Viper OpenLDAP backend
The complete Viper suite is named after its most important component -- Viper, a custom Perl-based backend for the OpenLDAP server. The backend builds on LDAP's great strenghts while introducing a number of features to solve deficiencies in LDAP when it comes to configuration management.
Generally speaking, the backend can be used in regular LDAP scenarios, in places where you want quick results on a platform that already has extra features and also lends itself to further customization.
However, its main role is serving as the backend for automatic system installations and configurations, where the hosts are Debian-based systems retrieving configuration data using various methods, such as HTTP preseeding, Debconf, Puppet etc.
A "rapid fire" list of Viper features includes search rewriting, default entries, default attributes, dynamic values, expansion to values of other fields or files, sub-searching, entry relocation, caching of computed values, smart ADD and MODIFY handling, and server-side Debconf prompting.
The implemented features allow LDAP to act as a powerful central data store, without any data duplication or need for batch scripts that would periodically re-generate config files -- all data needs to be specified only once, all components contact LDAP on the fly, and all data served by LDAP is current.
Having a custom backend also brings one other crucial benefit to the picture -- the ability to make provisions for existing components and their use of LDAP (DHCP server, puppet, Debconf) without requiring modifications to their source code.
Summary of features
Here's a more elaborate listing of Viper features, grouped by category:
Server data is kept in form of LDIF files in the filesystem (/var/lib/ldap/viper/), where each entry is represented by a directory structure (DN components) and a single file representing the final entry (leaf). It means that internally, on a low level, you can modify LDAP values like a filesystem operation (editing files and directories), that you can use symlinks and hardlinks, and that you can set file permissions to affect reading and writing.
Files have a mandatory suffix, so that they are distinguishable from the corresponding directories where the subentries are kept. Here's an example of disk structure:
s1:/var/lib/ldap/viper# find . . ./ou=dhcp.ldif ./ou=dhcp ./ou=dhcp/ou=eth1.ldif ./ou=dhcp/cn=s1.ldif ./ou=clients.ldif ./ou=defaults ./ou=defaults/ou=hosts ./ou=defaults/ou=hosts/cn=netcfg.get_domain.ldif ./ou=defaults/ou=hosts/cn=partman.confirm.ldif ./ou=defaults/ou=hosts/cn=preseed.early_command.ldif ./ou=defaults/ou=hosts/cn=console-keymaps-at.keymap.ldif ./ou=defaults/ou=hosts/cn=time.zone.ldif
Each file is expected to contain only one entry, and so within each file, exactly and only the first entry is relevant. Everything in there that is either another entry or just invalid data, after or within the first entry, is ignored.
There is a config option available to allow ADD operation to overwrite existing entries without throwing ALREADY_EXISTS error. See 'addoverwrites'
There is a config option available to allow ADD operation to ignore adds on existing entries without throwing ALREADY_EXISTS error. Applicable when addoverwrites=false. See 'addignoredups'
When entry is added somewhere under a Debconf tree
(configurable via a list of regexes), it is possible to check the list
of entry's variables and see if the entry should be relocated to other
parts of the tree. This is used to move keys up from their host-specific
location, to become site-wide or global default. See
AddRelocate functionality is debconf- and client- or server-specific, and requires that Debconf, running on client or server side, asks the regular question, and another question about where the answer should be saved. Thanks to Debconf's support for "transactions" and its capable frontends, both questions are usually asked on the same screen, practically without overhead. Both server-side prompter and client-side LDAP-enabled backend and frontend that implement this described scenario are available.
When a modify request is issued and the specific entry to modify is there, everything goes on as usual. But if the entry is not exactly there (i.e. it comes from a fallback), then it's possible to either return NO_SUCH_OBJECT (as if the entry was not found at all), or to copy & modify the fallback to the expected name, effectively creating the target entry in the process. See 'modifycopyonwrite'
Some of the entry's attributes may actually not be present in the entry, but come from so-called 'appended values' (see 'Searching' features below). If a modify request is issued and the modification is to be performed on an attribute that does really exist, everything goes on as usual. But if it is one of the appended attributes, we retrieve its values and add it to the entry as if it was there all along. Then we let the modify procedure continue, so the attribute gets copied over, modified, and saved directly into the entry.
