doc/proxy-protocol.txt

2014/06/14                                                        Willy Tarreau
                                                           HAProxy Technologies
                               The PROXY protocol
                                 Versions 1 & 2

Abstract

   The PROXY protocol provides a convenient way to safely transport connection
   information such as a client's address across multiple layers of NAT or TCP
   proxies. It is designed to require little changes to existing components and
   to limit the performance impact caused by the processing of the transported
   information.


Revision history

   2010/10/29 - first version
   2011/03/20 - update: implementation and security considerations
   2012/06/21 - add support for binary format
   2012/11/19 - final review and fixes
   2014/05/18 - modify and extend PROXY protocol version 2
   2014/06/11 - fix example code to consider ver+cmd merge
   2014/06/14 - fix v2 header check in example code, and update Forwarded spec


1. Background

Relaying TCP connections through proxies generally involves a loss of the
original TCP connection parameters such as source and destination addresses,
ports, and so on. Some protocols make it a little bit easier to transfer such
information. For SMTP, Postfix authors have proposed the XCLIENT protocol [1]
which received broad adoption and is particularly suited to mail exchanges.
For HTTP, there is the "Forwarded" extension [2], which aims at replacing the
omnipresent "X-Forwarded-For" header which carries information about the
original source address, and the less common X-Original-To which carries
information about the destination address.

However, both mechanisms require a knowledge of the underlying protocol to be
implemented in intermediaries.

Then comes a new class of products which we'll call "dumb proxies", not because
they don't do anything, but because they're processing protocol-agnostic data.
Both Stunnel[3] and Stud[4] are examples of such "dumb proxies". They talk raw
TCP on one side, and raw SSL on the other one, and do that reliably, without
any knowledge of what protocol is transported on top of the connection. Haproxy
running in pure TCP mode obviously falls into that category as well.

The problem with such a proxy when it is combined with another one such as
haproxy, is to adapt it to talk the higher level protocol. A patch is available
for Stunnel to make it capable of inserting an X-Forwarded-For header in the
first HTTP request of each incoming connection. Haproxy is able not to add
another one when the connection comes from Stunnel, so that it's possible to
hide it from the servers.

The typical architecture becomes the following one :


      +--------+      HTTP                      :80 +----------+
      | client |  --------------------------------> |          |
      |        |                                    | haproxy, |
      +--------+             +---------+            |  1 or 2  |
     /        /     HTTPS    | stunnel |  HTTP  :81 | listening|
    <________/    ---------> | (server | ---------> |  ports   |
                             |  mode)  |            |          |
                             +---------+            +----------+


The problem appears when haproxy runs with keep-alive on the side towards the
client. The Stunnel patch will only add the X-Forwarded-For header to the first
request of each connection and all subsequent requests will not have it. One
solution could be to improve the patch to make it support keep-alive and parse
all forwarded data, whether they're announced with a Content-Length or with a
Transfer-Encoding, taking care of special methods such as HEAD which announce
data without transfering them, etc... In fact, it would require implementing a
full HTTP stack in Stunnel. It would then become a lot more complex, a lot less
reliable and would not anymore be the "dumb proxy" that fits every purposes.

In practice, we don't need to add a header for each request because we'll emit
the exact same information every time : the information related to the client
side connection. We could then cache that information in haproxy and use it for
every other request. But that becomes dangerous and is still limited to HTTP
only.

Another approach consists in prepending each connection with a header reporting
the characteristics of the other side's connection. This method is simpler to
implement, does not require any protocol-specific knowledge on either side, and
completely fits the purpose since what is desired precisely is to know the
other side's connection endpoints. It is easy to perform for the sender (just
send a short header once the connection is established) and to parse for the
receiver (simply perform one read() on the incoming connection to fill in
addresses after an accept). The protocol used to carry connection information
across proxies was thus called the PROXY protocol.


2. The PROXY protocol header

This document uses a few terms that are worth explaining here :
  - "connection initiator" is the party requesting a new connection
  - "connection target" is the party accepting a connection request
  - "client" is the party for which a connection was requested
  - "server" is the party to which the client desired to connect
  - "proxy" is the party intercepting and relaying the connection
    from the client to the server.
  - "sender" is the party sending data over a connection.
  - "receiver" is the party receiving data from the sender.
  - "header" or "PROXY protocol header" is the block of connection information
    the connection initiator prepends at the beginning of a connection, which
    makes it the sender from the protocol point of view.

