Postfix TLS Support


WARNING

By turning on TLS support in Postfix, you not only get the ability to encrypt mail and to authenticate clients or servers. You also turn on thousands and thousands of lines of OpenSSL library code. Assuming that OpenSSL is written as carefully as Wietse's own code, every 1000 lines introduce one additional bug into Postfix.

What Postfix TLS support does for you

Transport Layer Security (TLS, formerly called SSL) provides certificate-based authentication and encrypted sessions. An encrypted session protects the information that is transmitted with SMTP mail or with SASL authentication.

Postfix version 2.2 introduces support for TLS as described in RFC 3207. TLS Support for older Postfix versions was available as an add-on patch. The section "Compatibility with Postfix < 2.2 TLS support" below discusses the differences between these implementations.

Topics covered in this document:

And last but not least, for the impatient:

How Postfix TLS support works

The diagram below shows the main elements of the Postfix TLS architecture and their relationships. Colored boxes with numbered names represent Postfix daemon programs. Other colored boxes represent storage elements.

Network->
smtpd(8)
 
<---seed---

<-session->

tlsmgr(8)
 
---seed--->

<-session->

smtp(8)
 
->Network
/
/
|
|
\
\
smtpd
session
key cache
PRNG
state
file
smtp
session
key cache

Building Postfix with TLS support

To build Postfix with TLS support, first we need to generate the make(1) files with the necessary definitions. This is done by invoking the command "make makefiles" in the Postfix top-level directory and with arguments as shown next.

NOTE: Do not use Gnu TLS. It will spontaneously terminate a Postfix daemon process with exit status code 2, instead of allowing Postfix to 1) report the error to the maillog file, and to 2) provide plaintext service where this is appropriate.

If you need to apply other customizations (such as Berkeley DB databases, MySQL, PosgreSQL, LDAP or SASL), see the respective Postfix README documents, and combine their "make makefiles" instructions with the instructions above:

% make tidy # if you have left-over files from a previous build
% make makefiles CCARGS="-DUSE_TLS \
    (other -D or -I options)" \
    AUXLIBS="-lssl -lcrypto \
    (other -l options for libraries in /usr/lib) \
    (-L/path/name + -l options for other libraries)"

To complete the build process, see the Postfix INSTALL instructions. Postfix has TLS support turned off by default, so you can start using Postfix as soon as it is installed.

SMTP Server specific settings

Topics covered in this section:

Server-side certificate and private key configuration

In order to use TLS, the Postfix SMTP server needs a certificate and a private key. Both must be in "pem" format. The private key must not be encrypted, meaning: the key must be accessible without password. Both certificate and private key may be in the same file.

Both RSA and DSA certificates are supported. Typically you will only have RSA certificates issued by a commercial CA. In addition, the tools supplied with OpenSSL will by default issue RSA certificates. You can have both at the same time, in which case the cipher used determines which certificate is presented. For Netscape and OpenSSL clients without special cipher choices, the RSA certificate is preferred.

In order for remote SMTP clients to check the Postfix SMTP server certificates, the CA certificate (in case of a certificate chain, all CA certificates) must be available. You should add these certificates to the server certificate, the server certificate first, then the issuing CA(s).

Example: the certificate for "server.dom.ain" was issued by "intermediate CA" which itself has a certificate issued by "root CA". Create the server.pem file with:

% cat server_cert.pem intermediate_CA.pem > server.pem

A Postfix SMTP server certificate supplied here must be usable as SSL server certificate and hence pass the "openssl verify -purpose sslserver ..." test.

A client that trusts the root CA has a local copy of the root CA certificate, so it is not necessary to include the root CA certificate here. Leaving it out of the "server.pem" file reduces the overhead of the TLS exchange.

If you want the Postfix SMTP server to accept remote SMTP client certificates issued by these CAs, append the root certificate to $smtpd_tls_CAfile or install it in the $smtpd_tls_CApath directory. When you configure trust in a root CA, it is not necessary to explicitly trust intermediary CAs signed by the root CA, unless $smtpd_tls_ccert_verifydepth is less than the number of CAs in the certificate chain for the clients of interest. With a verify depth of 1 you can only verify certificates directly signed by a trusted CA, and all trusted intermediary CAs need to be configured explicitly. With a verify depth of 2 you can verify clients signed by a root CA or a direct intermediary CA (so long as the client is correctly configured to supply its intermediate CA certificate).

RSA key and certificate examples:

/etc/postfix/main.cf:
    smtpd_tls_cert_file = /etc/postfix/server.pem
    smtpd_tls_key_file = $smtpd_tls_cert_file

Their DSA counterparts:

/etc/postfix/main.cf:
    smtpd_tls_dcert_file = /etc/postfix/server-dsa.pem
    smtpd_tls_dkey_file = $smtpd_tls_dcert_file

To verify a remote SMTP client certificate, the Postfix SMTP server needs to trust the certificates of the issuing certification authorities. These certificates in "pem" format can be stored in a single $smtpd_tls_CAfile or in multiple files, one CA per file in the $smtpd_tls_CApath directory. If you use a directory, don't forget to create the necessary "hash" links with:

# $OPENSSL_HOME/bin/c_rehash /path/to/directory 

The $smtpd_tls_CAfile contains the CA certificates of one or more trusted CAs. The file is opened (with root privileges) before Postfix enters the optional chroot jail and so need not be accessible from inside the chroot jail.

Additional trusted CAs can be specified via the $smtpd_tls_CApath directory, in which case the certificates are read (with $mail_owner privileges) from the files in the directory when the information is needed. Thus, the $smtpd_tls_CApath directory needs to be accessible inside the optional chroot jail.

When you configure Postfix to request client certificates (by setting $smtpd_tls_ask_ccert = yes), any certificates in $smtpd_tls_CAfile are sent to the client, in order to allow it to choose an identity signed by a CA you trust. If no $smtpd_tls_CAfile is specified, no preferred CA list is sent, and the client is free to choose an identity signed by any CA. Many clients use a fixed identity regardless of the preferred CA list and you may be able to reduce TLS negotiation overhead by installing client CA certificates mostly or only in $smtpd_tls_CApath. In the latter case you need not specify a $smtpd_tls_CAfile.

