lib/tsocket: avoid endless cpu-spinning in tstream_bsd_fde_handler()

lib/tsocket: avoid endless cpu-spinning in tstream_bsd_fde_handler()

There were some reports that strace output an LDAP server socket is in
CLOSE_WAIT state, returning EAGAIN for writev over and over (after a call to
epoll() each time).

In the tstream_bsd code the problem happens when we have a pending
writev_send, while there's no readv_send pending. In that case
we still ask for TEVENT_FD_READ in order to notice connection errors
early, so we try to call writev even if the socket doesn't report TEVENT_FD_WRITE.
And there are situations where we do that over and over again.

It happens like this with a Linux kernel:

    tcp_fin() has this:
        struct tcp_sock *tp = tcp_sk(sk);

        inet_csk_schedule_ack(sk);

        sk->sk_shutdown |= RCV_SHUTDOWN;
        sock_set_flag(sk, SOCK_DONE);

        switch (sk->sk_state) {
        case TCP_SYN_RECV:
        case TCP_ESTABLISHED:
                /* Move to CLOSE_WAIT */
                tcp_set_state(sk, TCP_CLOSE_WAIT);
                inet_csk_enter_pingpong_mode(sk);
                break;

It means RCV_SHUTDOWN gets set as well as TCP_CLOSE_WAIT, but
sk->sk_err is not changed to indicate an error.

    tcp_sendmsg_locked has this:
    ...
        err = -EPIPE;
        if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
                goto do_error;

        while (msg_data_left(msg)) {
                int copy = 0;

                skb = tcp_write_queue_tail(sk);
                if (skb)
                        copy = size_goal - skb->len;

                if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
                        bool first_skb;

    new_segment:
                        if (!sk_stream_memory_free(sk))
                                goto wait_for_space;

    ...

    wait_for_space:
                set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
                if (copied)
                        tcp_push(sk, flags & ~MSG_MORE, mss_now,
                                 TCP_NAGLE_PUSH, size_goal);

                err = sk_stream_wait_memory(sk, &timeo);
                if (err != 0)
                        goto do_error;

It means if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) doesn't
hit as we only have RCV_SHUTDOWN and sk_stream_wait_memory returns
-EAGAIN.

    tcp_poll has this:

        if (sk->sk_shutdown & RCV_SHUTDOWN)
                mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;

So we'll get EPOLLIN | EPOLLRDNORM | EPOLLRDHUP triggering
TEVENT_FD_READ and writev/sendmsg keeps getting EAGAIN.

So we need to always clear TEVENT_FD_READ if we don't
have readable handler in order to avoid burning cpu.
But we turn it on again after a timeout of 1 second
in order to monitor the error state of the connection.

And now that our tsocket_bsd_error() helper checks for POLLRDHUP,
we can check if the socket is in an error state before calling the
writable handler when TEVENT_FD_READ was reported.
Only on error we'll call the writable handler, which will pick
the error without calling writev().

BUG: https://bugzilla.samba.org/show_bug.cgi?id=15202

Signed-off-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
(cherry picked from commit e232ba946f00aac39d67197d9939bc923814479c)