The PROXY protocol's goal is to fill the server's internal structures with the
information collected by the proxy that the server would have been able to get
by itself if the client was connecting directly to the server instead of via a
proxy. The information carried by the protocol are the ones the server would
get using getsockname() and getpeername() :
  - address family (AF_INET for IPv4, AF_INET6 for IPv6, AF_UNIX)
  - socket protocol (SOCK_STREAM for TCP, SOCK_DGRAM for UDP)
  - layer 3 source and destination addresses
  - layer 4 source and destination ports if any

Unlike the XCLIENT protocol, the PROXY protocol was designed with limited
extensibility in order to help the receiver parse it very fast. Version 1 was
focused on keeping it human-readable for better debugging possibilities, which
is always desirable for early adoption when few implementations exist. Version
2 adds support for a binary encoding of the header which is much more efficient
to produce and to parse, especially when dealing with IPv6 addresses that are
expensive to emit in ASCII form and to parse.

In both cases, the protocol simply consists in an easily parsable header placed
by the connection initiator at the beginning of each connection. The protocol
is intentionally stateless in that it does not expect the sender to wait for
the receiver before sending the header, nor the receiver to send anything back.

This specification supports two header formats, a human-readable format which
is the only format supported in version 1 of the protocol, and a binary format
which is only supported in version 2. Both formats were designed to ensure that
the header cannot be confused with common higher level protocols such as HTTP,
SSL/TLS, FTP or SMTP, and that both formats are easily distinguishable one from
each other for the receiver.

Version 1 senders MAY only produce the human-readable header format. Version 2
senders MAY only produce the binary header format. Version 1 receivers MUST at
least implement the human-readable header format. Version 2 receivers MUST at
least implement the binary header format, and it is recommended that they also
implement the human-readable header format for better interoperability and ease
of upgrade when facing version 1 senders.

Both formats are designed to fit in the smallest TCP segment that any TCP/IP
host is required to support (576 - 40 = 536 bytes). This ensures that the whole
header will always be delivered at once when the socket buffers are still empty
at the beginning of a connection. The sender must always ensure that the header
is sent at once, so that the transport layer maintains atomicity along the path
to the receiver. The receiver may be tolerant to partial headers or may simply
drop the connection when receiving a partial header. Recommendation is to be
tolerant, but implementation constraints may not always easily permit this. It
is important to note that nothing forces any intermediary to forward the whole
header at once, because TCP is a streaming protocol which may be processed one
byte at a time if desired, causing the header to be fragmented when reaching
the receiver. But due to the places where such a protocol is used, the above
simplification generally is acceptable because the risk of crossing such a
device handling one byte at a time is close to zero.

The receiver MUST NOT start processing the connection before it receives a
complete and valid PROXY protocol header. This is particularly important for
protocols where the receiver is expected to speak first (eg: SMTP, FTP or SSH).
The receiver may apply a short timeout and decide to abort the connection if
the protocol header is not seen within a few seconds (at least 3 seconds to
cover a TCP retransmit).

The receiver MUST be configured to only receive the protocol described in this
specification and MUST not try to guess whether the protocol header is present
or not. This means that the protocol explicitly prevents port sharing between
public and private access. Otherwise it would open a major security breach by
allowing untrusted parties to spoof their connection addresses. The receiver
SHOULD ensure proper access filtering so that only trusted proxies are allowed
to use this protocol.

Some proxies are smart enough to understand transported protocols and to reuse
idle server connections for multiple messages. This typically happens in HTTP
where requests from multiple clients may be sent over the same connection. Such
proxies MUST NOT implement this protocol on multiplexed connections because the
receiver would use the address advertised in the PROXY header as the address of
all forwarded requests's senders. In fact, such proxies are not dumb proxies,
and since they do have a complete understanding of the transported protocol,
they MUST use the facilities provided by this protocol to present the client's
address.


2.1. Human-readable header format (Version 1)

This is the format specified in version 1 of the protocol. It consists in one
line of ASCII text matching exactly the following block, sent immediately and
at once upon the connection establishment and prepended before any data flowing
from the sender to the receiver :

  - a string identifying the protocol : "PROXY" ( \x50 \x52 \x4F \x58 \x59 )
    Seeing this string indicates that this is version 1 of the protocol.

  - exactly one space : " " ( \x20 )

  - a string indicating the proxied INET protocol and family. As of version 1,
    only "TCP4" ( \x54 \x43 \x50 \x34 ) for TCP over IPv4, and "TCP6"
    ( \x54 \x43 \x50 \x36 ) for TCP over IPv6 are allowed. Other, unsupported,
    or unknown protocols must be reported with the name "UNKNOWN" ( \x55 \x4E
    \x4B \x4E \x4F \x57 \x4E ). For "UNKNOWN", the rest of the line before the
    CRLF may be omitted by the sender, and the receiver must ignore anything
    presented before the CRLF is found. Note that an earlier version of this
    specification suggested to use this when sending health checks, but this
    causes issues with servers that reject the "UNKNOWN" keyword. Thus is it
    now recommended not to send "UNKNOWN" when the connection is expected to
    be accepted, but only when it is not possible to correctly fill the PROXY
    line.

