Data format for a read-only LDAP data store.  LDAP defines access to
records, each of them having n attributes.  Mandatory attributes are
"dn" and "objectClass".

The string table stores all strings, zero-terminated.  Binary records
are stored as empty string (single zero byte) followed by a
(uint32_t length,uint8_t data[]) tuple.

An Index is an array of uint32_t, each an offset inside the file to the
corresponding string.

Each Record is an array of uint32_t, each an offset inside the file to
the corresponding string.  Entries are in pairs, where the first
uint32_t points to the attribute name, the second points to the
attribute value.  Each record starts with a pair <number-of-attributes,0>.
The number of attributes equals the number of 64-bit pairs (including
this length pair itself).  The second pair is
<value-of-dn,value-of-objectClass>, the following pairs are all
<name-of-attribute,value-of-attribute>.

The Record Index is a table of offsets to the corresponding record.

All integers are stored LITTLE ENDIAN.

  const uint32_t magic = 0xfefe1da9;  /* 1da9 == "LDAP" ;-) */
  uint32_t attribute_count, record_count, indices_offset, size_of_string_table;
  char string_table[size_of_string_table];
  uint32_t attribute_names[attribute_count];
  uint32_t attribute_flags[attribute_count];  /* 1: match case insensitively */
  uint32_t records[record_count][];  /* in the same order as the records
                                        are physically on disk */
/* indices_offset points here */
  uint32_t record_index[record_count];
  struct {
    uint32_t index_type;  /* 0 == sorted array of pointers, rest reserved */
    uint32_t next;        /* offset of next index */
    /* for index_type==0: */
    uint32_t indexed_attribute; /* offset of attribute name */
    uint32_t record_offsets[record_count];
  }

The indices are at the end to make it possible to add more indices.
The next pointer is there to make extensions possible.



How do we do ACLs?

The goal is to reduce the number of ACLs that need to be checked.
We have a 0 dword reserved in each record.  The obvious use would be to
store a pointer to a list of permissions in each record.  The question is: do
we store the list of ACLs that is valid if you authenticate as that dn, or do
we store the list of ACLs that needs to be checked if anyone accesses
that dn?  I think it's better to store the permissions if anyone logs in
as that dn; the general user only has very simple access rules, so that
would keep the ACLs for the common case down.  On the down side we need
to store the permissions for the anonymous bind somewhere, too.  It also
means we optimize away the openldap "group member" indirection.

The question is: how do we store the ACLs in the database?  I suggest a
model where we store the auth ACLs first, then the read ACLs, then the
write ACLs.  That way you can stop evaluating at the first write ACL
when you only want to read.  And normally the bulk of the ACLs are for
writing.

So, for each dn and access type we need to keep a list of
(dn-pattern,attribute[]) that this dn has access to.

  uint32_t auth_count;

The syntax of the list should be:
  uint32_t attributes[];  /* offsets of attribute names in stringtab,
			     terminated by 0.  Empty list means: all */

Typische ACL:

access to dn="ou=(Fraktion-[^,]+),ou=Fraktionen,o=bundestag,c=de" attr=userPassword
        by self write
        by anonymous auth
        by group="cn=Gruppe A,ou=Administration,o=bundestag,c=de" write
        by group="cn=$1,ou=Administration,o=bundestag,c=de" write
        by * none