Sharing
Data Between Threads
A key advantage of multithreaded codes is that all threads see the same
memory, so data is already shared between threads. However, it often important
to coordinate access to this data, since failure to coordinate accesses could
cause data races that lead to incorrect results. POSIX provides a large number
of synchronization and data-sharing methods.
Protecting
Access Using Mutex Locks
A mutex lock is a mechanism
supported by the POSIX standard that can be acquired by only one thread at a
time. Other threads that attempt to acquire the same mutex must wait until it
is released by the thread that currently has it.
Before they can be used, mutex locks need to be
initialized to the appropriate state by a call to pthread_mutex_init() or, for statically defined mutexes, by assignment with the value PTHREAD_MUTEX_INITIALIZER. The call to pthread_mutex_init() takes an
optional parameter that points to attributes describing the type of mutex
required. Initialization through static assignment uses default parameters, as
does passing in a null pointer in the call to pthread_mutex_init().
Once a mutex is no longer needed, the resources it consumes can be freed
with a call to pthread_mutex_destroy(). Listing 5.19 shows examples of initializing and destroy-ing mutexes.
Listing 5.19 Creating
and Destroying Mutexes
#include
<pthread.h>
...
pthread_mutex_t
m1 = PTHREAD_MUTEX_INITIALIZER; pthread_mutex_t m2;
pthread_mutex_init(
&m2, 0 );
...
pthread_mutex_destroy(
&m1 );
pthread_mutex_destroy(
&m2 );
A thread can lock a mutex by calling pthread_mutex_lock(). Once it has finished with the mutex, the thread calls pthread_mutex_unlock(). If a thread calls pthread_ mutex_lock()
while another thread holds the mutex, the calling
thread will wait, or block, until the other thread releases
the mutex, allowing the calling thread to attempt to acquire the released mutex.
In many situations, it is not desirable for the
calling thread to wait for the mutex to be available. The call pthread_mutex_trylock() will attempt to acquire the mutex. If it succeeds, the function will
return the value of zero, and the calling thread will now be the owner of the
mutex. If the mutex is already locked by another thread, the function will
immediately return a nonzero value indicating the exact situation.
The code shown in Listing 5.20 shows a mutex lock
protecting the variable count against
simultaneous access by multiple threads. The variable count is declared as volatile to ensure
that it is read from memory on each access and written back to
memory after each access. Without the mutex lock,
there would be a data race between the two threads. Hence, it is very unlikely
that count would
end up with the correct value.
Listing 5.20 Mutex
Lock Avoiding Data Race
#include <pthread.h> #include <stdio.h>
pthread_mutex_t mutex; volatile int counter = 0;
void
* count( void * param )
{
for ( int i=0; i<100; i++ )
{
pthread_mutex_lock(
&mutex );
counter++;
printf(
"Count = %i\n", counter );
pthread_mutex_unlock(
&mutex );
}
}
int main()
{
pthread_t thread1, thread2; pthread_mutex_init( &mutex, 0 );
pthread_create( &thread1, 0, count, 0 ); pthread_create( &thread2, 0,
count, 0 ); pthread_join( thread1, 0 ); pthread_join( thread2, 0 );
pthread_mutex_destroy( &mutex );
return 0;
}
Mutex
Attributes
Mutexes can be shared between multiple processes. By default, mutexes
are private to a process. To create a mutex that can be shared between
processes, it is necessary to set up the attributes for pthread_mutex_init(), as shown in Listing 5.21.
Listing 5.21 Creating
a Mutex That Can Be Shared Between Processes
#include <pthread.h>
int main()
{
pthread_mutexattr_t
attributes;
pthread_mutex_t mutex;
pthread_mutexattr_init(
&attributes );
pthread_mutexattr_setpshared(
&attributes, PTHREAD_PROCESS_SHARED );
pthread_mutex_init( &mutex, &attributes
);
pthread_mutexattr_destroy(
&attributes );
...
}
The attributes structure pthread_mutexattr_t is initialized with default values by a call to pthread_mutexattr_init(). A call to pthread_mutex_setpshared() with a pointer to the attribute structure and the value PTHREAD_PROCESS_SHARED sets the attributes to cause a shared mutex to be created. By default,
mutexes are not shared between processes; calling pthread_mutex_setpshared() with the value PTHREAD_ PROCESS_PRIVATE
restores the attribute to the default.