  - exactly one space : " " ( \x20 )

  - the layer 3 source address in its canonical format. IPv4 addresses must be
    indicated as a series of exactly 4 integers in the range [0..255] inclusive
    written in decimal representation separated by exactly one dot between each
    other. Heading zeroes are not permitted in front of numbers in order to
    avoid any possible confusion with octal numbers. IPv6 addresses must be
    indicated as series of 4 hexadecimal digits (upper or lower case) delimited
    by colons between each other, with the acceptance of one double colon
    sequence to replace the largest acceptable range of consecutive zeroes. The
    total number of decoded bits must exactly be 128. The advertised protocol
    family dictates what format to use.

  - exactly one space : " " ( \x20 )

  - the layer 3 destination address in its canonical format. It is the same
    format as the layer 3 source address and matches the same family.

  - exactly one space : " " ( \x20 )

  - the TCP source port represented as a decimal integer in the range
    [0..65535] inclusive. Heading zeroes are not permitted in front of numbers
    in order to avoid any possible confusion with octal numbers.

  - exactly one space : " " ( \x20 )

  - the TCP destination port represented as a decimal integer in the range
    [0..65535] inclusive. Heading zeroes are not permitted in front of numbers
    in order to avoid any possible confusion with octal numbers.

  - the CRLF sequence ( \x0D \x0A )


The maximum line lengths the receiver must support including the CRLF are :
  - TCP/IPv4 :
      "PROXY TCP4 255.255.255.255 255.255.255.255 65535 65535\r\n"
    => 5 + 1 + 4 + 1 + 15 + 1 + 15 + 1 + 5 + 1 + 5 + 2 = 56 chars

  - TCP/IPv6 :
      "PROXY TCP6 ffff:f...f:ffff ffff:f...f:ffff 65535 65535\r\n"
    => 5 + 1 + 4 + 1 + 39 + 1 + 39 + 1 + 5 + 1 + 5 + 2 = 104 chars

  - unknown connection (short form) :
      "PROXY UNKNOWN\r\n"
    => 5 + 1 + 7 + 2 = 15 chars

  - worst case (optional fields set to 0xff) :
      "PROXY UNKNOWN ffff:f...f:ffff ffff:f...f:ffff 65535 65535\r\n"
    => 5 + 1 + 7 + 1 + 39 + 1 + 39 + 1 + 5 + 1 + 5 + 2 = 107 chars

So a 108-byte buffer is always enough to store all the line and a trailing zero
for string processing.

The receiver must wait for the CRLF sequence before starting to decode the
addresses in order to ensure they are complete and properly parsed. If the CRLF
sequence is not found in the first 107 characters, the receiver should declare
the line invalid. A receiver may reject an incomplete line which does not
contain the CRLF sequence in the first atomic read operation. The receiver must
not tolerate a single CR or LF character to end the line when a complete CRLF
sequence is expected.

Any sequence which does not exactly match the protocol must be discarded and
cause the receiver to abort the connection. It is recommended to abort the
connection as soon as possible so that the sender gets a chance to notice the
anomaly and log it.

If the announced transport protocol is "UNKNOWN", then the receiver knows that
the sender speaks the correct PROXY protocol with the appropriate version, and
SHOULD accept the connection and use the real connection's parameters as if
there were no PROXY protocol header on the wire. However, senders SHOULD not
use the "UNKNOWN" protocol when they are the initiators of outgoing connections
because some receivers may reject them. When a load balancing proxy has to send
health checks to a server, it SHOULD build a valid PROXY line which it will
fill with a getsockname()/getpeername() pair indicating the addresses used. It
is important to understand that doing so is not appropriate when some source
address translation is performed between the sender and the receiver.

An example of such a line before an HTTP request would look like this (CR
marked as "\r" and LF marked as "\n") :

    PROXY TCP4 192.168.0.1 192.168.0.11 56324 443\r\n
    GET / HTTP/1.1\r\n
    Host: 192.168.0.11\r\n
    \r\n

For the sender, the header line is easy to put into the output buffers once the
connection is established. Note that since the line is always shorter than an
MSS, the sender is guaranteed to always be able to emit it at once and should
not even bother handling partial sends. For the receiver, once the header is
parsed, it is easy to skip it from the input buffers. Please consult section 9
for implementation suggestions.


