New Replay Cache rst documentation

doc/rst_source/krb_basic/index.rst

@@ -7,6 +7,7 @@ Kerberos V5 concepts
 .. toctree::
    :maxdepth: 1
 
-   keytab_def.rst
-   stash_file_def.rst
-   date_format.rst
+   keytab_def
+   rcache_def
+   stash_file_def
+   date_format

doc/rst_source/krb_basic/rcache_def.rst

@@ -0,0 +1,97 @@
+.. _rcache_definition:
+
+replay cache
+============
+
+A replay cache or "rcache" keeps track of all athenticators recently
+presented to a given service.  If the duplicate authentication
+requests are detected in the rcache, the error message is sent to
+the application program.
+
+The replay cache interface, like the credential cache and
+:ref:`keytab_definition` interfaces, uses `type:value` strings to
+indicate the type of replay cache and any associated cache naming
+data to use.
+
+Background information
+----------------------
+
+Some Kerberos or GSSAPI services use a simple authentication mechanism
+where a message is sent containing an authenticator, and that
+establishes the encryption key that the client will use for talking to
+the service.  But nothing about that prevents an eavesdropper from
+recording the messages sent by the client, establishing a new
+connection, and re-sending or "replaying" the same messages; the
+replayed authenticator will establish the same encryption key for the
+new session, and the following messages will be decrypted and
+processed.  The attacker may not know what the messages say, and can't
+generate new messages under the same encryption key, but in some
+instances it may be harmful to the user (or helpful to the attacker)
+to cause the server to see the same messages again a second time.  For
+example, if the legitimate client sends "delete first message in
+mailbox", a replay from an attacker may delete another, different
+"first" message.  (Protocol design to guard against such problems has
+been discussed in :rfc:`4120#section-10`.)
+
+Even if one protocol uses further protection to verify that the client
+side of the connection actually knows the encryption keys (and thus is
+presumably a legitimate user), if another service uses the same
+service principal name, it may be possible to record an authenticator
+used with the first protocol and "replay" it against the second.
+
+The replay cache mitigates these attacks somewhat, by keeping track of
+authenticators that have been seen until their five-minute window
+expires.  Different authenticators generated by multiple connections
+from the same legitimate client will generally have different
+timestamps, and thus will not be considered the same.
+
+This mechanism isn't perfect.  If a message is sent to one application
+server but a man-in-the-middle attacker can prevent it from actually
+arriving at that server, the attacker could then use the authenticator
+(once!) against a different service on the same host.  This could be a
+problem if the message from the client included something more than
+authentication in the first message that could be useful to the
+attacker (which is uncommon; in most protocols the server has to
+indicate a successful authentication before the client sends
+additional messages), or if the simple act of presenting the
+authenticator triggers some interesting action in the service being
+attacked.
+
+Default rcache type
+-------------------
+
+There is currently only one implemented kind of replay cache, called
+**dfl**.  It stores data in one large file, occasionally rewriting it
+to purge old, expired entries.
+
+The default type can be overridden by **KRB5RCACHETYPE** environment
+variable.
+
+The placement of rcache file is determined by the following:
+
+#. The **KRB5RCACHEDIR** environment variable;
+
+#. If KRB5RCACHEDIR is unspecified, on UNIX, the library
+   will fall back to the environment variable **TMPDIR**, and then to
+   a temporary directory determined at configuration time such as
+   */tmp* or */var/tmp*; on Windows, it will check the environment
+   variables *TEMP* and *TMP*, and fall back to the directory C:\\.
+
+Performance issues
+------------------
+
+Several known minor performance issues that may occur when replay
+cache is enabled on the Kerberos system include: delays due to writing
+the authenticator data to disk slowing down response time for very
+heavily loaded servers, and delays during the rewrite that may be
+unacceptable to high-performance services.
+
+For use cases where replays are adequately defended against for all
+protocols using a given service principal name, or where performance
+or other considerations outweigh the risk of replays, the special
+replay cache type "none" can be specified::
+
+    KRB5RCACHETYPE=none
+
+It doesn't record any information about authenticators, and reports
+that any authenticator seen is not a replay.