libvirt users - May 2014 - Re: Live snapshots of a single block device

----- Original Message -----
> From: "Eric Blake" <eblake@redhat.com> To: "Andrew
Martin"
> <amartin@xes-inc.com>, libvirt-users@redhat.com Sent: Thursday, May
22, 2014
> 5:44:54 PM Subject: Re: [libvirt-users] Live snapshots of a single block
> device
>
> On 05/22/2014 04:31 PM, Andrew Martin wrote:
> > Hello,
> >
>
> [Can you convince your mailer to wrap long lines?]
Sorry, my client won't auto-wrap and I forgot to run my post through vim
before
sending. Will do now :)
>
> > I am working on a script to automatically create live snapshots of
running
> > VMs using qemu-kvm 1.4.0 and libvirt 1.0.2. If a VM has
>
> Any reason you aren't upgrading to newer versions?  Although these seem
> sufficient for what you are trying.
The VM hosts I'm using all run Ubuntu 12.04 and I haven't had time to
backport
and test newer versions yet. I've published the latest versions I built in
this
PPA: https://launchpad.net/~xespackages/+archive/virtualization

The main functionality I need is live snapshotting with snapshot-create-as and
blockpull, so I haven't had a good reason to invest time to
upgrade.
>
> The problem you are facing is that the snapshot code MUST have an action
for
> every disk; and if you don't specify the action directly in your
> <domainsnapshot> xml, then the action chosen is inherited from the
<domain>
> xml (if you used <disk snapshot='no'>) or ultimately from the
type of snapshot
> you are creating (here, since you didn't use <disk
snapshot='no'> in the
        > domain xml, and didn't specify vdb in the
<domainsnapshot>, libvirt
        > assumes from --disk-only that you want vdb to have an external
        > snapshot taken).
>
> > </domainsnapshot>
> >
> > What am I doing wrong - how can I tell snapshot-create-as to create an
> > external snapshot for a specific block device only (not all block
devices)?
>
> Use: --diskspec vda --diskspec vdb,snapshot=no
>
> which produces this subset of xml:
>
>   <disks> <disk name='vda'/> <disk
name='vdb' snapshot='no'/> </disks>
>
Ah, that makes sense, thanks so much for clearing this up!
> >
> > Also, while looking at the manpage, does the --live option do anything
> > different if used with the above command?
>
> --live only makes sense when mixed with memory snapshots (with --memspec);
but
> as you are doing a --disk-only snapshot, it doesn't help (I'm not
sure if it
> will error out as mutually exclusive or just be silently ignored,
> without reading the code).
I'm using this code in a script for creating live backups of VMs - would it
make
sense to include --memspec to capture the memory state so that an fsck on boot
isn't necessary? I remember you explained --quiese awhile back:
https://www.redhat.com/archives/libvirt-users/2013-February/msg00020.html

However I don't have qemu-guest-agent installed on the guests, so would
using
--memspec instead of --disk-only produce a snapshot that is consistent (since it
would "resume" with the same memory state) or would this not be ideal
for backups (e.g easy to restore/recover)? Can the memory image be written to
the same image file (qcow2) as the disk snapshot?

Thanks,

Andrew

On 05/23/2014 08:41 AM, Andrew Martin wrote:
>>
>> Any reason you aren't upgrading to newer versions?  Although these
seem
>> sufficient for what you are trying.
> The VM hosts I'm using all run Ubuntu 12.04 and I haven't had time
to backport
> and test newer versions yet. I've published the latest versions I built
in this
> PPA: https://launchpad.net/~xespackages/+archive/virtualization
One of the benefits of newer qemu is the ability to do point-in-time
snapshots without altering the backing chain in use by the active disk
(well, qemu 2.0 has some improvements, and unreleased qemu 2.1 has even
more in the pipeline; and it takes newer qemu to drive some of these
improvements).  Some of these new approaches may provide enough
additional efficiency over the older methods that they are worth
upgrading for - but if you can get things working to your needs with an
older version, great.
>> --live only makes sense when mixed with memory snapshots (with
--memspec); but
>> as you are doing a --disk-only snapshot, it doesn't help (I'm
not sure if it
>> will error out as mutually exclusive or just be silently ignored,
>> without reading the code).
> I'm using this code in a script for creating live backups of VMs -
would it make
> sense to include --memspec to capture the memory state so that an fsck on
boot
> isn't necessary? I remember you explained --quiese awhile back:
> https://www.redhat.com/archives/libvirt-users/2013-February/msg00020.html
--memspec and --quiesce are orthogonal, you only need one or the other
(in fact, they do NOT work together in the current implementation).
--memspec says to capture the domain memory state (migrate to file;
using --live additionally says to minimize the time the guest is paused
but the file may be larger), and when the guest pauses at the end of
that migration, then snapshot the disks.  By resuming the guest from
that migration stream and disk state, any pending I/O operations in
flight in the guest OS will still be ready to go when the guest starts
back up, no fsck needed.  --quiesce says to tell the guest to flush all
I/O, then captures disk state, all without stopping the guest.  But
--quiesce requires the guest to be running to work, while --memspec
forceably stops the guest (the migration algorithm must pause, even if
--live minimizes the pause to a fraction of a second).
> 
> However I don't have qemu-guest-agent installed on the guests, so would
using
> --memspec instead of --disk-only produce a snapshot that is consistent
(since it
> would "resume" with the same memory state) or would this not be
ideal
> for backups (e.g easy to restore/recover)? Can the memory image be written
to
> the same image file (qcow2) as the disk snapshot?
External checkpoints (those where --memspec was used) have more
information than disk-only snapshots (even if --quiesce was used to
minimze the need for an fsck on the disk contents).  Thus, --memspec
snapshots require no guest cooperation, at the expense of requiring more
disk space to hold the state.  On the other hand, if you are going to do
a --memspec snapshot, I _highly_ recommend that you snapshot ALL disks
at once, rather than trying to do just one disk at a time.  With
--disk-only (whether or not you use --quiesce), all you can recover is
the disk state by a fresh boot of the disk (--quiesce minimizes the
chance that the fresh boot needs an fsck).  But with --memspec, the
memory state is only consistent if it matches ALL disk state taken at
the same time as the memory state - if you omit a disk, then try to
revert to that memory state, your running guest will behave as if the
disk instantaneously corrupted data out of underneath the OS.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

