#!perl -w =head1 NAME dspam - dspam integration for qpsmtpd =head1 DESCRIPTION Uses dspam to classify messages. Use B, B, and B to train dspam. Adds the X-DSPAM-Result and X-DSPAM-Signature headers to messages. The latter is essential for training dspam and the former is useful to MDAs, MUAs, and humans. Adds a transaction note to the qpsmtpd transaction. The note is a hashref with at least the 'class' field (Spam,Innocent,Whitelisted). It will normally contain a probability and confidence rating. =head1 TRAINING DSPAM If you enable dspam rejection without training first, you will lose valid mail. The dspam false positive rate is high when untrained. The good news is; dspam learns very, very fast. The best method way to train dspam is to feed it two large equal sized corpuses of spam and ham from your mail server. The dspam authors suggest avoiding public corpuses. I train dspam as follows: =over 4 =item learn from SpamAssassin See the SPAMASSASSIN section. =item periodic training I have a script that searches the contents of every users maildir. Any read messages that have changed since the last processing run are learned as ham or spam. The ham message list consists of read messages in any folder not named like Spam, Junk, Trash, or Deleted. This catches messages that users have read and left in their inbox or filed away into subfolders. =item on-the-fly training The dovecot IMAP server has an antispam plugin that will train dspam when messages are moved to/from the Spam folder. =back =head1 CONFIG =head2 dspam_bin The path to the dspam binary. If yours is installed somewhere other than /usr/local/bin/dspam, set this. =head2 autolearn [ naughty | karma | spamassassin | any ] =over 4 =item naughty learn naughty messages as spam (see plugins/naughty) =item karma learn messages with negative karma as spam (see plugins/karma) =item spamassassin learn from spamassassins messages with autolearn=(ham|spam). See SPAMASSASSIN. =item any all of the above, and any future tests too! =back =head2 reject Set to a floating point value between 0 and 1.00 where 0 is no confidence and 1.0 is 100% confidence. If dspam's confidence is greater than or equal to this threshold, the message will be rejected. The default is 1.00. dspam reject .95 To only reject mail if dspam and spamassassin both think the message is spam, set I. =head2 reject_type reject_type [ perm | temp | disconnect ] By default, rejects are permanent (5xx). Set I to defer mail instead of rejecting it. Set I if you'd prefer to immediately disconnect the connection when a spam is encountered. This prevents the remote server from issuing a reset and attempting numerous times in a single connection. =head1 dspam.conf dspam must be configured and working properly. I had to modify the following settings on my system: =over 4 =item mysql storage =item Trust smtpd =item TrainingMode tum =item Tokenizer osb =item Preference "trainingMode=TOE" =item Preference "spamAction=deliver" =item Preference "signatureLocation=headers" =item TrainPristine off =item ParseToHeaders off =back Of those changes, the most important is the signature location. This plugin only supports storing the signature in the headers. If you want to train dspam after delivery (ie, users moving messages to/from spam folders), then the dspam signature must be in the headers. When using the dspam MySQL backend, use InnoDB tables. DSPAM training is dramatically slowed by MyISAM table locks and dspam requires a lot of training. InnoDB has row level locking and updates are much faster. =head1 DSPAM periodic maintenance Install this cron job to clean up your DSPAM