ldb speed

[tridge at samba.org]

Luke,

 > Renames may be less uncommon than you think. And it certainly is not
 > desirable to have to lock an entire subtree to do it.

In ldb+tdb write operations are serialised. At the start of a write
operation two locks are grabbed - a 'transaction' lock and a read lock
on the entire partition. During the commit phase of the write
operation the read lock is converted to a write lock, so reads can
only happen in parallel with writes before the commit phase starts.

Every read operation in ldb+tdb gets a read lock on the entire
partition. This gives us read parallelism, but not write parallelism.

I know this is a very different approach to most LDAP databases :)

Cheers, Tridge