[go-while]

Made with AI (Claude, via Claude Code) under my direction, like #376. Please review it with that in mind.

This is the follow-up announced in #376 and builds on it.
The first two commits are those of #376; the new change is the last commit, bc3424b.

Problem

A SIGTERM or SIGHUP that arrives while innd is busy (not waiting in select()), for instance while a Perl filter_art runs, is blocked by the signal mask.
xsignal_unmask() right before select() then runs the handler, which sets GotTerminate, but select() is not interrupted by it.
innd waits for the whole timeout or the next activity before it reaches the GotTerminate check.

Before signal masking, the handler ran at once but select() waited the same way, so this part is older than the idle case of #376.

Fix

When GotTerminate is set after the unmask, use a zero timeout, so that select() returns at once and the check right after it (moved there by #376) is reached:

xsignal_unmask();
if (GotTerminate) {
    tv.tv_sec = 0;
    tv.tv_usec = 0;
}
count = select(...);

A window of a few instructions between that test and select() remains.
Closing it completely would need pselect() or a self-pipe, which seemed too much here.
The news.pod entry of #376 gets one sentence for this case.

Testing

A Perl filter_innd.pl whose filter_art sleeps, an article fed with IHAVE, SIGTERM during the sleep:

innd	result
with #376 only	still running 10 s after the signal (and 12 s with a 4 s filter); exits only when something connects
with #376 and this fix	exits 0 once the filter returns: 4.1 s with a 3 s filter (3 s filter + the 1 s sleep in JustCleanup())

• Idle and loaded shutdowns are unchanged (still 1.0 s).
• make test passes.
• The end-to-end test is TestSIGTERMWhileBusy in tests/e2e/signal_test.go. It fails with innd: exit on SIGTERM and SIGHUP while waiting for activity #376 alone (Fedora 45) and passes with this fix (Fedora 45 and GitHub Actions).