pthread Tutorial

c Copyright 2005 by Peter C. Chapin August 23, 2005

Contents
1 Introduction 2 Creating and Destroying Threads 2.1 2.2 Creating Threads . . . . . . . . . . . . . . . . . . . . . . . . . . . Returning Results from Threads . . . . . . . . . . . . . . . . . . 2 3 3 5 8 8 10 13 14 16 16 17 18 18 19 20 20 23 23

3 Thread Synchronization 3.1 3.2 3.3 3.4 Mutual Exclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . Condition Variables . . . . . . . . . . . . . . . . . . . . . . . . . Semaphores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Reader/Writer Locks . . . . . . . . . . . . . . . . . . . . . . . . .

4 Thead Models 4.1 4.2 4.3 4.4 4.5 Boss/Worker Model . . . . . . . . . . . . . . . . . . . . . . . . . Pipeline Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . Background Task Model . . . . . . . . . . . . . . . . . . . . . . . Interface/Implementation Model . . . . . . . . . . . . . . . . . . General Comments . . . . . . . . . . . . . . . . . . . . . . . . . .

5 Thread Safety 5.1 5.2 5.3 Levels of Thread Safety . . . . . . . . . . . . . . . . . . . . . . . Writing Thread Safe Code . . . . . . . . . . . . . . . . . . . . . . Exception Safety vs Thread Safety . . . . . . . . . . . . . . . . . 1

6 Rules for Multithreaded Programming 6.1 Shared Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.1 6.1.2 6.1.3 6.1.4 6.2 What data is shared? . . . . . . . . . . . . . . . . . . . . What data is not shared? . . . . . . . . . . . . . . . . . . What type of simultaneous access causes a problem? . . . What type of simultaneous access is safe? . . . . . . . . .

24 24 24 25 25 25 26

What can I count on? . . . . . . . . . . . . . . . . . . . . . . . .

Legal
Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.1 or any later version published by the Free Software Foundation; with no Invariant Sections, with no Front-Cover Texts, and with no Back-Cover Texts. A copy of the license is A included in the le GFDL.txt distributed with the L TEX source of this document.

Introduction

This document is intended to be a short but useful tutorial on how to use Posix threads (pthreads). In this document I do not attempt to give a full description of all pthread features. Instead I hope to give you enough information to use pthreads in a basic, yet eective way. Please refer to a text on pthreads for the more esoteric details of the standard. In addition to talking about the pthread interface itself, I also spend time in this document discussing issues regarding concurrent programming in general. While they are not specic to pthreads, it is a must that you understand these issues if you are to use pthreadsor any thread libraryeectively. I will assume that you are compiling pthreads programs on a Unix system. However, you should be aware that the pthreads interface is not really specic to Unix. It is a standard application program interface that could potentially be implemented on many dierent systems. However, pthreads is the usual way multi-threaded support is typically oered in the Unix world. Although many systems support their own internal method of handling threads, virtually every Unix system that supports threads at all oers the pthreads interface. The pthreads API can be implemented either in the kernel of the operating system or in a library. It can either be preemptive or it can be non-preemptive.

A portable program based on pthreads should not make any assumptions about these matters. When you compile a program that uses pthreads, you will probably have to set special options on the compilers command line to indicate extra (or dierent) libraries and/or alternative code generating stratagies. Consult your compilers documentation for more information on this. Often you can indicate your desire to use pthreads by supplying the -pthread option at the end of the compiler command line. For example $ gcc -o myprog myprog.c -pthread This single option species that the pthreads library should be linked and also causes the compiler to properly handle the multiple threads in the code that it generates.

Creating and Destroying Threads

Clearly the rst step required in understanding how to build a multithreaded program is to understand how to create and destroy threads. There are a number of subtle issues associated with this topic. Normally one wants to not only create a thread but also to send that thread one or more parameters. In addition when a thread ends, one normally wants to be able to retieve one or more values that are returned from the thread. In this section I will describe how these things can be done with Posix threads.

2.1

Creating Threads

To create a new thread you need to use the pthread create() function. Listing 1 shows a skeleton program that creates a thread that does nothing and then waits for the thread to terminate. The pthread create() function gives back a thread identier that can be used in other calls. The second parameter is a pointer to a thread attribute object that you can use to set the threads attributes. The null pointer means to use default attributes. The third parameter is a pointer to the function the thread is to execute. The nal parameter is the argument to the function. By using void pointers here, any sort of data could potentially be sent to the thread function provided proper casts are applied. In the skeleton example I show how a single integer can be used as a thread argument. You should plan to collect the exit status of all the threads you create by calling pthread join() on each thread eventually. Alternatively you can create a detached thread. The exit status for such threads are thrown away. The problem 3

