Current scenario is epoll_wait over a couple of fds and a queue of possible incoming messages, I'd like the loop below epoll_wait to be executed on IO event or on new message.
Ways I know:
- Use a
time
msec timeout and check the queue first thing in the loop - Use the self-pipe trick from the queue code when messages become available
- Interrupt the syscall with a standard signal
- Use epoll_pwait and refine the previous point
None of the points posted above satisfy me enough and I was wondering if there are any other methods that I haven't found.
Reasons are:
- Signals are something to avoid on multithreaded code and are not very reliable
- Timeout one removes part of the benefit of the epoll, only waking with events
- Self-pipe trick looks the best approach for the moment, but still too much boilerplate
ideas?
You have enumerated the events that can wake up epoll, so the question really becomes: "How do I reduce the boilerplate for the self-pipe trick?"
The answer to that question really depends on your code, the language and what you are trying to do. I assume you have a thread that processes I/O and you want to do other work in that thread while there is no I/O ready. In the code that manages the epoll loop, it can have an internal handle that is exposed to other parts of the system as a "wake" function or a "submit work" function.
There are libraries that do this, for example boost.asio for C++. However, it isn't difficult to write your own if you're just targeting epoll, and the amount of actual boilerplate code should be minimal once you have a class/module/whatever that deals with the epoll loop.
You can use an eventfd which is effectively the same thing as the self-pipe trick except with fewer file descriptors and less boilerplate (glibc has convenience
eventfd_read/write
functions for instance).