Copy data through SSH tunnel over multiple hops

Question

We have two main environments in question:

Development and QA

Each environment has two servers:

• Jump Box
• Application server

In order to connect to the application server, you must connect to the jump box first, and then SSH to the Application server.

There are a few rules in place courtesy of the firewall:

• You MUST connect to the application server via the jump box
• The application server cannot connect to either jump boxes
• The jump boxes are on the same subnet, and CAN talk to each other.

Our Problem

We have a lot of content (670 GB) on the DEVELOPMENT APPLICATION SERVER, and we need to get this to the QA APPLICATION SERVER.

Copying this data to the jump boxes is not an option because they lack the required amount of space.

I did some research, and learned that we could potentially do a series of tunnels through these servers so that we can stream the data straight from one app server to the other via the tunnels. However, the problem that we cannot connect to the jump box from the application server.

Do we have any options? This is getting to be a desperate situation, and time is of the essence. We do not have time to download the data and re-upload it. Copying across the network on the servers will go quickly, as it is a gigabit connection.

Solution 1

By far, the easiest way is to just copy it via scp. Plus, this syntax actually works unlike some of the other suggestions.

You can't beat this syntax for ease. It allows you to recursively copy, rsync or what ever you'd like without the hassle of considering potentially complex pipes. This syntax is intuitively clear, will be more readily supportable by Sys Admins that follow you and does not make useless use of cat.

scp -3 devappserver:/path/to/copy/from qaappserver:/path/to/copy/to

From the scp man page: -3 Copies between two remote hosts are transferred through the local host. Without this option the data is copied directly between the two remote hosts. Note that this option disables the progress meter.

In the example below

• Your workstation is named MacBook-Pro.
• Dev Jump Box is named devjumpserver
• Dev Application Server is named devapplicationserver
  • Is on LAN DNS zone named .local
• QA Jump Box is named qajumpserver
• QA Application Server is named qaapplicationserver
  • Is on LAN DNZ zone named .local
• We'll perform a test copy of a 670GB /etc/hosts file ;-)
• Assumption is made that you have SSH public key authentication configured.

Here is an ~/.ssh/config file that sets up the direct access from your workstation to the application servers via the appropriate jump (aka bastion server).

MacBook-Pro:~ barrychapman$ cat ~/.ssh/config
Host *
  ServerAliveInterval 60
Host devapplicationsever
  HostName devapplicationserver.local
  ProxyCommand ssh -i ~/.ssh/id_rsa [email protected] -W %h:%p
  User barrychapman
Host qaapplicationserver
  HostName qaapplicationserver.local
  ProxyCommand ssh -i ~/.ssh/id_rsa [email protected] -W %h:%p
  User barrychapman

MacBook-Pro:~ barrychapman$

Testing for presence of file on target server, it won't be there.

MacBook-Pro:~ barrychapman$ ssh qaapplicationserver ls /tmp/hosts
ls: cannot access /tmp/hosts: No such file or directory
Killed by signal 1.
MacBook-Pro:~ barrychapman$

Now let's copy a file from Dev Application server to QA Application via your workstation.

MacBook-Pro:~ barrychapman$ scp -3 devapplicationserver:/etc/hosts qaapplicationserver:/tmp/
Killed by signal 1.
Killed by signal 1.
MacBook-Pro:~ barrychapman$

Now let's check for the presence of the copied file on the QA Application Server. It will be there this time.

MacBook-Pro:~ barrychapman$ ssh qaapplicationserver ls /tmp/hosts
/tmp/hosts
Killed by signal 1.
MacBook-Pro:~ barrychapman$ 

Note

When closing a ProxyCommand connection, you will see the warning message "Killed by signal 1". This is SSH tearing down the ProxyCommand connection and is nothing to be alarmed about. You can get rid of it by adding LogLevel Quiet to your bastion host config stanza.

Solution 2

PIPES!

If the internet is a series of tubes, Unix is a series of pipes -- something like:

cat ginormous-file | ssh user@host1 "cat | ssh user@host2 \"cat >out\" "

should work.

If you need to traverse more hosts, add more pipes (and more nested layers of \-escaped quotation) as needed. (Note however that if the pipeline/escaping gets so complex that you have to draw a diagram or resort to counting on your fingers to determine how many times you have to double up on escapes it's probably time to admit defeat and set up a proper VPN!)

Solution 3

OpenSSH v7.3 onward supports -J.

-J [user@]host[:port]
Connect to the target host by first making a ssh connection to the jump host and then establishing a TCP forwarding to the ultimate destination from there. Multiple jump hops may be specified separated by comma characters. This is a shortcut to specify a ProxyJump configuration directive.

For A → B → C, on A, just:

tar -cf - file1 file_n | pv | ssh -C -J userB@B:portB userB@C -p portC 'tar -C destDir -xvf -'

• Use as many hops as you want with ssh's -J option.
• Omit the tar's remote -C to leave the files on home folder.
• Send any files at once (text or binary).
• Increase speed by compressing the stream with or ssh's -C (or tar's -z). Particularly useful if the data is plain text (uncompressed).
• pv monitor the progress of data through a pipe. An alternative could be progress.

Inspired on Florian Fida and Dan Garthwaite's answers.

