Coda File System

protection database - lower edge

From: Peter Braam <braam_at_cs.cmu.edu>
Date: Sat, 14 Feb 1998 11:36:08 -0500 (EST)
John, 

Here is the lower end proposal.  I expect a number of people on the list
to make suggestions but nothing too essential will probably change.

I don't have the names here of the people acutally doing this project --
so I can only notify them on Wednesday.

Peter

The Protection Database -- lower edge
=====================================

This project aims to replace the protection database with a dbm type
database, and a clean interface.

The records in the protection database are described in detail in
Satya's paper "Integrating Security in a large Distributed System". 
A records contains (see sect 6.3):
  - a string identifying a user or group name

  - an integer, if positive making the name a user, if negative a
  group. 

  - if the record is a group name, the owner of the group is
  determined by the string in the group preceeding the first colon.

   - the remaining components of the records are three lists, the
   third of which is  different for group entries and user entries.
        - list 1: elements are groups this entry directly belongs to
        - list 2: the CPS (see Satya, sect 4)
	- list 3: for users, the list of groups owned by the id
                  for groups, the list of direct members of this group
      
  - an access control list, containing two sublists:
  the positive and negative right list. Each entry in each list is a pair 
  ( id right)	    
	 id is a user or group id defined in the pdb
	 a right +1: permit/deny modification of this entry by "id"
	         -1: permit/deny reading of this entry by "id"
	 negative rights will override positive rights.
   The access rights are applicable for modification of the pdb
   database using the yps (Yale Protection Server). 
	 
Please note that Coda user id's and group ids cannot be recycled.  One
would have to modify acls all over the filesystem.  It is therefore
necessary to maintain the maximum user/group ids so far allocated and
store that in the the pdb -- perhaps in a special header record stored
at the beginning of the database.

Task 0: define an in core representation of the records more or less
as follows:

struct pdbrecord {
       char *name;
       int32 *id;
       struct list_head member_of;
       struct list_head cps;
       union {
          struct list_head owned_groups;
	  struct list_head members;
       } l3
       struct acl *pdbacl;
}

Use the list.h routines found in the Linux 2.1 kernel (header
/usr/include/linux/list.h and list.c). They are excellent macros.

Task 2: define a serialization of this structure to a disk record.
The records will contain a header like:

struct pdbdiskrecord { 
       int namelen;
       int member_of_length;
       int cps_length;
       int l3_length;
       int acl_lengt;
} 
followed by data. 

These records must in principle be indexed by "id" (in network order),
packed in network order (they are shared by servers of different
endians), and stored in GNU dbm (gdbm) format.  (Install the gdbm rpm
and read the info file and man page coming with it).  We will also
want to search for records by name (I suspect). 

Taks 3: 

Routines will need to be written to accomplish the following: 
 - convert a record to an in core representation
 - convert an in core representation to the a disk record
 - store, delete or read a disk record by id
 - precompute the CPS in the format used by the server
 - lock the database (using flock "pdb.lock") so that the fileserver can be
 prohibited from reading the pdb while records are being looked up. 
 - we need an index table to find records by name (use a second dbm
 file) 
 - dump the database in readable format. 

This constitues the real bottom end of the interface.  From here up we 
get:

Task 4: 

 - given a cps (pdb) do a check for lookup or modification in
   the pdb.  The fileserver can _always_ look things up -- we may need
   to think about this, since it should prove its authority. 
 - create or remove a new user or group
 - add a member to a group
 - remove a member from a group
 - the owner of a group can change the ACL on the group
     - remove/add an entry in the ACL. 
 - user records can probably only be modified by System:Administrators
 - list membership etc, see the project 5 description. 


Pointers to source:
 - the coda-src/pdbstuff directory defines conversions between text files and 
   pdb databases.  Probably not very useful - this is exactly what we
   want to get rid of.
 - the libal directory contains a lot of routines to look up records
 etc. It might be good to understand how the server uses an ACL and a
 CPS to compute authorization. 

There are obscure things here such as internal and external CPS's: I
will help with this and make sure the fileserver can use the records
you guys produce. 

Prefix routines which are exported to the fileserver or other packages 
with PDB_, static ones with pdb_.  Try not to expose structs, but use
procedures to return relevant fields.  Write a simple test program to
test _all_ your functions. 

Please keep in touch and let me know about intermediary targets you
are reaching or are finding difficult.

- Peter -
Received on 1998-02-15 11:35:57