PPoossttffiixx SSMMTTPP AAcccceessss PPoolliiccyy DDeelleeggaattiioonn ------------------------------------------------------------------------------- PPuurrppoossee ooff PPoossttffiixx SSMMTTPP aacccceessss ppoolliiccyy ddeelleeggaattiioonn The Postfix SMTP server has a number of built-in mechanisms to block or accept mail at specific SMTP protocol stages. In addition, the Postfix SMTP server can delegate decisions to an external policy server (Postfix 2.1 and later). With this policy delegation mechanism, a simple greylist policy can be implemented with only a dozen lines of Perl, as is shown at the end of this document. A complete example can be found in the Postfix source code, in the directory examples/smtpd-policy. Another example of policy delegation is the SPF policy server at http:// www.openspf.org/Software. Policy delegation is now the preferred method for adding policies to Postfix. It's much easier to develop a new feature in few lines of Perl, Python, Ruby, or TCL, than trying to do the same in C code. The difference in performance will be unnoticeable except in the most demanding environments. On active systems a policy daemon process is used multiple times, for up to $max_use incoming SMTP connections. This document covers the following topics: * Policy protocol description * Simple policy client/server configuration * Advanced policy client configuration * Example: greylist policy server * Greylisting mail from frequently forged domains * Greylisting all your mail * Routine greylist maintenance * Example Perl greylist server PPrroottooccooll ddeessccrriippttiioonn The Postfix policy delegation protocol is really simple. The client sends a request and the server sends a response. Unless there was an error, the server must not close the connection, so that the same connection can be used multiple times. The client request is a sequence of name=value attributes separated by newline, and is terminated by an empty line. The server reply is one name=value attribute and it, too, is terminated by an empty line. Here is an example of all the attributes that the Postfix SMTP server sends in a delegated SMTPD access policy request: PPoossttffiixx vveerrssiioonn 22..11 aanndd llaatteerr:: request=smtpd_access_policy protocol_state=RCPT protocol_name=SMTP helo_name=some.domain.tld queue_id=8045F2AB23 sender=foo@bar.tld recipient=bar@foo.tld recipient_count=0 client_address=1.2.3.4 client_name=another.domain.tld reverse_client_name=another.domain.tld instance=123.456.7 PPoossttffiixx vveerrssiioonn 22..22 aanndd llaatteerr:: sasl_method=plain sasl_username=you sasl_sender= size=12345 ccert_subject=solaris9.porcupine.org ccert_issuer=Wietse+20Venema ccert_fingerprint=C2:9D:F4:87:71:73:73:D9:18:E7:C2:F3:C1:DA:6E:04 PPoossttffiixx vveerrssiioonn 22..33 aanndd llaatteerr:: encryption_protocol=TLSv1/SSLv3 encryption_cipher=DHE-RSA-AES256-SHA encryption_keysize=256 etrn_domain= PPoossttffiixx vveerrssiioonn 22..55 aanndd llaatteerr:: stress= PPoossttffiixx vveerrssiioonn 22..99 aanndd llaatteerr:: ccert_pubkey_fingerprint=68:B3:29:DA:98:93:E3:40:99:C7:D8:AD:5C:B9:C9:40 PPoossttffiixx vveerrssiioonn 33..00 aanndd llaatteerr:: client_port=1234 PPoossttffiixx vveerrssiioonn 33..11 aanndd llaatteerr:: policy_context=submission PPoossttffiixx vveerrssiioonn 33..22 aanndd llaatteerr:: server_address=10.3.2.1 server_port=54321 [empty line] Notes: * The "request" attribute is required. In this example the request type is "smtpd_access_policy". * The order of the attributes does not matter. The policy server should ignore any attributes that it does not care about. * When the same attribute name is sent more than once, the server may keep the first value or the last attribute value. * When an attribute value is unavailable, the client either does not send the attribute, sends the attribute with an empty value ("name="), or sends a zero value ("name=0") in the case of a numerical attribute. * The "recipient" attribute is available in the "RCPT TO" stage. It is also available in the "DATA" and "END-OF-MESSAGE" stages if Postfix accepted only one recipient for the current message. The DATA protocol state