tdb: Remove locking from tdb_traverse_read()
This restores the original intent of tdb_traverse_read() in
7dd31288a701d772e45b1960ac4ce4cc1be782ed
This is needed to avoid a deadlock with tdb_lockall() and the
transaction start, as ldb_tdb should take the allrecord lock during a
search (which calls tdb_traverse), and can otherwise deadlock against
a transaction starting in another process
We add a test to show that a transaction can now start while a read
traverse is in progress
This allows more operations to happen in parallel. The blocking point
is moved to the prepare commit.
This in turn permits a roughly doubling of unindexed search
performance, because currently ldb_tdb omits to take the lock due to
an unrelated bug, but taking the allrecord lock triggers the
above-mentioned deadlock.
This behaviour was added in
251aaafe3a9213118ac3a92def9ab2104c40d12a for
Solaris 10 in 2005. But the run-fcntl-deadlock test works also on Solaris 10,
see https://lists.samba.org/archive/samba-technical/2017-April/119876.html.
Signed-off-by: Andrew Bartlett <abartlet@samba.org>
Reviewed-by: Stefan Metzmacher <metze@samba.org>