Listing 1: Skeleton Thread Program

#include < p t h r e a d . h> /* * The function to be executed by the thread should take a * void * parameter and return a void * exit status code . */ void t h r e a d f u n c t i o n ( void a r g ) { // Cast the parameter into what is needed . i n t incoming = ( i n t ) a r g ; // Do whatever is necessary using * incoming as the argument . // The thread terminates when this function returns . return NULL; } i n t main ( void ) { pthread t thread ID ; void exit status ; int value ; // Put something meaningful into value . value = 42; // Create the thread , passing & value for the argument . p t h r e a d c r e a t e (& t h r e a d I D , NULL, t h r e a d f u n c t i o n , &v a l u e ) ; // The main program continues while the thread executes . // Wait for the thread to terminate . p t h r e a d j o i n ( t h r e a d I D , &e x i t s t a t u s ) ; // Only the main thread is running now . return 0 ; }

with detached threads is that, unless you make special arrangements, you are never sure when they complete. Usually you want to make sure all your threads have terminated cleanly before you end the process by returning from main(). If you want to kill a thread before its thread function returns normally, you can use pthread cancel(). However, there are diculties involved in doing that. You must be sure the thread has released any local resources that it has obtained before it actually dies. For example if a thread has dynamcially allocated memory and you cancel it before it can free that memory, your program will have a memory leak. This is dierent than when you kill an entire process. The operating system will typically clean up (certain) resources that are left dangling by the process. In particular, the entire address space of a process is recovered. However, the operating system will not do that for a thread since all the threads in a process share resources. For all the operating system knows, the memory allocated by one thread will be used by another thread. This situation makes canceling threads carelessly a bad idea. Exercies 1. Write a program that creates 10 threads. Have each thread execute the same function and pass each thread a unique number. Each thread should print Hello, World (thread n) ve times where n is replaced by the threads number. Use an array of pthread t objects to hold the various thread IDs. Be sure the program doesnt terminate until all the threads are complete. Try running your program on more than one machine. Are there any dierences in how it behaves?

2.2

Returning Results from Threads

The example in the last section illustrated how you can pass an argument into your thread function if necessary. In this section I will describe how to return results from thread functions. Note that the thread functions are declared to return a pointer to void. However, there are some pitfalls involved in using that pointer appropriately. The code below shows one attempt at returning an integer status code from a thread function.
void t h r e a d f u n c t i o n ( void ) { i n t code = DEFAULT VALUE; // Set the value of code as appropriate . return ( void ) code ; }

This method will only work on machines where integers can be converted to a pointer and then back to an integer without loss of information. On some machines such conversions are dangerous. In fact this method will fail in all cases where one attempts to return an object, such as a structure, that is larger than a pointer. In contrast, the code below doesnt ght the type system. It returns a pointer to an internal buer where the return value is stored. While the example shows an array of characters for the buer, one can easily imagine it being an array of any necessary type, or a single object such as an integer status code or a structure with many members.
void t h r e a d f u n c t i o n ( void ) { char b u f f e r [ 6 4 ] ; // Fill up the buffer with something good . return b u f f e r ; }

Alas, the code above fails because the internal buer is automatic and it vanishes as soon as the thread function returns. The pointer given back to the calling thread points at undened memory. This is another example of the classic dangling pointer error. In the next attempt the buer is made static so that it will continue to exist even after the thread function terminates. This gets around the dangling pointer problem.
void t h r e a d f u n c t i o n ( void ) { s t a t i c char b u f f e r [ 6 4 ] ; // Fill up the buffer with something good . return b u f f e r ; }

This method might be satisfactory in some cases, but it doesnt work in the common case of multiple threads running the same thread function. In such a situation the second thread will overwrite the static buer with its own data and destroy that left by the rst thread. Global data suers from this same problem since global data always has static duration. The version below is the most general and most robust.
void t h r e a d f u n c t i o n ( void ) { char b u f f e r = ( char ) m a l l o c ( 6 4 ) ;

// Fill up the buffer with something good . return b u f f e r ; }

This version allocates buer space dynamically. This approach will work correctly even if multiple threads execute the thread function. Each will allocate a dierent array and store the address of that array in a stack variable. Every thread has its own stack so automatic data objects are dierent for each thread. In order to receive the return value of a thread the higher level thread must join with the subordinate thread. This is shown in the main function of Listing 1. In particular
void e x i t s t a t u s ; // Wait for the thread to terminate . p t h r e a d j o i n ( t h r e a d I D , &e x i t s t a t u s ) ;

The pthread join() function blocks until the thread specied by its rst argument terminates. It then stores into the pointer pointed at by its second argument the value returned by the thread function. To use this pointer, the higher level thread must cast it into an appropriate type and dereference it. For example
char t h r e a d r e s u l t ; t h r e a d r e s u l t = ( char ) e x i t s t a t u s ; p r i n t f ( "I got %s back from the thread .\ n" , t h r e a d r e s u l t ) ; free ( exit status );

If the thread function allocated the space for the return value dynamically then it is essential for the higher level thread to free that space when it no longer needs the return value. If this isnt done the program will leak memory. Exercises 1. Write a program that computes the square roots of the integers from 0 to 99 in a separate thread and returns an array of doubles containing the results. In the meantime the main thread should display a short message to the user and then display the results of the computation when they are ready. 2. Imagine that the computations done by the program above were much more time consuming than merely calculating a few square roots. Imagine also that displaying the short message was also fairly time consuming. For example, perhaps the message needed to be fetched from a network 7

server as HTML and then rendered. Would you expect the multithreaded program to perform better than a single threaded program that, for example, did the calculations rst and then fetched the message? Explain.

Thread Synchronization

In order to eectively work together the threads in a program usually need to share information or coordinate their activity. Many ways to do this have been devised and such techniques usually go under the name of thread synchronization. In this section I will outline several common methods of thread synchronization and show how they can be done using Posix threads.

