DCC(8)             DCC -- Distributed Checksum Clearinghouse            DCC(8)


NAME

     DCC - Distributed Checksum Clearinghouse


DESCRIPTION

     The Distributed Checksum Clearinghouse or DCC is a cooperative, dis-
     tributed system intended to detect "bulk" mail or mail sent to many peo-
     ple.  It allows individuals receiving a single mail message to determine
     that many other people have received essentially identical copies of the
     message and so reject or discard the message.  It can identify some unso-
     licited bulk mail using "spam traps" and other detectors, but that is not
     its focus.

     Freely redistributable source for the server, client, and utilities is
     available at Rhyolite Software

   How the DCC Is Used
     The DCC can be viewed as a tool for end users to enforce their right to
     "opt-in" to streams of bulk mail by refusing bulk mail except from
     sources in a "whitelist."  Whitelists are the responsibility of DCC
     clients, since only they know which bulk mail they solicited.

     The only false positives (mail marked as "bulk" by a DCC server that is
     not) occur when one of the recipients of a message report it to a DCC
     server as having been received many times or when the "fuzzy" checksums
     of differing messages are the same.  The fuzzy checksums ignore aspects
     of messages in order to compute identical checksums for substantially
     identical messages.  The fuzzy checksums are designed to ignore only dif-
     ferences that do not affect meanings.

     It is not reasonable to worry about third parties reporting your incoming
     or outgoing mail to a DCC server as bulk unless you give them copies.  If
     you trust yourself and your correspondents to not report your mutual mail
     as bulk, then false positives are not a concern.

     A DCC server computes a lower bound on the total number of addresses to
     which a message has been sent by counting checksums reported by DCC
     clients.  Each client must decide which bulk messages are unsolicited and
     what degree of "bulkiness" is objectionable.  Client DCC software marks,
     rejects, or discards mail that is bulk according to local thresholds on
     target addresses from DCC servers and unsolicited according to local
     whitelists.  DCC servers are usually configured to receive reports from
     as many targets as possible, including sources that cannot be trusted to
     not exaggerate the number of copies of a message they see.  Any mail sent
     to what a client individually considers a "spam trap" can be seen as
     "definitely bulk," but not necessarily unsolicited by other DCC clients.
     An end user of a DCC client angry about receiving a message could report
     it with 10,000,000 separate DCC packets or with a single report claiming
     as many targets.  An unprincipled user could subscribe a "spam trap" to
     mailing lists such as those of the IETF or CERT.  Such abuses of the sys-
     tem area not problems, because much legitimate mail is "bulk."  You can-
     not reject bulk mail unless you have a whitelist of sources of legitimate
     bulk mail.

     The DCC can also be used by an Internet service provider to detect bulk
     mail coming from its own customers.  In such circumstances, the DCC
     client might be configured to only log bulk mail from unexpected (not
     white-listed) sources.  See the -N option for dccm(8) or dccifd(8).

   What the DCC Is
     A DCC server accumulates counts of cryptographically secure checksums of
     messages but not the messages themselves.  It exchanges reports of fre-
     quently seen checksums with other servers.  DCC clients send reports of
     checksums related to incoming mail to a nearby DCC server running
     dccd(8).  Each report from a client includes the number of recipients for
     the message.  A DCC server accumulates the reports and responds to
     clients the the current total number of recipients for each checksum.
     The client adds an SMTP header to incoming mail containing the total
     counts.  It then discards or rejects mail that is not "white-listed" and
     has counts that exceed local thresholds.

     A special value of the number of addressees is "MANY" and means it is
     certain that this message was bulk and might be unsolicited, perhaps be-
     cause it came from a locally blacklisted source or was addressed to an
     invalid address or "spam trap."  The special value "MANY" is merely the
     largest value that fits in the fixed sized field containing the count of
     addressees.  That "infinity" accumulated total can be reached with mil-
     lions of independent reports as well as with one or two.

     DCC servers share or flood reports of checksums that are seen frequently.
     Each server has its own threshold for determining "frequently," because a
     message sent to 50 addressees in a domain with 60 mailboxes is more like-
     ly to be unsolicited bulk advertising than a message sent to 100 ad-
     dressees in a domain with 600,000 mailboxes.

