I can't see how this patch is needed. fatal() calls exit(), which closes all process file descriptors. If exit() isn't closing file descriptors then that is a kernel bug AFAIK.

This is not a kernel bug. fatal() is designed to eventually call exit(), but
this bug is preventing sshd process from calling exit().

The child process calls fatal() when fork() failed at privsep_postauth().
(Please note that fork() is replaced with -1 for explanation.)

---------- privsep_postauth() in sshd.c ----------
static void
privsep_postauth(Authctxt *authctxt)
{
        u_int32_t rnd[256];

#ifdef DISABLE_FD_PASSING
        if (1) {
#else
        if (authctxt->pw->pw_uid == 0 || options.use_login) {
#endif
                /* File descriptor passing is broken or root login */
                use_privsep = 0;
                goto skip;
        }

        /* New socket pair */
        monitor_reinit(pmonitor);

        pmonitor->m_pid = -1; // Emulate the fork() failure.
        if (pmonitor->m_pid == -1)
                fatal("fork of unprivileged child failed");
(...snipped...)
---------- privsep_postauth() in sshd.c ----------

fatal() calls cleanup_exit(255). (Please note that dummy write(-1, "", $step)
lines are inserted for comparing with strace command.)

---------- fatal() in fatal.c ----------
void
fatal(const char *fmt,...)
{
        va_list args;

        va_start(args, fmt);
        do_log(SYSLOG_LEVEL_FATAL, fmt, args);
        va_end(args);
        write(-1, "", 0);
        cleanup_exit(255);
}
---------- fatal() in fatal.c ----------

cleanup_exit(255) will eventually call _exit(255).

---------- cleanup_exit() in sshd.c ----------
/* server specific fatal cleanup */
void
cleanup_exit(int i)
{
        static int in_cleanup;

        int is_privsep_child;

        write(-1, "", 1);
        /* cleanup_exit can be called at the very least from the privsep
           wrappers used for auditing.  Make sure we don't recurse
           indefinitely. */
        if (in_cleanup)
                _exit(i);
        write(-1, "", 2);
        in_cleanup = 1;

        if (the_authctxt)
                do_cleanup(the_authctxt);
        write(-1, "", 3);
        is_privsep_child = use_privsep && pmonitor != NULL && !mm_is_monitor();
        write(-1, "", 4);
        if (sensitive_data.host_keys != NULL)
                destroy_sensitive_data(is_privsep_child);
        write(-1, "", 5);
        packet_destroy_all(1, is_privsep_child);
#ifdef SSH_AUDIT_EVENTS
        /* done after do_cleanup so it can cancel the PAM auth 'thread' */
        write(-1, "", 6);
        if ((the_authctxt == NULL || !the_authctxt->authenticated) &&
            (!use_privsep || mm_is_monitor()))
                audit_event(SSH_CONNECTION_ABANDON);
#endif
        write(-1, "", 7);
        _exit(i);
}
---------- cleanup_exit() in sshd.c ----------

Did we reach the _exit(i) line? Let's check the strace command.

---------- strace log start ----------
[pid 17153] socketpair(PF_FILE, SOCK_STREAM, 0, [5, 6]) = 0
[pid 17153] fcntl(5, F_SETFD, FD_CLOEXEC) = 0
[pid 17153] fcntl(6, F_SETFD, FD_CLOEXEC) = 0
[pid 17153] sendto(4, "<82>Nov  1 11:11:08 sshd[17153]:"..., 73, MSG_NOSIGNAL, NULL, 0) = 73
[pid 17153] close(4)                    = 0
[pid 17153] write(4294967295, "", 0)    = -1 EBADF (Bad file descriptor)
[pid 17153] write(4294967295, "\0", 1)  = -1 EBADF (Bad file descriptor)
[pid 17153] write(4294967295, "\0G", 2) = -1 EBADF (Bad file descriptor)
[pid 17153] write(4294967295, "\0Go", 3) = -1 EBADF (Bad file descriptor)
[pid 17153] write(4294967295, "\0Got", 4) = -1 EBADF (Bad file descriptor)
[pid 17153] getuid()                    = 0
[pid 17153] write(5, "\0\0\0DR", 5)     = 5
[pid 17153] write(5, "\0\0\0/c1:bd:33:a5:66:d5:83:2d:0c:9"..., 67) = 67
[pid 17153] read(5, ^C <unfinished ...>
Process 17147 detached
Process 17153 detached
---------- strace log end ----------

We can see that the child process reached the write(-1, "", 4) line but
did not reach the write(-1, "", 5) line. This means that the child process
is sleeping at

        if (sensitive_data.host_keys != NULL)
                destroy_sensitive_data(is_privsep_child);

trying to read data from fd == 5.