Note, that unless client certificates are used to allow greater access to TLS authenticated clients, it is best to not ask for client certificates at all, as in addition to increased overhead some clients (notably in some cases qmail) are unable to complete the TLS handshake when client certificates are requested.

Example:

/etc/postfix/main.cf:
    smtpd_tls_CAfile = /etc/postfix/CAcert.pem
    smtpd_tls_CApath = /etc/postfix/certs

Server-side TLS activity logging

To get additional information about Postfix SMTP server TLS activity you can increase the loglevel from 0..4. Each logging level also includes the information that is logged at a lower logging level.

0 Disable logging of TLS activity.
1 Log TLS handshake and certificate information.
2 Log levels during TLS negotiation.
3 Log hexadecimal and ASCII dump of TLS negotiation process
4 Log hexadecimal and ASCII dump of complete transmission after STARTTLS

Use loglevel 3 only in case of problems. Use of loglevel 4 is strongly discouraged.

Example:

/etc/postfix/main.cf:
    smtpd_tls_loglevel = 0

To include information about the protocol and cipher used as well as the client and issuer CommonName into the "Received:" message header, set the smtpd_tls_received_header variable to true. The default is no, as the information is not necessarily authentic. Only information recorded at the final destination is reliable, since the headers may be changed by intermediate servers.

Example:

/etc/postfix/main.cf:
    smtpd_tls_received_header = yes

Enabling TLS in the Postfix SMTP server

By default, TLS is disabled in the Postfix SMTP server, so no difference to plain Postfix is visible. Explicitly switch it on using "smtpd_use_tls = yes".

Example:

/etc/postfix/main.cf:
    smtpd_use_tls = yes

With this, Postfix SMTP server announces STARTTLS support to SMTP clients, but does not require that clients use TLS encryption.

Note: when an unprivileged user invokes "sendmail -bs", STARTTLS is never offered due to insufficient privileges to access the server private key. This is intended behavior.

You can ENFORCE the use of TLS, so that the Postfix SMTP server announces STARTTLS and accepts no mail without TLS encryption, by setting "smtpd_enforce_tls = yes". According to RFC 2487 this MUST NOT be applied in case of a publicly-referenced Postfix SMTP server. This option is off by default and should only seldom be used.

Example:

/etc/postfix/main.cf:
    smtpd_enforce_tls = yes

TLS is sometimes used in the non-standard "wrapper" mode where a server always uses TLS, instead of announcing STARTTLS support and waiting for clients to request TLS service. Some clients, namely Outlook [Express] prefer the "wrapper" mode. This is true for OE (Win32 < 5.0 and Win32 >=5.0 when run on a port<>25 and OE (5.01 Mac on all ports).

It is strictly discouraged to use this mode from main.cf. If you want to support this service, enable a special port in master.cf and specify "-o smtpd_tls_wrappermode = yes" as an smtpd(8) command line option. Port 465 (smtps) was once chosen for this feature.

Example:

/etc/postfix/master.cf:
    smtps    inet  n       -       n       -       -       smtpd
      -o smtpd_tls_wrappermode=yes -o smtpd_sasl_auth_enable=yes

Client certificate verification

To receive a remote SMTP client certificate, the Postfix SMTP server must explicitly ask for one (any contents of $smtpd_tls_CAfile are also sent to the client as a hint for choosing a certificate from a suitable CA). Unfortunately, Netscape clients will either complain if no matching client certificate is available or will offer the user client a list of certificates to choose from. Additionally some MTAs (notably some versions of qmail) are unable to complete TLS negotiation when client certificates are requested, and abort the SMTP session. So this option is "off" by default. You will however need the certificate if you want to use certificate based relaying with, for example, the permit_tls_clientcerts feature.

Example:

/etc/postfix/main.cf:
    smtpd_tls_ask_ccert = no

You may also decide to REQUIRE a remote SMTP client certificate before allowing TLS connections. This feature is included for completeness, and implies "smtpd_tls_ask_ccert = yes".

Please be aware, that this will inhibit TLS connections without a proper client certificate and that it makes sense only when non-TLS submission is disabled (smtpd_enforce_tls = yes). Otherwise, clients could bypass the restriction by simply not using STARTTLS at all.

When TLS is not enforced, the connection will be handled as if only "smtpd_tls_ask_ccert = yes" is specified, and a warning is logged.

Example:

/etc/postfix/main.cf:
    smtpd_tls_req_ccert = no

A client certificate verification depth of 1 is sufficient if the certificate is directly issued by a CA listed in the CA file. The default value (5) should also suffice for longer chains (root CA issues special CA which then issues the actual certificate...)

Example:

/etc/postfix/main.cf:
    smtpd_tls_ccert_verifydepth = 5

Supporting AUTH over TLS only

Sending AUTH data over an unencrypted channel poses a security risk. When TLS layer encryption is required (smtpd_enforce_tls = yes), the Postfix SMTP server will announce and accept AUTH only after the TLS layer has been activated with STARTTLS. When TLS layer encryption is optional (smtpd_enforce_tls = no), it may however still be useful to only offer AUTH when TLS is active. To maintain compatibility with non-TLS clients, the default is to accept AUTH without encryption. In order to change this behavior, set "smtpd_tls_auth_only = yes".

Example:

/etc/postfix/main.cf:
    smtpd_tls_auth_only = no

Server-side TLS session cache

The Postfix SMTP server and the remote SMTP client negotiate a session, which takes some computer time and network bandwidth. By default, this session information is cached only in the smtpd(8) process actually using this session and is lost when the process terminates. To share the session information between multiple smtpd(8) processes, a persistent session cache can be used. You can specify any database type that can store objects of several kbytes and that supports the sequence operator. DBM databases are not suitable because they can only store small objects. The cache is maintained by the tlsmgr(8) process, so there is no problem with concurrent access. Session caching is highly recommended, because the cost of repeatedly negotiating TLS session keys is high.

Example:

/etc/postfix/main.cf:
    smtpd_tls_session_cache_database = btree:/etc/postfix/smtpd_scache

Cached Postfix SMTP server session information expires after a certain amount of time. Postfix/TLS does not use the OpenSSL default of 300s, but a longer time of 3600sec (=1 hour). RFC 2246 recommends a maximum of 24 hours.