     To keep a server's database of checksums from growing without bound,
     checksums are forgotten when they become old.  Checksums with large to-
     tals are kept longer.  See dbclean(8).

     DCC clients pick the nearest working DCC server using a small shared or
     memory mapped file, /var/lib/dcc/map. It contains server names, port numbers,
     passwords, recent performance measures, and so forth.  This file allows
     clients to use quick retransmission timeouts and to waste little time on
     servers that have temporarily stopped working or become unreachable.  The
     utility program cdcc(8) is used to maintain this file as well as to check
     the health of servers.

   X-DCC Headers
     The DCC includes several programs used by clients.  Dccm(8) uses the
     sendmail "milter" interface to query a DCC server, add header lines to
     incoming mail, and reject mail whose total checksum counts are high.  Dc-
     cm is intended to be run with SMTP servers using sendmail.

     Dccproc(8) adds header lines to mail presented by file name or stdin, but
     relies on other programs such as procmail to deal with mail with large
     counts.  Dccsight(8) is similar but deals with previously computed check-
     sums.

     Dccifd(8) is similar to dccproc but is not run separately for each mail
     message and so is far more efficient.  It receives mail messages via a
     socket somewhat like dccm, but with a simpler protocol that can be used
     by Perl scripts or other programs.

     DCC SMTP header lines are of the form:

       X-DCC-brand-Metrics: chost server-ID; bulk cknm1=count cknm2=count ...
     where
        brand   is the "brand name" of the DCC server, such as "RHYOLITE".
        chost   is the name or IP address of the DCC client that added the
                header line to the SMTP message.
        server-ID is the numeric ID of the DCC server that the DCC client con-
                tacted.
        bulk    is present if one or more checksum counts exceeded the DCC
                client's thresholds to make the message "bulky."
        cknm1,cknm2,... are types of checksums, and one of
                  IP           address of SMTP client
                  env_From     SMTP envelope value
                  From         SMTP header line


                  Message-ID   SMTP header line
                  Received     last Received: header line in the SMTP message
                  substitute   SMTP header line chosen by the DCC client, pre-
                               fixed with the name of the header
                  Body         SMTP body ignoring white-space
                  Fuz1         filtered or "fuzzy" body checksum
                  Fuz2         another filtered or "fuzzy" body checksum
                Counts for IP, env_From, From, Message-Id, Received, and
                substitute checksums are omitted by the DCC client if the
                server says it has no information.  Counts for Body, Fuz1, and
                Fuz2 are omitted if the message body is empty or contains too
                little of the right kind of information for the checksum to be
                computed.
        count   is the total number of recipients of messages with that check-
                sum reported directly or indirectly to the DCC server.  The
                special count "MANY" means that DCC client have claimed that
                the message is directed at millions of recipients.  "MANY" im-
                ples the message definitely bulk, but not necessarily unso-
                licited.  The special counts "OK" and "OK2" mean the checksum
                has been marked "good" or "half-good" by DCC servers.

     An example header line is:

       X-DCC-RHYOLITE-Metrics: calcite.rhyolite.com 101; Body=16 Fuz1=16 Fuz2=16

     DCC clients commonly accept any mail regardless of other checksum counts
     with at least one "OK" or at least two "OK2" counts among IP, env_from,
     and From checksum counts.  It is common to reject other mail with large
     (including "MANY") counts among Received, Body, Fuz1, and Fuz2 counts.
     It is generally not wise to reject mail based on the other counts.  For
     example, "MAILER-DAEMON" appears to send vast quantities of mail.

   Mailing lists
     Legitimate mailing list traffic differs from spam only in being solicited
     by recipients.  Each client should have a private whitelist.

     DCC whitelists can also mark mail as unsolicited bulk using blacklist en-
     tries for commonly forged marks such as "From: user@public.com".

     Systems that send many essentially identical copies of solicited mail
     such as "auto-responders," should be in the DCC servers whitelists be-
     cause their messages are often substantially identical and so "bulk."

