File: src/server/server.zig:789-794 (offending path),
src/server/server.zig:828-830 (correct pattern)
Severity: HIGH
Status: ✅ FIXED
self.loop.removeFd(pty.master) before
pty.deinit(), mirroring destroyPane.handlePtyEvent recognises POLLNVAL — drops the
registration immediately and, when the pane’s own master matches, sets
master = -1 before deinit() so nothing double-closes a possibly
recycled fd number while still releasing input_buf and reaping the child.Tests: pollOnce reports POLLNVAL for a closed registered fd — bug #364
(loop.zig, real kernel behaviour); handlePtyEvent drops registration on
POLLNVAL — bug #364; remain-on-exit unregisters pty before closing master —
bug #364 (pane survives with no live pty and no stale registration).
if (remain and has_other_panes) {
if (pane.pty) |*pty| {
pty.deinit(); // close(master) — fd still registered!
pane.pty = null;
}
The normal-exit path calls self.loop.removeFd(pty.master) first; this path
does not. Next poll() reports POLLNVAL on the dead fd — and nothing in
run()/handlePtyEvent checks NVAL (only IN/HUP/ERR are tested,
server.zig:559-561, 673-677). Level-triggered NVAL on every iteration →
permanent spin. Worse: if the fd number is recycled by another connection,
events misroute to a stale udata pane pointer.
Add self.loop.removeFd(pty.master) before pty.deinit() here, plus
defense-in-depth: treat .NVAL like .HUP in the run loop and lazily drop the
entry. Also see #365 for the killWindow twin.