Releases, locking and ldb

[Andrew Bartlett]

On Wed, 2017-06-28 at 00:00 +0200, Stefan Metzmacher wrote:
> Am 27.06.2017 um 20:43 schrieb Andrew Bartlett via samba-technical:
> > On Tue, 2017-06-27 at 22:06 +1200, Andrew Bartlett via samba-technical
> > wrote:
> > > On Tue, 2017-06-27 at 21:46 +1200, Andrew Bartlett via samba-technical
> > > wrote:
> > > > On Tue, 2017-06-27 at 08:24 +0200, Stefan Metzmacher via samba-
> > > > technical wrote:
> > > > > 
> > > > > The good thing is that due to LDB_MODULE_CHECK_VERSION()
> > > > > Samba and sssd are only runtime compatible with the version
> > > > > they're build against.
> > > > > 
> > > > > Would "ldb:tdb: Ensure we correctly decrement ltdb->read_lock_count"
> > > > > affect any old caller in any bad way?
> > > > 
> > > > I presume you mean some other hypothetical application beyond Samba and
> > > > SSSD?
> > > > 
> > > > The primary change is that during the ldb callbacks, the DB is locked,
> > > > rather than unlocked.  It is no longer possible to start a transaction
> > > > (write operation) during an ldb callback.  Instead, as good practice
> > > > would already dictate, you must start the transaction before the
> > > > search. 
> > > > 
> > > > > And what about "tdb: Remove locking from tdb_traverse_read()"?
> > > > 
> > > > I'm not sure about this one.  As we can't have one without the other I
> > > > don't know (and haven't studied) exactly what else it might impact. 
> > > > 
> > > > I don't think this without the ldb patches is actively harmful however.
> > > > 
> > > > > Whould something like this on the samba_dsdb module
> > > > > in 4.6 work around the problems (and keep it only as broken as it
> > > > > already is):
> > > > > - always call ldb_handle_use_global_event_context()
> > > > 
> > > > I don't think this case is worth worrying about.  It is rootdse
> > > > modified over LDAP in the single process model (only). 
> > > > 
> > > > > - overwrite read_lock/read_unlock() to be a no-op
> > > > 
> > > > No, this gets us to the point I was at during SambaXP, with random
> > > > LDB_ERR_BUSY / 'failed to upgrade hash locks', as we won't have a
> > > > global lock order.
> > > > 
> > > > I don't think it makes any difference to actually to have it pass the
> > > > read_lock()/read_unlock() down to sam.ldb.
> > > > 
> > > > I personally think we should just declare the incomparability, ideally
> > > > at build time, and move on.  (I tried to create an #ifdef hack to
> > > > detect this at build time, but wasn't able to build a reliable one
> > > > yet). 
> > > > 
> > > > I realise that everything that can happen will happen, but still: how
> > > > likely are they to be mixed in the real world?
> > > 
> > > I guess more likely than I care to admit.  Try the attached for a
> > > compile-time way to prevent this.
> > 
> > Can I please get some feedback on what I need to do to make progress on
> > this?
> 
> Can you prepare a branch on master with the required reverts?
> It should be possible to copy the lib/ldb lib/tdb directories
> to a 4.6 branch and let is pass autobuild.

Sure, I'll do that today.

> I think once I know what's exactly required to be reverted,
> I can try to think about a final solution.

Thanks,

Andrew Bartlett

-- 
Andrew Bartlett                       http://samba.org/~abartlet/
Authentication Developer, Samba Team  http://samba.org
Samba Developer, Catalyst IT          http://catalyst.net.nz/services/samba