3.1

Mutual Exclusion

When writing multithreaded programs it is frequently necessary to enforce mutually exclusive access to a shared data object. This is done with mutex objects. The idea is to associate a mutex with each shared data object and then require that every thread that wishes to use the shared data object rst lock the mutex before doing so. Here are the particulars 1. Declare an object of type pthread mutex t. 2. Initialize the object by calling pthread mutex init(). 3. Call pthread mutex lock() to gain exclusive access to the shared data object. 4. Call pthread mutex unlock() to release the exclusive access and allow another thread to use the shared data object. 5. Get rid of the object by calling pthread mutex destroy(). The program of Listing 2 demonstrates the basic approach. It is important to understand that if a thread attempts to lock the mutex while some other thread has it locked, the second thread is blocked until the rst releases the mutex with pthread mutex unlock(). Be sure to observe these points 1. No thread should attempt to lock or unlock a mutex that has not been initialized. 2. The thread that locks a mutex must be the thread that unlocks it.

Listing 2: Mutex Example

#include < p t h r e a d . h> #include < u n i s t d . h> pthread mutex t lock ; int shared data ; // Often shared data is more complex than just an int . void t h r e a d f u n c t i o n ( void a r g ) { int i ; f o r ( i = 0 ; i < 1 0 2 4 1 0 2 4 ; ++i ) { // Access the shared data here . p t h r e a d m u t e x l o c k (& l o c k ) ; s h a r e d d a t a ++; p t h r e a d m u t e x u n l o c k (& l o c k ) ; } return NULL; } i n t main ( void ) { pthread t thread ID ; void exit status ; int i; // Initialize the mutex before trying to use it . p t h r e a d m u t e x i n i t (& l o c k , NULL ) ; p t h r e a d c r e a t e (& t h r e a d I D , NULL, t h r e a d f u n c t i o n , NULL ) ; // Try to use the shared data . f o r ( i = 0 ; i < 1 0 ; ++i ) { sleep (1); p t h r e a d m u t e x l o c k (& l o c k ) ; p r i n t f ( "\ rShared integer s value = %d\n" , s h a r e d d a t a ) ; p t h r e a d m u t e x u n l o c k (& l o c k ) ; } p r i n t f ( "\n" ) ; p t h r e a d j o i n ( t h r e a d I D , &e x i t s t a t u s ) ; // Clean up the mutex when we are finished with it . p t h r e a d m u t e x d e s t r o y (& l o c k ) ; return 0 ; }

3. No thread should have the mutex locked when you destroy the mutex. 4. Any mutex that is initialized should eventually be destroyed, but only after any thread that uses it has either terminated or is no longer interesting in using it. In practice it is sometimes the case that threads are blocked on mutex objects when the program wishes to terminate. In such a situation it might make sense to pthread cancel() those threads before destroying the mutex objects they are blocked on. Coordinating this properly can be tricky, however. Notice that it is possible to assign special mutex attributes to a mutex object when it is created. This is done by creating a mutex attributes object, assigning attributes to the object, and then passing a pointer to the attributes object into pthread mutex init(). The program in Listing 2 just calls for default attributes by providing a NULL pointer instead. In many cases this is perfectly adequate. The use of mutex attribute objects is beyond the scope of this short document. Exercises 1. Enter the program in Listing 2 and try it out. Does it behave the way you expected? Try dierent values for the maximum loop index in the thread function and dierent sleep times in the main function. Try removing the call to sleep() entirely. Try the program on dierent machines. Can you explain what is happening? 2. Suppose you are building a C++ string class that you intend to use in a multithreaded program. You are worried about your string objects possibly getting corrupted if they are updated by more than one thread at a time. You consider adding a mutex as a member of each string and locking that mutex whenever any string method is called. Discuss the implications of this design. Be careful: this question is considerably trickier than it may appear!

3.2

Condition Variables

If you want one thread to signal an event to another thread, you need to use condition variables. The idea is that one thread waits until a certain condition is true. First it tests the condition and, if it is not yet true, calls pthread cond wait() to block until it is. At some later time another thread makes the condition true and calls pthread cond signal() to unblock the rst thread. Every call to pthread cond wait() should be done as part of a conditional statement. If you arent doing that, then you are most likely using condition variables incorrectly. For example 10

i f ( f l a g == 0 ) p t h r e a d c o n d w a i t ( . . . ) ;

Here Im waiting until the ag is not zero. You can test conditions of any complexity. For example
x = a + b (2 c ) ; i f ( x < 0 | | x > 9) pthread cond wait ( . . . ) ;

Here Im waiting until x is in the range from zero to nine inclusive where x is computed in some complex way. Note that pthread cond wait() is only called if the condition is not yet true. If the condition is already true, pthread cond wait() is not called. This is necessary because condition variables do not remember that they have been signaled. If you look at my examples, you will see that there is a serious race condition in them. Suppose the condition is not true. Then suppose that after the condition is tested but before pthread cond wait() is called, the condition becomes true. The fact that the condition is signaled (by some other thread) will be missed by pthread cond wait(). The rst thread will end up waiting on a condition that is already true. If the condition is never signaled again the thread will be stuck waiting forever. To deal with this problem, every time you use a condition variable you must also use a mutex to prevent the race condition. For example:
p t h r e a d m u t e x l o c k (&mutex ) ; i f ( f l a g == 0 ) p t h r e a d c o n d w a i t (& c o n d i t i o n , &mutex ) ; p t h r e a d m u t e x u n l o c k (&mutex ) ;

