Rationale

The rationale behind mq is to provide an efficient, priority-driven asynchronous IPC mechanism. When the AT&T System V UNIX message queues were first implemented, the reasoning was similar: the only form of IPC was half-duplex pipes and message queues were seen to overcome the performance limitations with these.

But arguably in modern systems there is little difference between the efficiency of the System V message queues, pipes, and UNIX domain sockets (if anything, the AT&T System V UNIX message queues tend to be slower than the rest). The fundamental performance bottleneck is however still there with mq as well: data must be first copied from the sender to the kernel and then from the kernel to the receiver. The bigger the message, the higher the overhead.

For realtime applications, mq offers some advantages:

1. Unlike the predecessors, mq provides an asynchronous notification mechanism.
2. Messages are prioritized. The queue always remains sorted such that the oldest message of the highest priority is always received first, regardless of the number of messages in the queue.
3. By default, the functions to send and receive messages are blocking calls. It is however possible to use non-blocking variants with mq. Furthermore, it is possible to specify timeouts to avoid non-deterministic blocking.
4. Resource limits can be enforced -- or perhaps more importantly, the availability of resources can be ensured as the internal data structures are preallocated.

Descriptors and Naming

Comparable to pipes and FIFOs (a.k.a. named pipes), all POSIX message queue operations are performed by using a descriptor. The used type is mqd_t, an abbreviation from a “message queue descriptor”. In the NetBSD implementation this is actually an ordinary file descriptor. This means that it is possible, but not portable, to monitor a message queue descriptor by using poll(2) or select(2).

Message queues are named by character strings that represent (absolute) pathnames. The used interface is analogous to the conventional file concepts. But unlike FIFOs and pipes, neither POSIX nor System V message queues are accessed by using open(2), read(2), or write(2). Instead, equivalents such as mq_open(), mq_close(), and mq_unlink() are used.

The standard does not specify whether POSIX message queues are exposed at the file system level. It can be argued that mq inherited an old problem with the System V message queues. Even if an implementation would have support for it, it is not portable to view message queues by ls(1), remove these with rm(1), or adjust the permissions with chmod(1).

Processes

When a new process is created or the program is terminated, message queues behave like files. More specifically, when fork(2) is called, files and message queues are inherited, and when the program terminates by calling exit(3) or _exit(2), both file descriptors and message queues are closed. However, the exec(3) family of functions behave somewhat differently for message queues and files: all message queues are closed when a process calls one of the exec() functions. In this respect POSIX message queues are closer to FIFOs than normal pipes.

Attributes

All message queues have an attribute associated with them. This is represented by the mq_attr structure:

struct mq_attr { 
	long	mq_flags; 
	long	mq_maxmsg; 
	long	mq_msgsize; 
	long	mq_curmsgs; 
};

The members in the structure are: flags set for the message queue (mq_flags), the maximum number of messages in the queue (mq_maxmsg), the maximum size of each message (mq_msgsize), and the number of queued messages (mq_curmsgs).

The overall resource requirements for a particular message queue are given by mq_maxmsg and mq_msgsize. These two can be specified when the queue is created by a call to mq_open(). The constraints are enforced through the lifetime of the queue: an error is returned if a message larger than mq_msgsize is sent, and if the message queue is already full, as determined by mq_maxmsg, the call to queue a message will either block or error out.

Although there are two functions, mq_getattr() and mq_setattr(), to retrieve and set attributes, resource limits cannot be changed once the queue has been created. In NetBSD the super user may however control the global resource limits by using few sysctl(7) variables.

Asynchronous Notification

Instead of blocking in the functions that receive messages, mq offers an asynchronous mechanism for a process to receive notifications that messages are available in the message queue. The function mq_notify() is used to register for notification. Either a signal or a thread can be used as the type of notification; see sigevent(3) for details.

Bear in mind that no notification is sent for an arrival of a message to a non-empty message queue. In other words, mq_notify() does not by itself ensure that a process will be notified every time a message arrives. Thus, after having called mq_notify(), an application may need to repeatedly call mq_receive() until the queue is empty. This requires that the message queue was created with the O_NONBLOCK flag; otherwise mq_receive() blocks until a message is again queued or the call is interrupted by a signal. This may be a limitation for some realtime applications.

Priorities

Each message has a priority, ranging from 0 to the implementation-defined MQ_PRIO_MAX. The POSIX standard enforces the minimum value of the maximum priority to be 32. All messages are inserted into a message queue according to the specified priority. High priority messages are sent before low priority messages. If the used priority is constant, mq follows the FIFO (First In, First Out) principle.

The basic rule of thumb with realtime prioritization is that low priority tasks should never unnecessarily delay high priority tasks. Priority inheritance is not however part of the provided API; the receiver process may run at low priority even when receiving high priority messages. To address this limitation and other potential realtime problems, the user may consider other functions from the POSIX Real-time Library (librt, -lrt). The process scheduling interface described in sched(3) can be mentioned as an example.