Solution 4

If I understand correctly, you have two jump servers (jump-qa and jump-dev) protecting two app servers (app-qa and app-dev); the jump servers can ssh to each other; no box other than the relevant jump server can ssh to the corresponding app server. The app servers can ssh to noone. A file is to be transferred from app-dev to app-qa. Both jump servers lack the space for an interim copy of the data.

You can solve this with ssh tunneling. We set up a connection to one remote app server, carrying a remote tunnel that connects back to an unused port on its jump server. We set up a second connection from one jump server to the other jump server, carrying a tunnel that picks up the dangling end of the remotely-forwarded port from tunnel one and sends it on to the other app server's ssh port.

Set up the tunnels (each of these commmands will need to be run in a separate window on jump-qa):

jump-qa% ssh app-qa -R 2345:localhost:2346
jump-qa% ssh jump-dev -L 2346:app-dev:22

You should now find that on app-qa, you can do telnet localhost 2345 and get app-dev's ssh banner. You may then copy the datafile:

app-qa% scp -P 2345 localhost:/path/on/app-dev/data.dat data.dat

View more solutions

Comments

[Alex_www]

Once connection is established, you may copy either way: $devel_host $ tar -cf - | ssh -t jumbox 'ssh app_serv "tar -xf -" ' or other way around tar.

[Mohit Gupta]

so we can connect from dev jump box to dev app server, but not the other way. If we have that connection established, we can copy either way? what is the best way to move this content to the other app server without first saving it on the jump boxes?

[Alex_www]

Exactly, "piping" "tar" is a simplest solution.

[MadHatter]

Does client have space for an interim copy?

[Mohit Gupta]

No, there is not enough space

[Mohit Gupta]

when you say client, what server are you referring to?

[MadHatter]

I had misunderstood. My current understanding is that there is no client; there are two jump servers, and two app servers.

[Mohit Gupta]

The first step works, however, when we try to ssh to the jump-dev via -L 2346:app-dev:22, it times out.

[MadHatter]

You said you can ssh from jump-qa to jump-dev. Is this not so? I also don't get what you mean by "via -L 2346:app-dev:22". Please make sure you have read and understood what I said about doing these two commands from two separate windows on jump-qa. They are not a pair of commands to be typed into the same window in sequence.

[Mohit Gupta]

jump-qa% ssh jump-dev -L 2346:app-dev:22 - when we execute that command (with our parameters) it times out. However, we are getting: Warning: remote port forwarding failed for listen port xxxx

[MadHatter]

What happens if you just do jump-qa% ssh jump-dev?

[Mohit Gupta]

You are right, it is failing from QA -> DEV but not from DEV -> QA. Is there a way we can operate within that?

[MadHatter]

Yes, the situation is symmetric. Change all the dev to qa, and vice-versa. The final copy will also become local to remote, instead of remote to local.

[Mohit Gupta]

Does it matter that ONLY port 22 is open between the jump servers? We are getting: Destination network unreachable

[MadHatter]

From where, doing what?

[MadHatter]

You have assured me that dev jump box can ssh to qa jump box. Is this not so?

[MadHatter]

I have five more minutes, then I have to go. Let us continue this discussion in chat

[voretaq7]

As Alex_www pointed out in his comment if you don't need the intermediate file for anything you can also just pipe the output of tar around. (Also in you don't need the cat in the intermediate stages of the pipeline - ssh is happy to eat stdin and relay it. The cat just makes me feel better and is a placeholder for other useful commands you might want to use, like tee.)

[MadHatter]

I think the OP's problem is there is no machine that both has the data and can make the connections, nor is there any one machine that can see everything and everyone (to act as connector for all the pipes).

[voretaq7]

@MadHatter If that's the case a combination of our two answers would work (connect to the dev server, forwarding a port back to the jump server's SSH port, then run the SSH pipeline over that port). This is of course making the solution increasingly disgusting to the point where people will start protesting outside your office with big signs that say VPN! NOW! on them...

[MadHatter]

Why, yes. Yes, they would. Hopefully!

[Mohit Gupta]

Cheers, this solved our problem. Merry Christmas!

[Andrew Wolfe]

What do you mean by "bastion host config stanza" ?? @BraveNewCurrency ?

[BraveNewCurrency]

I consider each "Host" line and the configs under it to be a stanza (like in poetry.) If you add "LogLevel Quiet" under your bastion "Host" line, it will only apply to that host.

[hello_there_andy]

Will doing this mean the intermediate server (in this case user@host1) will at some point have the full cat ginormous-file in storage at any point? Or is the data just sent directly to user@host2? Or is it somehow streamed? How is tar relevant to this? I guess it's relevant to the second to last question I asked. None of these questions are rhetorical btw...

[hello_there_andy]

In my case, the intermediate user@host1 server ran out of storage when scp'ing from from the user@host2. And I can't scp directly from user@host2 to user@local (I need to do two scps: user@host2 --> user@host1 --> user@local, each one manually). But I don't have a clue what my user@local's IP address is because it's a guest ubuntu VM in a Win10 host... I can ssh from user@local --> user@host1 --> user@host2, just not the other way (i.e. user@host1 --> user@local I can't do, since I do not know user@local's IP address, working on it...)

[Mohit Gupta]

This project is long gone, but this solution will undoubtedly serve me (and anyone else) very well in the future. Thank you for the addition! This is great to know.

[Federal Reserve]

I found this answer in 2022, It's god damn good, how can I didn't know this before.