These attributes are passed into the call to pthread_mutex_init() to set the attrib-utes of the initialized mutex. Once the attributes
have been used, they can be disposed of by a call to pthread_mutex_attr_destroy().
A mutex can have other attributes set using the
same mechanism:
n The type of mutex. This can be a normal mutex, a mutex that detects
errors such as multiple attempts to lock the mutex, or a recursive mutex that
can be locked multiple times and then needs to be unlocked the same number of
times.
n The protocol to follow when another thread is waiting for the mutex.
This can be the default of no change to thread priority, that the thread
holding the mutex inherits the priority of any higher-priority thread waiting
for the mutex, or that the thread gets the highest priority associated with the
mutexes held by it.
n The priority ceiling of the mutex. This is the priority that any
lower-priority thread will be elevated to while it holds the mutex.
The attributes governing the priority of any thread
holding the mutex are designed to avoid problems of priority inversion where a
higher-priority thread is waiting for a lower-priority thread to release the
mutex.
Using
Spin Locks
The critical difference between spin
locks and mutex locks is that a spin lock will keep spinning in a tight
loop and consuming processor resources until it finally acquires the lock.
Mutex locks will immediately put a thread to sleep when it cannot get the
mutex, or an adaptive mutex lock will
spin for a short time waiting for the lock to become free before going to
sleep.
The interface for spin locks is very similar to
that of mutex locks. The call pthread_spin_init() will
initialize a spin lock. The spin lock can be created as share-able between
processes or private to the process creating it. A spin lock that is private to
a process is created by passing the value PTHREAD_PROCESS_PRIVATE as a parameter to the call to pthead_spin_init(), and passing the value PTHREAD_PROCESS_SHARED creates a spin lock that can be shared between processes.
Multiple threads in the process that created the
lock will always be able to access it. However, if the spin lock is created to
be private to a process, the behavior of the lock is not defined if it is used
by other processes. The default is for the spin lock to be private to the
creating process.
The call pthread_spin_lock() will
spin until the lock is acquired, the call pthread_ spin_unlock() will
release the lock, and finally the call
pthread_spin_destroy() will release any resources used
by the lock. Listing 5.22 demonstrates the use of a spin lock. This example
places access to a local variable under the control of the spin lock; however,
in this example, the variable is not shared between threads, so it is not
actually necessary to use any locking.
Listing 5.22 Code Using a Spin Lock to Protect Access to a Variable
#include <pthread.h> pthread_spinlock_t lock;
void lockandunlock()
{
int i = 10000; while ( i>0 )
{
pthread_spin_lock(
&lock ); i--;
pthread_spin_unlock(
&lock );
}
}
int main()
{
pthread_spin_init(
&lock,
PTHREAD_PROCESS_PRIVATE
); lockandunlock();
pthread_spin_destroy(
&lock );
}
If the code is modified to create a spin lock that
is shared between multiple processes, only one process should initialize and
destroy the spin lock. The modification shown in Listing 5.23 will create a
spin lock that can be shared between processes.
Listing 5.23 Creating
a Spin Lock That Can Be Shared Between Processes
int main()
{
pthread_spin_init(&lock,PTHREAD_PROCESS_SHARED);
lockandunlock();
pthread_spin_destroy(&lock);
}
In addition, the call pthread_spin_trylock() will attempt to acquire the lock but will immediately return whether or
not the lock is successfully acquired. Since a spin-ning lock will be using
processor resources, it might be more useful to attempt to acquire the lock
and, if that fails, to complete some other task before repeating the test. This
utilizes the processor in useful work rather than just spinning. Listing 5.24
shows the earlier code modified to use pthread_spin_trylock() and keep a count of the number of times the thread fails to get the
lock.