What is fd == 5 connected with?
According to strace command, fd == 5 and fd == 6 are a socket pair
created by monitor_reinit() call in privsep_postauth().

---------- monitor_reinit() in monitor.c ----------
void
monitor_reinit(struct monitor *mon)
{
        int pair[2];

        monitor_socketpair(pair);

        mon->m_recvfd = pair[0];
        mon->m_sendfd = pair[1];
}
---------- monitor_reinit() in monitor.c ----------

destroy_sensitive_data() tried to write to fd == 5 and trying to read from
fd == 5, but there is no writers writing to fd == 6.

Dead lock caused by trying to I/O against wrong file descriptor.
This was why calling shutdown() seems to solve the problem.

Created attachment 2369 [details]
A patch which fixes this problem.

OK. I found the exact location.

fatal("fork of unprivileged child failed") calls cleanup_exit(255).
cleanup_exit(255) calls destroy_sensitive_data(1).
destroy_sensitive_data(1) calls mm_audit_destroy_sensitive_data(fp, pid, uid).

mm_audit_destroy_sensitive_data() uses global pmonitor->m_recvfd which now
references a socket pair.

---------- mm_audit_destroy_sensitive_data() in monitor_wrap.c ----------
void
mm_audit_destroy_sensitive_data(const char *fp, pid_t pid, uid_t uid)
{
        Buffer m;

        buffer_init(&m);
        buffer_put_cstring(&m, fp);
        buffer_put_int64(&m, pid);
        buffer_put_int64(&m, uid);

        mm_request_send(pmonitor->m_recvfd, MONITOR_REQ_AUDIT_SERVER_KEY_FREE, &m);
        mm_request_receive_expect(pmonitor->m_recvfd, MONITOR_ANS_AUDIT_SERVER_KEY_FREE,
                                  &m);
        buffer_free(&m);
}
---------- mm_audit_destroy_sensitive_data() in monitor_wrap.c ----------

Regarding privsep_preauth(), pmonitor = monitor_init() causes this problem
when fork() fails. Fortunately, timeout was previously configured via alarm()
which will eventually terminate the process even if fork() failed.

---------- privsep_preauth() in sshd.c ----------
static int
privsep_preauth(Authctxt *authctxt)
{
        int status;
        pid_t pid;

        /* Set up unprivileged child process to deal with network data */
        pmonitor = monitor_init();
        /* Store a pointer to the kex for later rekeying */
        pmonitor->m_pkex = &xxx_kex;

        pid = fork();
        if (pid == -1) {
                fatal("fork of unprivileged child failed");
(...snipped...)
---------- privsep_preauth() in sshd.c ----------

Regarding privsep_postauth(), monitor_reinit(pmonitor) causes this problem
when fork() fails. Unfortunately, timeout previously configured via alarm()
was disabled, which results in forever wait when fork() failed.

---------- privsep_postauth() in sshd.c ----------
static void
privsep_postauth(Authctxt *authctxt)
{
        u_int32_t rnd[256];

#ifdef DISABLE_FD_PASSING
        if (1) {
#else
        if (authctxt->pw->pw_uid == 0 || options.use_login) {
#endif
                /* File descriptor passing is broken or root login */
                use_privsep = 0;
                goto skip;
        }

        /* New socket pair */
        monitor_reinit(pmonitor);

        pmonitor->m_pid = fork();
        if (pmonitor->m_pid == -1)
                fatal("fork of unprivileged child failed");
(...snipped...)
---------- privsep_postauth() in sshd.c ----------

I confirmed that the patch in Comment 3 can fix this problem.

As I get no response, I set importance to "major" due to following impact.

This bug actually blocked an unattended ssh session (execution of batched job)
of an enterprise server. The worst case might be "sleeping forever while trying
to kdump over ssh" which means that "the kdump procedure is unable to reboot
after the kernel panic", severely affecting availability (down time) of the
enterprise servers.

look at this for next release

It's an issue of RHEL's downstream audit patches - https://bugzilla.redhat.com/show_bug.cgi?id=1028643 :

openssh-5.3p1-audit.patch:
+     is_privsep_child = use_privsep && pmonitor != NULL && !mm_is_monitor();

This hunk needs a change.

Set all RESOLVED bugs to CLOSED with release of OpenSSH 7.1