2.2. Binary header format (version 2)

Producing human-readable IPv6 addresses and parsing them is very inefficient,
due to the multiple possible representation formats and the handling of compact
address format. It was also not possible to specify address families outside
IPv4/IPv6 nor non-TCP protocols. Another drawback of the human-readable format
is the fact that implementations need to parse all characters to find the
trailing CRLF, which makes it harder to read only the exact bytes count. Last,
the UNKNOWN address type has not always been accepted by servers as a valid
protocol because of its imprecise meaning.

Version 2 of the protocol thus introduces a new binary format which remains
distinguishable from version 1 and from other commonly used protocols. It was
specially designed in order to be incompatible with a wide range of protocols
and to be rejected by a number of common implementations of these protocols
when unexpectedly presented (please see section 7). Also for better processing
efficiency, IPv4 and IPv6 addresses are respectively aligned on 4 and 16 bytes
boundaries.

The binary header format starts with a constant 12 bytes block containing the
protocol signature :

   \x0D \x0A \x0D \x0A \x00 \x0D \x0A \x51 \x55 \x49 \x54 \x0A

Note that this block contains a null byte at the 5th position, so it must not
be handled as a null-terminated string.

The next byte (the 13th one) is the protocol version and command.

The highest four bits contains the version. As of this specification, it must
always be sent as \x2 and the receiver must only accept this value.

The lowest four bits represents the command :
  - \x0 : LOCAL : the connection was established on purpose by the proxy
    without being relayed. The connection endpoints are the sender and the
    receiver. Such connections exist when the proxy sends health-checks to the
    server. The receiver must accept this connection as valid and must use the
    real connection endpoints and discard the protocol block including the
    family which is ignored.

  - \x1 : PROXY : the connection was established on behalf of another node,
    and reflects the original connection endpoints. The receiver must then use
    the information provided in the protocol block to get original the address.

  - other values are unassigned and must not be emitted by senders. Receivers
    must drop connections presenting unexpected values here.

The 14th byte contains the transport protocol and address family. The highest 4
bits contain the address family, the lowest 4 bits contain the protocol.

The address family maps to the original socket family without necessarily
matching the values internally used by the system. It may be one of :

  - 0x0 : AF_UNSPEC : the connection is forwarded for an unknown, unspecified
    or unsupported protocol. The sender should use this family when sending
    LOCAL commands or when dealing with unsupported protocol families. The
    receiver is free to accept the connection anyway and use the real endpoint
    addresses or to reject it. The receiver should ignore address information.

  - 0x1 : AF_INET : the forwarded connection uses the AF_INET address family
    (IPv4). The addresses are exactly 4 bytes each in network byte order,
    followed by transport protocol information (typically ports).

  - 0x2 : AF_INET6 : the forwarded connection uses the AF_INET6 address family
    (IPv6). The addresses are exactly 16 bytes each in network byte order,
    followed by transport protocol information (typically ports).

  - 0x3 : AF_UNIX : the forwarded connection uses the AF_UNIX address family
    (UNIX). The addresses are exactly 108 bytes each.

  - other values are unspecified and must not be emitted in version 2 of this
    protocol and must be rejected as invalid by receivers.

The transport protocol is specified in the lowest 4 bits of the the 14th byte :

  - 0x0 : UNSPEC : the connection is forwarded for an unknown, unspecified
    or unsupported protocol. The sender should use this family when sending
    LOCAL commands or when dealing with unsupported protocol families. The
    receiver is free to accept the connection anyway and use the real endpoint
    addresses or to reject it. The receiver should ignore address information.

  - 0x1 : STREAM : the forwarded connection uses a SOCK_STREAM protocol (eg:
    TCP or UNIX_STREAM). When used with AF_INET/AF_INET6 (TCP), the addresses
    are followed by the source and destination ports represented on 2 bytes
    each in network byte order.

  - 0x2 : DGRAM : the forwarded connection uses a SOCK_DGRAM protocol (eg:
    UDP or UNIX_DGRAM). When used with AF_INET/AF_INET6 (UDP), the addresses
    are followed by the source and destination ports represented on 2 bytes
    each in network byte order.

  - other values are unspecified and must not be emitted in version 2 of this
    protocol and must be rejected as invalid by receivers.

In practice, the following protocol bytes are expected :

  - \x00 : UNSPEC : the connection is forwarded for an unknown, unspecified
    or unsupported protocol. The sender should use this family when sending
    LOCAL commands or when dealing with unsupported protocol families. When
    used with a LOCAL command, the receiver must accept the connection and
    ignore any address information. For other commands, the receiver is free
    to accept the connection anyway and use the real endpoints addresses or to
    reject the connection. The receiver should ignore address information.