Listing 5.24 Counting
the Number of Times That the Spin Lock Fails to Be Acquired
void
lockandunlock()
{
int i = 0; int count = 0;
while ( i == 0 )
{
if ( pthread_spin_trylock(
&lock ) )
{
i++;
pthread_spin_unlock(
&lock );
}
else
{
count++;
}
}
printf( "Failed tries = %i\n", count );
}
Read-Write
Locks
Read-write locks allow multiple threads to simultaneously read a resource, but only a
single thread may update that
resource at any time. They share a similar initialization and destruction
syntax to mutex locks in that they take a set of attributes and can be
initial-ized either through a call to pthread_rwlock_init() or statically initialized by assign-ment of the value PTHREAD_RWLOC_INITIALIZER. Listing 5.25 shows the two methods.
Listing 5.25 Initializing
a Read-Write Lock
pthread_rwlock_t lock1, lock2;
...
pthread_rwlock_init(
&lock1, 0 ); lock2 = PTHREAD_RWLOCK_INITIALIZER;
...
If the attributes passed to the initialization
routine are zero, then the lock is initialized with the default attribute of
being private to the creating process. To create a read-write lock that is shared
between processes, it is necessary to create and use a set of attributes.
The call pthread_rwlockattr_init() initializes the attributes, while the call pthread_rwlockattr_setpshared()
sets the shared attribute to the desired value.
This set of attributes can then be passed into the pthread_rwlock_init() call to set the attributes for the read-write lock being created.
Listing 5.26 demonstrates this.
Listing 5.26 Creating a Read-Write Lock That Can Be Shared Between Processes
pthread_rwlockattr_t
attributes; pthread_rwlock_t lock;
...
pthread_rwlockattr_init(
&attr ); pthread_rwlockattr_setpshared( &attr, PTHREAD_PROCESS_SHARED
); pthread_rwlock_init( &lock, &attr );
pthread_rwlockattr_destroy( &attr );
...
pthread_rwlock_destroy( &lock );
Once the lock has been created,
it no longer references the attributes, so these can be either reused for a
different lock or destroyed. The call to destroy the attributes is pthread_rwlockattr_destroy(). The resources consumed by the read-write lock are
freed by the call to pthread_rwlock_destroy().
Read-write locks have a more
complex interface than mutex locks because they can be locked and unlocked for
either reading or writing. Hence, there are two pairs of lock and unlock calls.
The pairs pthread_rwlock_rdlock() and pthread_rwlock_ rdunlock()lock and unlock for reading, and
pthread_rwlock_wrlock() and
pthread_rwlock_wrunlock()lock and unlock for writing.
Listing 5.27 shows how the read-write lock might be used to
protect access to a shared resource.
Listing 5.27 Using
a Read-Write Lock to Protect Access to a Shared Resource
int readMatrix( int x,int y )
{
int result;
pthread_rwlock_rdlock(
&lock );
result = matrix[x,y];
pthread_rwlock_rdunlock(
&lock );
return result;
}
void updateMatrix( int x,int y,int value );
{
pthread_rwlock_wrlock(
&lock );
matrix[x,y] = value;
pthread_rwlock_wrunlock(
&lock );
}
The read-write lock is unnecessary in this short code snippet because
load or store accesses to aligned integer variables are atomic. The lock would
be critical if the updates and reads were of structures that required multiple
writes.
The behavior of calls to acquire the lock is for
the thread to block until the lock is acquired. However, there are also calls
to try to acquire the lock. These calls return immediately either with or
without having acquired the lock. To try to get a reader lock, the call is pthread_rwlock_tryrwlock(), and to try to acquire the lock as a writer, the call is pthread_rwlock_trywrlock(). Listing 5.28 shows an example of using these calls.
Listing 5.28 Updating
a Shared Value Only If the Read-Write Lock Is Acquired
void
typeUpdate( int value )
{
if ( pthread_rwlock_trywrlock(
&lock ) == 0 )
{
count += value;
pthread_rwlock_wrunlock(
&lock );
}
}
There is a further option of using a
timeout when acquiring the lock. The routines pthread_rwlock_timedrdlock() and
pthread_rwlock_timedwrlock() return 0 if the lock is acquired or an error code
if the lock has not been acquired by the absolute time specified. The routines
pthread_rwlock_timedrdlock_np() and pthread_ rwlock_timedwrlock_np() return an
error code if the lock has not be acquired by the relative time specified.
