waiting for sv, increment sv.
Another way of thinking about semaphores is that the semaphore variable, sv, is true when the critical section is available, is decremented by P(sv) so it’s false when the critical section is busy, and is incremented by V(sv) when the critical section is again available. Be aware that simply having a normal variable that you decrement and increment is not good enough, because you can’t express in C, C++, or almost any conventional programming language the need to make a single, atomic operation of the test to see whether the variable is true, and if so change the variable to make it false. This is what makes the semaphore operations special.
Linux Semaphore Facilities
Now that we’ve seen what semaphores are and how they work in theory, we can look at how the features are implemented in Linux. The interface is rather elaborate and offers far more facilities than are generally required. All the Linux semaphore functions operate on arrays of general semaphores rather than a single binary semaphore. At first sight, this just seems to make things more complicated, but in complex cases where a process needs to lock multiple resources, the ability to operate on an array of semaphores is a big advantage. In this chapter, we will concentrate on using single semaphores, since in most cases that’s all you will need to use.
The semaphore function definitions are
#include <sys/sem.h>
int semctl(int sem_id, int sem_num, int command, ...);
int semget(key_t key, int num_sems, int sem_flags);
int semop(int sem_id, struct sembuf *sem_ops, size_t num_sem_ops);
As we work through each function in turn, remember that these functions were designed to work for arrays of semaphore values, which makes their operation significantly more complex than would have been required for a single semaphore.
Notice that key acts very much like a filename in that it represents a resource that programs may use and cooperate in using if they agree on a common name. Similarly, the identifier returned by semget and used by the other shared memory functions is very much like the FILE * file stream returned by fopen in that it’s a value used by the process to access the shared file. And, just as with files, different processes will have different semaphore identifiers, though they refer to the same semaphore. This use of a key and identifiers is common to all of the IPC facilities discussed here, although each facility uses independent keys and identifiers.
Semget
The semget function creates a new semaphore or obtains the semaphore key of an existing semaphore.
int semget(key_t key, int num_sems, int sem_flags);
The first parameter, key, is an integral value used to allow unrelated processes to access the same semaphore. All semaphores are accessed indirectly by the program supplying a key, for which the system then generates a semaphore identifier. The semaphore key is used only with semget. All other semaphore functions use the semaphore identifier returned from semget.
There is a special semaphore key value, IPC_PRIVATE (usually 0), that is intended to create a semaphore that only the creating process could access. This rarely has any useful purpose, which is fortunate, because on some Linux systems the manual pages list as a bug the fact that IPC_PRIVATE may not prevent other processes from accessing the semaphore. The num_sems parameter is the number of semaphores required. This is almost always 1. The sem_flags parameter is a set of flags, very much like the flags to the open function. The lower nine bits are the permissions for the semaphore, which behave like file permissions. In addition, these can be bitwise ORed with the value IPC_CREAT to create a new semaphore. It’s not an error to have the IPC_CREAT flag set and give the key of an existing semaphore. The IPC_CREAT flag is silently ignored if it is not required. We can use IPC_CREAT and IPC_EXCL together to ensure that we obtain a new, unique semaphore. It will return an error if the semaphore already exists.
The semget function returns a positive (nonzero) value on success; this is the semaphore identifier used in the other semaphore functions. On error, it returns –1.
semop
The function semop is used for changing the value of the semaphore:
int semop(int sem_id, struct sembuf *sem_ops, size_t num_sem_ops);
The first parameter, sem_id, is the semaphore identifier, as returned from semget. The second parameter,
sem_ops, is a pointer to an array of structures, each of which will have at least the following members:
struct sembuf {
short sem_num;
short sem_op;
short sem_flg;
}
The first member, sem_num, is the semaphore number, usually 0 unless you’re working with an array of semaphores. The sem_op member is the value by which the semaphore should be changed. (You can change a semaphore by amounts other than 1.) In general, only two values are used, –1, which is our P operation to wait for a semaphore to become available, and +1, which is our V operation to signal that a semaphore is now available. The final member, sem_flg, is usually set to SEM_UNDO. This causes the operating system to track the changes made to the semaphore by the current process and, if the process terminates without releasing the semaphore, allows the operating system to automatically release the semaphore if it was held by this process. It’s good practice to set sem_flg to SEM_UNDO, unless you specifically require different behavior. If you do decide you need a value other than SEM_UNDO, it’s important to be consistent, or you can get very confused as to whether the kernel is attempting to “tidy up” your semaphores when your process exits. All actions called for by semop are taken together to avoid a race condition implied by the use of multiple semaphores. You can find full details of the processing of semop in the manual pages.
Semctl
The semctl function allows direct control of semaphore information:
int semctl(int sem_id, int sem_num, int command, ...);
The first parameter, sem_id, is a semaphore identifier, obtained from semget. The sem_num parameter is the semaphore number. You use this when you’re working with arrays of semaphores. Usually, this is 0, the first and only semaphore. The command parameter is the action to take, and a fourth parameter, if present, is a union semun, which according to the X/OPEN specification, must have at least the following members:
union semun {
int val;
struct semid_ds *buf;
unsigned short *array;
}
Many versions of Linux have a definition of the semun union in a header file (usually sem.h), though X/Open does say that you have to declare your own. If you do find that you need to declare your own, check the manual pages for semctl to see if there is a definition given. If there is, we suggest you use exactly the definition given in your manual, even if it differs from that above. There are many different possible values of command allowed for semctl. Only the two that we describe here are commonly used. For full details of the semctl function, you should consult the manual page. The two common values of command are
❑ SETVAL: Used for initializing a semaphore to a known value. The value required is passed as
the val member of the union semun. This is required to set the semaphore up before it’s used for the first time.
❑ IPC_RMID: Used for deleting a semaphore identifier when it’s no longer required. The semctl function returns different values depending on the command parameter. For SETVAL and IPC_RMID it returns 0 for success and –1 on error.