also applies to email that is received with BDAT commands (Postfix 3.4 and later). * The "recipient_count" attribute (Postfix 2.3 and later) is non-zero only in the "DATA" and "END-OF-MESSAGE" stages. It specifies the number of recipients that Postfix accepted for the current message. The DATA protocol state also applies to email that is received with BDAT commands (Postfix 3.4 and later). * The remote client or local server IP address is an IPv4 dotted quad in the form 1.2.3.4 or it is an IPv6 address in the form 1:2:3::4:5:6. * The remote client or local server port is a decimal number in the range 0- 65535. * For a discussion of the differences between reverse and verified client_name information, see the reject_unknown_client_hostname discussion in the postconf(5) document. * An attribute name must not contain "=", null or newline, and an attribute value must not contain null or newline. * The "instance" attribute value can be used to correlate different requests regarding the same message delivery. These requests are sent over the same policy connection (unless the policy daemon terminates the connection). Once Postfix sends a query with a different instance attribute over that same policy connection, the previous message delivery is either completed or aborted. * The "size" attribute value specifies the message size that the client specified in the MAIL FROM command (zero if none was specified). With Postfix 2.2 and later, it specifies the actual message size after the client sends the END-OF-MESSAGE. * The "sasl_*" attributes (Postfix 2.2 and later) specify information about how the client was authenticated via SASL. These attributes are empty in case of no SASL authentication. * The "ccert_*" attributes (Postfix 2.2 and later) specify information about how the client was authenticated via TLS. These attributes are empty in case of no certificate authentication. As of Postfix 2.2.11 these attribute values are encoded as xtext: some characters are represented by +XX, where XX is the two-digit hexadecimal representation of the character value. With Postfix 2.6 and later, the decoded string is an UTF-8 string without non- printable ASCII characters. * The "encryption_*" attributes (Postfix 2.3 and later) specify information about how the connection is encrypted. With plaintext connections the protocol and cipher attributes are empty and the keysize is zero. * The "etrn_domain" attribute is defined only in the context of the ETRN command, and specifies the ETRN command parameter. * The "stress" attribute is either empty or "yes". See the STRESS_README document for further information. * The "policy_context" attribute provides a way to pass information that is not available via other attributes (Postfix version 3.1 and later). The following is specific to SMTPD delegated policy requests: * Protocol names are ESMTP or SMTP. * Protocol states are CONNECT, EHLO, HELO, MAIL, RCPT, DATA, END-OF-MESSAGE, VRFY or ETRN; these are the SMTP protocol states where the Postfix SMTP server makes an OK/REJECT/HOLD/etc. decision. The DATA protocol state also applies to email that is received with BDAT commands (Postfix 3.4 and later). The policy server replies with any action that is allowed in a Postfix SMTPD access(5) table. Example: action=defer_if_permit Service temporarily unavailable [empty line] This causes the Postfix SMTP server to reject the request with a 450 temporary error code and with text "Service temporarily unavailable", if the Postfix SMTP server finds no reason to reject the request permanently. In case of trouble the policy server must not send a reply. Instead the server must log a warning and disconnect. Postfix will retry the request at some later time. SSiimmppllee ppoolliiccyy cclliieenntt//sseerrvveerr ccoonnffiigguurraattiioonn The Postfix delegated policy client can connect to a TCP socket or to a UNIX- domain socket. Examples: inet:127.0.0.1:9998 unix:/some/where/policy unix:private/policy The first example specifies that the policy server listens on a TCP socket at 127.0.0.1 port 9998. The second example specifies an absolute pathname of a UNIX-domain socket. The third example specifies a pathname relative to the Postfix queue directory; use this for policy servers that are spawned by the Postfix master daemon. To create a policy service that listens on a UNIX-domain socket called "policy", and that runs under control of the Postfix spawn(8) daemon, you would use something like this: 1 /etc/postfix/master.cf: 2 policy unix - n n - 0 spawn 3 user=nobody argv=/some/where/policy-server 4 5 /etc/postfix/main.cf: 6 smtpd_recipient_restrictions = 7 ... 