How to run ssh over an existing TCP connection

I want to be able to SSH to a number linux devices at once, behind different NATs. I can’t configure the network that they are on. However, I’m having trouble getting ssh to go over an existing connection.

I have full control over both my client and the devices. Here’s the process so far: On my client, I first run

socat TCP-LISTEN:5001,pktinfo,fork EXEC:./create_socket.sh,fdin=3,fdout=4,nofork

Contents of ./create_socket.sh:

ssh -N -M -S "~/sockets/${SOCAT_PEERADDR}" -o "ProxyCommand=socat - FD:3!!FD:4" "root@${SOCAT_PEERADDR}"

On the device, I’m running

socat TCP:my_host:4321 TCP:localhost:22

However, nothing comes in or out of FD:3!!FD:4, I assume because the ProxyCommand is a subprocess. I’ve also tried setting fdin=3,fdout=3 and changing ./create_socket.sh to:

ssh -N -M -S "~/sockets/${SOCAT_PEERADDR}" -o "ProxyUseFdpass=yes" -o "ProxyCommand=echo 3" "root@${host}"

This prints an error:

mm_receive_fd: no message header
proxy dialer did not pass back a connection

I believe this is because the fd should be sent in some way using sendmsg, but the fd doesn’t originate from the subprocess anyways. I’d like to make it as simple as possible, and this feels close to workable.

Answer

You want to turn the client/server model on its head and make a generic server to spawn a client on-demand and in-response-to an incoming unauthenticated TCP connection from across a network boundary, and then tell that newly-spawned client to use that unauthenticated TCP session. I think that may have security considerations that you haven’t thought of. If a malicious person spams connections to your computer, your computer will spawn a lot of SSH instances to connect back and these processes can take up a lot of local system resources while authenticating. You’re effectively trying to set up SSH to automatically connect to an untrusted (unverified) remote-initiated machine across a network boundary. I can’t stress how dangerous that could be for your client computer. Using the wrong options could expose any credentials you have or even give a malicious person full access to your machine.

It’s also worth noting that the scenario you’re asking to do, building a tunnel between multiple devices to multiplex additional connections across an untrusted network boundary, is exactly the purpose of VPN software. Yes, SSH can build tunnels. VPN software can build tunnels better. The concept would be that you’d run a VPN server on your client machine. The VPN server will create a new (virtual) network interface which represents only your devices. The devices would connect to the VPN server and be assigned an IP address. Then, from the client machine, you’d just initiate SSH to the device’s VPN address and it will be routed over the virtual network interface and arrive at the device and be handled by its SSH daemon server. Then you don’t need to muck around with socat or SSH options for port forwarding. And you’d get all the tooling and tutorials that exist around VPNs. I strongly encourage you to look at VPN software.

If you really want to use SSH, then I strongly encourage you to learn about securing SSH servers. You’ve stated that the devices are across network boundaries (NAT) and that your client system is unprotected. I’m not going to stop you from shooting yourself in the foot but it would be very easy to spectacularly do so in the situation you’ve stated. If you’re in a work setting, you should talk to your system administrators to discuss firewall rules, bastion hosts, stuff like that.

Yes, you can do what you’ve stated. I strongly advise caution though. I advise it strongly enough that I won’t suggest anything which would work with that as stated. I will suggest a variant with the same concepts but more authentication.

First, you’ve effectively set up your own SSH bounce server but without any of the common tooling compatible with SSH servers. So that’s the first thing I’d fix: use SSH server software to authenticate incoming tunnel requests by using ssh client software to initiate the connection from the device instead of socat. ssh already has plenty of capabilities to create tunnels in both directions and you get authentication bundled with it (with socat, there’s no authentication). The devices should be able to authenticate using encryption keys (ssh calls these identities). You’ll need to connect once manually from the device to verify and authorize the remote encryption key fingerprint. You’ll also need to copy the public key file (NOT the private key file) to your client machine and add it to your authorized_keys files. You can ask for help on that separately if you need it.

A second issue is that you appear to be using fd3 and fd4. I don’t know why you’re doing that. If anything, you should be using fd0 and fd1 since these are stdin and stdout, respectively. But you don’t even need to do that if you’re using socat to initiate a connection. Just use - where stdin and stdout are meant. It should be completely compatible with -o ProxyCommand without specifying any file descriptors. There’s an example at the end of this answer.