database. http://dspam.git.sourceforge.net/git/gitweb.cgi?p=dspam/dspam;a=tree;f=contrib/dspam_maintenance;hb=HEAD =head1 SPAMASSASSIN DSPAM can be trained by SpamAssassin. This relationship between them requires attention to several important details: =over 4 =item 1 dspam must be listed B spamassassin in the config/plugins file. Because SA runs first, I set the SA reject_threshold up above 100 so that all spam messages will be used to train dspam. Once dspam is trained and errors are rare, I plan to run dspam first and reduce the SA load. =item 2 For I to work, autolearn must be enabled and configured in SpamAssassin. SA autolearn will determine if a message is learned by dspam. The settings to pay careful attention to in your SA local.cf file are I and I. Make sure they are set to conservative values that will yield no false positives. If you are using I and I, messages that exceed the SA threshholds will cause dspam to reject them. Again I say, make sure the SA autolearn threshholds are set high enough to avoid false positives. =back =head1 MULTIPLE RECIPIENT BEHAVIOR For messages with multiple recipients, the user that dspam is running as will be the dspam username. When messages have a single recipient, the recipient address is used as the dspam username. For dspam to trust qpsmtpd with modifying the username, you B add the username that qpsmtpd is running to to the dspamd.conf file. ie, (Trust smtpd). =head1 CHANGES 2012-06 - Matt Simerson - added karma & naughty learning support - worked around the DESTROY bug in dspam_process =head1 AUTHOR 2012 - Matt Simerson =cut use strict; use warnings; use lib 'lib'; use Qpsmtpd::Constants; use Qpsmtpd::DSN; use IO::Handle; use Socket qw(:DEFAULT :crlf); sub register { my ($self, $qp) = (shift, shift); $self->log(LOGERROR, "Bad parameters for the dspam plugin") if @_ % 2; $self->{_args} = { @_ }; $self->{_args}{reject} = 1 if ! defined $self->{_args}{reject}; $self->{_args}{reject_type} ||= 'perm'; $self->{_args}{dspam_bin} ||= '/usr/local/bin/dspam'; $self->get_dspam_bin() or return DECLINED; $self->register_hook('data_post', 'data_post_handler'); } sub get_dspam_bin { my $self = shift; my $bin = $self->{_args}{dspam_bin}; if ( ! -e $bin ) { $self->log(LOGERROR, "error, dspam CLI binary not found: install dspam and/or set dspam_bin"); return; }; if ( ! -x $bin ) { $self->log(LOGERROR, "error, no permission to run $bin"); return; }; return $bin; }; sub data_post_handler { my $self = shift; my $transaction = shift || $self->qp->transaction; return (DECLINED) if $self->is_immune(); if ( $transaction->data_size > 500_000 ) { $self->log(LOGINFO, "skip, too big (" . $transaction->data_size . ")" ); return (DECLINED); }; my $user = $self->select_username( $transaction ); my $bin = $self->{_args}{dspam_bin}; my $filtercmd = "$bin --user $user --mode=tum --process --deliver=summary --stdout"; $self->log(LOGDEBUG, $filtercmd); my $response = $self->dspam_process( $filtercmd, $transaction ); if ( ! $response->{result} ) { $self->log(LOGWARN, "error, no dspam response. Check logs for errors."); return (DECLINED); }; $transaction->notes('dspam', $response); $self->attach_headers( $response, $transaction ); $self->autolearn( $response, $transaction ); return $self->log_and_return( $transaction ); }; sub select_username { my ($self, $transaction) = @_; my $recipient_count = scalar $transaction->recipients; $self->log(LOGDEBUG, "Message has $recipient_count recipients"); if ( $recipient_count > 1 ) { $self->log(LOGINFO, "multiple recipients ($recipient_count), ignoring user prefs"); return getpwuid($>); }; # use the recipients email address as username. This enables user prefs my $username = ($transaction->recipients)[0]->address; return lc($username); }; sub assemble_message { my ($self, $transaction) = @_; my $message = "X-Envelope-From: " . $transaction->sender->format . "\n" . $transaction->header->as_string . "\n\n"; $transaction->body_resetpos; while (my $line = $transaction->body_getline) { $message .