A
timespec structure is used to pass the timing information into the function.
For rel-ative timing, this structure needs to be initialized with the duration
of the wait time; for absolute time, the structure can be initialized with the
current time using a call to clock_gettime(). Listing 5.29 demonstrates how
the timeout can be set to elapse in five seconds.
Listing 5.29 Acquiring
a Read-Write Lock with a Timeout
#include
<time.h>
void
timeout_lock()
{
struct
timespec now;
clock_gettime(
CLOCK_REALTIME, &now
); now.tv_sec += 5;
if ( pthread_rwlock_timedrdlock(
&lock, now ) == 0 )
{
...
pthread_rwlock_rdunlock( &lock );
}
}
Alternatively, the code could directly call the relative timeout, as
shown in Listing 5.30.
Listing 5.30 Setting
a Relative Timeout
#include <time.h>
void timeout_lock()
{
struct
timespec now;
now.tv_sec =
5; now.tv_nsec = 0;
if ( pthread_rwlock_timedrdlock_np(
&lock, now ) == 0 )
{
...
pthread_rwlock_rdunlock( &lock );
}
}
Barriers
There are situations where a program needs to wait until an entire group
of threads has completed its work before further progress can be made. This is
a barrier.
A barrier is created by a call to pthread_barrier_init(). The call to initialize the barrier takes the following:
n A pointer to the barrier to be initialized.
n An optional attributes structure, this structure determines whether the
barrier is private to a process or shared across processes.
n The number of threads that need to reach the barrier before any threads
are released.
The resources consumed by a barrier can be released by a call to pthread_barrier_ destroy().
Each thread calls pthread_barrier_wait() when it reaches the barrier. This call will return when the appropriate
number of threads has reached the barrier. The code in Listing 5.31
demonstrates using a barrier to cause the threads in an application to wait
until all the threads have been created.
Listing 5.31 Creating
and Using a Barrier
#include
<pthread.h>
#include
<stdio.h>
pthread_barrier_t
barrier;
void
* work( void* param)
{
int id=(int)param;
printf( "Thread arrived %i\n", id );
pthread_barrier_wait(
&barrier );
printf( "Thread departed %i\n", id );
}
int
main()
{
pthread_t threads[10];
pthread_barrier_init(
&barrier, 0, 10 );
for ( int i=0; i<10; i++ )
{
pthread_create( &threads[i], 0, work, (void*)i );
}
for ( int i=0; i<10; i++ )
{
pthread_join( threads[i], 0 );
}
pthread_barrier_destroy(
&barrier );
}
The output from the program would show all the
threads arriving at the barrier and then all the threads departing from the
barrier. Without the barrier, the arrivals and departures of all the threads
would be mixed.
Semaphores
A semaphore is a counting and
signaling mechanism. One use for it is to allow threads access to a specified
number of items. If there is a single item, then a semaphore is essen-tially
the same as a mutex, but it is more commonly useful in a situation where there
are multiple items to be managed. Semaphores can also be used to signal between
threads or processes, for example, to tell another thread that there is data
present in a queue. There are two types of semaphores: named and unnamed
semaphores.
An unnamed semaphore is initialized with a
call to sem_init(). This
function takes three parameters. The first parameter is a pointer to the
semaphore. The next is an inte-ger to indicate whether the semaphore is shared
between multiple processes or private to a single process. The final parameter
is the value with which to initialize the semaphore. A semaphore created by a
call to sem_init() is
destroyed with a call to sem_destroy().
The code shown in Listing 5.32 initializes a semaphore with a count of
10. The mid-dle parameter of the call to sem_init() is zero, and this makes the semaphore private to the thread; passing
the value one rather than zero would enable the semaphore to be shared between
multiple processes.
