Re: [Qemu-devel] Online image backup

[Lorenzo Mancini <lmancini@develer.com>]

> Matteo Bertini wrote:
>> Hello everyone,
>> 
>> is there any support for an online image backup? Or some direction to
>> have it?
>> 
>> Imagine I have an emulated server. How can I make an online backup of
>> the server state?
>> 
>> If I'm right in the actual design this is very hard.
>> 
>> Or there are some image internals update rules that permit a consistent
>> copy possible?
> 
> It should not be difficult provided the VM is using the QEMU qcow2 image
> format (doing a backup can be seen as a kind of snapshot).

Fabrice, that's just part of the problem.  A full automated remote 
backup service for qemu VMs should work like this:

  1. send a "savevm state" command to qemu;
  2. wait for savevm completion;
  3. perform an *online* copy (cp, rsync, whatever) of the .qcow2 image, 
while it's still running in qemu;
  4. ...repeat for all running VMs.

Now suppose you lose the host machine and must revert to backup.  A 
recovery scenario would be:

  1. copy the .qcow2 image on the new host machine;
  2. start qemu with the "-loadvm state" commandline switch.


The weak spot of this configuration is the third step of backup: if you 
perform an online copy (and that's your only real choice, since you 
don't want to powerdown the VMs every night just to backup them), you 
are copying a file while it's always modifying, and chances are high 
that the copy will be corrupted (i.e. different from any state the 
original went through), since you can't perform the copy with a single read.

However, if the .qcow2 file format is designed in such a way that even 
while copying it online, the resulting copy will certainly be corrupted 
but at least the states inside of it will be loadvm-able, that saves the 
day.


If that's not the case, one could fallback to perform the online copy 
between a couple of "stop" and "cont" commands, so you're sure that the 
underlying .qcow2 image is frozen while you're reading it (...is it?), 
but again that's suboptimal since you force the VMs to be down for some 
time.


Thanks in advance!

-- 
Lorenzo Mancini