openssh unix dev - Apr 2021 - Restoring block mode on exiting

Hello,

OpenSSH incorrectly restores the standard mode (blocking mode) on standard
output upon exiting. This causes the next shell scripts commands to
potentially fail in EAGAIN.
The reproducer is:

#!/bin/sh
(
ssh localhost true
cat /dev/zero
) | sleep 30

Restoring the blocking modes happens with the duped file descriptors and
too late.

The changes causing this problem was introduced in
https://github.com/openssh/openssh-portable/commit/4d5456c7de108e17603a0920c4d15bca87244921

The 2 possible fixes can be found in
https://github.com/openssh/openssh-portable/pull/244 (adhoc patch) and
https://github.com/openssh/openssh-portable/pull/246 (more robust one).

I think it's worth fixing the issue in this, that or any other way.
I'll be happy to adopt any of those patches to the state when they become
suitable for upstream.

Many thanks in advance!

-- 
Dmitry Belyavskiy

Hi Dmitry,

On 23/4/21 1:06 am, Dmitry Belyavskiy wrote:
> OpenSSH incorrectly restores the standard mode (blocking mode) on standard
> output upon exiting.
I agree that SSH should *restore* the blocking mode.? I've only spent a 
small
amount of timelooking at it and I think SSH currently does not do that, and
your patch also does not.

Your patch sets blocking mode if SSH cleared it.? I think it would be more
correct to restore it tothe state that it was when SSH took over the file
because we shouldn't assume.

In channel_register_fds, don't record if we set non-blocking mode, record
what mode it currentlyhas:

    c->nonblock = (get_nonblock(rfd) << 2) | (get_nonblock(wfd)
<< 1) |
    get_nonblock(efd);

Restore in channel_close_fds:

    channel_close_fd(ssh, &c->sock, 0);
    if (rfd != sock) channel_close_fd(ssh, &c->rfd, c->nonblock &
4);
    if (wfd != sock) channel_close_fd(ssh, &c->wfd, c->nonblock &
2);
    if (efd != sock) channel_close_fd(ssh, &c->efd, c->nonblock &
1);

In channel_close_fd:

    if (fd != -1) (nonblock?set_nonblock:unset_nonblock)(fd);

Regards,

David