   White and Blacklists
     DCC server and client whitelist files share a common format.  Server
     files are always named whitelist and one is required to be in the DCC
     home directory with the other server files.  Client whitelist files are
     commonly named whiteclnt in the DCC home directory or a subdirectory
     specified with the -U option for dccm(8).  They specify mail that should
     not be reported to a DCC server or that is unsolicited bulk.

     A DCC whitelist file contains blank lines, comments starting with "#",
     and lines of the forms:
           include pathname
           log all-grey
           log no-grey
           count ip hostname
           count env_From 821-path
           count env_To 821-path
           count From 822-mailbox
           count substitute header string
           count Message-ID <string>
           count Received string
           count hex_type hex_cksum


     where
           include can occur only in the main file.
           pathname should be absolute or relative to the DCC home directory.
           log can only be in a DCC client whitelist file and turns on or off
               the logging of messages affected by greylisting in per-user log
               directories.  In a per-user log file, it only affects logging
               for that mail recipient.
           count is null and assumed to be the same as on the previous line or
               one of
                   MANY  indicating millions of targets have received messages
                         with that checksum.
                   OK    if the message is OK.
                   OK2   if it is "half OK."  Two OK2 checksums associated
                         with a message are generally equivalent to an OK.
           hostname is an
                   address   IPv4 or IPv6.
                   block     of 2 to 1024 IPv4 or IPv6 addresses in the stan-
                             dard form xxx.yyy.zzz.www/mm with mm limited for
                             server whitelists to 16 for IPv4 or 112 for IPv6.
                   name      that will be converted to one or more IP address-
                             es.
           821-path is an RFC 821 address.
           822-mailbox is an RFC 822 address with optional name.
           header is the name of an SMTP header such as "Sender" or the name
               of one of two SMTP envlope values, "HELO" or "Mail_Host" for
               the sendmail resolved host name from the 821-path in the mes-
               sage's 821-path.
           hex_type is the string hex followed by a blank and one of the pre-
               ceding checksum types or body, Fuz1, or Fuz2.
           hex_cksum is a string of four hexadecimal numbers obtained from a
               DCC log file.

     A DCC server never shares or floods reports containing checksums marked
     in its whitelist with OK or OK2 to other servers.  A DCC client does not
     report or ask its server about messages with a checksum marked OK or OK2
     in the client whitelist.  This is intended to allow a DCC client to keep
     private mail so private that even its checksums are not disclosed.

     Checksums of the IP address of the SMTP client sending a mail message are
     practically unforgeable, because it is impractical for an SMTP client to
     "spoof" its address or pretend to use some other IP address.  That would
     make the IP address of the sender useful for white-listing, except that
     the IP address of the SMTP client is often not available to users of
     dccproc(8).  In addition, legitimate mail relays make whitelist entries
     for IP addresses of little use.  For example, the IP address from which a
     message arrived might be that of a local relay instead of the home ad-
     dress of a white-listed mailing list.

     Envelope and header From values can be forged, so whitelist entries for
     their checksums are not completely reliable.

     Checksums of env_To values are never sent to DCC servers.  They are valid
     in only whiteclnt files and used only by dccm(8),  dccifd(8),  and other
     DCC clients with access to the envelope Rcpt To value.  They are another
     mechanism used by DCC clients to protect the privacy of some mail.

   Greylists
     The DCC server, dccd(8),  can be used to maintain a greylist database for
     some DCC clients including dccm(8) and dccifd(8).  Greylisting involves
     temporarily refusing mail from unfamiliar SMTP clients and is unrelated
     to Distributed Checksum Clearinghouses.
     See http://projects.puremagic.com/greylisting/

   Privacy
     Because sending mail is a less private act than receiving it, and because
     sending bulk mail is usually not private at all and cannot be very pri-
     vate, the DCC tries first to protect the privacy of mail recipients, and
     second the privacy of senders of mail that is not bulk.

