[cyberax]

I had this problem, and I solved it by farming the known-blocking syscalls to a separate thread pool. Then the calling thread can just abandon the wait. To make it a bit better, you can also use bounded timeouts (~1-2 seconds) with retries for some calls like recvfrom() via SO_TIMEOUT so that the termination time becomes bounded.

This is probably the cleanest solution that is portable.

[naasking]

Then those thread pool threads have to be careful not to take locks and other scarce resources that aren't cleaned up then, no?

[cyberax]

Sure. But if you only use it to run blocking requests, it's probably going to be pretty trivial. It won't be super-high performance because of all the cache bounces, but if you're not using something like epoll, then you probably are not really focused on performance anyway.