= $line; }; $message = join(CRLF, split /\n/, $message); return $message . CRLF; }; sub parse_response { my $self = shift; my $response = shift or do { $self->log( LOGDEBUG, "missing dspam response!" ); return; }; # example DSPAM results: # user@example.com; result="Spam"; class="Spam"; probability=1.0000; confidence=1.00; signature=N/A # smtpd; result="Innocent"; class="Innocent"; probability=0.0023; confidence=1.00; signature=4f8dae6a446008399211546 #return $self->parse_response_regexp( $response ); # probably slower my ($user, $result, $class, $prob, $conf, $sig) = split /; /, $response; (undef, $result) = split /=/, $result; (undef, $class ) = split /=/, $class; (undef, $prob ) = split /=/, $prob; (undef, $conf ) = split /=/, $conf; (undef, $sig ) = split /=/, $sig; $result = substr($result, 1, -1); # strip off quotes $class = substr($class, 1, -1); return { class => $class, result => $result, probability => $prob, confidence => $conf, signature => $sig, }; }; sub parse_response_regexp { my ($self, $response) = @_; my ($result, $class, $prob, $conf, $sig) = $response =~ / result=\"(Spam|Innocent)\";\s class=\"(Spam|Innocent)\";\s probability=([\d\.]+);\s confidence=([\d\.]+);\s signature=(.*) /x; return { class => $class, result => $result, probability => $prob, confidence => $conf, signature => $sig, }; }; sub dspam_process { my ( $self, $filtercmd, $transaction ) = @_; my $response = $self->dspam_process_backticks( $filtercmd ); #my $response = $self->dspam_process_open2( $filtercmd, $transaction ); #my $response = $self->dspam_process_fork( $filtercmd ); return $self->parse_response( $response ); }; sub dspam_process_fork { my ( $self, $filtercmd, $transaction ) = @_; # yucky. This method (which forks) exercises a bug in qpsmtpd. When the # child exits, the Transaction::DESTROY method is called, which deletes # the spooled file from disk. The contents of $self->qp->transaction # needed to spool it again are also destroyed. Don't use this. my $message = $self->assemble_message( $transaction ); my $in_fh; if (! open($in_fh, '-|')) { # forks child for writing open(my $out_fh, "|$filtercmd") or die "Can't run $filtercmd: $!\n"; print $out_fh $message; close $out_fh; exit(0); }; my $response = <$in_fh>; close $in_fh; chomp $response; $self->log(LOGDEBUG, $response); return $response; }; sub dspam_process_backticks { my ( $self, $filtercmd ) = @_; my $transaction = $self->qp->transaction; my $message = $self->temp_file(); open my $fh, '>', $message; print $fh "X-Envelope-From: " . $transaction->sender->format . CRLF . $transaction->header->as_string . CRLF . CRLF; $transaction->body_resetpos; while (my $line = $transaction->body_getline) { print $fh $line; }; close $fh; my ($line1) = split /[\r|\n]/, `$filtercmd < $message`; $self->log(LOGDEBUG, $line1); return $line1; }; sub dspam_process_open2 { my ( $self, $filtercmd, $transaction ) = @_; my $message = $self->assemble_message( $transaction ); # not sure why, but this is not as reliable as I'd like. What's a dspam # error -5 mean anyway? use FileHandle; use IPC::Open3; my ($read, $write, $err); use Symbol 'gensym'; $err = gensym; my $pid = open3($write, $read, $err, $filtercmd); print $write $message; close $write; #my $response = join('', <$dspam_out>); # get full response my $response = <$read>; # get first line only waitpid $pid, 0; my $child_exit_status = $? >> 8; #$self->log(LOGINFO, "exit status: $child_exit_status"); if ( $response ) { chomp $response; $self->log(LOGDEBUG, $response); }; my $err_msg = <$err>; if ( $err_msg ) { $self->log(LOGDEBUG, $err_msg ); }; return $response; }; sub log_and_return { my $self = shift; my $transaction = shift || $self->qp->transaction; my $d = $self->get_dspam_results( $transaction ) or return DECLINED; if ( ! $d->{class} ) { $self->log(LOGWARN, "skip, no dspam class detected"); return DECLINED; }; my $status = "$d->{class}, $d->{confidence} c."; my $reject = $self->{_args}{reject} or do { $self->log(LOGINFO, "skip, reject disabled ($status)"); return DECLINED; }; if ( $reject eq 'agree' ) { return $self->reject_agree( $transaction ); }; if ( $d->{class} eq 'Innocent' ) { $self->log(LOGINFO, "pass, $status"); return DECLINED; }; if ( $self->qp->connection->relay_client ) { $self->log(LOGINFO, "skip, allowing spam, user authenticated ($status)"); return DECLINED; }; if ( $d->{probability} <= $reject ) { $self->log(LOGINFO, "pass, $d->{class} probability is too low ($d->{probability} < $reject)"); return DECLINED; }; if ( $d->{confidence} != 1 ) { $self->log(LOGINFO, "pass, $d->{class} confidence is too low ($d->{confidence})"); return DECLINED; }; # dspam is more than $reject percent sure this message is spam $self->log(LOGINFO, "fail, $d->{class}, ($d->{confidence} confident)"); my $deny = $self->get_reject_type(); return Qpsmtpd::DSN->media_unsupported($deny, 'dspam says, no spam please'); } sub reject_agree { my ($self, $transaction ) = @_; my $sa = $transaction->notes('spamassassin' ); my $d = $transaction->notes('dspam' ); my $status = "$d->{class}, $d->{confidence} c"; if ( ! $sa->{is_spam} ) { $self->log(LOGINFO, "pass, cannot agree, SA results missing ($status)"); return DECLINED; }; if ( $d->{class} eq 'Spam' ) { if ( $sa->{is_spam} eq 'Yes' ) { $self->adjust_karma( -2 ); $self->log(LOGINFO, "fail, agree, $status"); my $reject = $self->get_reject_type(); return ($reject, 'we agree, no spam please'); }; $self->log(LOGINFO, "fail, disagree, $status"); return DECLINED; }; if ( $d->{class} eq 'Innocent' ) { if ( $sa->{is_spam} eq 'No' ) { if ( $d->{confidence} > .9 ) { $self->adjust_karma( 1 ); }; $self->log(LOGINFO, "pass, agree, $status"); return DECLINED; }; $self->log(LOGINFO, "pass, disagree, $status"); return DECLINED; }; $self->log(LOGINFO, "pass, other $status"); return DECLINED; }; sub get_dspam_results { my $self = shift; my $transaction = shift || $self->qp->transaction; if ( $transaction->notes('dspam') ) { return $transaction->notes('dspam'); }; my $string = $transaction->header->get('X-DSPAM-Result') or do { $self->log(LOGWARN, "get_dspam_results: failed to find the header"); return; }; my @bits = split /,\s+/, $string; chomp @bits; my $class = shift @bits; my %d; foreach (@bits) { my ($key,$val) = split /=/, $_; $d{$key} = $val; }; $d{class} = $class; my $message = $d{class}; if ( defined $d{probability} && defined $d{confidence} ) { $message .= ", prob: $d{probability}, conf: $d{confidence}"; }; $self->log(LOGDEBUG, $message); $transaction->notes('dspam', \%d); return \%d; }; sub attach_headers { my ($self, $r, $transaction) = @_; $transaction ||= $self->qp->transaction; my $header_str = "$r->{result}, probability=$r->{probability}, confidence=$r->{confidence}"; $self->log(LOGDEBUG, $header_str); my $name = 'X-DSPAM-Result'; $transaction->header->delete($name) if $transaction->header->get($name); $transaction->header->add($name, $header_str, 0); # the signature header is required if you intend to train dspam later. # In dspam.conf, set: Preference "signatureLocation=headers" $transaction->header->add('X-DSPAM-Signature', $r->{signature}, 0); }; sub train_error_as_ham { my $self = shift; my $transaction = shift; my $user = $self->select_username( $transaction ); my $dspam_bin = $self->{_args}{dspam_bin} || '/usr/local/bin/dspam'; my $cmd = "$dspam_bin --user $user --mode=toe --source=error --class=innocent --deliver=summary --stdout"; my $response = $self->dspam_process( $cmd, $transaction ); if ( $response ) { $transaction->notes('dspam', $response); } else { $transaction->notes('dspam', { class => 'Innocent', result => 'Innocent', confidence=>1 } ); }; }; sub train_error_as_spam { my $self = shift; my $transaction = shift; my $user = $self->select_username( $transaction ); my $dspam_bin = $self->{_args}{dspam_bin} || '/usr/local/bin/dspam'; my $cmd = "$dspam_bin --user $user --mode=toe --source=error --class=spam --deliver=summary --stdout"; my $response = $self->dspam_process( $cmd, $transaction ); if ( $response ) { $transaction->notes('dspam', $response); } else { $transaction->notes('dspam', { class => 'Spam', result => 'Spam', confidence=>1 } ); }; }; sub autolearn { my ( $self, $response, $transaction ) = @_; defined $self->{_args}{autolearn} or return; if ( $self->{_args}{autolearn} ne 'any' && $self->{_args}{autolearn} ne 'karma' && $self->{_args}{autolearn} ne 'naughty' && $self->{_args}{autolearn} ne 'spamassassin' ) { $self->log(LOGERROR, "bad autolearn setting! Read 'perldoc plugins/dspam' again!"); return; }; # only train once. $self->autolearn_naughty( $response, $transaction ) and return; $self->autolearn_karma( $response, $transaction ) and return; $self->autolearn_spamassassin( $response, $transaction ) and return; }; sub autolearn_naughty { my ( $self, $response, $transaction ) = @_; my $learn = $self->{_args}{autolearn} or return; if ( $learn ne 'naughty' && $learn ne 'any' ) { $self->log(LOGDEBUG, "skipping naughty autolearn"); return; }; if ( $self->connection->notes('naughty') && $response->{result} eq 'Innocent' ) { $self->log(LOGINFO, "training naughty FN message as spam"); $self->train_error_as_spam( $transaction ); return 1; }; $self->log(LOGDEBUG, "falling through naughty autolearn"); return; }; sub autolearn_karma { my ( $self, $response, $transaction ) = @_; my $learn = $self->{_args}{autolearn} or return; return if ( $learn ne 'karma' && $learn ne 'any' ); my $karma = $self->connection->notes('karma'); return if ! defined $karma; if ( $karma < -2 && $response->{result} eq 'Innocent' ) { $self->log(LOGINFO, "training bad karma ($karma) FN as spam"); $self->train_error_as_spam( $transaction ); return 1; }; if ( $karma > 2 && $response->{result} eq 'Spam' ) { $self->log(LOGINFO, "training good karma ($karma) FP as ham"); $self->train_error_as_ham( $transaction ); return 1; }; return; }; sub autolearn_spamassassin { my ( $self, $response, $transaction ) = @_; my $learn = $self->{_args}{autolearn} or return; return if ( $learn ne 'spamassassin' && $learn ne 'any' ); my $sa = $transaction->notes('spamassassin' ); if ( ! $sa || ! $sa->{is_spam} ) { if ( ! $self->connection->notes('naughty') ) { $self->log(LOGERROR, "SA results missing"); # SA skips naughty }; return; }; if ( ! $sa->{autolearn} ) { $self->log(LOGERROR, "SA autolearn unset"); return; }; if ( $sa->{is_spam} eq 'Yes' && $sa->{autolearn} eq 'spam' && $response->{result} eq 'Innocent' ) { $self->log(LOGINFO, "training SA FN as spam"); $self->train_error_as_spam( $transaction ); return 1; } elsif ( $sa->{is_spam} eq 'No' && $sa->{autolearn} eq 'ham' && $response->{result} eq 'Spam' ) { $self->log(LOGINFO, "training SA FP as ham"); $self->train_error_as_ham( $transaction ); return 1; }; return; };