Example:

/etc/postfix/main.cf:
    smtpd_tls_session_cache_timeout = 3600s

Server access control

Postfix TLS support introduces three additional features for Postfix SMTP server access control:

permit_tls_clientcerts

Allow the remote SMTP client SMTP request if the client certificate passes verification, and if its fingerprint is listed in the list of client certificates (see relay_clientcerts discussion below).

permit_tls_all_clientcerts

Allow the remote client SMTP request if the client certificate passes verification.

check_ccert_access type:table

If the client certificate passes verification, use its fingerprint as a key for the specified access(5) table.

The permit_tls_all_clientcerts feature must be used with caution, because it can result in too many access permissions. Use this feature only if a special CA issues the client certificates, and only if this CA is listed as trusted CA. If other CAs are trusted, any owner of a valid client certificate would be authorized. The permit_tls_all_clientcerts feature can be practical for a specially created email relay server.

It is however recommended to stay with the permit_tls_clientcerts feature and list all certificates via $relay_clientcerts, as permit_tls_all_clientcerts does not permit any control when a certificate must no longer be used (e.g. an employee leaving).

Example:

/etc/postfix/main.cf:
    smtpd_recipient_restrictions = 
        ... 
        permit_tls_clientcerts 
        reject_unauth_destination
        ...

The Postfix list manipulation routines give special treatment to whitespace and some other characters, making the use of certificate names impractical. Instead we use the certificate fingerprints as they are difficult to fake but easy to use for lookup. Postfix lookup tables are in the form of (key, value) pairs. Since we only need the key, the value can be chosen freely, e.g. the name of the user or host.

Example:

/etc/postfix/main.cf:
    relay_clientcerts = hash:/etc/postfix/relay_clientcerts

/etc/postfix/relay_clientcerts:
    D7:04:2F:A7:0B:8C:A5:21:FA:31:77:E1:41:8A:EE:80 lutzpc.at.home

Server-side cipher controls

To influence the Postfix SMTP server cipher selection scheme, you can give cipherlist string. A detailed description would go to far here; please refer to the OpenSSL documentation. If you don't know what to do with it, simply don't touch it and leave the (openssl-)compiled in default!

DO NOT USE " to enclose the string, specify just the string!!!

Example:

/etc/postfix/main.cf:
    smtpd_tls_cipherlist = DEFAULT

If you want to take advantage of ciphers with EDH, DH parameters are needed. Instead of using the built-in DH parameters for both 1024bit and 512bit, it is better to generate "own" parameters, since otherwise it would "pay" for a possible attacker to start a brute force attack against parameters that are used by everybody. For this reason, the parameters chosen are already different from those distributed with other TLS packages.

To generate your own set of DH parameters, use:

% openssl gendh -out /etc/postfix/dh_1024.pem -2 -rand /var/run/egd-pool 1024
% openssl gendh -out /etc/postfix/dh_512.pem -2 -rand /var/run/egd-pool 512

Examples:

/etc/postfix/main.cf:
    smtpd_tls_dh1024_param_file = /etc/postfix/dh_1024.pem
    smtpd_tls_dh512_param_file = /etc/postfix/dh_512.pem

Miscellaneous server controls

The smtpd_starttls_timeout parameter limits the time of Postfix SMTP server write and read operations during TLS startup and shutdown handshake procedures.

Example:

/etc/postfix/main.cf:
    smtpd_starttls_timeout = 300s

SMTP Client specific settings

Topics covered in this section:

TLS support in the LMTP delivery agent

In Postfix 2.3, the smtp(8) and lmtp(8) delivery agents have been merged into a single dual-purpose program. As a result the lmtp(8) delivery agent is no longer the poor cousin of the more extensively used smtp(8). Specifically, as of Postfix 2.3, all the TLS features described below apply equally to SMTP and LMTP, after replacing the smtp_ prefix of the each parameter name with lmtp_.

The LMTP delivery agent can communicate with LMTP servers listening on unix-domain sockets. When server certificate verification is enabled and the server is listening on a unix-domain socket, the $myhostname parameter is used to set the TLS verification nexthop and hostname. Note, mere encryption of LMTP traffic over unix-domain sockets is futile, the channel is already secure. TLS is only useful in this context, when it is mandatory and at least one of the server ar client is authenticating the other. The eNULL TLS ciphers may be appropriate in this context when available on both client and server, these provide authentication without encryption.

Client-side certificate and private key configuration

During TLS startup negotiation the Postfix SMTP client may present a certificate to the remote SMTP server. The Netscape client is rather clever here and lets the user select between only those certificates that match CA certificates offered by the remote SMTP server. As the Postfix SMTP client uses the "SSL_connect()" function from the OpenSSL package, this is not possible and we have to choose just one certificate. So for now the default is to use _no_ certificate and key unless one is explicitly specified here.

Both RSA and DSA certificates are supported. You can have both at the same time, in which case the cipher used determines which certificate is presented.

It is possible for the Postfix SMTP client to use the same key/certificate pair as the Postfix SMTP server. If a certificate is to be presented, it must be in "pem" format. The private key must not be encrypted, meaning: it must be accessible without password. Both parts (certificate and private key) may be in the same file.

In order for remote SMTP servers to verify the Postfix SMTP client certificates, the CA certificate (in case of a certificate chain, all CA certificates) must be available. You should add these certificates to the client certificate, the client certificate first, then the issuing CA(s).

Example: the certificate for "client.example.com" was issued by "intermediate CA" which itself has a certificate of "root CA". Create the client.pem file with:

% cat client_cert.pem intermediate_CA.pem > client.pem 

A Postfix SMTP client certificate supplied here must be usable as SSL client certificate and hence pass the "openssl verify -purpose sslclient ..." test.

A server that trusts the root CA has a local copy of the root CA certificate, so it is not necessary to include the root CA certificate here. Leaving it out of the "client.pem" file reduces the overhead of the TLS exchange.

