RFC: dbwrap_ctdb and empty vs deleted records

Ralph Böhme slow at samba.org
Mon Jul 25 09:05:03 UTC 2016


On Mon, Jul 25, 2016 at 10:57:38AM +0200, Michael Adam wrote:
> On 2016-07-23 at 13:47 +0200, Ralph Böhme wrote:
> > You need exactly the order of operations you described:
> > 
> > - lock byterange
> > - unlock -> this creates the tombstone record
> > - lock byterange again
> 
> Yeah, this is exactly what I meant with the above.
> (Long vac interval to make sure the tombstone rec
>  does not get cleaned in the meantime between
>  the operations.)

no need to worry about that here, as both ops are fired without delay
and default vac interval is 10 seconds iirc.

> > I've added a brl test and with that am able to reproduce the deadlock
> > when 925625b52886d40b50fc is reverted.
> 
> Oh, very cool! thanks for that.
> I think we could bring this into master right now as a
> first step, but one question: To my understanding, the
> patch does not explicitly but only implicitly test the
> ctdb/smbd interaction.

yes.

> It is written such that it should always pass.

yes.

> Hence we do not need to skip it if clustering is not active...

yes. There's no need to skip it, but I added it as another explicit
hint as to what this is all about as an explicit docoumentation that
this is not really testing brl per se, but something special.

> > My patch does not deadlock, so it seems it's ok. Updated patchset
> > that contains the test attached.
> 
> Will look again.

that would be great.

Thanks!
-Ralph



More information about the samba-technical mailing list