Switch from polling on FIFOs to signal

[geohot]

First off, remove randomness unless there's a real justification for it. In general, unless you absolutely need it, randomness shouldn't be used.

Secondly, with uids being pids, I think the signalling solution is obvious. The polling process can sleep until the publishing process decides it's time to wake it up. I don't know the nuance here, but usleep can be exited by EINTR, though I think the signal handler might have to be configured to do that.

You could get more fine grained with the poll, where the subscriber says if it's blocked on a given socket to avoid spurious wakeups (right now it wakes on all subscribed sockets), but this is a perf optimization for later and maybe not needed at all.

Right now this segfaults with python demo.py, and I don't know why. In general, this software needs a much better test suite, it underlies all of system stability and has to deal with many different orders of events.

[pd0wm]

Sounds like nanosleep is the way to go here. It is designed to be interrupted, has nice exit codes and even stores the remaining time when interrupted in a new struct.

http://man7.org/linux/man-pages/man2/nanosleep.2.html