If you want the Postfix SMTP client to accept remote SMTP server certificates issued by these CAs, append the root certificate to $smtp_tls_CAfile or install it in the $smtp_tls_CApath directory. When you configure trust in a root CA, it is not necessary to explicitly trust intermediary CAs signed by the root CA, unless $smtp_tls_scert_verifydepth is less than the number of CAs in the certificate chain for the servers of interest. With a verify depth of 1 you can only verify certificates directly signed by a trusted CA, and all trusted intermediary CAs need to be configured explicitly. With a verify depth of 2 you can verify servers signed by a root CA or a direct intermediary CA (so long as the server is correctly configured to supply its intermediate CA certificate).

RSA key and certificate examples:

/etc/postfix/main.cf:
    smtp_tls_cert_file = /etc/postfix/client.pem
    smtp_tls_key_file = $smtp_tls_cert_file

Their DSA counterparts:

/etc/postfix/main.cf:
    smtp_tls_dcert_file = /etc/postfix/client-dsa.pem
    smtp_tls_dkey_file = $smtpd_tls_cert_file

To verify a remote SMTP server certificate, the Postfix SMTP client needs to trust the certificates of the issuing certification authorities. These certificates in "pem" format can be stored in a single $smtp_tls_CAfile or in multiple files, one CA per file in the $smtp_tls_CApath directory. If you use a directory, don't forget to create the necessary "hash" links with:

# $OPENSSL_HOME/bin/c_rehash /path/to/directory 

The $smtp_tls_CAfile contains the CA certificates of one or more trusted CAs. The file is opened (with root privileges) before Postfix enters the optional chroot jail and so need not be accessible from inside the chroot jail.

Additional trusted CAs can be specified via the $smtp_tls_CApath directory, in which case the certificates are read (with $mail_owner privileges) from the files in the directory when the information is needed. Thus, the $smtp_tls_CApath directory needs to be accessible inside the optional chroot jail.

The choice between $smtp_tls_CAfile and $smtpd_tls_CApath is a space/time tradeoff. If there are many trusted CAs, the cost of preloading them all into memory may not pay off in reduced access time when the certificate is needed.

Example:

/etc/postfix/main.cf:
    smtp_tls_CAfile = /etc/postfix/CAcert.pem
    smtp_tls_CApath = /etc/postfix/certs

Client-side TLS activity logging

To get additional information about Postfix SMTP client TLS activity you can increase the loglevel from 0..4. Each logging level also includes the information that is logged at a lower logging level.

0 Disable logging of TLS activity.
1 Log TLS handshake and certificate information.
2 Log levels during TLS negotiation.
3 Log hexadecimal and ASCII dump of TLS negotiation process
4 Log hexadecimal and ASCII dump of complete transmission after STARTTLS

Example:

/etc/postfix/main.cf:
    smtp_tls_loglevel = 0

Client-side TLS session cache

The remote SMTP server and the Postfix SMTP client negotiate a session, which takes some computer time and network bandwidth. By default, this session information is cached only in the smtp(8) process actually using this session and is lost when the process terminates. To share the session information between multiple smtp(8) processes, a persistent session cache can be used. You can specify any database type that can store objects of several kbytes and that supports the sequence operator. DBM databases are not suitable because they can only store small objects. The cache is maintained by the tlsmgr(8) process, so there is no problem with concurrent access. Session caching is highly recommended, because the cost of repeatedly negotiating TLS session keys is high. Future Postfix SMTP servers may limit the number of sessions that a client is allowed to negotiate per unit time.

Example:

/etc/postfix/main.cf:
    smtp_tls_session_cache_database = btree:/etc/postfix/smtp_scache

Cached Postfix SMTP client session information expires after a certain amount of time. Postfix/TLS does not use the OpenSSL default of 300s, but a longer time of 3600s (=1 hour). RFC 2246 recommends a maximum of 24 hours.

Example:

/etc/postfix/main.cf:
    smtp_tls_session_cache_timeout = 3600s

TLS security levels

The security properties of TLS communications channels are application specific. While the TLS protocol supports mutual authentication between client and server, not all of its security features are applicable in every case.

For example, while mutual TLS authentication between browsers and web servers is possible, it is not practical or even useful for web-servers that serve the public to verify the identity of every potential user. In practice, most HTTPS transactions are asymmetric: the browser verifies the HTTPS server's identity, but the user remains anonymous. Much of the security policy is up to client, if the client chooses to not verify the server's name, the server is not aware of this. There are many fascinating issues related to browser security, but we shall not dwell on them here. Rather our goal is to understand the security features of TLS in conjuction with SMTP.

An important SMTP-specific observation is that a public MX host is even more at the mercy of the SMTP client. Not only can it not enforce due care in the client's use of TLS, but it cannot even enforce the use of TLS, because TLS support in SMTP clients is still the exception rather than the rule. One cannot in practice limit access to one's MX hosts to just TLS enabled clients. Such a policy would result in a vast reduction in one's ability to communicate by email with the world at-large.

One may be tempted to try to enforce TLS for mail from specific sending organizations, but this too runs into obstacles. One such obstacle is that we don't know who is (allegedly) sending mail until we see the MAIL FROM: SMTP command, and at that point if TLS is not already in use, a potentially sensitive sender address has already been leaked in the clear. Another obstacle is that mail from the sender to the recipient may be forwarded, and the forwarding organization may not have any security arrangements with the final destination. Also bounces need to be protected, these can only be identified by the IP address and HELO name of the connecting client, and it is difficult to keep track of all the potential IP addresses or HELO names of the outbound email servers of the sending organization.

Consequently, TLS security for mail delivery to public MX hosts is almost entirely the client's responsibility. The server is largely a passive enabler of TLS security as implemented in the client. While the server's security role is more significant when it is an MSA that only handles outbound mail from trusted clients, in what follows we focus exclusively on the client security policy.

On the SMTP client, there are further complications. When delivering mail to a given domain, in contrast to HTTPS one rarely uses the domain name directly as the target host of the SMTP session. More typically, one uses (unauthenticated) MX lookups to obtain the domain's SMTP server hostname(s). When, as is current practice, the client verifies the insecurely obtained MX hostname, it is subject to a DNS man-in-the-middle attack.

If on the other hand, the client verifies the original domain name, an SMTP server that receives mail for multiple domains needs to list all its email domain names in its certificate, and generate a new certificate each time a new domain is added. At least some CAs set fairly low limits (20 for one prominent CA) on the number of names that server certificates can contain. This approach does not scale.

It is regrettably the case, that TLS secure-channels (fully authenticated and immune to man-in-the-middle attacks) impose constraints on the sending and receiving sites that preclude ubiquitous deployment. One needs to manually configure this type of security for each destination domain, and in many cases implement non-default transport(5) table entries for additional domains hosted at a common secured destination. With Postfix 2.3, we make secure-channel configurations substantially easier to configure, but they will never be the norm. For the generic domain with which you have made no specific security arrangements, this security level is not a good fit.

Historical note: while the documentation of these issues and many of the related features are new with Postfix 2.3, the issue was well understood before Postfix 1.0, when Lutz Jaenicke was designing the first unofficial Postfix TLS patch. See, his original post http://thread.gmane.org/gmane.ietf.apps-tls/304/focus=304 and the first response http://thread.gmane.org/gmane.ietf.apps-tls/304/focus=305. The problem is not even unique to SMTP or even TLS, similar issues exist for secure connections via aliases for HTTPS and Kerberos. SMTP merely uses indirect naming (via MX records) more frequently.

We are now ready to describe the security levels available to a TLS-enabled SMTP client:

No TLS
TLS is disabled even if supported by the server. Mail is sent unencrypted and TLS is not used to authenticate the server. This corresponds to the none security level in the new (Postfix 2.3 and later) TLS policy table and to the NONE keyword in the obsolete per-site table.
Opportunistic TLS
TLS is used if supported by the server, otherwise mail is sent unencrypted. Even if encryption is used, mail delivery continues even if the server certificate is untrusted or bears the wrong name. Starting with Postfix 2.3, if the TLS handshake fails, rather than give up on mail delivery, the connection is retried with TLS disabled. Trying an unencrypted connection makes it possible to deliver mail to sites with non-interoperable server TLS implementations. This security level corresponds to the may level in the new (Postfix 2.3 and later) TLS policy table and to the MAY keyword in the obsolete per-site table.
Mandatory TLS encryption
TLS encryption is required, mail delivery is deferred if the server does not support TLS. Even though TLS encryption is always used, mail delivery continues even if the server certificate is untrusted or bears the wrong name. This corresponds to the encrypt security level in the new (Postfix 2.3 and later) TLS policy table and to the MUST_NOPEERMATCH keyword in the obsolete per-site table.
Mandatory TLS verification
TLS encryption is required, mail is deferred if the server does not support TLS or certificate verification fails. The server certificate trust chain must be rooted at a CA trusted by the client (present in smtp_tls_CAfile or smtp_tls_CApath), and the server hostname (possibly derived from an insecure MX lookup) must be listed in the certificate. This corresponds to the verify security level in the new (Postfix 2.3 and later) TLS policy table and to the MUST keyword in the obsolete per-site table.
Secure-channel TLS
TLS encryption is required, mail is deferred if the server does not support TLS or certificate verification fails. The server certificate trust chain must be rooted at a CA trusted by the client (present in smtp_tls_CAfile or smtp_tls_CApath), and the nexthop destination domain taken from the email address or the client's transport(5) table, must be listed in the certificate (the policy tables supports per-site overrides of the expected certificate names). This corresponds to the secure security level in the new (Postfix 2.3 and later) TLS policy table. There is no direct equivalent in the obsolete per-site table, but it is possible to harden Postfix 2.2.9 and later against DNS attacks, and thereby achieve equivalent security.

The above security levels can be configured in main.cf as global settings for all destinations, or as per-destination settings via the policy table or obsolete per-site table. The per-destination table syntax is described separately, below we explain the key main.cf settings that come into play at each of TLS security levels.

No TLS
Setting smtp_use_tls = no (default) and smtp_enforce_tls = no (default) disables the use of TLS for all destinations not otherwise specified in the per-destination tables.
Opportunistic TLS
Setting smtp_use_tls = yes and smtp_enforce_tls = no (default) enables the use of opportunistic TLS for all destinations not otherwise specified via the per-destination tables. Since sending in the clear is acceptable, demanding stronger than default TLS security parameters merely reduces inter-operability, for this reason, Postfix 2.3 ignores any non-default settings of smtp_tls_cipherlist or (new with 2.3) smtp_tls_protocols at this security level.
Mandatory TLS encryption
Setting smtp_enforce_tls = yes and smtp_tls_enforce_peername = no enforces the use of TLS encryption for all destinations not otherwise specified in the per-destination tables. At this security level and higher smtp_tls_cipherlist and (new with Postfix 2.3) smtp_tls_protocols define the sufficiently secure ciphers and SSL/TLS protocol versions. The documentation for both parameters includes useful interoperability and security guidelines.
Mandatory TLS verification
Setting smtp_enforce_tls = yes and smtp_tls_enforce_peername = yes (default) enforces verified TLS encryption for all destinations not otherwise specified in the per-destination tables. At this security level (new with Postfix 2.3) smtp_tls_verify_cert_match defines how the SMTP server hostname is verified.
Secure-channel TLS
This security level is only available via the new with Postfix 2.3 policy table. At this security level (new with Postfix 2.3) smtp_tls_secure_cert_match defines how the next-hop destination name is verified.
The following use cases motivate the new policy table design.
Opportunistic TLS:

The client will negotiate a TLS session when possible, at the maximum cryptographic strength supported by the server. The configuration settings are:

/etc/postfix/main.cf:
    smtp_use_tls = yes
Mandatory Encryption for selected domains:

In this case the client encrypts all traffic to example.com (sent via the corresponding MX hosts) and to the non-MX destination example.net (on port 587). All sessions are encrypted via medium or high strength ciphers (128 bit or better). Traffic to example.net will only use TLSv1 and HIGH strength ciphers.

/etc/postfix/main.cf:
    smtp_tls_cipherlist = HIGH:MEDIUM:!aNULL:+RC4:@STRENGTH
    smtp_tls_policy_maps = hash:/etc/postfix/tls_policy

/etc/postfix/tls_policy:
    example.com	      encrypt
    [example.net]:587 encrypt cipherlist=HIGH:!aNULL:@STRENGTH protocols=TLSv1
Verified TLS for selected domains:

In this case the client encrypts all traffic to example.com domain. The peer hostname is verified, but is vulnerable to DNS response forgery. All mandatory encryption uses high strength (better than 128 bit) ciphers.

/etc/postfix/main.cf:
    smtp_tls_CAfile = /etc/postfix/trusted-CAs.pem
    smtp_tls_cipherlist = HIGH:!aNULL:@STRENGTH
    smtp_tls_policy_maps = hash:/etc/postfix/tls_policy

/etc/postfix/tls_policy:
    example.com       verify
Secure-channel TLS with MX lookups:

The client will encrypt all traffic and verify the destination name immune from forged DNS responses. MX lookups are still used to find the SMTP servers for example.com, but these are not used when checking the names in the server certificate(s). Rather the requirement is that the MX hosts for example.com have trusted certificates with a subject name of example.com or a sub-domain (see the documentation for the smtp_tls_secure_cert_match parameter).

The related domains example.co.uk and example.co.jp are hosted on the same MX hosts as the primary example.comdomain, and traffic to these is secured by verifying the primary example.com domain in the server certificates. This frees the server administrator from needing the CA to sign certificates that list all the secondary domains. The downside is that clients that want secure channels to the secondary domains need explicit transport table entries. Given that secure channels require at the very least a corresponding policy table entry, adding the transport table entry is not onerous.

/etc/postfix/main.cf:
    smtp_tls_CAfile = /etc/postfix/trusted-CAs.pem
    smtp_tls_cipherlist = HIGH:MEDIUM:!aNULL:+RC4:@STRENGTH
    transport_maps = hash:/etc/postfix/transport
    smtp_tls_policy_maps = hash:/etc/postfix/tls_policy

/etc/postfix/transport:
    example.co.uk   smtp:example.com
    example.co.jp   smtp:example.com

/etc/postfix/tls_policy:
    example.com     secure
Secure-channel TLS without MX lookups:

In this case traffic to example.com and its related domains is sent to a single logical gateway (ideally, to avoid a single point of failure, its name resolves to one or more load-balancer addresses, or to the combined addresses of multiple hosts). All the physical hosts reachable via the gateway's IP addresses have the gateway name listed in their certificates. This secure-channel configuration can also be implemented via a hardened variant of the MUST policy in the obsolete per-site table.

/etc/postfix/main.cf:
    smtp_tls_CAfile = /etc/postfix/trusted-CAs.pem
    smtp_tls_cipherlist = HIGH:MEDIUM:!aNULL:+RC4:@STRENGTH
    transport_maps = hash:/etc/postfix/transport
    smtp_tls_policy_maps = hash:/etc/postfix/tls_policy

/etc/postfix/transport:
    example.com     smtp:[tls.example.com]
    example.co.uk   smtp:[tls.example.com]
    example.co.jp   smtp:[tls.example.com]

/etc/postfix/tls_policy:
    [tls.example.com] secure match=tls.example.com

Enabling TLS in the Postfix SMTP client

By default, TLS is disabled in the Postfix SMTP client, so no difference to plain Postfix is visible. If you enable TLS, the Postfix SMTP client will send STARTTLS when TLS support is announced by the remote SMTP server.

When the server accepts the STARTTLS command, but the subsequent TLS handshake fails, and no other server is available, the Postfix SMTP client defers the delivery attempt, and the mail stays in the queue. After a handshake failure, the communications channel is in an indeterminate state and cannot be used for non-TLS deliveries.

Example:

/etc/postfix/main.cf:
    smtp_use_tls = yes

Requiring TLS encryption

You can ENFORCE the use of TLS, so that the Postfix SMTP client will not deliver mail over unencrypted connections. In this mode, the remote SMTP server hostname must match the information in the remote server certificate, and the server certificate must be issued by a CA that is trusted by the Postfix SMTP client (see smtp_tls_CAfile and smtp_tls_CApath). If the remote server certificate doesn't verify or the remote SMTP server hostname doesn't match, and no other server is available, the delivery attempt is deferred and the mail stays in the queue.

The remote SMTP server hostname is verified against any DNS names in the SubjectAlternativeName certificate extension. If no DNS names are specified, the subject CN (CommonName) is used. Verification may be turned off with the smtp_tls_enforce_peername option which is discussed below.

Enforcing TLS in this manner corresponds to a policy setting of verify in the policy table or MUST in the obsolete per-site table. With the new policy table approach, you choose names other the peer hostname to check against the server certificate.

Enforcing the use of TLS is useful if you know that you will only connect to servers that support RFC 2487 and that present server certificates that meet the above requirements. An example would be a client that sends all email to a central mailhub that offers the necessary STARTTLS support. To harden server certificate verification against DNS forgery, see below.

Example:

/etc/postfix/main.cf:
    smtp_enforce_tls = yes

Server certificate verification

When TLS encryption is enforced for all connections, the Postfix TLS client also by default enforces server certificate verification. The relevant configuration parameter is smtp_tls_enforce_peername which defaults to yes.

When TLS is required (smtp_enforce_tls = yes), setting smtp_tls_enforce_peername to "no" disables SMTP server certificate checks. In this case, the mail delivery will proceed regardless of whether the signing CAs are trusted or the subject name(s) listed in the certificate.

Despite the potential for eliminating "man-in-the-middle" and other attacks, mandatory certificate trust chain and subject name verification is not viable as a default Internet mail delivery policy. A significant number of TLS enabled MTAs use self-signed certificates, or certificates that are signed by a private certificate authority. On a machine that delivers mail to the Internet, if you set smtp_enforce_tls = yes, you should probably also set smtp_tls_enforce_peername = no.

You can use the TLS policy table or obsolete per-site table to enable full peer verification for specific destinations that are known to have verifiable TLS server certificates.

Example:

/etc/postfix/main.cf:
    smtp_enforce_tls = yes
    smtp_tls_enforce_peername = no

TLS policy table

As described in the security levels section above, Postfix 2.3 introduces a new more flexible TLS policy table. Read the description of the obsolete Postfix 2.2 per-site table.