8 reject_unauth_destination 9 check_policy_service unix:private/policy 10 ... 11 policy_time_limit = 3600 12 # smtpd_policy_service_request_limit = 1 NOTES: * Lines 2-3: this creates the service called "policy" that listens on a UNIX- domain socket. The service is implemented by the Postfix spawn(8) daemon, which executes the policy server program that is specified with the aarrggvv attribute, using the privileges specified with the uusseerr attribute. * Line 2: specify a "0" process limit instead of the default "-", to avoid "connection refused" and other problems when you increase the smtpd process limit. * Line 8: reject_unauth_destination is not needed here if the mail relay policy is specified with smtpd_relay_restrictions (available with Postfix 2.10 and later). * Lines 8, 9: always specify "check_policy_service" AFTER "reject_unauth_destination" or else your system could become an open relay. * Line 11: this increases the time that a policy server process may run to 3600 seconds. The default time limit of 1000 seconds is too short; the policy daemon needs to run long as the SMTP server process that talks to it. See the spawn(8) manpage for more information about the transport_time_limit parameter. Note: the "policy_time_limit" parameter will not show up in "postconf" command output before Postfix version 2.9. This limitation applies to many parameters whose name is a combination of a master.cf service name (in the above example, "policy") and a built-in suffix (in the above example: "_time_limit"). * Line 12: specify smtpd_policy_service_request_limit to avoid error-recovery delays with policy servers that cannot maintain a persistent connection. * With Solaris < 9, or Postfix < 2.10 on any Solaris version, use TCP sockets instead of UNIX-domain sockets: 1 /etc/postfix/master.cf: 2 127.0.0.1:9998 inet n n n - 0 spawn 3 user=nobody argv=/some/where/policy-server 4 5 /etc/postfix/main.cf: 6 smtpd_recipient_restrictions = 7 ... 8 reject_unauth_destination 9 check_policy_service inet:127.0.0.1:9998 10 ... 11 127.0.0.1:9998_time_limit = 3600 12 # smtpd_policy_service_request_limit = 1 Configuration parameters that control the client side of the policy delegation protocol: * smtpd_policy_service_default_action (default: 451 4.3.5 Server configuration problem): The default action when an SMTPD policy service request fails. Available with Postfix 3.0 and later. * smtpd_policy_service_max_idle (default: 300s): The amount of time before the Postfix SMTP server closes an unused policy client connection. * smtpd_policy_service_max_ttl (default: 1000s): The amount of time before the Postfix SMTP server closes an active policy client connection. * smtpd_policy_service_request_limit (default: 0): The maximal number of requests per policy connection, or zero (no limit). Available with Postfix 3.0 and later. * smtpd_policy_service_timeout (default: 100s): The time limit to connect to, send to or receive from a policy server. * smtpd_policy_service_try_limit (default: 2): The maximal number of attempts to send an SMTPD policy service request before giving up. Available with Postfix 3.0 and later. * smtpd_policy_service_retry_delay (default: 1s): The delay between attempts to resend a failed SMTPD policy service request. Available with Postfix 3.0 and later. * smtpd_policy_service_policy_context (default: empty): Optional information that is passed in the "policy_context" attribute of an SMTPD policy service request (originally, to share the same SMTPD service endpoint among multiple check_policy_service clients). Available with Postfix 3.1 and later. Configuration parameters that control the server side of the policy delegation protocol: * transport_time_limit ($command_time_limit): The maximal amount of time the policy daemon is allowed to run before it is terminated. The transport is the service name of the master.cf entry for the policy daemon service. In the above examples, the service name is "policy" or "127.0.0.1:9998". AAddvvaanncceedd ppoolliiccyy cclliieenntt ccoonnffiigguurraattiioonn The previous section lists a number of Postfix main.cf parameters that control time limits and other settings for all policy clients. This is sufficient for simple configurations. With more complex configurations it becomes desirable to have different settings per policy client. This is supported with Postfix 3.0 and later. The following example shows a "non-critical" policy service with a short timeout, and with "DUNNO" as default action when the service is unvailable. The "DUNNO" action causes Postfix to ignore the result. 1 /etc/postfix/main.cf: 2 mua_recipient_restrictions = 3 ... 4 reject_unauth_destination 5 check_policy_service { inet:host:port, 6 timeout=10s, default_action=DUNNO 7 policy_context=submission } 8 ... Instead of a server endpoint, we now have a list enclosed in {}. * Line 5: The first item in the list is the server endpoint. This supports the exact same "inet" and "unix" syntax as described earlier. * Line 6-7: The remainder of the list contains per-client settings. These settings override global main.cf parameters, and have the same name as those parameters, without the "smtpd_policy_service_" prefix. Inside the list, syntax is similar to what we already know from main.cf: items separated by space or comma. There is one difference: yyoouu mmuusstt eenncclloossee aa sseettttiinngg iinn ppaarreenntthheesseess,, aass iinn ""{{ nnaammee == vvaalluuee }}"",, iiff yyoouu wwaanntt ttoo hhaavvee ssppaaccee oorr ccoommmmaa wwiitthhiinn aa vvaalluuee oorr aarroouunndd ""=="". This comes in handy when different policy servers require different default actions with different SMTP status codes or text: 1 /etc/postfix/main.cf: 2 smtpd_recipient_restrictions = 3 ... 4 reject_unauth_destination 5 check_policy_service { 6 inet:host:port1, 7 { default_action = 451 4.3.5 See http://www.example.com/ support1 } 8 } 9 ... EExxaammppllee:: ggrreeyylliisstt ppoolliiccyy sseerrvveerr Greylisting is a defense against junk email that is described at http:// www.greylisting.org/. The idea was discussed on the postfix-users mailing list one year before it was popularized. The file examples/smtpd-policy/greylist.pl in the Postfix source tree implements a simplified greylist policy server. This server stores a time stamp for every (client, sender, recipient) triple. By default, mail is not accepted until a time stamp is more than 60 seconds old. This stops junk mail with randomly selected sender addresses, and mail that is sent through randomly selected open proxies. It also stops junk mail from spammers that change their IP address frequently. Copy examples/smtpd-policy/greylist.pl to /usr/libexec/postfix or whatever location is appropriate for your system. In the greylist.pl Perl script you need to specify the location of the greylist database file, and how long mail will be delayed before it is accepted. The default settings are: $database_name="/var/mta/greylist.db"; $greylist_delay=60; The /var/mta directory (or whatever you choose) should be writable by "nobody", or by whatever username you configure below in master.cf for the policy service. Example: # mkdir /var/mta # chown nobody /var/mta Note: DO NOT create the greylist database in a world-writable directory such as /tmp or /var/tmp, and DO NOT create the greylist database in a file system that may run out of space. Postfix can survive "out of space" conditions with the mail queue and with the mailbox store, but it cannot survive a corrupted greylist database. If the file becomes corrupted you may not be able to receive mail at all until you delete the file by hand. The greylist.pl Perl script can be run under control by the Postfix master daemon. For example, to run the script as user "nobody", using a UNIX-domain socket that is accessible by Postfix processes only: 1 /etc/postfix/master.cf: 2 greylist unix - n n - 0 spawn 3 user=nobody argv=/usr/bin/perl /usr/libexec/postfix/greylist.pl 4 5 /etc/postfix/main.cf: 6 greylist_time_limit = 3600 7 smtpd_recipient_restrictions = 8 ... 