Atomicity of on-the-fly updates of ACL entries (using ldapmodify)?

[Romain Komorn <rkomorn@dofc.org>]

How atomic are changes made to database entries with the following type of 
LDIFs:

dn: olcDatabase={1}bdb,cn=config
changetype: modify
add: olcAccess
olcAccess: {100}to <what> by <who> read
olcAccess: {65000}to <what> by <who> none

Does ldapmodify simply add/remove attributes on the fly, or does it get 
the whole entry, then update a new one in? This may become significant 
if there are a large number of ACLs. If the update is not atomic, is there 
a potential for problems if the list of ACLs becomes large (ie: 200 
<whats> with an average of 10 <whos> each) ?

Background: The aim here is to leverage nss_ldap and olcAccess entries to 
manage access to servers while maintaining control of which accounts 
appear on which servers via OpenLDAP, rather than filters on the client 
side. This is to prevent servers that have been misconfigured or 
(maliciously) modified from being able to gain read access to user 
entries other than the ones explicitly allowed by the directory. IMO, a 
client server being able to pick who should be allowed to log on defeats 
the purpose of centrally managed logins.

Romain Komorn