Listing 5.32 Creating and Initializing a Semaphore
#include <semaphore.h>
int main()
{
sem_t
semaphore;
sem_init(
&semaphore, 0, 10 );
...
sem_destroy(
&semaphore );
}
A named
semaphore is opened rather than initialized. The process for doing this is
similar to opening a file. The call to sem_open() returns a pointer to a semaphore. The first parameter to the call is
the name of the semaphore. The name must conform to the naming conventions for
files on the operating system and must start with a single slash sign and
contain no further slash signs. The next parameter is the set of flags. There
are three combinations of flags that can be passed to the sem_open() call. If no flags are passed, the function will return a pointer to the
existing named semaphore if it exists and if the semaphore has the appropriate
permissions to be shared with the calling process. If the O_CREAT flag is passed, the semaphore will be created; if it does not exist or
if it does exist, a pointer will be returned to the existing version. The flag O_EXCL can be passed with the O_CREAT flag.
This will successfully return a semaphore only if that semaphore does not
already exist.
Creating a semaphore requires two
additional parameters: the permissions that the semaphore should be created
with and the initial value for the semaphore. Listing 5.33 shows an example of
opening a semaphore with an initial value of 10, with read and write
permissions for the user, the group, and all users.
Listing 5.33 Opening
a Named Semaphore
#include <semaphore.h>
int main()
{
sem_t *
semaphore;
semaphore = sem_open(
"/my_semaphore", O_CREAT, 0777, 10 );
...
Listing 5.34 Closing
and Unlinking a Named Semaphore
sem_close(
semaphore );
sem_unlink(
"/my_semaphore" );
}
The semaphore is used through a combination of
three methods. The function sem_wait() will
attempt to decrement the semaphore. If the semaphore is already zero,
the calling thread will wait until the semaphore
becomes nonzero and then return, hav-ing decremented the semaphore. The call sem_trywait() will return immediately either having decremented the semaphore or if
the semaphore is already zero. The call to sem_post() will increment the semaphore. One more call,
sem_getvalue(), will write
the current value of the semaphore into an integer
variable. The code in Listing 5.35 shows a semaphore used in the same way as a
mutex might be, to protect the increment of the variable count. On Solaris, the semaphore functions are defined in the real-time
library, so code needs to be linked with this library using -lrt.
Listing 5.35 Using
a Semaphore as a Mutex
int
main()
{
sem_t
semaphore;
int count = 0;
sem_init(
&semaphore, 0, 1 ); sem_wait( &semaphore );
count++;
sem_post(
&semaphore ); sem_destroy( &semaphore );
}
Another property of semaphores that is not fully
exploited when using them as mutex locks is signaling
between threads. Semaphores can be used to signal that one task has been
completed or to ensure that two tasks will be executed in a predetermined
order. Consider the code shown in Listing 5.36.
Listing 5.36 Two
Threads Executing Two Functions in a Nondeterministic Order
#include <pthread.h> #include <stdio.h>
void
*func1( void * param )
{
printf( "Thread 1\n" );
}
void *func2( void * param )
{
printf( "Thread 2\n" );
}
int main()
{
pthread_t threads[2];
pthread_create( &threads[0], 0, func1, 0 ); pthread_create(
&threads[1], 0, func2, 0 ); pthread_join( threads[0], 0 ); pthread_join(
threads[1], 0 );
}
At runtime, the code can print either "Thread
1" or "Thread
2" first, depending on which thread
gets to the printf()
statement first. Semaphores can be used to ensure that the threads execute in a
specific order. Suppose we want to ensure that the output is always "Thread
1" before "Thread
2"; then we need to make the second
thread wait until the first thread completes before the second thread produces
its output. Listing 5.37 shows how a semaphore can be used to ensure this
ordering.
Listing 5.37 Using
a Semaphore to Enforce a Deterministic Ordering on Two Threads
#include <pthread.h>
#include <stdio.h>
#include <semaphore.h>
sem_t
semaphore;
void *func1( void * param )
{
printf( "Thread 1\n" );
sem_post(
&semaphore );
}
void *func2( void * param )
{
sem_wait(
&semaphore ); printf( "Thread 2\n" );
}
int main()
{
pthread_t threads[2];
sem_init(
&semaphore, 0, 1 );
pthread_create( &threads[0], 0, func1, 0 );
pthread_create( &threads[1], 0, func2, 0 );
pthread_join( threads[0], 0 );
pthread_join( threads[1], 0 );
sem_destroy(
&semaphore );
}
The code creates a semaphore. Once the first thread
completes its task, it signals the semaphore that the second thread can now
perform its task. The logic of the second thread will cause it to wait at the
semaphore until the first thread signals it, or if it does not reach the
semaphore before the first thread completes its task, the second will not even
wait at the semaphore. This use of a single semaphore ensures that the second
thread always executes the printf()
statement after the first thread has completed its printf()
statement.