The thread that signals this condition will use the same mutex to gain exclusive access to the ag. Thus there is no way that the signalling could occur between the test of the ag and the waiting on the condition. For the above to work, pthread cond wait() needs to wait on the condition and unlock the mutex as an atomic action. It does this, but it needs to know which mutex to unlock. Hence the need for the second parameter of pthread cond wait(). When the condition is signaled, pthread cond wait() will lock the mutex again so that the pthread mutex unlock() in the above example is appropriate regardless of which branch of the if is taken. Here is how the signalling thread might look
p t h r e a d m u t e x l o c k (&mutex ) ; flag = 1; p t h r e a d m u t e x u n l o c k (&mutex ) ; p t h r e a d c o n d s i g n a l (& c o n d i t i o n ) ;

Before setting the ag, and thus making the condition true, the signalling thread locks the mutex to make sure the waiting thread cant get caught in a race condition. 11

There is a futher subtlty regarding the use of condition variables. Under certain conditions the wait function might return even though the condition variable has not actually been signaled. For example, if the Unix process in general receives a signal, the thread blocked in pthread cond wait() might be elected to process the signal handling function. If system calls are not restarting (the default in many cases) the pthread cond wait() call might return with an interrupted system call error code1 . This has nothing to do with the state of the condition so proceeding as if the condition is true would be inappropriate. The solution to this problem is to simply retest the condition after pthreadcond wait() returns. This is most easily done using a while loop. For example
p t h r e a d m u t e x l o c k (&mutex ) ; while ( f l a g == 0 ) p t h r e a d c o n d w a i t (& c o n d i t i o n , &mutex ) ; p t h r e a d m u t e x u n l o c k (&mutex ) ;

Of course this assumes you want to ignore any spurious returns from the wait function. In a more complex application you might want to process the error codes in various ways depending on the situation. The pthread cond signal function releases only one thread at a time. In some cases it is desirable to release all threads waiting on a condition. This can be accomplished using pthread cond broadcast. For example
p t h r e a d m u t e x l o c k (&mutex ) ; flag = 1; p t h r e a d m u t e x u n l o c k (&mutex ) ; p t h r e a d c o n d b r o a d c a s t (& c o n d i t i o n ) ;

Exercises 1. The code fragments above do not constitute a complete program. Write a complete program that illustrates how condition variables work. Create a thread that waits on a condition. In the main function make the condition true. Print messages and add delays (or wait for user input) at various places so you can verify that the thread is, in fact, waiting for the condition as appropriate. Verify that the thread does not wait if the condition is already true when it is rst tested. 2. In the text above, when a condition is signaled the signalling thread releases the mutex before calling pthread cond signal. However, it is also possible swap those operations as shown below.
p t h r e a d m u t e x l o c k (&mutex ) ; flag = 1; p t h r e a d c o n d s i g n a l (& c o n d i t i o n ) ; p t h r e a d m u t e x u n l o c k (&mutex ) ;
1 Of course this assumes you are dealing with an actual kernel thread. If the thread is purely a user mode thread such unexpected returns wont occur.

12

Does this result in the same behavior as before? Are any race conditions introduced (or xed) by this change? How does this approach impact application performance?

3.3

Semaphores

A semaphore are essentially gloried integer counters. They support two primary operations. One operation, called down or wait, attempts to decrement the counter. The other operation, called up or signal, attempts to increment the counter. What makes semaphores special is that if a thread tries to wait on a zero semaphore it is blocked instead. Later when another thread signals the semaphore the blocked thread is activated while the semaphore remains at zero. In eect, the signalling causes the semaphore to be incremented but then the thread that was blocked trying to do a wait is allowed to proceed, causing the semaphore to be immediately decremented again. If multiple threads are blocked waiting on a semaphore then the system chooses one to unblock. Exactly how this choice is made is generally system dependent. You can not assume that it will be in FIFO order2 . However, the order in which the threads are unblocked is not normally a concern. A semaphore with an initial value of one can be used like a mutex. When a thread wishes to enter its critical section and access a shared data structure, it does a wait operation on the semaphore. If no other thread is in its critical section, the semaphore will have its initial value of one and the wait will return immediately. The semaphore will then be zero. If another thread tries to wait on the semaphore during this time it will be blocked. When the rst thread is nished executing its critical section it does a signal operation on the semaphore. This will unblock one waiting thread or, if there are no waiting threads, increment the semaphore back to its initial value of one. A semaphore used in this way is called a binary semaphore because it has exactly two states. However, because semaphores are integers they can take on values larger than one. Thus they are often used to count scarce resources. For example a thread might wait on a semaphore to eective reserve one item of a resource. If there are no items left, the semaphore will be zero and the wait operation will block. When a thread is nished using an item of a resource it signals the semaphore to either increment the count of available items or to allow a blocked thread to access the now free item. A semaphore used in this way is called a counting semaphore. The Posix semaphore API is not really part of the normal pthread API. Instead Posix standardizes semaphores under a dierent API. Traditional Unix systems
2 If threads have dierent priorities, normally the highest priority thread is allowed to go rst.