9 reject_unauth_destination 10 check_policy_service unix:private/greylist 11 ... 12 # smtpd_policy_service_request_limit = 1 Notes: * Lines 2-3: this creates the service called "greylist" that listens on a UNIX-domain socket. The service is implemented by the Postfix spawn(8) daemon, which executes the greylist.pl script that is specified with the aarrggvv attribute, using the privileges specified with the uusseerr attribute. * Line 2: specify a "0" process limit instead of the default "-", to avoid "connection refused" and other problems when you increase the smtpd process limit. * Line 3: Specify "greylist.pl -v" for verbose logging of each request and reply. * Line 6: this increases the time that a greylist server process may run to 3600 seconds. The default time limit of 1000 seconds is too short; the greylist daemon needs to run long as the SMTP server process that talks to it. See the spawn(8) manpage for more information about the transport_time_limit parameter. * Line 9: reject_unauth_destination is not needed here if the mail relay policy is specified with smtpd_relay_restrictions (available with Postfix 2.10 and later). Note: the "greylist_time_limit" parameter will not show up in "postconf" command output before Postfix version 2.9. This limitation applies to many parameters whose name is a combination of a master.cf service name (in the above example, "greylist") and a built-in suffix (in the above example: "_time_limit"). * Line 12: specify smtpd_policy_service_request_limit to avoid error-recovery delays with policy servers that cannot maintain a persistent connection. With Solaris < 9, or Postfix < 2.10 on any Solaris version, use inet: style sockets instead of unix: style, as detailed in the "Policy client/server configuration" section above. 1 /etc/postfix/master.cf: 2 127.0.0.1:9998 inet n n n - 0 spawn 3 user=nobody argv=/usr/bin/perl /usr/libexec/postfix/greylist.pl 4 5 /etc/postfix/main.cf: 6 127.0.0.1:9998_time_limit = 3600 7 smtpd_recipient_restrictions = 8 ... 9 reject_unauth_destination 10 check_policy_service inet:127.0.0.1:9998 11 ... 12 # smtpd_policy_service_request_limit = 1 GGrreeyylliissttiinngg mmaaiill ffrroomm ffrreeqquueennttllyy ffoorrggeedd ddoommaaiinnss It is relatively safe to turn on greylisting for specific domains that often appear in forged email. At some point in cyberspace/time a list of frequently forged MAIL FROM domains could be found at http://www.monkeys.com/anti-spam/ filtering/sender-domain-validate.in. 1 /etc/postfix/main.cf: 2 smtpd_recipient_restrictions = 3 reject_unlisted_recipient 4 ... 5 reject_unauth_destination 6 check_sender_access hash:/etc/postfix/sender_access 7 ... 8 smtpd_restriction_classes = greylist 9 greylist = check_policy_service unix:private/greylist 10 11 /etc/postfix/sender_access: 12 aol.com greylist 13 hotmail.com greylist 14 bigfoot.com greylist 15 ... etcetera ... NOTES: * Line 9: On Solaris < 9, or Postfix < 2.10 on any Solaris version, use inet: style sockets instead of unix: style, as detailed in the "Example: greylist policy server" section above. * Line 5: reject_unauth_destination is not needed here if the mail relay policy is specified with smtpd_relay_restrictions (available with Postfix 2.10 and later). * Line 6: Be sure to specify "check_sender_access" AFTER "reject_unauth_destination" or else your system could become an open mail relay. * Line 3: With Postfix 2.0 snapshot releases, "reject_unlisted_recipient" is called "check_recipient_maps". Postfix 2.1 understands both forms. * Line 3: The greylist database gets polluted quickly with bogus addresses. It helps if you protect greylist lookups with other restrictions that reject unknown senders and/or recipients. GGrreeyylliissttiinngg aallll yyoouurr mmaaiill If you turn on greylisting for all mail you may want to make exceptions for mailing lists that use one-time sender addresses, because each message will be delayed due to greylisting, and the one-time sender addresses can pollute your greylist database relatively quickly. Instead of making exceptions, you can automatically whitelist clients that survive greylisting repeatedly; this avoids most of the delays and most of the database pollution problem. 