go-ethereum/cmd/clef/docs/setup.md

Setting up Clef

This document describes how Clef can be used in a more secure manner than executing it from your everyday laptop, in order to ensure that the keys remain safe in the event that your computer should get compromised.

Qubes OS

Background

The Qubes operating system is based around virtual machines (qubes), where a set of virtual machines are configured, typically for different purposes such as:

• personal
  • Your personal email, browsing etc
• work
  • Work email etc
• vault
  • a VM without network access, where gpg-keys and/or keepass credentials are stored.

A couple of dedicated virtual machines handle externalities:

• sys-net provides networking to all other (network-enabled) machines
• sys-firewall handles firewall rules
• sys-usb handles USB devices, and can map usb-devices to certain qubes.

The goal of this document is to describe how we can set up clef to provide secure transaction signing from a vault vm, to another networked qube which runs Dapps.

Setup

There are two ways that this can be achieved: integrated via Qubes or integrated via networking.

1. Qubes Integrated

Qubes provides a facility for inter-qubes communication via qrexec. A qube can request to make a cross-qube RPC request to another qube. The OS then asks the user if the call is permitted.

A policy-file can be created to allow such interaction. On the target domain, a service is invoked which can read the stdin from the client qube.

This is how Split GPG is implemented. We can set up Clef the same way:

Server

On the target qubes, we need to define the RPC service.

qubes.Clefsign:

#!/bin/bash

SIGNER_BIN="/home/user/tools/clef/clef"
SIGNER_CMD="/home/user/tools/gtksigner/gtkui.py -s $SIGNER_BIN"

# Start clef if not already started
if [ ! -S /home/user/.clef/clef.ipc ]; then
	$SIGNER_CMD &
	sleep 1
fi

# Should be started by now
if [ -S /home/user/.clef/clef.ipc ]; then
    # Post incoming request to HTTP channel
	curl -H "Content-Type: application/json" -X POST -d @- http://localhost:8550 2>/dev/null
fi

This RPC service is not complete (see notes about HTTP headers below), but works as a proof-of-concept. It will forward the data received on stdin (forwarded by the OS) to Clef's HTTP channel.

It would have been possible to send data directly to the /home/user/.clef/.clef.ipc socket via e.g nc -U /home/user/.clef/clef.ipc, but the reason for sending the request data over HTTP instead of IPC is that we want the ability to forward HTTP headers.

To enable the service:

sudo cp qubes.Clefsign /etc/qubes-rpc/
sudo chmod +x /etc/qubes-rpc/ qubes.Clefsign

This setup uses gtksigner, which is a very minimal GTK-based UI that works well with minimal requirements.

Client

On the client qube, we need to create a listener which will receive the request from the Dapp, and proxy it.

qubes-client.py:

"""
This implements a dispatcher which listens to localhost:8550, and proxies
requests via qrexec to the service qubes.EthSign on a target domain
"""

import http.server
import socketserver,subprocess

PORT=8550
TARGET_DOMAIN= 'debian-work'

class Dispatcher(http.server.BaseHTTPRequestHandler):
    def do_POST(self):
        post_data = self.rfile.read(int(self.headers['Content-Length']))
        p = subprocess.Popen(['/usr/bin/qrexec-client-vm',TARGET_DOMAIN,'qubes.Clefsign'],stdin=subprocess.PIPE, stdout=subprocess.PIPE)
        output = p.communicate(post_data)[0]
        self.wfile.write(output)


with socketserver.TCPServer(("",PORT), Dispatcher) as httpd:
    print("Serving at port", PORT)
    httpd.serve_forever()

Testing

To test the flow, if we have set up debian-work as the target, we can do

$ cat newaccnt.json 
{ "id": 0, "jsonrpc": "2.0","method": "account_new","params": []}

$ cat newaccnt.json| qrexec-client-vm debian-work qubes.Clefsign

A dialog should pop up first to allow the IPC call:

Followed by a GTK-dialog to approve the operation:

To test the full flow, we use the client wrapper. Start it on the client qube:

[user@work qubes]$ python3 qubes-client.py 

Make the request over http (client qube):

[user@work clef]$ cat newaccnt.json | curl -X POST -d @- http://localhost:8550

And it should show the same popups again.

Pros and cons

The benefits of this setup are:

• This is the qubes-os intended model for inter-qube communication,
• and thus benefits from qubes-os dialogs and policies for user approval

However, it comes with a couple of drawbacks:

• The qubes-gpg-client must forward the http request via RPC to the target qube. When doing so, the proxy will either drop important headers, or replace them.
  • The Host header is most likely localhost
  • The Origin header must be forwarded
  • Information about the remote ip must be added as a X-Forwarded-For. However, Clef cannot always trust an XFF header, since malicious clients may lie about XFF in order to fool the http server into believing it comes from another address.
• Even with a policy in place to allow RPC calls between caller and target, there will be several popups:
  • One qubes-specific where the user specifies the target vm
  • One clef-specific to approve the transaction

2. Network integrated

The second way to set up Clef on a qubes system is to allow networking, and have Clef listen to a port which is accessible from other qubes.

USBArmory

The USB armory is an open source hardware design with an 800 Mhz ARM processor. It is a pocket-size computer. When inserted into a laptop, it identifies itself as a USB network interface, basically adding another network to your computer. Over this new network interface, you can SSH into the device.

Running Clef off a USB armory means that you can use the armory as a very versatile offline computer, which only ever connects to a local network between your computer and the device itself.

Needless to say, while this model should be fairly secure against remote attacks, an attacker with physical access to the USB Armory would trivially be able to extract the contents of the device filesystem.