> >> --live only makes sense when mixed with memory snapshots (with
--memspec);
> >> but
> >> as you are doing a --disk-only snapshot, it doesn't help
(I'm not sure if
> >> it
> >> will error out as mutually exclusive or just be silently ignored,
> >> without reading the code).
> > I'm using this code in a script for creating live backups of VMs -
would it
> > make
> > sense to include --memspec to capture the memory state so that an fsck
on
> > boot
> > isn't necessary? I remember you explained --quiese awhile back:
> >
https://www.redhat.com/archives/libvirt-users/2013-February/msg00020.html
> 
> --memspec and --quiesce are orthogonal, you only need one or the other
> (in fact, they do NOT work together in the current implementation).
> --memspec says to capture the domain memory state (migrate to file;
> using --live additionally says to minimize the time the guest is paused
> but the file may be larger), and when the guest pauses at the end of
> that migration, then snapshot the disks.  By resuming the guest from
> that migration stream and disk state, any pending I/O operations in
> flight in the guest OS will still be ready to go when the guest starts
> back up, no fsck needed.  --quiesce says to tell the guest to flush all
> I/O, then captures disk state, all without stopping the guest.  But
> --quiesce requires the guest to be running to work, while --memspec
> forceably stops the guest (the migration algorithm must pause, even if
> --live minimizes the pause to a fraction of a second).
Would calling snapshot-create-as with --disk-only and --quiese still succeed
on a VM that is stopped (since the disk would already be consistent, it would
just call "qemu-img create")?
> 
> > 
> > However I don't have qemu-guest-agent installed on the guests, so
would
> > using
> > --memspec instead of --disk-only produce a snapshot that is consistent
> > (since it
> > would "resume" with the same memory state) or would this not
be ideal
> > for backups (e.g easy to restore/recover)? Can the memory image be
written
> > to
> > the same image file (qcow2) as the disk snapshot?
> 
> External checkpoints (those where --memspec was used) have more
> information than disk-only snapshots (even if --quiesce was used to
> minimze the need for an fsck on the disk contents).  Thus, --memspec
> snapshots require no guest cooperation, at the expense of requiring more
> disk space to hold the state.  On the other hand, if you are going to do
> a --memspec snapshot, I _highly_ recommend that you snapshot ALL disks
> at once, rather than trying to do just one disk at a time.  With
> --disk-only (whether or not you use --quiesce), all you can recover is
> the disk state by a fresh boot of the disk (--quiesce minimizes the
> chance that the fresh boot needs an fsck).  But with --memspec, the
> memory state is only consistent if it matches ALL disk state taken at
> the same time as the memory state - if you omit a disk, then try to
> revert to that memory state, your running guest will behave as if the
> disk instantaneously corrupted data out of underneath the OS.
It sounds like --quiese would work better for my use case, since I'm only
concerned with keeping the disk state consistent and prefer to not stop
the guest (even for a fraction of a second). The qemu-kvm 1.4.0 package
I built includes a version of qemu-guest-agent 1.4.0, so I will start
testing after installing it in a guest and adding the XML to the config:
http://wiki.libvirt.org/page/Qemu_guest_agent

Thanks!

Andrew