Thanks, that is a very good point about threads. I remember that testing tkinter leaves an internal TK or TCL thread running in the background. And I understand Free BSD delivers process signals to arbitrary threads (not necessarily the main thread).

But the FreeBSD 3.5 buildbot looks like it ran test_io in a separate process (-j4 specified). Anyway it is another possible explanation of the hang (signal delivered to wrong thread, so write() call not interrupted).

Perhaps it is better to use signal.pthread_kill() instead of os.kill() in my proposal.