There is a config option available to ignore MODIFY requests which do not result in a different entry. (Useful with Debconf which treats all invoked questions as modified, and submits them back to the directory as modifications, even if their value did not change during the process). See 'modifysmarts'
When a search is made, it is possible to regex match a combination of the input params (base, scope, deref, size, time, filter, attrOnly), and if all of them match a specification, then the params can be arbitrarily rewritten, i.e. you could replace search base with another base if certain filter is used. See 'searchsubst'
Search rewriting is completely free-form, and it is possible to rewrite
searches to a completely different suffix, as long as both are located
in the same base directory.
This is a legitimate feature of the rewrite model and is officially used to rewrite incoming DHCP search queries to appropriate places with appropriate parameters.
When a specific search base is requested, and it does not exist in the searched location, it is possible to fallback to a chain of default entries. See 'searchfallback'
When an entry is found (directly or through fallbacks), it can be appended with attributes from other entries. The default entries to look up can be specified with the entry's seeAlso attribute, or fixed in a config file. See 'entryappend'
When an entry is read, it is possible to cause attribute value expansion into values of other entries' attributes. See 'exp'
When an entry is read, it is possible to cause attribute value expansion into values of a complete, valid sub-search. See 'find'
When an entry is read, it is possible to cause attribute value expansion into values of on-disk files, ALWAYS relative to server data dir. See 'file'
When an entry is read, it is possible to cause Perl evaluation of contents within attributes. ** WARNING ** DANGEROUS ** DISABLED BY DEFAULT **. See 'perl'
When search results are returned to the client, each entry will contain a DN of its actual location on the disk (if the entry is coming from a fallback, its DN may be something completely different from what was requested). That's more useful (and easier) than changing the DN to its would-be value. Experience also shows that the calling programs do not care about the DN and are not confused by this, so there are only beneficial aspects to it.
All of the above features are tunable from the config file; non-trivial options do not have a default and so they must be present in the config file to get configured in the first place, otherwise they effectively stay disabled.
Some config directives have short names, but you can call them using any longer form. For example, directive 'exp' can be, for clarity, written as 'expand' or 'expandVal'.
All directives from the config file are registered on the stack. The stack can be saved to file, loaded from file for processing, or reset. Saving and loading the stack allows for quicker and less error-prone approach in repetitive blocks, especially if they repeat for each suffix you configure. (Different suffixes can load each other's files as long as they're configured with the same base directory). 'Clean' can be called on purpose to remove all on-disk stack files.
OpenLDAP locking is such that only one Viper function may be executing at a time (even if you have multiple Perl suffixes configured, only one is executing at a time), so no custom locking is necessary. But we still do proper locking where necessary ourselves to keep things under control and cover scenarios with multiple LDAP servers in the same tree.
Viper backend is called from OpenLDAP's back_perl handler, which is severely limited when it comes to user and access control. If you need any bind control beyond bind DN and password, or any access control based on bind DN, chances are you'll have to extend back_perl.c.
Using scripts/viper.pl (or anything else that resembles the approach shown there) allows Viper to run standalone, directly under the Perl interpreter, yet producing the exact same results as if it was running under slapd. This feature is best suited for running Viper under perl profiler or debugger, but can come handy in other slick scenarios.
Some of the dynamically produced values might take a long time to compute. Viper supports a generic cache mechanism which allows caching of any overlay's result, with variable validity period, specified as time interval, number of accesses, or number of LDAP operations. Examples for time-based expiry: 5 (implies seconds), 5s, 10m, 2h, 2d, 1w. To specify expiry in number of accesses, use specifier 'u', such as 20u. Note that 20u implies the value will be rebuilt on 20th access, giving you cached value 19 times. To specify expiry in number of LDAP operations, use specifier 'o', such as 1o. Expiry setting of 1o is by far best overall setting that minimizes or eliminates risk of stale data getting served, while still showing massive performance improvement (multiple times). See 'overlayconfig'
Similar to the above, it's also possible to cache LDAP entry reads from disk. Cache specifications and explanation are the same as we just discussed, except that the performance gains are modest compared to the overlay cache (in the <= 20% range). See 'cacheread'