Re: [PATCH v3 1/6] block: add bitmap-populate job

[Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>]

23.06.2020 00:44, Eric Blake wrote:
> On 6/19/20 11:16 PM, Vladimir Sementsov-Ogievskiy wrote:
>> 19.06.2020 22:56, Eric Blake wrote:
>>> From: John Snow <jsnow@redhat.com>
>>>
>>> This job copies the allocation map into a bitmap. It's a job because
>>> there's no guarantee that allocation interrogation will be quick (or
>>> won't hang), so it cannot be retrofitted into block-dirty-bitmap-merge.
>>>
>>> It was designed with different possible population patterns in mind,
>>> but only top layer allocation was implemented for now.
>>>
>>> Signed-off-by: John Snow <jsnow@redhat.com>
>>> Signed-off-by: Eric Blake <eblake@redhat.com>
>>> ---
> 
>>> +{ 'struct': 'BlockDirtyBitmapPopulate',
>>> +  'base': 'BlockDirtyBitmap',
>>> +  'data': { 'job-id': 'str',
>>> +            'pattern': 'BitmapPattern',
>>> +            '*on-error': 'BlockdevOnError',
>>> +            '*auto-finalize': 'bool',
>>> +            '*auto-dismiss': 'bool' } }
>>> +
>>
>> Peter said about a possibility of populating several target bitmaps simultaneously.
>>
>> What about such a generalized semantics:
>>
>> Merge all sources to each target
>>
>> @targets: list of bitmaps to be populated by the job
>> { 'struct': 'BlockDirtyBitmapPopulate',
>>    'data': { <common job fields>,
>>              'targets': ['BlockDirtyBitmap'],
>>              'sources': ['BitmapPopulateSource'] } }
> 
> We still need the 'pattern' argument (the idea being that if we have: Base <- Active, we want to be able to merge in the allocation map of Active into bitmaps stored in Base as part of a commit operation, whether that is active commit of a live guest or offline commit while the guest is offline).  Having an array for 'targets' to merge into is fine, but for 'sources', it's less a concern about selecting from multiple sources, and more a concern about selecting the allocation pattern to be merged in (libvirt wants to merge the same allocation pattern into each bitmap in Base).  Generalizing things to allow the merge of more than one source at once might not hurt, but I'm not sure we need it yet.
> 
> But there are other patterns that we may want to support: an all-ones pattern, or maybe a pattern that tracks known-zeros instead of allocation.
> 
>>
>>
>> @bitmap: specify dirty bitmap to be merged to target bitamp(s)
>> @node: specify a node name, which allocation-map is to be merged to target bitmap(s)
>> { 'alternate': 'BitmapPopulateSource',
>>    'data': { 'bitmap': 'BlockDirtyBitmap',
>>              'node': 'str' } }
> 
> This design is clever in that it lets us merge in both existing bitmaps and using a node-name for merging in an allocation map instead of a bitmap; but it limits us to only one pattern.

Ah, yes, we can't discriminate by type node-name from 'all-ones' or something like this.

>Better might be something where we supply a union (hmm, we've had proposals in the past for a default value to the discriminator to allow it to be optional, so I'll proceed as if we will finally implement that):
> 
> { 'enum': 'BitmapPattern', 'data': [ 'bitmap', 'allocation-top' ] }
> { 'union': 'BitmapPopulateSource',
>    'base': { '*pattern': 'BitmapPattern' },
>    'discriminator': { 'name': 'pattern', 'default': 'bitmap' },
>    'data': { 'bitmap': 'BitmapPopulateSource',
>              'allocation-top': { 'node': 'str' } } }

Yes, this is better, of course.

> 
> so that you can then do:
> 
> { "execute": "block-dirty-bitmap-populate",
>    "arguments": { "targets": [ { "node": "base", "name": "b1" },
>                                { "node": "base", "name": "b2" } ],
>          "sources": [ { "pattern": "allocation-top", "node": "top" } ]
>    } }
> 
> to merge in the allocation information of top into multiple bitmaps of base at once, or conversely, do:
> 
> { "execute": "block-dirty-bitmap-populate",
>    "arguments": { "targets": [ { "node": "base", "name": "b1" } ],
>          "sources": [ { "pattern": "bitmap",
>                         "node": "top", "name": "b1" } ]
>    } }
> { "execute": "block-dirty-bitmap-populate",
>    "arguments": { "targets": [ { "node": "base", "name": "b2" } ],
>          "sources": [ { "node": "top", "name": "b2" } ]
>    } }
> 
> and of course, wrap this in a "transaction" to ensure that it all succeeds or fails as a unit, rather than messing up one bitmap if another fails, while also allowing future extension for additional patterns.
> 
>>
>>
>> - so, we can merge several bitmaps together with several allocation maps into several target bitmaps.
>> (I remember, we also said about a possibility of starting several populating jobs, populating into
>>   same bitmap, I think it may be substituted by one job with several sources. Still, it's not hard to
>>   allow to use target bitmaps in a several jobs simultaneously and this is not about the QAPI interface)
>>
>> Will this simplify things in libvirt?
> 
> Peter, in your preliminary experiments with block-dirty-bitmap-populate, did you ever need to start more than one job to a single bitmap destination, or was it merely starting multiple jobs because you had multiple destinations but always just a single source?
> 
> 


-- 
Best regards,
Vladimir