select() and poll() inconsistency in writability test
A pipe without readers will yield EPIPE/SIGPIPE when trying to write() into. A broken or shutdown TCP connection will behave the same modulo networking errors (which are returned only once as it seems).
A pipe without readers will set POLLERR in pipe_poll() and this turns up as being writable according to select() so it won't wait. However, a TCP connection no longer open [for writing] does not set POLLERR/POLLOUT in tcp_poll(): so it is not yet writable according to select() causing it to wait. This is not consistent behavior.
Shouldn't the kernel set POLLOUT in both "EPIPE" cases instead of POLLERR/nothing as it does now?
Signature · 1 line
Frank