     DCC clients necessarily disclose some information about mail they have
     received.  The DCC database contains checksums of mail bodies, header
     lines, and source addresses.  While it contains significantly less infor-
     mation than is available by "snooping" on Internet links, it is important
     that the DCC database be treated as containing sensitive information and
     to not put the most private information in the DCC database.  Given the
     contents of a message, one might determine whether that message has been
     received by a system that subscribes to the DCC.  Guesses about the
     sender and addressee of a message can also be validated if the checksums
     of the message have been sent to a DCC server.

     Because the DCC is distributed, organizations can operate their own DCC
     servers, and configure them to share or "flood" only the checksums of
     bulk mail that is not in local whitelists.

     DCC clients should not report the checksums of messages known to be pri-
     vate to a DCC server.  For example, checksums of messages local to a sys-
     tem or that are otherwise known a priori to not be unsolicited bulk
     should not be sent to a remote DCC server.  This can accomplished by
     adding entries for the sender to the client's local whitelist file.
     Client whitelist files can also include entries for email recipients
     whose mail should not be reported to a DCC server.

     Additional privacy protections are provided by the thresholds at which
     DCC servers exchange or flood reports.  These thresholds are primarily
     intended to reduce the traffic among DCC servers using the observation
     that the vast majority of messages are sent to a handful of addressees
     and so are useless to other DCC servers.  A DCC server's peer reporting
     thresholds also ensure that checksums shared with peer DCC servers are
     "bulk" and so intrinsically not private.

   Security
     Whenever considering security, one must first consider the risks.  The
     worst DCC security problems are unauthorized commands to a DCC service,
     denial of the DCC service, and corruption of DCC data.  The worst that
     can be done with remote commands to a DCC server is to turn it off or
     otherwise cause it to stop responding.  The DCC is designed to fail
     gracefully, so that a denial of service attack would at worst allow de-
     livery of mail that would otherwise be rejected.  Corruption of DCC data
     might at worst cause mail that is already somewhat "bulk" by virtue of
     being received by two or more people to appear have higher recipient num-
     bers.  Since all DCC users must "white-list" all sources of legitimate
     bulk mail, this is also not a concern.  Such security risks should be ad-
     dressed, but only with defenses that don't cost more than the possible
     damage from an attack..

     The DCC must contend with senders of unsolicited bulk mail who resort to
     unlawful actions to express their displeasure at having their advertising
     blocked.  Because the DCC protocol is based on UDP, an unhappy advertiser
     could try to flood a clearinghouse server with packets supposedly from
     subscribers or non-subscribers.  DCC servers defend against that attack
     by rate-limit requests from non-subscribers.

     Also because of the use of UDP, clients must be protected against forged
     answers to their queries.  Otherwise an unsolicited bulk mail advertiser
     could send a stream of "not spam" answers to an SMTP client while simul-
     taneously sending mail that would otherwise be rejected.  This is not a
     problem for authenticated clients of the DCC because they share a secret
     with the DCC.  Unauthenticated DCC clients do not share any secrets with
     the DCC, except for unique and unpredictable bits in each query or report
     sent to the DCC.  Therefore, DCC servers cryptographically sign answers
     to unauthenticated clients with bits from the corresponding queries.
     This protects against attackers that do not have access to the stream of
     packets from the DCC client.

     The passwords or shared secrets used in the DCC client and server pro-
     grams are "cleartext" for several reasons.  In any shared secret authen-
     tication system, at least one party must know the secret or keep the se-
     cret in cleartext.  You could encrypt the secrets in a file, but because
     they are used by programs, you would need a cleartext copy of the key to
     decrypt the file somewhere in the system, making such a scheme more ex-
     pensive but no more secure than a file of cleartext passwords.  Asymmet-
     ric systems such as that used in UNIX allow one party to not know the se-
     crets, but they must be and are designed to be computationally expensive
     when used in applications like the DCC that involve thousands or more au-
     thentication checks per second.  Moreover, because of "dictionary at-
     tacks," asymmetric systems are now little more secure than keeping pass-
     words in cleartext.  An adversary can compare the hash values of combina-
     tions of common words with /etc/passwd hash values to look for bad pass-
     words.  Worse, by the nature of a client/server protocol like that used
     in the DCC or a UNIX shell login, clients must have the cleartext pass-
     word.  Since it is among the more numerous and much less secure clients
     that adversaries would seek files of DCC passwords, it would be a waste
     to complicate the DCC server with an asymmetric system like that used by
     UNIX.