13

support shared memory, message queues, and semaphores as part of what is called System V Interprocess Communication (System V IPC). Posix also provides shared memory, message queues, and semaphores as a package that competes with, or replaces, the older standard. The functionality of the two systems is similar although the details of the two APIs are dierent. Note that Posix semaphores, like System V IPC semaphores, can be used to synchronize two or more separate processes. This is dierent than pthread mutexes. A mutex can only be used by threads in the same process. Because Posix semaphores can be used for interprocess communication, they have the option of being named. One process can create a semaphore under a particular name and other processes can open that semaphore by name. In this tutorial, however, I will focus only on sychronizing threads in the same process. Thus here I will only go over what that requires and skip matters of semaphore naming. The skeleton program in Listing 3 shows how to initialize, clean up, and use a Posix semaphore. For brevity the skeleton program does not show the threads being created or joined nor does it show any error handling. See the manual pages for the various functions for more information on error returns. Another dierence between a pthread mutex and a semaphore is that, unlike a mutex, a semaphore can be signaled in a dierent thread than the thread that does the wait operation. This is necessary when using a semaphore to count instances of a scarce resource. The skeleton program in Listing 3 is using a semaphore like a mutex. I did this to simplify the listing so that the functions used to manipulate the semaphore would be clear. In a later version of this tutorial I may show an example of how to use a semaphore to count scarce resources. Exercises 1. I need something here.

3.4

Reader/Writer Locks

Blah, blah, blah. . . Exercises 1. I need something here.

14

Listing 3: Semaphore Example

#include <semaphore . h> int shared ; s em t b i n a r y s e m ;

// Used like a mutex .

void t h r e a d f u n c t i o n ( void a r g ) { // ... s e m w a i t (& b i n a r y s e m ) ; // Used shared resource . s e m s i g n a l (& b i n a r y s e m ) ; // ... } void main ( void ) { // ... s e m i n i t (& b inary s em , 1 ) ; // Give semaphore initial count . // Start threads here . // ... s e m w a i t (& b i n a r y s e m ) ; // Use shared resource . s e m s i g n a l (& b i n a r y s e m ) ; // ... // Join with threads here . s e m d e s t r o y (& b i n a r y s e m ) ; // ... return 0 ; } // Decrements count . // Increments count .

15

Thead Models

In this section I will describe some ways that threads can be used in real programs. The goal is to give you a feeling for the kind of design ideas that lend themselves to a threaded solution. It is usually possible to build a single threaded program that solves the same problem, but in some cases the single threaded solution is awkward and dicult to manage. Be aware, however, that single threaded solutions are often the most appropriate. There can be a signicant amount of overhead associated with synchronizing threads; multiple threads, poorly applied, can easily result in a slower and more confusing program.

4.1

Boss/Worker Model

Here the idea is to have a single boss thread that creates work and several worker threads that process the work. Typically the boss thread creates a certain number of workers right awayeven before any work has arrived. The worker threads form a thread pool and are all programed to block immediately. When the boss thread generates some work, it arranges to have one worker thread unblock to handle it. Should all workers be busy the boss thread might 1. Queue up the work to be handled later as soon as a worker is free. 2. Create more worker threads. 3. Block until a worker is free to take the new work. If no work has arrived recently and there are an excessive number of worker threads in the thread pool, the boss thread might terminate a few of them to recover resources. In any case, since creating and terminate threads is relatively expensive (compared to, say, blocking on a mutex) it is generally better to avoid creating a thread for each unit of work produced. You have already seen this model in action many times. Consider a bank. When you arrive you have work that needs doing. You get in a queue and wait for a free teller (worker thread). When a teller is available that teller handles your work while other tellers are handling other work at the same time. Should someone in line have an unusually complicated transaction, it wont hold up the line. Only one teller will be tied up dealing with the large work item. The other tellers will be available to handle other peoples work normally. Thus the response time is reasonable even when some work items are very time consuming. A web server is another excellent example of where the boss/worker model can be used. The boss thread listens to the network for incoming connections. When a connection is made, the boss thread directs a worker thread to handle that 16

connection. The boss thread then returns to listening to the network again. In this way many connections can be handled at once. If a particularly time consuming connection is active, it wont prevent the program for dealing with other connections as well. This model works the best when the work items are independent of each other. If the work items depend on each other or have to be processed in a particular sequence the worker threads have to talk to each other and the overall eciency of this model is much reduced. Also, if you run a boss/worker program on a single processor machine it is important that servicing a work item involves a fair amount of blocking. If the work items are all 100% CPU bound then there wont be any performance enhancement. A single thread servicing all the items in sequence would be just as fast as having multiple threads servicing several items at once. However, if servicing an item requires a lot of blocking, then another thread can use the CPU while the rst is blocked and the overall performance is better (often drastically so).

4.2

Pipeline Model

Many programs take some input, transform it in several stages, and then output the result. Instead of having a single thread perform each step in sequence you could have a separate thread handling each stage. The result is much like an assembly line. The data ows from one worker to another and each worker performs their particular operation on the data. By the time the data reaches the end of the line it has been fully transformed into the desired output. Usually writing a single threaded program to process sequential data in stages is fairly straightforward. However, a multithreaded pipeline has the potential to outperform the single threaded solution. In general, if there are N stages to the pipeline there can be N items being operated on at once by the multithreaded program and the result will be N times faster. In practice it rarely works out this well. To obtain its full eciency the time required for every stage must be identical and the processing of one stage cant in any way slow down the processing of the others. If the program runs on a single processor the operations being done in each stage must block frequently so that the CPU can excute another stage while the blocked stages are waiting. To balance the load between the stages, the programmer might need to use proling tools to nd out which stages are taking shorter or longer amounts of time. Stages that are very short can be combined with the stage on either side while stages that are very long can be split into multiple stages (ideally with blocking operations divided evenly between the stages). Getting this balance just right is dicult yet without it the multithreaded solution to the pipeline model will hardly be any faster than the single threaded solution. In fact, because of locking overhead, it may even be slower3 .
3 The

buers between the stages must be careful to avoid race conditions and overow/un-

17

4.3

Background Task Model

Many programs have tasks that they would like to complete in the background. For example a program might want to backup its data les every 10 minutes or update a special status window every 5 seconds. It is awkward to program such things with only a single thread. The program must remember to check the time regularly and call the background function whenever an appropriate amount of time has elapsed. Since that might happen at any point in the programs execution, the programs logic must be littered with calls to functions that are largely unrelated to the main ow of the program. With multiple threads, however, this model is quite easy to program. A background thread can be created when the program initializes itself. That thread can sleep for a certain xed time and then, when it wakes up, perform the necessary background operation. The thread would then just loop back and sleep again until the next time interval has expired. This can happen independently of the main programs logic. The main complication involved with programming this model is in making sure that the background access is properly synchronized with the main access. In this approach the threads are used in a manner similar to the way interrupt service routines are used. They provide background services that the main program does not have to explicity invoke. Many useful tasks can be eectively handled in this way.

4.4

Interface/Implementation Model

Most graphical environments are event driven. Each action taken by the user is a separate event that the program must handle. Examples of events include mouse clicks, menu selections, keystrokes, and so forth. Typically the program contains a single function that is called by the system whenever an event happens. That function must process the event and then return before the system calls the function with the next event. If the event handling function does not return quickly enough events will back up and the user will precieve the program as unresponsive and sluggish. In an extreme case the program will even appear to be dead. To avoid this the program can use multiple threads. If handling an event is going to be time consuming and dicult (and involve a lot of blocking), the event handling function can just create a thread to deal with it and then return at once. This gives the event handling function the opportunity to handle additional events while past events are being processed by other threads. The result is that the programs interface remains responsive even if some of the operations requested are very time consuming.
derow conditions. This involves a signicant amount of locking activity.

18

It is not necessary for an entire program to be organized this way. Internal modules in a program can use the same trick. When a function is called in such a module, the function might create a thread to carry out the requested operation and then return at once. The calling program will see the function as very quick and responsive even though the actual work requested hasnt really been completed when the function returns. The diculty with this model is that eventually the user of an interface will usually need know for sure when certain operations requested in the past have actually completed. Some way of coordinating that information must be provided. Also it is dicult for the program to report errors eectively with this model because an error might occur long after the operation was requested and apparently handled. Many operating systems themselves use this model extensively. For example, when a program writes to a le, the le is typically not put onto the disk at once. Instead it is just put into a cache (faster) and written to disk latter when the system is less busy. In eect, the operating system writes to disk in a separate thread that is independent of the thread that actually requested the write operation in the rst place.

4.5

General Comments

In general, multithreaded programs work best if the threads are as independent as possible. The less the threads have to talk with each other the better. Whenever threads have to synchronize or share data there will be locking overhead and time spent waiting for other threads. Time blocked while waiting for another thread is time wasted. Such a thread is not doing anything useful. In contrast, if a thread is blocked waiting for I/O it is doing something that the program needs done. Such blocking is good because it allows other threads to get the CPU. But if a thread waits for another thread then it is not accomplishing anything that the program needs. The more threads interact with each other the more time they will spend waiting for each other and the more inecient the program will be. It is easy to understand this idea when you think about working with another person. If you and your partner can do two largely independent activities you can both work without getting in each others way and you can get twice as much work done. But if you try to work too closely then one of you will simply be waiting for the other and the work wont get done any more quickly than it would by a single person alone. Consider what would happen if you decided to type a paper with your partner but that you and your partner had to alternate keystrokes on the keyboard. To type Hello rst you type H then your partner types e then you type l and so on. Obviously this would be very inecient. You would spend more time getting the alternation right than you would actually typing keys. The exact same issues arise in multithreaded programs. 19

An improperly written multithreaded program can be slowersometimes a lot slowerthan its single threaded equivalent. If two tasks are very independent, they can often be handled by two entirely separate processes. Large software systems are often composed of many executable les, each taking care of a single aspect of the system. At this level the system is multithreaded even if the individual programs are not. However, multiple processes can nd it dicult to share information eectively. Putting multiple threads into a single process makes the parallelism more ne grained and allows the threads to interact more closely and share more resources. This can be a good thing. But if taken to extreme it causes ineciencies as I described above. A good multithreaded program will strike the right balance between sharing and independence. That balance is often dicult to nd.