  - \x11 : TCP over IPv4 : the forwarded connection uses TCP over the AF_INET
    protocol family. Address length is 2*4 + 2*2 = 12 bytes.

  - \x12 : UDP over IPv4 : the forwarded connection uses UDP over the AF_INET
    protocol family. Address length is 2*4 + 2*2 = 12 bytes.

  - \x21 : TCP over IPv6 : the forwarded connection uses TCP over the AF_INET6
    protocol family. Address length is 2*16 + 2*2 = 36 bytes.

  - \x22 : UDP over IPv6 : the forwarded connection uses UDP over the AF_INET6
    protocol family. Address length is 2*16 + 2*2 = 36 bytes.

  - \x31 : UNIX stream : the forwarded connection uses SOCK_STREAM over the
    AF_UNIX protocol family. Address length is 2*108 = 216 bytes.

  - \x32 : UNIX datagram : the forwarded connection uses SOCK_DGRAM over the
    AF_UNIX protocol family. Address length is 2*108 = 216 bytes.


Only the UNSPEC protocol byte (\x00) is mandatory. A receiver is not required
to implement other ones, provided that it automatically falls back to the
UNSPEC mode for the valid combinations above that it does not support.

The 15th and 16th bytes is the address length in bytes in network endien order.
It is used so that the receiver knows how many address bytes to skip even when
it does not implement the presented protocol. Thus the length of the protocol
header in bytes is always exactly 16 + this value. When a sender presents a
LOCAL connection, it should not present any address so it sets this field to
zero. Receivers MUST always consider this field to skip the appropriate number
of bytes and must not assume zero is presented for LOCAL connections. When a
receiver accepts an incoming connection showing an UNSPEC address family or
protocol, it may or may not decide to log the address information if present.

So the 16-byte version 2 header can be described this way :

    struct proxy_hdr_v2 {
        uint8_t sig[12];  /* hex 0D 0A 0D 0A 00 0D 0A 51 55 49 54 0A */
        uint8_t ver_cmd;  /* protocol version and command */
        uint8_t fam;      /* protocol family and address */
        uint16_t len;     /* number of following bytes part of the header */
    };

Starting from the 17th byte, addresses are presented in network byte order.
The address order is always the same :
  - source layer 3 address in network byte order
  - destination layer 3 address in network byte order
  - source layer 4 address if any, in network byte order (port)
  - destination layer 4 address if any, in network byte order (port)

The address block may directly be sent from or received into the following
union which makes it easy to cast from/to the relevant socket native structs
depending on the address type :

    union proxy_addr {
        struct {        /* for TCP/UDP over IPv4, len = 12 */
            uint32_t src_addr;
            uint32_t dst_addr;
            uint16_t src_port;
            uint16_t dst_port;
        } ipv4_addr;
        struct {        /* for TCP/UDP over IPv6, len = 36 */
             uint8_t  src_addr[16];
             uint8_t  dst_addr[16];
             uint16_t src_port;
             uint16_t dst_port;
        } ipv6_addr;
        struct {        /* for AF_UNIX sockets, len = 216 */
             uint8_t src_addr[108];
             uint8_t dst_addr[108];
        } unix_addr;
    };

The sender must ensure that all the protocol header is sent at once. This block
is always smaller than an MSS, so there is no reason for it to be segmented at
the beginning of the connection. The receiver should also process the header
at once. The receiver must not start to parse an address before the whole
address block is received. The receiver must also reject incoming connections
containing partial protocol headers.

A receiver may be configured to support both version 1 and version 2 of the
protocol. Identifying the protocol version is easy :

    - if the incoming byte count is 16 or above and the 13 first bytes match
      the protocol signature block followed by the protocol version 2 :

           \x0D\x0A\x0D\x0A\x00\x0D\x0A\x51\x55\x49\x54\x0A\x02

    - otherwise, if the incoming byte count is 8 or above, and the 5 first
      characters match the ASCII representation of "PROXY" then the protocol
      must be parsed as version 1 :

           \x50\x52\x4F\x58\x59

    - otherwise the protocol is not covered by this specification and the
      connection must be dropped.

If the length specified in the PROXY protocol header indicates that additional
bytes are part of the header beyond the address information, a receiver may
choose to skip over and ignore those bytes, or attempt to interpret those
bytes.

