Re: [Qemu-devel] KVM call agenda for June 28

[Kevin Wolf]

Am 05.07.2011 20:18, schrieb Marcelo Tosatti:
> On Tue, Jul 05, 2011 at 04:37:08PM +0100, Stefan Hajnoczi wrote:
>> On Tue, Jul 5, 2011 at 3:32 PM, Marcelo Tosatti <address@hidden> wrote:
>>> On Tue, Jul 05, 2011 at 04:39:06PM +0300, Dor Laor wrote:
>>>> On 07/05/2011 03:58 PM, Marcelo Tosatti wrote:
>>>>> On Tue, Jul 05, 2011 at 01:40:08PM +0100, Stefan Hajnoczi wrote:
>>>>>> On Tue, Jul 5, 2011 at 9:01 AM, Dor Laor<address@hidden>  wrote:
>>>>>>> I tried to re-arrange all of the requirements and use cases using this 
>>>>>>> wiki
>>>>>>> page: http://wiki.qemu.org/Features/LiveBlockMigration
>>>>>>>
>>>>>>> It would be the best to agree upon the most interesting use cases 
>>>>>>> (while we
>>>>>>> make sure we cover future ones) and agree to them.
>>>>>>> The next step is to set the interface for all the various verbs since 
>>>>>>> the
>>>>>>> implementation seems to be converging.
>>>>>>
>>>>>> Live block copy was supposed to support snapshot merge.  I think the
>>>>>> current favored approach is to make the source image a backing file to
>>>>>> the destination image and essentially do image streaming.
>>>>>>
>>>>>> Using this mechanism for snapshot merge is tricky.  The COW file
>>>>>> already uses the read-only snapshot base image.  So now we cannot
>>>>>> trivally copy the COW file contents back into the snapshot base image
>>>>>> using live block copy.
>>>>>
>>>>> It never did. Live copy creates a new image were both snapshot and
>>>>> "current" are copied to.
>>>>>
>>>>> This is similar with image streaming.
>>>>
>>>> Not sure I realize what's bad to do in-place merge:
>>>>
>>>> Let's suppose we have this COW chain:
>>>>
>>>>   base <-- s1 <-- s2
>>>>
>>>> Now a live snapshot is created over s2, s2 becomes RO and s3 is RW:
>>>>
>>>>   base <-- s1 <-- s2 <-- s3
>>>>
>>>> Now we've done with s2 (post backup) and like to merge s3 into s2.
>>>>
>>>> With your approach we use live copy of s3 into newSnap:
>>>>
>>>>   base <-- s1 <-- s2 <-- s3
>>>>   base <-- s1 <-- newSnap
>>>>
>>>> When it is over s2 and s3 can be erased.
>>>> The down side is the IOs for copying s2 data and the temporary
>>>> storage. I guess temp storage is cheap but excessive IO are
>>>> expensive.
>>>>
>>>> My approach was to collapse s3 into s2 and erase s3 eventually:
>>>>
>>>> before: base <-- s1 <-- s2 <-- s3
>>>> after:  base <-- s1 <-- s2
>>>>
>>>> If we use live block copy using mirror driver it should be safe as
>>>> long as we keep the ordering of new writes into s3 during the
>>>> execution.
>>>> Even a failure in the the middle won't cause harm since the
>>>> management will keep using s3 until it gets success event.
>>>
>>> Well, it is more complicated than simply streaming into a new
>>> image. I'm not entirely sure it is necessary. The common case is:
>>>
>>> base -> sn-1 -> sn-2 -> ... -> sn-n
>>>
>>> When n reaches a limit, you do:
>>>
>>> base -> merge-1
>>>
>>> You're potentially copying similar amount of data when merging back into
>>> a single image (and you can't easily merge multiple snapshots).
>>>
>>> If the amount of data thats not in 'base' is large, you create
>>> leave a new external file around:
>>>
>>> base -> merge-1 -> sn-1 -> sn-2 ... -> sn-n
>>> to
>>> base -> merge-1 -> merge-2
>>>
>>>>>
>>>>>> It seems like snapshot merge will require dedicated code that reads
>>>>>> the allocated clusters from the COW file and writes them back into the
>>>>>> base image.
>>>>>>
>>>>>> A very inefficient alternative would be to create a third image, the
>>>>>> "merge" image file, which has the COW file as its backing file:
>>>>>> snapshot (base) ->  cow ->  merge
>>>
>>> Remember there is a 'base' before snapshot, you don't copy the entire
>>> image.
>>
>> One use case I have in mind is the Live Backup approach that Jagane
>> has been developing.  Here the backup solution only creates a snapshot
>> for the period of time needed to read out the dirty blocks.  Then the
>> snapshot is deleted again and probably contains very little new data
>> relative to the base image.  The backup solution does this operation
>> every day.
>>
>> This is the pathalogical case for any approach that copies the entire
>> base into a new file.  We could have avoided a lot of I/O by doing an
>> in-place update.
>>
>> I want to make sure this works well.
> 
> This use case does not fit the streaming scheme that has come up. Its a
> completly different operation.
> 
> IMO it should be implemented separately.

I agree, this is a case for a live commit operation.

Kevin