On 07/09/2011 09:52 AM, Matthias Bolte wrote:
Anyway, you decided to add an snapshot_mode attribute to the disk
element and exposed the VMX values there. I'm not sure that this is a
good idea as scsi0:0.mode affects two aspects.
scsi0:0.mode can basically have three different modes
- persistent, the default, a disk with this mode will take part in
snapshots and changes to the disk's content persist domain power
cycles and snapshot restoring.
- independent-persistent, a disk with this mode will not take part in
snapshots, but changes to the disk's content persist domain power
cycles and snapshot restoring.
- independent-nonpersistent, a disk with this mode will be not take
part in snapshots and changes to the disk's content don't persist
domain power cycles and snapshot restoring. This is realized by
writing all changes into an additional .vmdk instead of the original
.vmdk. This additional .vmdk is automatically deleted on domain power
cycles and snapshot restoring.
There are two additional but outdated modes undoable and nonpersistent
that aren't supported anymore.
So the two aspects scsi0:0.mode affects is snapshot and the
persistence of changes. I think it makes more sense to use two
attributes for the disk element to expose this.
<disk ... snapshot='yes|no' persistent='yes|no'>
- snapshot=yes persistent=yes maps to scsi0:0.mode=persistent
- snapshot=yes persistent=no is unsupported for ESX
- snapshot=no persistent=yes maps to scsi0:0.mode=independent-persistent
- snapshot=no persistent=no maps to scsi0:0.mode=independent-nonpersistent
Hmm, this indeed seems like it might be reasonable to represent both
aspects in XML. See also
https://www.redhat.com/archives/libvir-list/2011-May/msg00315.html.
At stake is whether a disk has a snapshot taken by default, and whether
a disk is treated as temporary for the life of the domain (qemu has a
-snapshot command line option that treats all disks as temporary, but
with better per-volume snapshot abilities, libvirt could certainly offer
the same fine-tuning of per-disk as esx appears to offer).
So yes, I'll need to fold something like this into my v2 proposal for
snapshot handling.
The more important question is: Is this an general concept or is it
too ESX specific?
It's sounding generic enough that it will be worth getting it right in
the XML.
Eric is currently discussing/designing an extension to libvirt's
snapshot/checkpoint capabilities. At the moment libvirt supports
checkpointing a complete domain including RAM image and all storage
volumes, but it doesn't support snapshotting of single storage volumes
or a subset of the storage volumes of a domain.
https://www.redhat.com/archives/libvir-list/2011-June/msg00761.html
Eric suggest to extend the<domainsnapshot> and virStorageVol* APIs to
allow to include only a subset of the domain's storage volumes in a
checkpoint. This approach allows to specifiy for each checkpoint which
storage volumes to include. ESX allows something similar with the
independent modes, but you cannot define this on a per snapshot basis,
but have to decided this before. Eric's approach is more flexible but
doesn't work for ESX. I wonder if we could add the snapshot
attribute/subelement to the disk element. This allows to set the
independent mode for ESX and allows to define a preset for other
hypervisors like QEMU that will support Eric's more flexible approach.
So when you don't explicitly define which disk to include in a
checkpoint in the<domainsnapshot> XML then the snapshot setting from
the the<domain> XML apply. If there are no presets for snapshot it
defaults to yes. For QEMU you could override the snapshot setting from
the<domain> XML in the<domainsnapshot> XML, for ESX you either don't
specify this in the<domainsnapshot> XML or have to match the settings
from the<domain> XML due to the way ESX works.
Yes, having a per-disk default in the<domain> XML (applicable to both
qemu and ESX), as well as a per-disk override in the<domainsnapshot>
(here, qemu can take advantage, but ESX would have to fail if the
override is not identical to the domain defaults). does make sense.
The persistence setting might be more ESX specific, but I think
libvirt could realize this for QEMU too, when the domain is using
qcow2 images with a base image. In that case libvirt could clear the
qcow2 image when the domain is restarted to realize persistent=no. I
might be incorrect here as I'm not the QEMU expert here.
You're exactly right - qemu can implement per-disk persistent=no by
doing a qcow2 wrapper around just the disks that should be reverted when
the VM stops running. There might be some interactions with migration
to worry about, though.
On a second thought we might want to use negative word so we don't add
subelements for the defaults, for example
<disk type='file' device='disk'>
<source file='[datastore] test1/test1.vmdk'/>
<target dev='sda' bus='scsi'/>
<independent/>
<nonpersistent/>
<address type='drive' controller='0' bus='0' unit='0'/>
</disk>
Now we're doing a bit of bike-shedding - I think there's definitely
consensus that this has to be in the XML somewhere, but whether as an
attribute or as a sub-element still remains to be decided.