1 /etc/postfix/main.cf: 2 smtpd_recipient_restrictions = 3 reject_unlisted_recipient 4 ... 5 reject_unauth_destination 6 check_sender_access hash:/etc/postfix/sender_access 7 check_policy_service unix:private/policy 8 ... 9 10 /etc/postfix/sender_access: 11 securityfocus.com OK 12 ... NOTES: * Line 7: On Solaris < 9, or Postfix < 2.10 on any Solaris version, use inet: style sockets instead of unix: style, as detailed in the "Example: greylist policy server" section above. * Line 5: reject_unauth_destination is not needed here if the mail relay policy is specified with smtpd_relay_restrictions (available with Postfix 2.10 and later). * Lines 6-7: Be sure to specify check_sender_access and check_policy_service AFTER reject_unauth_destination or else your system could become an open mail relay. * Line 3: The greylist database gets polluted quickly with bogus addresses. It helps if you precede greylist lookups with restrictions that reject unknown senders and/or recipients. RRoouuttiinnee ggrreeyylliisstt mmaaiinntteennaannccee The greylist database grows over time, because the greylist server never removes database entries. If left unattended, the greylist database will eventually run your file system out of space. When the status file size exceeds some threshold you can simply rename or remove the file without adverse effects; Postfix automatically creates a new file. In the worst case, new mail will be delayed by an hour or so. To lessen the impact, rename or remove the file in the middle of the night at the beginning of a weekend. EExxaammppllee PPeerrll ggrreeyylliisstt sseerrvveerr This is the Perl subroutine that implements the example greylist policy. It is part of a general purpose sample policy server that is distributed with the Postfix source as examples/smtpd-policy/greylist.pl. # # greylist status database and greylist time interval. DO NOT create the # greylist status database in a world-writable directory such as /tmp # or /var/tmp. DO NOT create the greylist database in a file system # that can run out of space. # $database_name="/var/mta/greylist.db"; $greylist_delay=60; # # Auto-whitelist threshold. Specify 0 to disable, or the number of # successful "come backs" after which a client is no longer subject # to greylisting. # $auto_whitelist_threshold = 10; # # Demo SMTPD access policy routine. The result is an action just like # it would be specified on the right-hand side of a Postfix access # table. Request attributes are available via the %attr hash. # sub smtpd_access_policy { my($key, $time_stamp, $now); # Open the database on the fly. open_database() unless $database_obj; # Search the auto-whitelist. if ($auto_whitelist_threshold > 0) { $count = read_database($attr{"client_address"}); if ($count > $auto_whitelist_threshold) { return "dunno"; } } # Lookup the time stamp for this client/sender/recipient. $key = lc $attr{"client_address"}."/".$attr{"sender"}."/".$attr{"recipient"}; $time_stamp = read_database($key); $now = time(); # If new request, add this client/sender/recipient to the database. if ($time_stamp == 0) { $time_stamp = $now; update_database($key, $time_stamp); } # The result can be any action that is allowed in a Postfix access(5) map. # # To label the mail, return ``PREPEND headername: headertext'' # # In case of success, return ``DUNNO'' instead of ``OK'', so that the # check_policy_service restriction can be followed by other restrictions. # # In case of failure, return ``DEFER_IF_PERMIT optional text...'', # so that mail can still be blocked by other access restrictions. # syslog $syslog_priority, "request age %d", $now - $time_stamp if $verbose; if ($now - $time_stamp > $greylist_delay) { # Update the auto-whitelist. if ($auto_whitelist_threshold > 0) { update_database($attr{"client_address"}, $count + 1); } return "dunno"; } else { return "defer_if_permit Service temporarily unavailable"; } }