Difference between revisions of "POSIX Threads Explained, Part 3"
Latest revision as of 08:09, December 31, 2014
Improve efficiency with condition variables
Condition variables explained
I ended my previous article by describing a particular dilemma how does a thread deal with a situation where it is waiting for a specific condition to become true? It could repeatedly lock and unlock a mutex, each time checking a shared data structure for a certain value. But this is a waste of time and resources, and this form of busy polling is extremely inefficient. The best way to do this is to use the
pthread_cond_wait() call to wait on a particular condition to become true.
It's important to understand what pthread_cond_wait() does -- it's the heart of the POSIX threads signalling system, and also the hardest part to understand.
First, let's consider a scenario where a thread has locked a mutex, in order to take a look at a linked list, and the list happens to be empty. This particular thread can't do anything -- it's designed to remove a node from the list, and there are no nodes available. So, this is what it does.
While still holding the mutex lock, our thread will call
pthread_cond_wait(&mycond,&mymutex). The pthread_cond_wait() call is rather complex, so we'll step through each of its operations one at a time.
The first thing pthread_cond_wait() does is simultaneously unlock the mutex mymutex (so that other threads can modify the linked list) and wait on the condition mycond (so that pthread_cond_wait() will wake up when it is "signalled" by another thread). Now that the mutex is unlocked, other threads can access and modify the linked list, possibly adding items.
At this point, the pthread_cond_wait() call has not yet returned. Unlocking the mutex happens immediately, but waiting on the condition mycond is normally a blocking operation, meaning that our thread will go to sleep, consuming no CPU cycles until it is woken up. This is exactly what we want to happen. Our thread is sleeping, waiting for a particular condition to become true, without performing any kind of busy polling that would waste CPU time. From our thread's perspective, it's simply waiting for the pthread_cond_wait() call to return.
Now, to continue the explanation, let's say that another thread (call it "thread 2") locks mymutex and adds an item to our linked list. Immediately after unlocking the mutex, thread 2 calls the function pthread_cond_broadcast(&mycond). By doing so, thread 2 will cause all threads waiting on the mycond condition variable to immediately wake up. This means that our first thread (which is in the middle of a pthread_cond_wait() call) will now wake up.
Now, let's take a look at what happens to our first thread. After thread 2 called pthread_cond_broadcast(&mymutex) you might think that thread 1's pthread_cond_wait() will immediately return. Not so! Instead, pthread_cond_wait() will perform one last operation: relock mymutex. Once pthread_cond_wait() has the lock, it will then return and allow thread 1 to continue execution. At that point, it can immediately check the list for any interesting changes.
Stop and review!
One more miscellaneous file before we get to the biggie. Here's dbug.h:
We use this code to handle unrecoverable errors in our work crew code.
The work crew code
Speaking of the work crew code, here it is:
Now it's time for a quick code walkthrough. The first struct defined is called
wq, and contains a data_control and a queue header. The data_control structure will be used to arbitrate access to the entire queue, including the nodes in the queue. Our next job is to define the actual work nodes. To keep the code lean to fit in this article, all that's contained here is a job number.
Next, we create the cleanup queue. The comments explain how this works. OK, now let's skip the threadfunc(), join_threads(), create_threads() and initialize_structs() calls, and jump down to main(). The first thing we do is initialize our structures -- this includes initializing our data_controls and queues, as well as activating our work queue.
Now it's time to initialize our threads. If you look at our create_threads() call, everything will look pretty normal -- except for one thing. Notice that we are allocating a cleanup node, and initializing its threadnum and TID components. We also pass a cleanup node to each new worker thread as an initial argument. Why do we do this?
Because when a worker thread exits, it'll attach its cleanup node to the cleanup queue, and terminate. Then, our main thread will detect this addition to the cleanup queue (by use of a condition variable) and dequeue the node. Because the TID (thread id) is stored in the cleanup node, our main thread will know exactly which thread terminated. Then, our main thread will call pthread_join(tid), and join with the appropriate worker thread. If we didn't perform such bookkeeping, our main thread would need to join with worker threads in an arbitrary order, possibly in the order that they were created. Because the threads may not necessarily terminate in this order, our main thread could be waiting to join with one thread while it could have been joining with ten others. Can you see how this design decision can really speed up our shutdown code, especially if we were to use hundreds of worker threads?
