libvirt users - Jan 2014 - Looks like blockpull does not accept a subset of the entire chain of backing files

Hello

If I'm not terribly mistaken, looks like libvirt 1.2.1 does not provide
ability of merging only a subset of the entire chain of backing files.

So, if I have a chain like this:

                 root <- a <-b <- c <- d <- active

... and  I'd like to obtain a chain like this:

                  root <-  c <- d <- active

... looks like it's not supported, since I'm trying the command

/                   //virsh blockpull domain --path /path/to/c
--bandwidth 30 --base /path/to/root//          (1)
/
... but it complains because /--path /path/to/c/ is not recognized,
since only /--path /path/to/active/ is acceptable:

/                   //virsh blockpull domain --path /path/to/active
--bandwidth 30 --base /path/to/root//          (2)
/
So, command (1) fails but command (2) succeeds.

The point is: How could I obtain the results I'm trying to achieve via
command (1) ?

I'm new to libvirt, but the article below made me think that what I'm
trying to do would be possible:
     
http://kashyapc.fedorapeople.org/virt/lc-2012/snapshots-handout.html  (
see section on /blockpull/ )

Thoughts?


For your information, my environment is:

    Compiled against library: libvirt 1.2.1
    Using library: libvirt 1.2.1
    Using API: QEMU 1.2.1
    Running hypervisor: QEMU 1.7.0
    Running against daemon: 1.2.1


Thanks

-- 
Richard Gomes
http://rgomes.info
http://www.linkedin.com/in/rgomes
mobile: +44(77)9955-6813
inum <http://www.inum.net/>: +883(5100)0800-9804
sip:rgomes@ippi.fr

On 01/29/2014 07:51 AM, Richard Gomes wrote:
> Hello
> 
> If I'm not terribly mistaken, looks like libvirt 1.2.1 does not provide
> ability of merging only a subset of the entire chain of backing files.
Correct - qemu doesn't provide full functionality for merging in all
directions.  And even though upstream qemu has been adding more pieces
(for example, they recently added the ability to commit the active
image), libvirt still has to be taught to exercise that.
> 
> So, if I have a chain like this:
> 
>                  root <- a <-b <- c <- d <- active
> 
> ... and  I'd like to obtain a chain like this:
> 
>                   root <-  c <- d <- active
Are you trying to squash a and b into root (commit direction, in which
case, the existing block-commit command works just fine for this
scenario) or into c (pull direction, in which case, qemu doesn't support
it yet)?
> The point is: How could I obtain the results I'm trying to achieve via
> command (1) ?
In the current blockpull implementation, qemu refuses to pull into
anything but the active image (true even for your use of qemu 1.7).
I've been asking for the ability to pull into intermediate images, but
it isn't there yet.
> 
> I'm new to libvirt, but the article below made me think that what
I'm
> trying to do would be possible:
That said, it IS possible to fake the same effect, using a series of
block-rebase and raw qemu-img commands (not heavily tested, so you may
still have to tweak the actions you use).  Also, block mirroring doesn't
yet work with persistent disks (because qemu doesn't yet have a way to
resume an operation across restarts), so you temporarily have to make
your domain transient (virsh undefine $dom); you can redefine it at the
end of the sequence.

Start from root <- a <- b <- c <- d <- active

Create c1 with 'cp c c1' for this tree:
root <- a <- b \- c <- d <- active
                - c1

Merge a and b into c1 with 'qemu-img rebase -b base c1' for this tree:
root \- a <- b <- c <- d <- active
      ------------c1

Create d1 with 'cp d d1' for this tree:
root \- a <- b <- c \- d <- active
      ------------c1 - d1

Use 'qemu-img rebase -u -b c1 d1' for this tree:
root \- a <- b <- c <- d <- active
      ------------c1 <- d1

Create a blank destination active file wrapping d1, with 'qemu-img
create -f qcow2 -o backing_file=d1,backing_fmt=qcow2 active1' for this tree:
root \- a <- b <- c <- d <- active
      ------------c1 <- d1 <- active1

Set up a block mirroring, and pivot over to the new chain once it is
stable, with 'virsh blockcopy $dom active --dest active1 --shallow
--reuse-external --wait --pivot', for this tree:
root \------------c1 <- d1 <- active1
      - a <- b <- c <- d <- active

Clean up - a, b, c, d, and active are now no longer in use, and both
qemu and libvirt see the shorter chain (albeit with new file names),
with virtually no guest downtime.

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

Hello Eric,

Absolutely enlightening!