     The DCC protocol is vulnerable to dictionary attacks to recover pass-
     words.  An adversary could capture some DCC packets, and then check to
     see if any of the 100,000 to 1,000,000 passwords in so called "cracker
     dictionaries" applied to a packet generated the same signature.  This is
     a concern only if DCC passwords are poorly chosen, such as any combina-
     tion of words in an English dictionary.  There are ways to prevent this
     vulnerability regardless of how badly passwords are chosen, but they are
     computationally expensive and require additional network round trips.
     Since DCC passwords are created and typed into files once and do not need
     to be remembered by people, it is cheaper and quite easy to simply choose
     good passwords that are not in dictionaries.

   Reliability
     It is better to fail to filter unsolicited bulk mail than to fail to de-
     liver legitimate mail, so DCC clients fail in the direction of assuming
     that mail is legitimate or even white-listed.

     A DCC client sends a report or other request and waits for an answer.  If
     no answer arrives within a reasonable time, the client retransmits.
     There are many things that might result in the client not receiving an
     answer, but the most important is packet loss.  If the client's request
     does not reach the server, it is easy and harmless for the client to re-
     transmit.  If the client's request reached the server but the server's
     response was lost, a retransmission to the same server would be misunder-
     stood as a new report of another copy of the same message unless it is
     detected as a retransmission by the server.  The DCC protocol includes
     transactions identifiers for this purpose.  If the client retransmitted
     to a second server, the retransmission would be misunderstood by the sec-
     ond server as a new report of the same message.

     Each request from a client includes a timestamp to aid the client in mea-
     suring the round trip time to the server and to let the client pick the
     closest server.  Clients monitor the speed of all of the servers they
     know including those they are not currently using, and use the quickest.

   Client and Server-IDs
     Servers and clients use numbers or IDs to identify themselves.  ID 1 is
     reserved for anonymous, unauthenticated clients.  All other IDs are asso-
     ciated with a pair of passwords in the ids file, the current and next or
     previous and current passwords.  Clients included their client IDs in
     their messages.  When they are not using the anonymous ID, they digitally
     sign their messages to servers with the first password associated with
     their client-ID.  Servers treat messages with signatures that match nei-
     ther of the passwords for the client-ID in their own ids file as if the
     client had used the anonymous ID.

     Each server has a unique server-ID less than 32768.  Servers use their
     IDs to identify checksums that they flood to other servers.  Each server
     expects local clients sending administrative commands to use the server's
     ID and sign administrative commands with the associated password.

     Server-IDs must be unique among all systems that share reports by "flood-
     ing."  All servers must be told of the IDs all other servers whose re-
     ports can be received in the local /var/lib/dcc/flod file described in
     dccd(8).  However, server-IDs can be mapped during flooding between inde-
     pendent DCC organizations.

     Passwd-IDs are server-IDs that should not be assigned to servers but used
     to specify passwords used in the inter-server flooding protocol.  They
     are used in publicly readable configuration files to specify passwords in
     private files.

     The client identified by a client-ID might be a single computer with a
     single IP address, a single but multi-homed computer, or many computers.
     Client-IDs are not used to identify checksum reports, but the organiza-
     tion operating the client.  A client-ID need only be unique among clients
     using a single server.  A single client can use different client-IDs for
     different servers, each client-ID authenticated with a separate password.

     An obscure but important part of all of this is that the inter-server
     flooding algorithm depends on server-IDs and timestamps attached to re-
     ports of checksums.  The inter-server flooding mechanism requires cooper-
     ating DCC servers to maintain reasonable clocks ticking in UTC.  Clients
     include timestamps in their requests, but as long as their timestamps are
     unlikely to be repeated, they need not be very accurate.

