I have a chunk of shared memory containing some data. Reading / writing this data needs to be synchronized between different processes. At this moment I'm using unnamed POSIX semaphores. Since my main concern is speed, I'd like to know if semaphores are optimal here? Do they end with a syscall for example (userspace -> kernel space), or force context switches? What would you suggest?
Some info on unnamed semaphores: http://linux.die.net/man/3/sem_init
Since distinct processes have distinct address spaces, any synchronization feature will have to go through the kernel... except if you manage to synchronize through a chunk of shared memory. And you so happen to have such a chunk.
Therefore, I recommend using a pthreads mutex (see pthread_mutex_lock()) located at some place within your chunk of shared memory. This will work on a recent enough Linux system (glibc-2.3.2 or later, kernel 2.6.x). With a mutex, the kernel will be invoked only upon contention (two processes competing for the same lock), in which case going through the kernel is the right thing to do.