Thanks a lot :)

Richard Gomes
http://rgomes.info
http://www.linkedin.com/in/rgomes
mobile: +44(77)9955-6813
inum <http://www.inum.net/>: +883(5100)0800-9804
sip:rgomes@ippi.fr

On 29/01/14 16:46, Eric Blake wrote:
> On 01/29/2014 07:51 AM, Richard Gomes wrote:
>> Hello
>>
>> If I'm not terribly mistaken, looks like libvirt 1.2.1 does not
provide
>> ability of merging only a subset of the entire chain of backing files.
> Correct - qemu doesn't provide full functionality for merging in all
> directions.  And even though upstream qemu has been adding more pieces
> (for example, they recently added the ability to commit the active
> image), libvirt still has to be taught to exercise that.
>
>> So, if I have a chain like this:
>>
>>                  root <- a <-b <- c <- d <- active
>>
>> ... and  I'd like to obtain a chain like this:
>>
>>                   root <-  c <- d <- active
> Are you trying to squash a and b into root (commit direction, in which
> case, the existing block-commit command works just fine for this
> scenario) or into c (pull direction, in which case, qemu doesn't
support
> it yet)?
>
>> The point is: How could I obtain the results I'm trying to achieve
via
>> command (1) ?
> In the current blockpull implementation, qemu refuses to pull into
> anything but the active image (true even for your use of qemu 1.7).
> I've been asking for the ability to pull into intermediate images, but
> it isn't there yet.
>
>> I'm new to libvirt, but the article below made me think that what
I'm
>> trying to do would be possible:
> That said, it IS possible to fake the same effect, using a series of
> block-rebase and raw qemu-img commands (not heavily tested, so you may
> still have to tweak the actions you use).  Also, block mirroring
doesn't
> yet work with persistent disks (because qemu doesn't yet have a way to
> resume an operation across restarts), so you temporarily have to make
> your domain transient (virsh undefine $dom); you can redefine it at the
> end of the sequence.
>
> Start from root <- a <- b <- c <- d <- active
>
> Create c1 with 'cp c c1' for this tree:
> root <- a <- b \- c <- d <- active
>                 - c1
>
> Merge a and b into c1 with 'qemu-img rebase -b base c1' for this
tree:
> root \- a <- b <- c <- d <- active
>       ------------c1
>
> Create d1 with 'cp d d1' for this tree:
> root \- a <- b <- c \- d <- active
>       ------------c1 - d1
>
> Use 'qemu-img rebase -u -b c1 d1' for this tree:
> root \- a <- b <- c <- d <- active
>       ------------c1 <- d1
>
> Create a blank destination active file wrapping d1, with 'qemu-img
> create -f qcow2 -o backing_file=d1,backing_fmt=qcow2 active1' for this
tree:
> root \- a <- b <- c <- d <- active
>       ------------c1 <- d1 <- active1
>
> Set up a block mirroring, and pivot over to the new chain once it is
> stable, with 'virsh blockcopy $dom active --dest active1 --shallow
> --reuse-external --wait --pivot', for this tree:
> root \------------c1 <- d1 <- active1
>       - a <- b <- c <- d <- active
>
> Clean up - a, b, c, d, and active are now no longer in use, and both
> qemu and libvirt see the shorter chain (albeit with new file names),
> with virtually no guest downtime.
>

On 01/29/2014 08:21 PM, Richard Gomes wrote:
> Hello
> 
> If I'm not terribly mistaken, looks like libvirt 1.2.1 does not provide
> ability of merging only a subset of the entire chain of backing files.
> 
> So, if I have a chain like this:
> 
>                  root <- a <-b <- c <- d <- active
> 
> ... and  I'd like to obtain a chain like this:
> 
>                   root <-  c <- d <- active
> 
> ... looks like it's not supported, since I'm trying the command
> 
> /                   //virsh blockpull domain --path /path/to/c
> --bandwidth 30 --base /path/to/root//          (1)
> /
> ... but it complains because /--path /path/to/c/ is not recognized,
> since only /--path /path/to/active/ is acceptable:
> 
> /                   //virsh blockpull domain --path /path/to/active
> --bandwidth 30 --base /path/to/root//          (2)
> /
> So, command (1) fails but command (2) succeeds.
> 
> The point is: How could I obtain the results I'm trying to achieve via
> command (1) ?
> 
> I'm new to libvirt, but the article below made me think that what
I'm
> trying to do would be possible:
>      
> http://kashyapc.fedorapeople.org/virt/lc-2012/snapshots-handout.html  (
> see section on /blockpull/ )
I still have to make to time to submit a V2 of that that patch[1] with
reflecting latest upstream status.

