qpsmtpd/plugins/dspam

#!perl -w

=head1 NAME

dspam - dspam integration for qpsmtpd

=head1 DESCRIPTION

Uses dspam to classify messages. Use B<spamassassin>, B<karma>, and B<naughty>
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<reject agree>.

=head2 reject_type

 reject_type [ perm | temp | disconnect ]

By default, rejects are permanent (5xx). Set I<reject_type temp> to
defer mail instead of rejecting it.

Set I<reject_type disconnect> 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<after> 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<autolearn spamassassin> 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<bayes_auto_learn_threshold_spam>
and I<bayes_auto_learn_threshold_nonspam>. Make sure they are set to
conservative values that will yield no false positives.

If you are using I<autolearn spamassassin> and I<reject>, 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<must> 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 = shift;
    my $qp = 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";
    $self->dspam_process($cmd, $transaction);
    return;
}

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";
    $self->dspam_process($cmd, $transaction);
    return;
}

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;
    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->is_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->is_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;
    };

    if (   $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;
}