The invocation from the device side might look like this (put it into a script file):

IDENTITY=/home/WavesAtParticles/.ssh/tunnel.id_rsa # on device
REMOTE_SOCKET=/home/WavesAtParticles/.ssh/$(hostname).sock # on client
REMOTEUSER=WavesAtParticles # on client
REMOTEHOST=remotehost # client hostname or IP address accessible from device
while true
do
  echo "$(date -Is) connecting"
  #
  # Set up your SSH tunnel. Check stderr for known issues.
  ssh 
    -i "${IDENTITY}" 
    -R "${REMOTE_SOCKET}:127.0.0.1:22" 
    -o ExitOnForwardFailure=yes 
    -o PasswordAuthentication=no 
    -o IdentitiesOnly=yes 
    -l "${REMOTEUSER}" 
    "${REMOTEHOST}" 
      "sleep inf" 
  2> >(
    read -r line
    if echo "${line}" | grep -q "Error: remote port forwarding failed"
    then
      ssh 
        -i "${IDENTITY}" 
        -o PasswordAuthentication=no 
        -o IdentitiesOnly=yes 
        -l "${REMOTEUSER}" 
        "${REMOTEHOST}" 
          "rm ${REMOTE_SOCKET}" 
      2>/dev/null # convince me this is wrong
      echo "$(date -Is) removed stale socket"
    fi
    #
    # Re-print stderr to the terminal
    >&2 echo "${line}" # the stderr line we checked
    >&2 cat - # and any unused stderr messages
  )

  echo "disconnected"
  sleep 30
done

Remember, copying and pasting is bad in terms of shell scripts. At a minimum, I recommend you read man ssh and man ssh_config, and to check the script against shellcheck.net. The intent of the script is:

1. In a loop, have your device (re)connect to your client to maintain your tunnel.
2. If the connection drops or fails, then reconnect every 30 seconds.
3. Run ssh with the following parameters:
   • -i "${IDENTITY}": specify a private key to use for authentication.
   • -R "${REMOTE_SOCKET}:127.0.0.1:22": specify a connection request forwarder which accept connections on the Remote side /home/WavesAtParticles/$(hostname).sock then forward them to the local side by connecting to 127.0.0.1:22.
   • -o ExitOnForwardFailure=yes: if the remote side fails to set up the connection forwarder, then the local side should emit an error and die (and we check for this error in a subshell).
   • -o PasswordAuthentication=no: do not fall back to a password request, particularly since the local user isn’t here to type it in
   • -o IdentitiesOnly=yes: do not use any default identity nor any identity offered by any local agent. Use only the one specified by -i.
   • -l "${REMOTEUSER}": log in as the specified user.
   • remotehost, eg your client machine that you want a device to connect to.
4. Sleep forever

5. If the connection failed because of a stale socket, then work around the issue by:

   • Log in separately
   • Delete the (stale) socket
   • Print today’s date indicating when it was deleted
   • Loop again

   There’s an option which is intended to make this error-handling redundant: StreamLocalBindUnlink. However the option does not correctly work and has a bug open for years. I imagine that’s because there really aren’t many people who use ssh to forward over unix domain sockets. It’s annoying but not difficult to workaround.

Using a unix domain socket should limit connectivity to whoever can reach the socket file (which should be only you and root if it’s placed in your ${HOME}/.ssh directory and the directory has correct permissions). I don’t know if that’s important for your case or not.

On the other hand you can also simplify this a lot if you’re willing to open a TCP port on 127.0.0.1 for each device. But then any other user on the same system can also connect. You should specifically listen on 127.0.0.1 which would then only accept connections from the same host to prevent external machines from reaching the forwarding port. You’d change the ${REMOTE_SOCKET} variable to, for example, 127.0.0.1:4567 to listen on port 4567 and only accept local connections. So you’d lose the named socket capability and permit any other user on the client machine to connect to your device, but gain a much simpler tunnel script (because you can remove the whole bit about parsing stderr to remove a stale socket file).

As long as your device is online (can reach your workstation’s incoming port) and is running that script, and the authentication is valid, then the tunnel should also be online or coming-online. It will take some time to recover after a loss (and restore) of network connectivity, though. You can tune that with ConnectTimeout, TCPKeepAlive, and ServerAliveInterval options and the sleep 30 part of the loop. You could run it in a tmux session to keep it going even when you don’t have a login session running. You could also run it as a system service on the device to bring it online even after recovering from a power failure.

Then from your client, you can connect in reverse:

ssh -o ProxyCommand='socat - unix-connect:/home/WavesAtParticles/remotehost.sock' -l WavesAtParticles .

In this invocation, you’ll start ssh. It will then set up the proxycommand using socat. It will take its stdin/stdout and relay it through a connected AF_UNIX socket at the path provided. You’ll need to update the path for the remote host you expect. But there’s no need to specify file descriptors at all. If ssh complains:

2019/08/26 18:09:52 socat[29914] E connect(5, AF=1 "/home/WavesAtParticles/remotehost.sock", 7): Connection refused
ssh_exchange_identification: Connection closed by remote host

then the tunnel is currently down and you should investigate the remotehost device’s connectivity.

If you use the remote forwarding option with a TCP port listening instead of a unix domain socket, then the client-through-tunnel-to-remote invocation becomes even easier: ssh -p 4567 WavesAtParticles@localhost.

Again, you’re trying to invert the client/server model and I don’t think that’s a very good idea to do with SSH.