Thread Safety

Typically when a complicated object is operated on, it goes through several intermediate, invalid states before the operation is complete. As an analogy consider what a surgeon does when he operates on a person. Although the purpose of the operation is to increase the patients health, the surgeon performs several steps that would greatly decrease the patients health if they were left incomplete. Imagine what would happen if a surgeon cut open a patients chest and then decided to go on vacation for three weeks! Similarly a function that operates on an object will often temporary put that object into an unusable state while it performs the update. Once the update is complete, the function (if written correctly) will leave the object is a fully functional state again. Should another thread try to use an object while it is in an unusable state (often called an inconsistent state) the object will not respond properly and the result will be undened. Keeping this from happening is the essential problem of thread safety. The problem doesnt come up in a single threaded program because there is no possibility of another thread trying to access the object while the rst thread is updating it. (Unless exceptions are a possibility. In that case the updating thread might abort the update and then later try to access the incompletely updated object. This causes pretty much the same sort of problems to occur).

5.1

Levels of Thread Safety

One of the problems people have discussing thread safety is that there are many dierent levels of safety one might want to talk about. Just saying that a piece of code is thread safe doesnt really say all that much. Yet most people have certain natural expectations about thread safety. Sometimes those expectations

20

are reasonable and valid, but sometimes they are not. Here are some of those expectations. Reading an objects value with multiple threads is not normally expected to be a problem. Problems only occur when an object is updated since it is only then that it has to be modied and run the risk of entering inconsistent states. However some objects have internal state that gets modied even when its value is read (think about an object that has an internal cache). If two threads try to read such an object there might be problems unless the read operations on that object have been designed to handle the multiple threads properly. Updating two independent objects, even of the same type, is not normally expected to be a problem. It is usually assumed that objects that appear to be independent are, in fact, independent and thus the inconsistent states of one such object have no impact on the other. However some objects share information behind the scenes (static class data, global data, etc) that causes them to be linked internally even when they do not appear to be linked from a logical point of view. In that case, modifying two independent objects might cause a problem anyway. Consider:
void f ( ) { std : : s t r i n g x ; // Modify x. } } void g ( ) { std : : s t r i n g y ; // Modify y.

If one thread is in function f() modifying the string x and another is in function g() moding string y, will there be a problem? Most of the time you can assume that the two apparently independent objects can be simultaneously modied without a problem. But it is possible, depending on how std::string is implemented, that the two objects share some data internally and that simultaneous modications will cause a problem. In fact, even if one of the functions merely reads the value of the string, there might be a problem if they share internal data that is being updated by the other function. Functions that acquire resources, even if from a common pool of resources, are not normally exepected to be a problem. Consider:
void f ( ) { char p = new char [ 5 1 2 ] ; // Use the array p. } } void g ( ) { char p = new char [ 5 1 2 ] ; // Use the array p.

21

If one thread is in function f() and another thread is in function g(), both threads might try to allocate memory simultaneously by invoking the new operator. In a multi-threaded environment, it is safe to assume that new has been written to work correctly in this case even though both invocations of new are trying to take memory from the same pool of memory. Internally new will synchronize the threads so that each call will return a unique allocation and the internal memory pool will not be corrupted. Similar comments can be made about functions that open les, make network connections, and perform other resource allocation tasks. However if the resource allocation functions are not designed with threads in mind then they may well fail if invoked by multiple threads at once. What people typically expect to cause problems is when a program tries to access (update or read) an object while another thread is updating that same object. Global objects are particularly prone to this problem. Local objects are much less so. For example:
std : : s t r i n g x ; void f ( ) { std : : s t r i n g y ; // Modify x and y. }

If two threads enter function f() at the same time, they will get dierent versions of the string y. This is because every thread has its own stack and local objects are allocated on the threads stack. Thus every thread has its own, independent copy of the local objects. As a result, manipulating y inside f() will not cause a problem (assuming that manipulating independent objects is safe). However, since there is only one copy of the global x that both threads will be touching, there could be a problem caused by those operations. Local objects are not immune to problems since any function can start a new thread and pass a pointer to a local object as a parameter to that thread. For example
void f ( ) { std : : s t r i n g x ; s t a r t t h r e a d ( s o m e f u n c t i o n , &x ) ; s t a r t t h r e a d ( s o m e f u n c t i o n , &x ) ; }

Here I assume there is a library function named start thread() that accepts a pointer to a thread function (dened elsewhere) and a pointer to a parameter 22

for that function. In this case I start two threads executing some function(), giving both of them a pointer to the string x. If some function() tries to modify that string then two threads will be modifying the same object and problems are likely. Note that this case is particularly insidious because some function() has no particular reason to expect that it will be given the same parameter twice. Thus it is unlikely to have any protection to handle such a case.

5.2

Writing Thread Safe Code

In theory the only way to control the actions of a thread is to use synchronization primitives such as mutexes or semaphores. In languages that provide threads as a part of the language, synchronization primitives of some kind are normally provided by the language itself. In other cases, such as with C, they are library functions, such as the Posix API, that interact with the operating system. Normally you should write code that meets the usual expectations that people have about thread safe code. If you are implementing a class, make sure that multiple simultaneous reads on an object are safe. If you do update internal data behind the callers back, you will probably have to protect those updates yourself. Also make sure the simultaneous writes to independent objects are safe. If you do make use of shared data, you will probably have to protect updates to that shared data yourself. If you write a function that manages a shared resource for multiple threads from a common pool, you will probably have to protect that shared resource from corruption by multiple, simultaneous requests. However, in general, you probably dont have to bother protecting every single object against simultaneous updates. Let the calling program worry about that case. Such total safety is usually very expensive in terms of runtime eciency and not normally necessary or even appropriate.