Now that we've started our worker threads (and they're off performing their threadfunc(), which we'll get to in a bit), our main thread begins inserting items into the work queue. First, it locks wq's control mutex, and then allocates 16000 work packets, inserting them into the queue one-by-one. After this is done, pthread_cond_broadcast() is called, so that any sleeping threads are woken up and able to do the work. Then, our main thread sleeps for two seconds, and then deactivates the work queue, telling our worker threads to terminate. Then, our main thread calls the join_threads() function to clean up all the worker threads.
Time to look at threadfunc(), the code that each worker thread executes. When a worker thread starts, it immediately locks the work queue mutex, gets one work node (if available) and processes it. If no work is available, pthread_cond_wait() is called. You'll notice that it's called in a very tight while() loop, and this is very important. When you wake up from a pthread_cond_wait() call, you should never assume that your condition is definitely true -- it will probably be true, but it may not. The while loop will cause pthread_cond_wait() to be called again if it so happens that the thread was mistakenly woken up and the list is empty.
If there's a work node, we simply print out its job number, free it, and exit. Real code would do something more substantial. At the end of the while() loop, we lock the mutex so we can check the active variable as well as checking for new work nodes at the top of the loop. If you follow the code through, you'll find that if wq.control.active is 0, the while loop will be terminated and the cleanup code at the end of threadfunc() will begin.
The worker thread's part of the cleanup code is pretty interesting. First, it unlocks the work_queue, since pthread_cond_wait() returns with the mutex locked. Then, it gets a lock on the cleanup queue, adds our cleanup node (containing our TID, which the main thread will use for its pthread_join() call), and then it unlocks the cleanup queue. After that, it signals any cq waiters (pthread_cond_signal(&cq.control.cond)) so that the main thread will know that there's a new node to process. We don't use pthread_cond_broadcast() because it's not necessary -- only one thread (the main thread) is waiting for new entries in the cleanup queue. Our worker thread prints a shutdown message, and then terminates, waiting to be pthread_joined() by the main thread when it calls join_threads().
If you want to see a simple example of how condition variables should be used, take a look at the join_threads() function. While we still have worker threads in existence, join_threads() loops, waiting for new cleanup nodes in our cleanup queue. If there is a new node, we dequeue the node, unlock the cleanup queue (so that other cleanup nodes can be added by our worker threads), join with our new thread (using the TID stored in the cleanup node), free the cleanup node, decrement the number of threads "out there", and continue.
Wrapping it up
We've reached the end of the "POSIX threads explained" series, and I hope that you're now ready to begin adding multithreaded code to your own applications. For more information, please see the Resources section, which also contains a tarball of all the sources used in this article. I'll see you next series!
Daniel Robbins is best known as the creator of Gentoo Linux and author of many IBM developerWorks articles about Linux. Daniel currently serves as Benevolent Dictator for Life (BDFL) of Funtoo Linux. Funtoo Linux is a Gentoo-based distribution and continuation of Daniel's original Gentoo vision.
ARM RebuildARM systems will use new stage3's that are not compatible with earlier versions.
ABI X86 64 and 32Funtoo Linux has new 32-bit compatibility libraries inherited from Gentoo. Learn about them here.
Pre-built kernels!Funtoo stage3's are now starting to offer pre-built kernels for ease of install. read more....
Browse all our Linux-related articles, below:
- Funtoo Filesystem Guide, Part 1
- Funtoo Filesystem Guide, Part 2
- Funtoo Filesystem Guide, Part 3
- Funtoo Filesystem Guide, Part 4
- Funtoo Filesystem Guide, Part 5
- Learning Linux LVM, Part 1
- Learning Linux LVM, Part 2
- Linux Fundamentals, Part 1
- Linux Fundamentals, Part 1/pt-br
- Linux Fundamentals, Part 2
- Linux Fundamentals, Part 3
- Linux Fundamentals, Part 4
- LVM Fun
- Making the Distribution, Part 1
- Making the Distribution, Part 2
- Making the Distribution, Part 3
- Maximum Swappage
- Partition Planning Tips
- Partitioning in Action, Part 1
- Partitioning in Action, Part 2
- POSIX Threads Explained, Part 1
- POSIX Threads Explained, Part 2
- POSIX Threads Explained, Part 3
- Prompt Magic
- The Gentoo.org Redesign, Part 1
- The Gentoo.org Redesign, Part 2
- The Gentoo.org Redesign, Part 3
- The Gentoo.org Redesign, Part 4
- Traffic Control