process_wait can hang if reusing FD in rapid succession.

[ioquatix]

It looks like EPoll can hang when executing system. If process_wait is called in rapid succession and hits the same file descriptor, the internal state can be reused, which may assume the epoll is currently armed when it is not, leading to a hang.

1. IO_Event_Selector_EPoll_Waiting_register(..., 0, pidfd, ...)
2. epoll_ctl ADD.
3. epoll_wait -> status
4. IO_Event_Selector_EPoll_Waiting_register(..., 0, pidfd, ...) but we hit epoll_descriptor->registered_events == epoll_descriptor->waiting_events branch (return).
5. Hang

This is because the argument for VALUE io is 0 in both cases, so we assume the IO is the same object, and thus the registration has not changed. By providing _pid as the io object, we prevent this from happening. I'll think about a better solution, as this feels like a hack; we can probably clear the state after process_wait instead, on all cases, which prevents any re-use. _pid is unlikely to wrap around rapidly so the chance of hitting this bug is greatly reduced, but not zero, with the current proposed change.

Fixes socketry/async#351.

Types of Changes

• Bug fix.

Contribution

• I added tests for my changes.
• I tested my changes locally.
• I agree to the Developer's Certificate of Origin 1.1.

[mezbahalam]

You are an extraordinary wizard @ioquatix