5.3

Exception Safety vs Thread Safety

Both thread and exception safety share a number of common issues. Both are concerned with objects that are in an inconsistent state. Both have to think about resources (although in dierent ways... exception safety is concerned with resource leaks, thread safety is concerned with resource corruption). Both have several levels of safety that could be dened along with some common expectations about what is and is not safe. However, there is one important dierence between the exception safety and thread safety. Exceptions occur synchronously with the programs execution while threads are asynchronous. In other words, exceptions occur, in theory, only at certain well dened times. Although it is not always clear which operations might throw an exception and which might not, in theory it is possible to precisely dene exactly when an exception might happen and when it cant

23

happen. As a result it is often possible to make a function exception safe by just reorganizing it. In contrast there is no way to control when two threads might clash. Reorganzing a function is rarely helpful when it comes to making it thread safe. This dierence makes thread related errors very dicult to reproduce and especially hard to manage.

Rules for Multithreaded Programming

In this section Ill try to summarize a few rules of thumb that one should keep in mind when building a multithreaded program. Although using multiple threads can provide elegant and natural solutions to some programming problems, they can also introduce race conditions and other subtle, dicult to debug problems. Many of these problems can be avoided by following a few simple rules.

6.1

Shared Data

As I described in Section 3, when two threads try to access the same data object there is a potential for problems. Normally modifying an object requires several steps. While those steps are being carried out the object is typically not in a well formed state. If another thread tries to access the object during that time, it will likely get corrupt information. The entire program might have undened behavior afterwards. This must be avoided. 6.1.1 What data is shared?

1. Static duration data (data that lasts as long as the program does). This includes global data and static local data. The case of static local data is only signicant if two (or more) threads execute the function containing the static local at the same time. 2. Dynamically allocated data that has had its address put into a static variable. For example, if a function uses malloc() or new to allocate an object and then places the address of that object into a variable that is accessible by more than one thread, the dynamically allocated object will then be accessible by more than one thread. 3. The data members of a class object that has two (or more) of its member functions called by dierent threads at the same time.

24

6.1.2

What data is not shared?

1. Local variables. Even if two threads call the same function they will have dierent copies of the local variables in that function. This is because the local variables are kept on the stack and every thread has its own stack. 2. Function parameters. In languages like C the parameters to functions are also put on the stack and thus every thread will have its own copy of those as well. Since local variables and function parameters cant be shared they are immune to race conditions. Thus you should use local variables and function parameters whenever possible. Avoid using global data. 6.1.3 What type of simultaneous access causes a problem?

1. Whenever one thread tries to update an object, no other threads should be allowed to touch the object (for either reading or writing). Mutual exclusion should be enforced with some sort of mutex-like object (or by some other suitable means). 6.1.4 What type of simultaneous access is safe?

1. If multiple threads only read the value of an object, there should be no problem. Be aware, however, that complicated objects often update internal information even when, from the outside, they are only being read. Some objects maintain a cache or keep track of usage statistics internally even for reads. Simultaneous reads on such an object might not be safe. 2. If one thread writes to one object while another thread touches a totally independent object, there should be no problem. Be aware, however, that many functions and objects do share some data internally. What appears to be two separate objects might really be using a shared data structure behind the scenes. 3. Certain types of objects are updated in an uninterruptable way. Thus simultaneous reads and writes to such objects are safe because it is impossible for the object to be in an inconsistent state during the update. Such updates are said to be atomic. The bad news is that the types that support atomic updates are usually very simple (for example: int) and there is no good way to know for sure exactly which types they are. The C standard provides the type sig atomic t for this purpose. It is dened in <signal.h> and is some kind of integer. Simultaneous updates to an object declared to be volatile sig atomic t are safe. Mutexes are not necessary in this case.

25

6.2

What can I count on?

Unless a function is specically documented as being thread-safe, you should assume that it is not. Many libraries make extensive use of static data internally and unless those libraries were designed with multiple threads in mind that static data is probably not being properly protected with mutexes. Similarly you should regard the member functions of a class as unsafe for multiple threads unless it has been specically documented to be otherwise. It is easy to see that there might be problems if two threads try to manipulate the same object. However, even if two threads try to manipulate dierent objects there could still be problems. For various reasons, many classes use internal static data or try to share implementation details among objects that appear to be distinct from the outside. You can count on the following: 1. The API functions of the operating system are thread-safe. 2. The Posix thread standard requires that most functions in the C standard library be thread-safe. There are a few exceptions which are documented as part of the standard. 3. Under Windows the C standard library is totally thread safe provided you use the correct version of the library and you initialize it properly. 4. The thread safety of the C++ standard library is vauge and very much dependent on the compiler you are using. The SGI criteria for thread safety of the standard template library is gaining ground as the defacto standard. It is not universal. If you use a non thread-safe function in one of your functions, your function will be rendered non thread-safe as well. However, you are free to use a non thread-safe function in a multithreaded program provided it is never called by two or more threads at the same time. You can either arrange to use such functions in only one thread or protect calls to such functions with mutexes. Keep in mind that many functions share data internally with other functions. If you try to protect calls to a non thread-safe function with a mutex you must also protect calls to all the other related functions with the same mutex. This is often dicult.

26