An extension of this ordering mechanism is the
producer-consumer configuration of threads, as shown in Listing 5.38. The
semaphore in this instance contains the number of items waiting in the queue to
be processed. If there are no items in the queue, the con-sumer thread can
sleep until an item is placed in the queue by the producer. The code uses the
semaphore as a signaling mechanism between the two threads, not as a mecha-nism
that ensures mutual exclusion to accesses to the queue. The code that
manipulates the queue is omitted, but this code would ensure that multiple
threads can safely simul-taneously access the queue data structure.
Listing 5.38 Using
a Semaphore in a Producer-Consumer System
#include
<pthread.h>
#include
<stdio.h>
#include
<semaphore.h>
sem_t semaphore;
pthread_mutex_t
mutex = PTHREAD_MUTEX_INITIALIZER;
int queue[200]; int queueLength;
void
*producer( void * param)
{
for ( int i=0; i<500; i++ )
{
// Add item to queue pthread_mutex_lock( &mutex
); queue[ queueLength++ ] = i; pthread_mutex_unlock( &mutex ); // Signal
semaphore
sem_post(
&semaphore );
}
}
void *consumer(void * param)
{
for ( int i=0; i<500; i++ )
{
int item;
// Wait if nothing in queue
if (queueLength==0) { sem_wait(&semaphore);
}
pthread_mutex_lock( &mutex ); item = queue[
--queueLength ]; pthread_mutex_unlock( &mutex ); printf( "Received
%i\n", item);
}
}
int main()
{
pthread_t threads[2];
sem_init(
&semaphore, 0, 0 );
pthread_create( &threads[0], 0, producer, 0 );
pthread_create( &threads[1], 0, consumer, 0 ); pthread_join( threads[0], 0
);
pthread_join( threads[1], 0 );
sem_destroy(
&semaphore );
}
Controlling access to a finite number of elements
is another situation where a sema-phore is useful. This could be a real physical
constraint, such as only sufficient spaces in a list exist or only a finite
amount of memory has been reserved. Or it could be a throt-tling feature. For
example, in the producer-consumer, we might want to limit the queue length to
avoid stacking up too much work for the consumer. Listing 5.39 shows the
modified version of the code.
Listing 5.39 Producer-Consumer
Modified So That the Producer Thread Can Be Throttled
#include <pthread.h>
#include <stdio.h>
#include <semaphore.h>
sem_t
semaphore;
sem_t limit;
pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
int queue[10]; int queueLength;
void
*producer( void * param)
{
for ( int i=0; i<500; i++ )
{
// Wait for space
sem_wait(
&limit );
// Add item to queue pthread_mutex_lock( &mutex
); queue[ queueLength++ ] = i; pthread_mutex_unlock( &mutex ); // Signal
semaphore
sem_post(
&semaphore );
}
}
void
*consumer(void * param)
{
for ( int i=0; i<500; i++ )
{
int item;
// Wait if nothing in queue
if (queueLength==0) { sem_wait(&semaphore);
}
pthread_mutex_lock( &mutex ); item = queue[
--queueLength ]; pthread_mutex_unlock( &mutex ); printf( "Received
%i\n", item);
sem_post(
&limit );
}
}
int
main()
{
pthread_t threads[2];
sem_init(
&semaphore, 0, 0 );
sem_init(
&limit, 0, 10 );
pthread_create( &threads[0], 0, producer, 0 );
pthread_create( &threads[1], 0, consumer, 0 ); pthread_join( threads[0], 0
);
pthread_join( threads[1], 0 );
sem_destroy(
&limit );
sem_destroy(
&semaphore );
}
The modifications introduce a second semaphore, limit. This semaphore is initialized with a value of 10. Before the producer
threads adds an item to the queue, it calls sem_wait(), which will decrement the value of the semaphore. Every time the
con-sumer thread removes an item from the queue, it calls sem_post(), which will increase the value of the semaphore. When the semaphore
reaches zero, the producer thread will call sem_wait() and will not return from the call until the consumer thread has removed
an item from the list. This will stop the producer thread from adding more
items to the queue before the consumer thread has had the chance to deal with
those items already there.
