What causes "poll() on talk fds failed" in kovidgoyal/kitty?

Check dmesg/system logs for fd exhaustion; raise ulimit -n if the process is near its limit.

poll() on talk fds failed

Error message

poll() on talk fds failed

What it means

perror emitted when poll() over the 'talk' file descriptors (kitty's internal peer/remote-control IPC sockets) returns a negative value that is not EAGAIN or EINTR. This means the multiplexed IPC event loop in child-monitor.c could not wait on its sockets, usually due to a bad/stale fd or EBADF/EFAULT after a peer died. kitty logs it and continues; it indicates IPC plumbing trouble rather than a terminal failure.

Source

Thrown at kitty/child-monitor.c:2271

            for (size_t k = 0; k < talk_data.num_peers; k++) {
                Peer *p = talk_data.peers + k;
                if (p->fd_array_idx) {
                    if (fds[p->fd_array_idx].revents & POLLIN) read_from_peer(self, p);
                    if (fds[p->fd_array_idx].revents & POLLOUT) write_to_peer(p);
                    if (fds[p->fd_array_idx].revents & POLLHUP) {
                        // try to read and write nonetheless these functions will set the failed flags.
                        if (!p->read.finished) read_from_peer(self, p);
                        if (p->write.used) write_to_peer(p);
                    }
                    if (fds[p->fd_array_idx].revents & POLLNVAL) {
                        p->read.finished = true;
                        p->write.failed = true;
                        p->write.used = 0;
                    }
                }
            }
        } else if (ret < 0) {
            if (errno != EAGAIN && errno != EINTR) perror("poll() on talk fds failed");
        }
    }
end:
    free_loop_data(&talk_data.loop_data);
    for (size_t i = 0; i < talk_data.num_peers; i++) free_peer(talk_data.peers + i);
    free(talk_data.peers);
    return 0;
}

static void
send_response_to_peer(id_type peer_id, const char *msg, size_t msg_sz, bool is_async_response) {
    bool wakeup = false;
    talk_mutex(lock);
    for (size_t i = 0; i < talk_data.num_peers; i++) {
        Peer *peer = talk_data.peers + i;
        if (peer->id == peer_id) {
            peer->waiting_for_async_response = is_async_response;
            if (peer->num_of_unresponded_messages_sent_to_main_thread) peer->num_of_unresponded_messages_sent_to_main_thread--;

View on GitHub (pinned to 6d5d0c4406)

Solutions

Check dmesg/system logs for fd exhaustion; raise ulimit -n if the process is near its limit.
Update kitty — stale-peer fd handling in the talk loop has been fixed in newer releases.
Reproduce with KITTY_DEBUG / --debug-ipc if available and report the errno value printed after the message.
If tied to a kitten, restart the kitten rather than the whole kitty session; the main loop recovers.

Defensive patterns

Strategy: fallback

Prevention

Keep peer kittens and the kitty main binary on the same version.
Avoid spawning hundreds of concurrent peer kittens that die abruptly.
Monitor fd usage (ls /proc/$(pidof kitty)/fd | wc -l) in automation.

When it happens

Trigger: poll() returning -1 with errno such as EBADF (a closed talk fd still in the poll set), EFAULT (fds array out of reach), or ENOMEM, while running kitty with remote control (--listen-on), single-instance sockets, or peer kittens talking to the main process.

Common situations: A peer kitten or socket client crashes mid-connection leaving a stale fd; fd exhaustion on systems with many kittens/windows; bugs in older kitty versions where dead peer fds were not removed from the poll set.

Related errors

AI-assisted analysis of kovidgoyal/kitty@6d5d0c4406 (2026-08-27). Data as JSON: /api/errors/afa2e9f92bff11c5. Report an issue: GitHub.