HAProxy community

How to get HAProxy to route TCP based on SNI (using openssl s_client to test)?

I want to use HAProxy to terminate TLS-encrypted TCP connnections and to pass the unencrypted TCP traffic to various backends based on the Server Name Indication used to initiate the TLS connection.

I have 3 services running on a backend server, each on a different port (5001, 5002, 5003). HAProxy binds to port 5000. I’d like to route connections to the first 2 servies by name or to the third if there is not a match. I am initiating the connecton to HAProxy using openssl s_client. However, in the logs I can see that the connections are only ever routed to the default server, i.e. the SNI seems to be ignored.

DNS:

A record     demo.sni.example.com     1.2.3.4
CNAME        1.sni.example.com        pointing to demo.sni.example.com   
CNAME        2.sni.example.com        pointing to demo.sni.example.com   

i.e. I’d like the following routing to occur:

SNI = 1.sni.example.com:5000 -> 1.2.3.4:5001
SNI = 2.sni.example.com:5000 -> 1.2.3.4:5002
anything else on port 5000 -> 1.2.3.4:5003

haproxy.cfg:

global
  log stdout  format raw  local0 info

defaults
  timeout client 30s
  timeout server 30s
  timeout connect 5s
  option tcplog

frontend tcp-proxy
  bind :5000 ssl crt combined-cert-key.pem
  mode tcp
  log global
  tcp-request inspect-delay 5s
  tcp-request content accept if { req_ssl_hello_type 1 }

  use_backend bk_sni_1 if { req.ssl_sni -i 1.sni.example.com }
  use_backend bk_sni_2 if { req.ssl_sni -i 2.sni.example.com }
  default_backend bk_default

backend bk_sni_1
  mode tcp
  log global
  balance roundrobin
  server server1 1.2.3.4:5001 check

backend bk_sni_2
  mode tcp
  log global
  balance roundrobin
  server server1 1.2.3.4:5002 check

backend bk_default
  mode tcp
  log global
  balance roundrobin
  server server1 1.2.3.4:5003 check

combined-cert-key.pem is a self-signed certificate file plus key where the CN is the IP of the server (1.2.3.4) and there are SANs of all DNS values and the IP.

Connections initiated using openssl s_client:

I have tried connecting via DNS (A & CNAME records, as well as IP):

echo test | openssl s_client -connect demo.sni.example.com:5000 -servername 1.sni.example.com
echo test | openssl s_client -connect 1.sni.example.com:5000 -servername 1.sni.example.com
echo test | openssl s_client -connect 1.2.3.4:5000 -servername 1.sni.example.com

However, all connections are routed to the default backend bk_default.

What is causing HAProxy to not recognise the SNI servername? (I am using the latest HAProxy docker image: https://hub.docker.com/_/haproxy)

The answer is to use ssl_fc-sni, instead of req.ssl_sni. The former is for SSL-terminated sessions, whereas the latter is for sessions where TCP is passed straight through.

Correct, also please remove tcp-request rules and make sure you have dedicated certificates for each of those names, otherwise - if this is HTTP and browsers are the clients - you will continue to route to wrong backends when browser keepalive HTTP within the TLS sessions (there is only one SNI value in the client hello - if your TLS sessions covers multiple transactions they may intent do go to different backends if the certificate permits that).