A small fraction of servers offer STARTTLS but the negotiation consistently fails. With Postfix 2.3, so long as encryption is not enforced, the delivery is immediately retried with TLS disabled. You no longer need to explicitly disable TLS for the problem destinations. As soon as their TLS software or configuration is repaired, encryption will be used.

The new policy table is specified via the smtp_tls_policy_maps parameter. This lists optional lookup tables with the Postfix SMTP client TLS security policy by next-hop destination. It supersedes the obsolete smtp_tls_per_site parameter. When $smtp_tls_policy_maps is not empty, the smtp_tls_per_site parameter is ignored (a warning is written to the logs if it is also non-empty).

The TLS policy table is indexed by the full next-hop destination, which is either the recipient domain, or the verbatim next-hop specified in the transport table, $local_transport, $virtual_transport, $relay_transport or $default_transport. This includes any enclosing square brackets and any non-default destination server port suffix. The LMTP socket type prefix (inet: or unix:) is not included in the lookup key.

Only the next-hop domain, or $myhostname with LMTP over unix-domain sockets, is used as the nexthop name for certificate verification. The port and any enclosing square brackets are used in the table lookup key, but are not used for server name verification.

The lookup result is a security level, followed by an optional list of whitespace and/or comma separated name=value attributes that override related main.cf settings. The TLS security levels are described above. Below, we describe the level names and attribute syntax:

none
No TLS. No additional attributes are supported at this level.
may
Opportunistic TLS. No additional attributes are supported at this level.
encrypt
Mandatory TLS encryption. At this level and higher the optional cipherlist attribute operrides the main.cf smtp_tls_cipherlist parameter and the optional protocols keyword overrides the main.cf smtp_tls_protocols parameter. In the policy table, multiple protocols must be separated by colons, as attribute values may not contain whitespace or commas.
verify
Mandatory TLS verification. The optional match attribute overrides the main.cf smtp_tls_verify_cert_match parameter. In the policy table, multiple match patterns and strategies must be separated by colons.
secure
Secure-channel TLS. The optional match attribute overrides the main.cf smtp_tls_secure_cert_match parameter. In the policy table, multiple match patterns and strategies must be separated by colons.

Example:

/etc/postfix/main.cf:
    smtp_tls_policy_maps = hash:/etc/postfix/tls_policy
/etc/postfix/tls_policy:
    example.edu             none
    example.mil             may
    example.gov             encrypt
            protocols=SSLv3:TLSv1
            cipherlist=HIGH+aDSS:MEDIUM+aDSS:+RC4:@STRENGTH
    example.com             verify     
            match=hostname:dot-nexthop
            protocols=SSLv3:TLSv1
            cipherlist=kEDH+HIGH:!aNULL:@STRENGTH
    example.net             secure match=.example.net
    [mail.example.org]:587  secure match=nexthop

Note: The hostname strategy if listed in a non-default setting of smtp_tls_secure_cert_match or in the match attribute in the policy table can render the secure level vulnerable to DNS forgery. Do not use the hostname strategy for secure-channel configurations in environments where DNS security is not assured.

Obsolete per-site TLS policy support

This section describes an obsolete per-site TLS policy mechanism. Unlike the new policy table mechanism it implements TLS policy lookup by a potentially untrusted server hostname, and lacks control over what names can appear in server certificates. Because of this, the obsolete mechanism is typically vulnerable to false DNS hostname information in MX or CNAME records. These attacks can be eliminated only with great difficulty. Additionally, the new policy table makes secure channel configurations easier and provides more control over the cipher and protocol selection for sessions with mandatory encryption.

Starting with Postfix 2.3, the underlying TLS enforcement levels are common to the obsolete per-site table and the new policy table. The main.cf smtp_tls_cipherlist and smtp_tls_protocols parameters control the TLS ciphers and protocols for mandatory encryption regardless of which table is used. The smtp_tls_verify_cert_match parameter determines the match strategy for the obsolete MUST keyword in the same way as for the verify level in the new policy.

With Postfix < 2.3, the smtp_tls_cipherlist is applied also for opportunistic TLS sessions, and should be used with care, or not at all. Setting cipherlist restrictions that are incompatible with a remote SMTP server render that server unreachable, TLS handshakes are always attempted and always fail.

When smtp_tls_policy_maps is empty (default) and smtp_tls_per_site is not empty, the per-site table is searched for a policy that matches the following information:

remote SMTP server hostname
This is simply the DNS name of the server that the Postfix SMTP client connects to; this name may be obtained from other DNS lookups, such as MX lookups or CNAME lookups.
next-hop destination
This is normally the domain portion of the recipient address, but it may be overruled by information from the transport(5) table, from the relayhost parameter setting, or from the relay_transport setting. When it's not the recipient domain, the next-hop destination can have the Postfix-specific form "[name]", [name]:port", "name" or "name:port".

When both the hostname lookup and the next-hop lookup succeed, the host policy does not automatically override the next-hop policy. Instead, precedence is given to either the more specific or the more secure per-site policy as described below.

The smtp_tls_per_site table uses a simple "name whitespace value" format. Specify host names or next-hop destinations on the left-hand side; no wildcards are allowed. On the right hand side specify one of the following keywords:

NONE
No TLS. This overrides a less specific MAY lookup result from the alternate host or next-hop lookup key, and overrides the global smtp_use_tls, smtp_enforce_tls, and smtp_tls_enforce_peername settings.
MAY
Opportunistic TLS. This has less precedence than a more specific result (including NONE) from the alternate host or next-hop lookup key, and has less precedence than the more specific global "smtp_enforce_tls = yes" or "smtp_tls_enforce_peername = yes".
MUST_NOPEERMATCH
Mandatory TLS encryption. This overrides a less secure NONE or a less specific MAY lookup result from the alternate host or next-hop lookup key, and overrides the global smtp_use_tls, smtp_enforce_tls and smtp_tls_enforce_peername settings.
MUST
Mandatory TLS verification. This overrides a less secure NONE and MUST_NOPEERMATCH or a less specific MAY lookup result from the alternate host or next-hop lookup key, and overrides the global smtp_use_tls, smtp_enforce_tls and smtp_tls_enforce_peername settings.

