x/crypto/ssh: TCP/IP port forwarding expects IP addresses

[tt]

(*ssh.Client).ListenTCP expects an IP address (via *net.TCPAddr) and therefore (*ssh.Client).Listen attempts to resolve addresses.

However, section 7.1 of RFC 4254 states:

The 'address to bind' and 'port number to bind' specify the IP
address (or domain name) and port on which connections for forwarding
are to be accepted. Some strings used for 'address to bind' have
special-case semantics.

• "" means that connections are to be accepted on all protocol
  families supported by the SSH implementation.

• "0.0.0.0" means to listen on all IPv4 addresses.

• "::" means to listen on all IPv6 addresses.

• "localhost" means to listen on all protocol families supported by
  the SSH implementation on loopback addresses only ([RFC3330] and
  [RFC3513]).

• "127.0.0.1" and "::1" indicate listening on the loopback
  interfaces for IPv4 and IPv6, respectively.

There are two consequences of the current interface:

1. You can only provide resolvable names. This prohibits two of the strings with special-case semantics from working ("", reported in x/crypto/ssh: listening on remote address blocks indefinitely when address doesn't contain a host part #33227, and "::").

2. Resolution happens client side. This changes the meaning of the string "localhost" from being "all protocol families supported by the SSH implementation on loopback addresses only" to being only one of those and may provide a different result for other names (AWS hostnames resolving to internal addresses inside a data center comes to mind).

Outside of defining a new public interface, I think the least breaking change would be to extract an unexported listenTCP function taking a string address and call this from Listen which can then drop resolution but of course if you're relying on that behavior, it will still be surprising.

I'm happy to submit a pull request but I'd appreciate some thoughts on how to best evolve the interface into something that both supports the scope of the RFC and doesn't disregard current users.

[tt]

I should mention that it's of course possible to treat "", "::" and "localhost" separately with even less potential breakage but this doesn't solve the problem of client side resolution generally which was my intention for filing this issue.

[toothrot]

/cc @hanwen @FiloSottile

[nya3jp]

+1. I'm particularly interested in localhost case so that I don't need to create two listeners for IPv4/IPv6.

[luizfeliperj]

+1 The address should be resolved by the remote host, as is where the listen() call will happen, not the originating ssh client.
Please fix this!

[luizfeliperj]

Just adding some more information, supposing it is a multi-homed server, with multiple IP addresses, calling net. Listen () using a hostname needs to resolve the name within the remote server to figure out which interface it should bind. The current implementation doesn't allow this.

[tiger5226]

wow...this is so insurmountable....

SSH Protocol : "channel forward message address is a string"
Golang: "Noooooo we think it should be a Golang IP, which we will convert to a string :)"

All the above, plus I would add that a domain might not be "resolvable" from the client side in instances where a remote forward tcpip-forward is created. "resolvable" should not be a requirement of a domain client side.

[rahmir-fabrice]

+1 This should be resolved on the remote host and treated as a string client side to allow the remote host to preform the lookup on what hostname qualifies to access the connection

[eslym]

or just simply not to handle forwarded-tcpip channel before any ListenTCP called, so we can implement our own handling for this case

update:
i read the source code, the handleForwards does only called once ListenTCP or ListenUnix is called (at least in the latest version)

[gopherbot]

Change https://go.dev/cl/599995 mentions this issue: ssh: allow to bind to a hostname in remote forwarding

[drakkan]

Hello, can you please confirm that the above CL fixes the issue? Thank you