The information in those bytes will be arranged in Type-Length-Value (TLV
vectors) in the following format.  The first byte is the Type of the vector.
The second two bytes represent the length in bytes of the value (not included
the Type and Length bytes), and following the length field is the number of
bytes specified by the length.

        struct {
            uint8_t type;
            uint8_t length_hi;
            uint8_t length_lo;
            uint8_t value[0];
        } tlv;


3. Implementations

Haproxy 1.5 implements version 1 of the PROXY protocol on both sides :
  - the listening sockets accept the protocol when the "accept-proxy" setting
    is passed to the "bind" keyword. Connections accepted on such listeners
    will behave just as if the source really was the one advertised in the
    protocol. This is true for logging, ACLs, content filtering, transparent
    proxying, etc...

  - the protocol may be used to connect to servers if the "send-proxy" setting
    is present on the "server" line. It is enabled on a per-server basis, so it
    is possible to have it enabled for remote servers only and still have local
    ones behave differently. If the incoming connection was accepted with the
    "accept-proxy", then the relayed information is the one advertised in this
    connection's PROXY line.

  - Haproxy 1.5 also implements version 2 of the PROXY protocol as a sender. In
    addition, a TLV with limited, optional, SSL information has been added.

Stunnel added support for version 1 of the protocol for outgoing connections in
version 4.45.

Stud added support for version 1 of the protocol for outgoing connections on
2011/06/29.

Postfix added support for version 1 of the protocol for incoming connections
in smtpd and postscreen in version 2.10.

A patch is available for Stud[5] to implement version 1 of the protocol on
incoming connections.

Support for the protocol in the Varnish cache is being considered [6].

Exim added support for version 1 and version 2 of the protocol for incoming
connections on 2014/05/13, and will be released as part of version 4.83.

The protocol is simple enough that it is expected that other implementations
will appear, especially in environments such as SMTP, IMAP, FTP, RDP where the
client's address is an important piece of information for the server and some
intermediaries. In fact, several proprietary deployments have already done so
on FTP and SMTP servers.

Proxy developers are encouraged to implement this protocol, because it will
make their products much more transparent in complex infrastructures, and will
get rid of a number of issues related to logging and access control.


4. Architectural benefits
4.1. Multiple layers

Using the PROXY protocol instead of transparent proxy provides several benefits
in multiple-layer infrastructures. The first immediate benefit is that it
becomes possible to chain multiple layers of proxies and always present the
original IP address. for instance, let's consider the following 2-layer proxy
architecture :

         Internet
          ,---.                     | client to PX1:
         (  X  )                    | native protocol
          `---'                     |
            |                       V
         +--+--+      +-----+
         | FW1 |------| PX1 |
         +--+--+      +-----+       | PX1 to PX2: PROXY + native
            |                       V
         +--+--+      +-----+
         | FW2 |------| PX2 |
         +--+--+      +-----+       | PX2 to SRV: PROXY + native
            |                       V
         +--+--+
         | SRV |
         +-----+

Firewall FW1 receives traffic from internet-based clients and forwards it to
reverse-proxy PX1. PX1 adds a PROXY header then forwards to PX2 via FW2. PX2
is configured to read the PROXY header and to emit it on output. It then joins
the origin server SRV and presents the original client's address there. Since
all TCP connections endpoints are real machines and are not spoofed, there is
no issue for the return traffic to pass via the firewalls and reverse proxies.
Using transparent proxy, this would be quite difficult because the firewalls
would have to deal with the client's address coming from the proxies in the DMZ
and would have to correctly route the return traffic there instead of using the
default route.


4.2. IPv4 and IPv6 integration

The protocol also eases IPv4 and IPv6 integration : if only the first layer
(FW1 and PX1) is IPv6-capable, it is still possible to present the original
client's IPv6 address to the target server eventhough the whole chain is only
connected via IPv4.


4.3. Multiple return paths

When transparent proxy is used, it is not possible to run multiple proxies
because the return traffic would follow the default route instead of finding
the proper proxy. Some tricks are sometimes possible using multiple server
addresses and policy routing but these are very limited.

Using the PROXY protocol, this problem disappears as the servers don't need
to route to the client, just to the proxy that forwarded the connection. So
it is perfectly possible to run a proxy farm in front of a very large server
farm and have it working effortless, even when dealing with multiple sites.

This is particularly important in Cloud-like environments where there is little
choice of binding to random addresses and where the lower processing power per
node generally requires multiple front nodes.

The example below illustrates the following case : virtualized infrastructures
are deployed in 3 datacenters (DC1..DC3). Each DC uses its own VIP which is
handled by the hosting provider's layer 3 load balancer. This load balancer
routes the traffic to a farm of layer 7 SSL/cache offloaders which load balance
among their local servers. The VIPs are advertised by geolocalised DNS so that
clients generally stick to a given DC. Since clients are not guaranteed to
stick to one DC, the L7 load balancing proxies have to know the other DCs'
servers that may be reached via the hosting provider's LAN or via the internet.
The L7 proxies use the PROXY protocol to join the servers behind them, so that
even inter-DC traffic can forward the original client's address and the return
path is unambiguous. This would not be possible using transparent proxy because
most often the L7 proxies would not be able to spoof an address, and this would
never work between datacenters.

