I write network library which offers blocking socket operations (connect, read, write, ...) with timeouts, using boost::asio asynchronous operations as implementation. It is actually some sort of wrapper around boost::asio. Implementation does something like this (read, for example):
do_read() { timer.async_wait(&handle_timer) socket.async_read(..., handle_read); wait_for_all_handlers(); deal_with_ec(); }
handle_timer() { socket.cancel(); } handle_read() { timer.cancel(); }
I have one io_service for all the sockets/timers with one thread doing io_service::run(). Code is multithreaded, so many threads run that functions like do_read() simultaneously (for different sockets, of course).
I have two questions:
1. How to wait for completion of all handlers? Currently I use shared_ptr with mutex+condvar (for each library call), it works fine, but seems to be too heavy-weight.
2. There is a chance (with small timeouts), that timer expires before async_read() is started, so sock.cancel() is no-op in handle_timer(), and timer will be disabled. Currently, I just ignore this. I can use another mutex (or strand) to protect from this, but it is also heavy-weight.
It probably isn't an answer to your questions, but perhaps you'll find the following approach useful: http://lists.boost.org/Archives/boost/2007/04/120339.php