Condition
Variables
Condition
variables enable threads to communicate
state changes. Using them requires both
a mutex and a condition variable, together with the additional state that
threads need to check.
A condition variable is
initialized with a call to pthread_cond_init(), which
takes the address of the condition variable together with any attributes.
Condition variables are destroyed with a call to pthread_cond_destroy(), passing the address of the condition variable.
The default for condition
variables is to be private to a process. Attributes can be used to produce a
condition variable shared between processes. Listing 5.40 demonstrates using
attributes to create a shared condition variable.
Listing 5.40 Creating
a Condition Variable That Can Be Shared Between Processes
#include <pthread.h>
pthread_cond_t
CV;
int main()
{
pthread_condattr_t
CVA; pthread_condattr_init( &CVA );
pthread_condattr_setpshared(
&CVA, PTHREAD_PROCESS_SHARED );
pthread_cond_init( &CV, &CVA
);
pthread_condattr_destroy(
&CVA );
...
pthread_cond_destroy( &CV );
}
The
condition variable requires an actual variable to be monitored. A producer-consumer
is a good scenario to use for an example. We will use the variable length to denote the length of the
queue. The condition variable is used to wake the consumer thread when the
length of the queue is greater than zero. Listing 5.41 shows the initial-ization
code.
Listing 5.41 Creating
Threads and Condition Variable for Producer-Consumer Example
#include <pthread.h> #include <stdio.h>
pthread_cond_t
cv; pthread_mutex_t mutex; int length;
int
queue[200];
...
int
main()
{
pthread_t threads[2];
pthread_cond_init(
&cv, 0 );
pthread_mutex_init(
&mutex, 0 );
length = 0;
pthread_create( &threads[0], 0, producer, 0 );
pthread_create( &threads[1], 0, consumer, 0 ); pthread_join( threads[1], 0
);
pthread_join( threads[0], 0 ); pthread_mutex_destroy( &mutex ); pthread_cond_destroy( &cv );
}
Listing 5.42 shows the code for the producer
thread. The producer thread will obtain the mutex and then increment the length
of the queue before using the condition vari-able to signal to waiting threads
that there is an item in the queue. This signal will wake one of the waiting
threads. After the signal has completed, the mutex can be released.
Listing 5.42 Code
for Producer Thread
void
* producer( void* param )
{
for (int i=0; i<200; i++)
{
pthread_mutex_lock( &mutex ); queue[ length++ ]
= i;
pthread_cond_signal(
&cv );
pthread_mutex_unlock( &mutex );
}
}
If there
are no threads waiting on the condition variable, the call to pthread_cond_ signal() has no
effect. Hence, it is really necessary to make the call only if the queue was empty before the item was
added. It is only in this situation where the consumer thread might have been
waiting for items to be placed in the queue. In situations where there are
already items in the queue, the consumer thread will not have stopped.
It is
also possible to use a broadcast to signal to all waiting threads that there is
an item in the queue. The function pthread_cond_broadcast() wakes
all the threads waiting on the condition variable. This is illustrated in the
version of the producer thread shown in Listing 5.43.
Listing 5.43 Broadcasting
the Arrival of a New Item to All Waiting Threads
void * producer( void* param )
{
item_t * item;
for( int i=0; i<200; i++ )
{
pthread_mutex_lock( &mutex ); queue[ length++ ]
= i ;
pthread_cond_broadcast(
&cv );
pthread_mutex_unlock( &mutex );
}
}
There is no advantage to using broadcast in a
situation where there is only a single task to perform, since it will incur the
overhead of waking all the threads and then send all but one of them back to
sleep. It is useful in the situation where there are multiple independent tasks
to complete and each woken thread is able to identify an independent item of
work.
Listing 5.44 shows the code for the consumer
thread. This is slightly more complex than the code for the producer thread.
The consumer thread is placed in a while(true) loop. In this loop, the first thing it needs to do is to acquire the
mutex in order to get access to the variable length, which, in this example, is the proxy for the queue.