                               Internet

            DC1                  DC2                  DC3
           ,---.                ,---.                ,---.
          (  X  )              (  X  )              (  X  )
           `---'                `---'                `---'
             |    +-------+       |    +-------+       |    +-------+
             +----| L3 LB |       +----| L3 LB |       +----| L3 LB |
             |    +-------+       |    +-------+       |    +-------+
       ------+------- ~ ~ ~ ------+------- ~ ~ ~ ------+-------
       |||||   ||||         |||||   ||||         |||||    ||||
      50 SRV   4 PX        50 SRV   4 PX        50 SRV    4 PX


5. Security considerations

Version 1 of the protocol header (the human-readable format) was designed so as
to be distinguishable from HTTP. It will not parse as a valid HTTP request and
an HTTP request will not parse as a valid proxy request. Version 2 add to use a
non-parsable binary signature to make many products fail on this block. The
signature was designed to cause immediate failure on HTTP, SSL/TLS, SMTP, FTP,
and POP. It also causes aborts on LDAP and RDP servers (see section 6). That
makes it easier to enforce its use under certain connections and at the same
time, it ensures that improperly configured servers are quickly detected.

Implementers should be very careful about not trying to automatically detect
whether they have to decode the header or not, but rather they must only rely
on a configuration parameter. Indeed, if the opportunity is left to a normal
client to use the protocol, he will be able to hide his activities or make them
appear as coming from someone else. However, accepting the header only from a
number of known sources should be safe.


6. Validation

The version 2 protocol signature has been sent to a wide variety of protocols
and implementations including old ones. The following protocol and products
have been tested to ensure the best possible behaviour when the signature was
presented, even with minimal implementations :

  - HTTP :
    - Apache 1.3.33    : connection abort        => pass/optimal
    - Nginx 0.7.69     : 400 Bad Request + abort => pass/optimal
    - lighttpd 1.4.20  : 400 Bad Request + abort => pass/optimal
    - thttpd 2.20c     : 400 Bad Request + abort => pass/optimal
    - mini-httpd-1.19  : 400 Bad Request + abort => pass/optimal
    - haproxy 1.4.21   : 400 Bad Request + abort => pass/optimal
  - SSL :
    - stud 0.3.47      : connection abort        => pass/optimal
    - stunnel 4.45     : connection abort        => pass/optimal
    - nginx 0.7.69     : 400 Bad Request + abort => pass/optimal
  - FTP :
    - Pure-ftpd 1.0.20 : 3*500 then 221 Goodbye  => pass/optimal
    - vsftpd 2.0.1     : 3*530 then 221 Goodbye  => pass/optimal
  - SMTP :
    - postfix 2.3      : 3*500 + 221 Bye         => pass/optimal
    - exim 4.69        : 554 + connection abort  => pass/optimal
  - POP :
    - dovecot 1.0.10   : 3*ERR + Logout          => pass/optimal
  - IMAP :
    - dovecot 1.0.10   : 5*ERR + hang            => pass/non-optimal
  - LDAP :
    - openldap 2.3     : abort                   => pass/optimal
  - SSH :
    - openssh 3.9p1    : abort                   => pass/optimal
  - RDP :
    - Windows XP SP3   : abort                   => pass/optimal

This means that most protocols and implementations will not be confused by an
incoming connection exhibiting the protocol signature, which avoids issues when
facing misconfigurations.


7. Future developments

It is possible that the protocol may slightly evolve to present other
information such as the incoming network interface, or the origin addresses in
case of network address translation happening before the first proxy, but this
is not identified as a requirement right now. Some deep thinking has been spent
on this and it appears that trying to add a few more information open a pandora
box with many information from MAC addresses to SSL client certificates, which
would make the protocol much more complex. So at this point it is not planned.
Suggestions on improvements are welcome.


8. Contacts and links

Please use w@1wt.eu to send any comments to the author.

The following links were referenced in the document.