Apart from what Eric said in painstakingly-detailed way, here's some old
notes from a previous discussion with Eric on mailing lists/IRC. Eric,
please point out if something is wrong below.


- "blockpull" can only pull into the *active* layer.

    - Support to pull into an intermediate/arbitrary layer is in
      progress.

    - NOTES:

       - Pulling into intermediate images requires opening the
         intermediate file read/write, as well as ensuring that any
         reads done in the active layer get correct data.

       - Case of 'reading data' - In the chain base <- snap1 <-
active
         if you pull base into snap1, reading active has to know whether
         the data comes from base or from snap1 . But if you pull into
         active, you already have the destination read/write, and once
         the data is pulled, you don't have to worry about keeping the
         chain consistent.

       - Case of 'writing data' -- If you pull into the active layer,
         then a write eliminates the need to pull that cluster from the
         backing file, thanks to copy-on-write semantics. But, if you
         pull into an intermediate layer, a write at the active layer
         doesn't affect the fac t that you still have to pull into the
         intermediate layer. Even though the active layer won't use that
         cluster from the intermediate layer.

- At the moment, virsh blockpull cannot be done 'offline' (although,
  'qemu-img' can do that.)


[1] http://www.redhat.com/archives/libvir-list/2013-January/msg01903.html


-- 
/kashyap

Hi Kashyap,

Thanks a lot for clarifying.

If there's an open issue for this improvement, please let me so that I
will subscribe to it.

Cheers :)

Richard Gomes
http://rgomes.info
http://www.linkedin.com/in/rgomes
mobile: +44(77)9955-6813
inum <http://www.inum.net/>: +883(5100)0800-9804
sip:rgomes@ippi.fr

On 31/01/14 18:31, Kashyap Chamarthy wrote:
> On 01/29/2014 08:21 PM, Richard Gomes wrote:
>> Hello
>>
>> If I'm not terribly mistaken, looks like libvirt 1.2.1 does not
provide
>> ability of merging only a subset of the entire chain of backing files.
>
>> So, if I have a chain like this:
>>
>>                  root <- a <-b <- c <- d <- active
>>
>> ... and  I'd like to obtain a chain like this:
>>
>>                   root <-  c <- d <- active
>>
>> ... looks like it's not supported, since I'm trying the command
>>
>> /                   //virsh blockpull domain --path /path/to/c
>> --bandwidth 30 --base /path/to/root//          (1)
>> /
>> ... but it complains because /--path /path/to/c/ is not recognized,
>> since only /--path /path/to/active/ is acceptable:
>>
>> /                   //virsh blockpull domain --path /path/to/active
>> --bandwidth 30 --base /path/to/root//          (2)
>> /
>> So, command (1) fails but command (2) succeeds.
>>
>> The point is: How could I obtain the results I'm trying to achieve
via
>> command (1) ?
>>
>> I'm new to libvirt, but the article below made me think that what
I'm
>> trying to do would be possible:
>>      
>> http://kashyapc.fedorapeople.org/virt/lc-2012/snapshots-handout.html  (
>> see section on /blockpull/ )
> I still have to make to time to submit a V2 of that that patch[1] with
> reflecting latest upstream status.
>
> Apart from what Eric said in painstakingly-detailed way, here's some
old
> notes from a previous discussion with Eric on mailing lists/IRC. Eric,
> please point out if something is wrong below.
>
>
> - "blockpull" can only pull into the *active* layer.
>
>     - Support to pull into an intermediate/arbitrary layer is in
>       progress.
>
>     - NOTES:
>
>        - Pulling into intermediate images requires opening the
>          intermediate file read/write, as well as ensuring that any
>          reads done in the active layer get correct data.
>
>        - Case of 'reading data' - In the chain base <- snap1
<- active
>          if you pull base into snap1, reading active has to know whether
>          the data comes from base or from snap1 . But if you pull into
>          active, you already have the destination read/write, and once
>          the data is pulled, you don't have to worry about keeping the
>          chain consistent.
>
>        - Case of 'writing data' -- If you pull into the active
layer,
>          then a write eliminates the need to pull that cluster from the
>          backing file, thanks to copy-on-write semantics. But, if you
>          pull into an intermediate layer, a write at the active layer
>          doesn't affect the fac t that you still have to pull into the
>          intermediate layer. Even though the active layer won't use
that
>          cluster from the intermediate layer.
>
> - At the moment, virsh blockpull cannot be done 'offline'
(although,
>   'qemu-img' can do that.)
>
>
> [1] http://www.redhat.com/archives/libvir-list/2013-January/msg01903.html
>
>