Re: [Xen-devel] [PATCH v11 20/27] Support colo mode for qemu disk

[Wei Liu <wei.liu2@xxxxxxxxxx>]

On Mon, Mar 07, 2016 at 10:10:07AM +0800, Wen Congyang wrote:
> On 03/05/2016 04:30 AM, Konrad Rzeszutek Wilk wrote:
> > On Fri, Mar 04, 2016 at 05:52:09PM +0000, Ian Jackson wrote:
> >> Changlong Xie writes ("[PATCH v11 20/27] Support colo mode for qemu disk"):
> >>> +Enable COLO HA for disk. For better understanding block replication on
> >>> +QEMU, please refer to:
> >>> +http://wiki.qemu.org/Features/BlockReplication
> >>
> >> Sorry, I missed this link on my first pass.  I still think that at the
> >> very least this needs something more user-facing (ie, how should one
> >> set this up).
> >>
> >> But, I'm kind of worried that qemu is the wrong place to be doing
> >> this.
> >>
> >> How can this be made to work with PV guests ?
> > 
> > QEMU can also serve PV guests (qdisk).
> > 
> > I think your question is more of - what about making this work with
> > PV block backend?
> 
> I don't know how to work with PV block backend. It is one reason that
> why we only support pure HVM now.
> For PV block backend, there is also other problem. For exampe resuming
> it in the secondary side is very slow, because we need to disconnect and
> reconnect.
> 

Supporting PV guest is certainly going to be non-trivial. And I don't
think we would ever ask you to actually implement that.

The point is to have a story that when other people want to implement
COLO for PV-aware guests (PVHVM, PV and PVH), they are not crippled by
existing interfaces.

Currently the disk spec seems to be designed exclusively for QEMU. This
is not very desirable, but at least it wouldn't stop people from either
reusing them or inventing new parameters.

Furthermore, I think coming up with a story for PV-aware guests (PVHVM,
PV and PVH) is also non-trivial. For one the disk replication logic is
not implemented in PV block backend,  we're not sure how feasible to
replicate thing in QEMU into kernel, but we're quite sure it is not
going to be trivial technically and politically. The uncertainty is too
big to come up with a clear idea what it would look like.

Wei.

> Thanks
> Wen Congyang
> 
> >>
> >> What if an HVM guest has PV-on-HVM drivers ?  In this case there might
> >> be two relevant qemus, one for the qdisk Xen PV block backend, and one
> >> for the emulated IDE.
> > 
> > In both cases QEMU would use the same underlaying API to actually write/read
> > out the blocks. That API would then use NBD, etc to replicate writes.
> > 
> > Maybe a little ASCII art?
> > 
> >     qdisk  ide
> >       \    /
> >            \  /
> >            block API
> >             |
> >            QCOW2
> >             |
> >            NBD
> > 
> > Or such?
> > 
> >>
> >> I don't understand how discrepant writes are detected.  Surely they
> >> might occur and should trigger a resynch ?
> >>
> >> Ian.
> > 
> > 
> > .
> > 
> 
> 
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel