Re: [Qemu-devel] [PATCH v12 2/3] quorum: implement bdrv_add_child() and bdrv_del_child()

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Max Reitz (mreitz@redhat.com) wrote:
> On 29.03.2016 18:03, Dr. David Alan Gilbert wrote:
> > * Max Reitz (mreitz@redhat.com) wrote:
> >> On 29.03.2016 17:54, Dr. David Alan Gilbert wrote:
> >>> * Max Reitz (mreitz@redhat.com) wrote:
> >>>> On 29.03.2016 17:50, Dr. David Alan Gilbert wrote:
> >>>>> * Eric Blake (eblake@redhat.com) wrote:
> >>>>>> On 03/29/2016 09:38 AM, Max Reitz wrote:
> >>>>>>> On 17.03.2016 10:56, Wen Congyang wrote:
> >>>>>>>> On 03/17/2016 05:48 PM, Dr. David Alan Gilbert wrote:
> >>>>>>>
> >>>>>>> [...]
> >>>>>>>
> >>>>>>>>> The children.0 notation is really confusing in the way that Berto
> >>>>>>>>> describes; I hit this a couple of months ago and it really doesn't
> >>>>>>>>> make sense.
> >>>>>>>>
> >>>>>>>> Do you mean: read from children.1 first, and then read from children.0 in
> >>>>>>>> fifo mode? Yes, the behavior is very strange.
> >>>>>>>
> >>>>>>> So is this intended or is it not? In
> >>>>>>> http://lists.nongnu.org/archive/html/qemu-block/2016-03/msg00526.html
> >>>>>>> you said that it is.
> >>>>>>>
> >>>>>>> I myself would indeed say it is very strange. If I were a user, I would
> >>>>>>> not expect this behavior. And as I developer, I think that how a BDS's
> >>>>>>> child is used by its parent should solely depend on its role (e.g.
> >>>>>>> whether it is "children.0" or "children.1").
> >>>>>>
> >>>>>> It sounds like the argument here, and in Max's thread on
> >>>>>> query-block-node-tree, is that we DO have cases where order matters, and
> >>>>>> so we need a way for the hot-add operation to explicitly specify where
> >>>>>> in the list a child is inserted (whether it is being inserted as the new
> >>>>>> primary image, or explicitly as the last resort, or somewhere in the
> >>>>>> middle).  An optional parameter, that defaults to appending, may be ok,
> >>>>>> but we definitely need to consider how the order of children is affected
> >>>>>> by hot-add.
> >>>>>
> >>>>> Certainly in the COLO case the two children are not identical; and IMHO we need
> >>>>> to get away from thinking about ordering and start thinking about functional
> >>>>> namingd - children.0/children.1 doesn't suggest the fact they behave
> >>>>> differently.
> >>>>
> >>>> To me it does. If quorum is operating in a mode call "FIFO" I would
> >>>> expect some order on the child nodes, and if the child nodes are
> >>>> actually numbered in an ascending order, that is an obvious order.
> >>>
> >>> I don't understand why it's called 'FIFO'.
> >>
> >> Because in that mode quorum successively reads from all of its children
> >> and returns the first successful result. So the First successful Input
> >> is the one that becomes quorum's Output (there isn't much of a
> >> successive output, so it doesn't make much sense to call that the First
> >> Output, though...).
> >>
> >> I didn't name it, though. *waves hands defensively* :-)
> > 
> > But that description doesn't make sense for what COLO uses it for.
> > 
> > They have, on the primary host:
> >    0) Local disk
> >    1) an NBD connection to the secondary
> > 
> > So in theory a read should always happen from (0) and writes should
> > go to both.
> 
> Well that's the way it works, isn't it?
> 
> I didn't mention what happens with writes, but those are indeed
> distributed to all of quorum's children. And as long as the local disk
> doesn't fail, data is always read from it alone.

I guess so, but it seems to be odd to name something after an ordering
when you never expect it to actually perform the read from anything other
than the first;  and certainly for fault tolerance stuff I think it's
important to define the failure modes.

Dave

> All you need to do is make sure that the local disk is the first node in
> whatever order FIFO is supposed to use.
> 
> Max
> 



--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK