thread_do used to use pthread_exit(). However that resulted in
hard times of testing for memory leaks. The reason is that on pthread_exit()
not all memory is freed bt pthread (probably for future threads or false
belief that the application is terminating). For these reasons a simple return
is used.
Interestingly using pthread_exit() results in much more memory being allocated.
This is needed only in the tests. The reason is that if you call thpool_destroy()
and then exit immediately, maybe the program will exit before all the threads
had the time to deallocate. In that way it is impossible to check for memory
leaks.
In production you don't have to worry about this since if you call exit(),
immediately after you destroyed the pool, the threads will be freed
anyway by the OS. If you anyway destroy the pool in the middle of your
program it doesn't matter again since the program will not exit immediately
and thus threads will have more than enough time to terminate.
Notice: As of 11-Dec-2015 wait() doesn't use polling anymore. Instead a conditional variable is being used so in theory there should not be any CPU overhead.
Normally wait() will spike CPU usage to full when called. This is normal as long as it doesn't last for more than 1 second. The reason this happens is that wait() goes through various phases of polling (what is called smart polling).
- Initially there is no interval between polling and hence the 100% use of your CPU.
- After that the polling interval grows exponentially.
- Finally after x seconds, if there is still work, polling falls back to a very big interval.
The reason wait() works in this way, is that the function is mostly used when someone wants to wait for some calculation to finish. So if the calculation is assumed to take a long time then we don't want to poll too often. Still we want to poll fast in case the calculation is a simple one. To solve these two problems, this seemingly awkward behaviour is present.