(1) poll is not required to wake on such events leading to hangs
if infinite timeouts are used.
This can be observed in macOS: (a) syncterm -iC (b) connect via telnet to some system (c) use the menu to disconnect from the system Note: socket must be idle to reproduce.
(2) If a blocked recv/send receives an error in response to a close
from another thread, the context of the socket has been lost: the
descriptor is free to be reallocated under the thread performing
I/O by the system. Any retry logic that may be written in response
to the error condition could lead to operations being performed
on a valid but incorrect resource.
(3) Processes that inherited the socket via fork() will still have
a valid descriptor to the socket. i.e., the socket does not
globally shutdown via a call to close(); however, shutdown()
acts on the connection by triggering a TCP FIN. This will lead
to the actual shutdown of the socket by the remote connection
as well.
Solution:
Use shutdown() ahead of close(). This will initiate a proper shutdown
sequence on the socket while leaving the descriptor open. poll() will
wake, and a subsequent read or write will yield the desired EOF/EPIPE.