Listing 5.44 Code
for Consumer Thread
void * consumer( void* param)
{
for( int i=0; i<200; i++ )
{
pthread_mutex_lock(&mutex); while (length==0)
{
pthread_cond_wait(
&cv, &mutex );
}
int item = queue[ --length ];
pthread_mutex_unlock(&mutex);
printf( "Received %i\n", item );
}
}
The consumer thread needs to wait on the condition
variable only when there are no items in the queue. If there are items in the
queue, the consumer thread can immediately remove one and process it. Once the
consumer thread has decremented the queue, it can release the mutex and process
the item.
If the queue is empty, the consumer thread will
need to wait to be signaled by the producer thread. It does this by calling pthread_cond_wait() while still holding the mutex. This call will release the mutex while
the thread is waiting, but when signaled, the thread will wake up holding the
mutex again. The call to pthread_cond_wait() needs to
be placed in a loop. The thread will be woken when it is signaled that length is greater than zero, but it may also be signaled when length does not meet these crite-ria. Therefore, the thread needs to loop
calling pthread_cond_wait() until
the condi-tion, in this instance the value of the variable length, is met.
An example of a thread being woken up when the
condition is not true is when there are multiple threads waiting on the
condition variable and all the threads are woken by a broadcast signal. If
there is one item of work and two threads are woken, the first thread will get
the item of work. When the second thread wakes, it will discover that there is
no work for it. Hence, the second thread will appear to have suffered a
spurious wake-up.
There is one problem that should be avoided when
coding threads that wait on con-dition variables: the lost wake-up problem. Listing 5.45 shows an example.
Listing 5.45 Potential
“Lost Wake-Up” Issue
void
* consumer( void* param )
{
for( int i=0; i<200; i++ )
{
int item;
int go = 0;
pthread_mutex_lock( &mutex );
pthread_cond_wait(
&cv, &mutex );
if (length > 0)
{
item = queue[ --length ]; go = 1;
}
pthread_mutex_unlock(&mutex); if (go)
{
printf( "Received %i\n", item );
}
}
}
In this version of the code, the consumer threads wait on the condition
variable for each iteration of the loop. If the condition variable is signaled
before the consumer thread reaches the wait call, then the signal is lost, and
the consumer variable will wait until the next signal. If no further work is
produced by the producer thread, the con-sumer thread will wait indefinitely,
even though it has work waiting. This problem is compounded if the producer
thread is set to signal only when new work was added to an empty queue; in this
instance, the consumer thread will never get signaled, and the producer thread
will keep adding work to the queue.
Condition variables have a method to provide a
timeout when waiting to be signaled by the condition variable. The call is pthread_cond_timedwait(), which takes the timeout period, specified as an absolute time, as well
as the condition variable and mutex. This call will return either holding the
mutex lock or with an error code indicat-ing the reason for the return. The
code in Listing 5.46 illustrates using this function call to count the number
of minutes waited until the condition variable was signaled.
Listing 5.46 Using
a Timeout to Count Elapsed Minutes
#include <time.h> #include <errno.h>
void * consumer( void* param )
{
for( int i=0; i<200; i++ )
{
int seconds = 0; pthread_mutex_lock( &mutex );
while ( length == 0 )
{
struct
timespec now;
now.tv_sec =
time( 0 ) + 1;
now.tv_nsec =
0;
if ( pthread_cond_timedwait(
&cv, &mutex, &now ) == ETIMEDOUT )
{
seconds++;
}
}
int item = queue[ --length ];
if
( seconds ) { printf( "%i seconds waited\n", seconds ); }
pthread_mutex_unlock( &mutex );
printf( "Received %i\n", item );
}
}
The code uses pthread_cond_timedwait() to wait in units of one second. Every time the call fails to acquire
the mutex, a count of seconds waited is incremented. If the producer thread is
modified so that it sleeps between producing each item, then it is pos-sible to
see the timeout of the consumer thread.
Related Topics
Privacy Policy, Terms and Conditions, DMCA Policy and Compliant
Copyright © 2018-2024 BrainKart.com; All Rights Reserved. Developed by Therithal info, Chennai.