[PATCHES] Convert gencache to dbwrap to enable mutexes

[Christof Schmitt]

On Tue, Jun 24, 2014 at 12:43:54PM -0700, Christof Schmitt wrote:
> On Sat, Jun 21, 2014 at 10:20:44AM +0200, Stefan (metze) Metzmacher wrote:
> > Hi Christof,
> > 
> > >> While having a look at it I may found a problem with mutexes and ctdb.
> > >>
> > >>         /* only pass through specific flags */
> > >>         tdb_flags &= TDB_SEQNUM|TDB_VOLATILE;
> > >>
> > >> in db_open_ctdb() seems to completely useless (or wrong).
> > >> ctdbd_db_attach() is called before and gets the raw value of tdb_flags.
> > >> It has only has effect for
> > >>
> > >>         db_ctdb->wtdb = tdb_wrap_open(db_ctdb, db_path, hash_size,
> > >>                                       lpcfg_tdb_flags(lp_ctx, tdb_flags),
> > >>                                       O_RDWR, 0);
> > >>
> > >> So I think we can just remove tdb_flags &= TDB_SEQNUM|TDB_VOLATILE; above
> > >> and let lpcfg_tdb_flags() remove TDB_MUTEX_LOCKING when adding TDB_NOMMAP.
> > > 
> > > The two patches at the end of the changeset implement those changes.
> > 
> > Thanks! I'll try an autobuild on a machine with mutex support next week.
> 
> I just discovered another issue. gencache opens databases with CLEAR_IF_FIRST.
> That is not allowed with mutexes and it also triggers a warning in dbwrap when
> transactions are being used.
> 
> The easiest option would be using dbwrap only for gencache_notrans.tdb and
> removing CLEAR_IF_FIRST for that database. For gencache.tdb, i am not sure how
> to handle the case when tdb_check fails without CLEAR_IF_FIRST.

Sorry for the confusion. CLEAR_IF_FIRST has to be set for mutexes. The
main issue is that transactions are not allowed in dbwrap if
CLEAR_IF_FIRST is set. My idea here is to adapt dwrap to only warn if
the database has been opened through db_open which allows the clustered
usage. Databases opened directly with dbwrap_local_open should still
allow transactions and CLEAR_IF_FIRST.

gencache.tdb is not consistently opened with CLEAR_IF_FIRST, only if the
check fails. That can probably also cause issues when the db is
recreated with mutexes, but the next caller tries without
CLEAR_IF_FIRST.  To avoid this issue, I will change the patch to only
convert gencache_notrans to dbwrap.

Christof