   Installation Considerations
     DCC clients on a computer share information about which servers are cur-
     rently working and their speeds in a shared memory segment.  This segment
     also contains server host names, IP addresses, and the passwords needed
     to authenticate known clients to servers.  That generally requires that
     dccm(8),  dccproc(8),  dccifd(8),  and cdcc(8) execute with an UID that
     can write to the DCC home directory and its files.  The sendmail inter-
     face, dccm, is a daemon that can be started by an "rc" or other script
     already running with the correct UID.  The other two, dccproc and cdcc
     need to be set-UID because they are used by end users.  They relinquish
     set-UID privileges when not needed.

     Files that contain cleartext passwords including the shared file used by
     clients must be readable only by "owner."

     The data files required by a DCC can be in a single "home" directory, of-
     ten /var/lib/dcc. Distinct DCC servers can run on a single computer, provided
     they use distinct UDP port numbers and home directories.  It is possible
     and convenient for the DCC clients using a server on the same computer to
     use the same home directory as the server.

     The DCC source distribution includes sample control files.  They should
     be modified appropriately and then copied to the DCC home directory.
     Files that contain cleartext passwords must not be publicly readable.

     The DCC source includes "feature" m4 files to configure sendmail to use
     dccm(8) to check a DCC server about incoming mail.

     See also the INSTALL.txt or INSTALL.html file.

   Client Installation
     Installing a DCC client starts with obtaining or compiling program bina-
     ries for the client server data control tool, cdcc(8).  Installing the
     sendmail DCC interface, dccm(8),  or dccproc(8),  the general or proc-
     mail(1) interface is the main part of the client installation.  Connect-
     ing the DCC to sendmail with dccm is most powerful, but requires adminis-
     trative control of the system running sendmail.

     As noted above, cdcc and dccproc should be set-UID to a suitable UID.
     Root or 0 is thought to be safe for both, because they are careful to re-
     lease privileges except when they need them to read or write files in the
     DCC home directory.  A DCC home directory should be created, often in
     /var/lib/dcc. It must be owned and writable by the UID to which cdcc is set.

     After the DCC client programs have been obtained, contact the operator(s)
     of the chosen DCC server(s) to obtain each server's hostname, port num-
     ber, and a client-ID and corresponding password.  No client-IDs or pass-
     words are needed touse DCC servers that allow anonymous clients.  Use the
     load or add commands of cdcc to create a map file in the DCC home direc-
     tory.  It is usually necessary to create a client whitelist file of the
     format described above.  To accommodate users sharing a computer but not
     ideas about what is solicited bulk mail, the client whitelist file can be
     any valid path name and need not be in the DCC home directory.

     If dccm is chosen, arrange to start it with suitable arguments before
     sendmail is started.  See the homedir/dcc_conf file and the misc/rcDCC
     script in the DCC source.  The procmail DCCM interface, dccproc(8),  can
     be run manually or by a procmailrc(5) rule.

   Server Installation
     The DCC server, dccd(8),  also requires that the DCC home directory ex-
     ist.  It does not use the client shared or memory mapped file of server
     addresses, but it requires other files.  One is the ids file of client-
     IDs,  server-IDs, and corresponding passwords.  Another is a flod file of
     peers that send and receive floods of reports of checksums with large
     counts.  Both files are described in dccd(8).

     The server daemon should be started when the system is rebooted, probably
     before sendmail.  See the misc/rcDCC and misc/start-dccd files in the DCC
     source.

     The database should be cleaned regularly with dbclean(8) such as by run-
     ning the crontab job that is in the misc directory.


SEE ALSO

     cdcc(8),  dbclean(8),  dcc(8),  dccd(8),  dccifd(8),  dccm(8),
     dccproc(8),  dblist(8),  dccsight(8),  sendmail(8).


HISTORY

     The Distributed Checksum Clearinghouse is based on an idea of Paul Vixie
     with code designed and written at Rhyolite Software starting in 2000.
     This describes version 1.2.16.

                               November 19, 2003                             8

Man(1) output converted with man2html modified for the DCC $Date 2001/04/29 03:22:18 $