[1] http://www.postfix.org/XCLIENT_README.html
[2] http://tools.ietf.org/html/rfc7239
[3] http://www.stunnel.org/
[4] https://github.com/bumptech/stud
[5] https://github.com/bumptech/stud/pull/81
[6] https://www.varnish-cache.org/trac/wiki/Future_Protocols


9. Sample code

The code below is an example of how a receiver may deal with both versions of
the protocol header for TCP over IPv4 or IPv6. The function is supposed to be
called upon a read event. Addresses may be directly copied into their final
memory location since they're transported in network byte order. The sending
side is even simpler and can easily be deduced from this sample code.

  struct sockaddr_storage from; /* already filled by accept() */
  struct sockaddr_storage to;   /* already filled by getsockname() */
  const char v2sig[12] = "\x0D\x0A\x0D\x0A\x00\x0D\x0A\x51\x55\x49\x54\x0A";

  /* returns 0 if needs to poll, <0 upon error or >0 if it did the job */
  int read_evt(int fd)
  {
      union {
          struct {
              char line[108];
          } v1;
          struct {
              uint8_t sig[12];
              uint8_t ver_cmd;
              uint8_t fam;
              uint16_t len;
              union {
                  struct {  /* for TCP/UDP over IPv4, len = 12 */
                      uint32_t src_addr;
                      uint32_t dst_addr;
                      uint16_t src_port;
                      uint16_t dst_port;
                  } ip4;
                  struct {  /* for TCP/UDP over IPv6, len = 36 */
                       uint8_t  src_addr[16];
                       uint8_t  dst_addr[16];
                       uint16_t src_port;
                       uint16_t dst_port;
                  } ip6;
                  struct {  /* for AF_UNIX sockets, len = 216 */
                       uint8_t src_addr[108];
                       uint8_t dst_addr[108];
                  } unx;
              } addr;
          } v2;
      } hdr;

      int size, ret;

      do {
          ret = recv(fd, &hdr, sizeof(hdr), MSG_PEEK);
      } while (ret == -1 && errno == EINTR);

      if (ret == -1)
          return (errno == EAGAIN) ? 0 : -1;

      if (ret >= 16 && memcmp(&hdr.v2, v2sig, 12) == 0 &&
          (hdr.v2.ver_cmd & 0xF0) == 0x20) {
          size = 16 + hdr.v2.len;
          if (ret < size)
              return -1; /* truncated or too large header */

          switch (hdr.v2.ver_cmd & 0xF) {
          case 0x01: /* PROXY command */
              switch (hdr.v2.fam) {
              case 0x11:  /* TCPv4 */
                  ((struct sockaddr_in *)&from)->sin_family = AF_INET;
                  ((struct sockaddr_in *)&from)->sin_addr.s_addr =
                      hdr.v2.addr.ip4.src_addr;
                  ((struct sockaddr_in *)&from)->sin_port =
                      hdr.v2.addr.ip4.src_port;
                  ((struct sockaddr_in *)&to)->sin_family = AF_INET;
                  ((struct sockaddr_in *)&to)->sin_addr.s_addr =
                      hdr.v2.addr.ip4.dst_addr;
                  ((struct sockaddr_in *)&to)->sin_port =
                      hdr.v2.addr.ip4.dst_port;
                  goto done;
              case 0x21:  /* TCPv6 */
                  ((struct sockaddr_in6 *)&from)->sin6_family = AF_INET6;
                  memcpy(&((struct sockaddr_in6 *)&from)->sin6_addr,
                      hdr.v2.addr.ip6.src_addr, 16);
                  ((struct sockaddr_in6 *)&from)->sin6_port =
                      hdr.v2.addr.ip6.src_port;
                  ((struct sockaddr_in6 *)&to)->sin6_family = AF_INET6;
                  memcpy(&((struct sockaddr_in6 *)&to)->sin6_addr,
                      hdr.v2.addr.ip6.dst_addr, 16);
                  ((struct sockaddr_in6 *)&to)->sin6_port =
                      hdr.v2.addr.ip6.dst_port;
                  goto done;
              }
              /* unsupported protocol, keep local connection address */
              break;
          case 0x00: /* LOCAL command */
              /* keep local connection address for LOCAL */
              break;
          default:
              return -1; /* not a supported command */
          }
      }
      else if (ret >= 8 && memcmp(hdr.v1.line, "PROXY", 5) == 0) {
          char *end = memchr(hdr.v1.line, '\r', ret - 1);
          if (!end || end[1] != '\n')
              return -1; /* partial or invalid header */
          *end = '\0'; /* terminate the string to ease parsing */
          size = end + 2 - hdr.v1.line; /* skip header + CRLF */
          /* parse the V1 header using favorite address parsers like inet_pton.
           * return -1 upon error, or simply fall through to accept.
           */
      }
      else {
          /* Wrong protocol */
          return -1;
      }

  done:
      /* we need to consume the appropriate amount of data from the socket */
      do {
          ret = recv(fd, &hdr, size, 0);
      } while (ret == -1 && errno == EINTR);
      return (ret >= 0) ? 1 : -1;
  }