The precedences between global (main.cf) and per-site TLS policies can be summarized as follows:

Closing a DNS loophole with obsolete per-site TLS policies

For a discussion of hardened TLS secure-channels in Postfix 2.3 and later see the introduction to security levels and the policy table documentation. What follows is applicable with the obsolete Postfix 2.2 per-site table.

As long as no secure DNS lookup mechanism is available, false hostnames in MX or CNAME responses can change the server hostname that Postfix uses for TLS policy lookup and server certificate verification. Even with a perfect match between the server hostname and the server certificate, there is no guarantee that Postfix is connected to the right server. To avoid this loophole take the following steps:

Example:

/etc/postfix/main.cf:
    smtp_tls_per_site = hash:/etc/postfix/tls_per_site
    relayhost = [msa.example.net]:587

/etc/postfix/tls_per_site:
    # relayhost exact nexthop match
    [msa.example.net]:587       MUST

    # TLS should not be used with the example.org MX hosts.
    example.org                 NONE

    # TLS should not be used with the host smtp.example.com.
    smtp.example.com            NONE

Discovering servers that support TLS

As we decide on a "per site" basis whether or not to use TLS, it would be good to have a list of sites that offered "STARTTLS". We can collect it ourselves with this option.

If the smtp_tls_note_starttls_offer feature is enabled and a server offers STARTTLS while TLS is not already enabled for that server, the Postfix SMTP client logs a line as follows:

postfix/smtp[pid]: Host offered STARTTLS: [hostname.example.com]

Example:

/etc/postfix/main.cf:
    smtp_tls_note_starttls_offer = yes

Server certificate verification depth

When verifying a remote SMTP server certificate, a verification depth of 1 is sufficient if the certificate is directly issued by a CA specified with smtp_tls_CAfile or smtp_tls_CApath. The default value of 5 should also suffice for longer chains (root CA issues special CA which then issues the actual certificate...)

Example:

/etc/postfix/main.cf:
    smtp_tls_scert_verifydepth = 5

Client-side cipher controls

To influence the Postfix SMTP client cipher selection scheme, you can give cipherlist string. A detailed description would go to far here; please refer to the OpenSSL documentation. If you don't know what to do with it, simply don't touch it and leave the (openssl-)compiled in default!

DO NOT USE " to enclose the string, specify just the string!!!

Example:

/etc/postfix/main.cf:
    smtp_tls_cipherlist = DEFAULT

Miscellaneous client controls

The smtp_starttls_timeout parameter limits the time of Postfix SMTP client write and read operations during TLS startup and shutdown handshake procedures. In case of problems the Postfix SMTP client tries the next network address on the mail exchanger list, and defers delivery if no alternative server is available.

Example:

/etc/postfix/main.cf:
    smtp_starttls_timeout = 300s

TLS manager specific settings

The security of cryptographic software such as TLS depends critically on the ability to generate unpredictable numbers for keys and other information. To this end, the tlsmgr(8) process maintains a Pseudo Random Number Generator (PRNG) pool. This is queried by the smtp(8) and smtpd(8) processes when they initialize. By default, these daemons request 32 bytes, the equivalent to 256 bits. This is more than sufficient to generate a 128bit (or 168bit) session key.

Example:

/etc/postfix/main.cf:
    tls_daemon_random_bytes = 32

In order to feed its in-memory PRNG pool, the tlsmgr(8) reads entropy from an external source, both at startup and during run-time. Specify a good entropy source, like EGD or /dev/urandom; be sure to only use non-blocking sources (on OpenBSD, use /dev/arandom when tlsmgr(8) complains about /dev/urandom timeout errors). If the entropy source is not a regular file, you must prepend the source type to the source name: "dev:" for a device special file, or "egd:" for a source with EGD compatible socket interface.

Examples (specify only one in main.cf):

/etc/postfix/main.cf:
    tls_random_source = dev:/dev/urandom
    tls_random_source = egd:/var/run/egd-pool

By default, tlsmgr(8) reads 32 bytes from the external entropy source at each seeding event. This amount (256bits) is more than sufficient for generating a 128bit symmetric key. With EGD and device entropy sources, the tlsmgr(8) limits the amount of data read at each step to 255 bytes. If you specify a regular file as entropy source, a larger amount of data can be read.

Example:

/etc/postfix/main.cf:
    tls_random_bytes = 32

In order to update its in-memory PRNG pool, the tlsmgr(8) queries the external entropy source again after a pseudo-random amount of time. The time is calculated using the PRNG, and is between 0 and the maximal time specified with tls_random_reseed_period. The default maximal time interval is 1 hour.

Example:

/etc/postfix/main.cf:
    tls_random_reseed_period = 3600s

The tlsmgr(8) process saves the PRNG state to a persistent exchange file at regular times and when the process terminates, so that it can recover the PRNG state the next time it starts up. This file is created when it does not exist. Its default location is under the Postfix configuration directory, which is not the proper place for information that is modified by Postfix. Instead, the file location should probably be on the /var partition (but not inside the chroot jail).

Examples:

/etc/postfix/main.cf:
    tls_random_exchange_name = /etc/postfix/prng_exch
    tls_random_prng_update_period = 3600s

Getting started, quick and dirty

The following steps will get you started quickly. Because you sign your own Postfix public key certificate, you get TLS encryption but no TLS authentication. This is sufficient for testing, and for exchanging email with sites that you have no trust relationship with. For real authentication, your Postfix public key certificate needs to be signed by a recognized Certificate Authority, and Postfix needs to be configured with a list of public key certificates of Certificate Authorities, so that Postfix can verify the public key certificates of remote hosts.

In the examples below, user input is shown in bold font, and a "#" prompt indicates a super-user shell.

Reporting problems

When reporting a problem, please be thorough in the report. Patches, when possible, are greatly appreciated too.

Please differentiate when possible between:

Compatibility with Postfix < 2.2 TLS support

Postfix version 2.2 TLS support is based on the Postfix/TLS patch by Lutz Jänicke, but differs in a few minor ways.

The smtp_tls_per_site limitations were removed by the end of the Postfix 2.2 support cycle.

Credits