Bugfix for tdb transactions

[tridge at samba.org]

Hi Volker,

 > Ok, another argument against the change in semantics: If you
 > put this into ctdb, i.e. changing the locking rules, then we
 > will have to import the same changes into Samba 3.4,
 > otherwise you will get screwups between those two if you run
 > them together.

I guess what you mean is that the new code (with Rusty's patch) is
doing a commit, and the old code (with the bug) does a tdb open. In
that case the recovery code runs without a lock, so the bug can still
happen. Is that the scenario you mean?

To cover that case we'd need both changes. We'd need to get a write
lock in the recovery code (which has got to be the right thing to do),
and also move the unlock of the GLOBAL_LOCK till after the recovery
marker is removed (ie. the patch you sent).

Cheers, Tridge