Re: bitmap migration bug with -drive while block mirror runs

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

01.10.2019 12:54, Kevin Wolf wrote:
> Am 01.10.2019 um 10:57 hat Vladimir Sementsov-Ogievskiy geschrieben:
>> 01.10.2019 3:09, John Snow wrote:
>>> Hi folks, I identified a problem with the migration code that Red Hat QE
>>> found and thought you'd like to see it:
>>>
>>> https://bugzilla.redhat.com/show_bug.cgi?id=1652424#c20
>>>
>>> Very, very briefly: drive-mirror inserts a filter node that changes what
>>> bdrv_get_device_or_node_name() returns, which causes a migration problem.
>>>
>>>
>>> Ignorant question #1: Can we multi-parent the filter node and
>>> source-node? It looks like at the moment both consider their only parent
>>> to be the block-job and don't have a link back to their parents otherwise.
>>>
>>>
>>> Otherwise: I have a lot of cloudy ideas on how to solve this, but
>>> ultimately what we want is to be able to find the "addressable" name for
>>> the node the bitmap is attached to, which would be the name of the first
>>> ancestor node that isn't a filter. (OR, the name of the block-backend
>>> above that node.)
>>
>>
>> Better would be to migrate by node-name only.. But am I right that
>> node-names are different on source and destination? Or this situation
>> changed?
> 
> Traditionally, I think migration assumes that frontends (guest devices)
> must match exactly, but backends may and usually will differ.
> 
> Of course, dirty bitmaps are a backend feature that isn't really related
> to guest devices, so this doesn't really work out any more in your case.
> BlockBackend names are unusable for this purpose (especially as we're
> moving towards anonymous BlockBackends everywhere), which I guess
> essentially means node-name is the only option left.
> 
> Is bitmap migration something that must be enabled explicitly or does
> it happen automatically? If it's explicit, then making an additional
> requirement (matching node-names) shouldn't be a problem.

The problem is that mirror filter is already in Rhel qemu and we have this bug,
but libvirt is not yet prepared to match node-name on migration, or am I wrong?

> 
>>> A simple way to do this might be a "child_unfiltered" BdrvChild role
>>> that simply bypasses the filter that was inserted and serves no real
>>> purpose other than to allow the child to have a parent link and find who
>>> it's """real""" parent is.
>>>
>>> Because of flushing, reopen, sync, drain &c &c &c I'm not sure how
>>> feasible this quick idea might be, though.
>>>
>>>
>>> - Corollary fix #1: call error_setg if the bitmap node name that's about
>>> to go over the wire is an autogenerated node: this is never correct!
>>>
>>> (Why not? because the target is incapable of matching the node-name
>>> because they are randomly generated AND you cannot specify node-names
>>> with # prefixes as they are especially reserved!
>>>
>>> (This raises a related problem: if you explicitly add bitmaps to nodes
>>> with autogenerated names, you will be unable to migrate them.))
>>>
>>
>> In other words, we need a well defined way to match nodes on source and destination,
>> keeping in mind filters, to migrate bitmaps correctly.
>>
>> Hm, did you thought about bitmaps in filters? It's not a problem to create bitmap in
>> mirror-top filter during mirror job:)
>>
>> Or what about bitmaps in Quorum children? Or what about bitmap in qcow2 file child bs?
>>
>> If node-names are different on source and destination, what is the same? Top blk name
>> and bdrv-children names (I recently saw Max's idea to check node "path" in iotest).
> 
> blk_name has to be assumed to be "". The BdrvChild path changes when
> filters are inserted (and inserting filters on the destination that
> aren't present on the source, or vice versa, sounds like something that
> should just work).

Skipping filters is not a problem until we don't want bitmaps in filters.

> 
> So both parts of this are not great ways for addressing nodes.

Not great, yes..

> 
>> So, actually node is migration-addressable, if path <blk-name>/root[/child-name] to the
>> defines this node directly (we must not have children with same name for some node in
>> the path).
>>
>> And I think it's a correct way to define node in migration stream - by path.
> 
> I'm afraid node-name is the only thing that could possibly work reliably
> for identifying nodes.
> 

Can we do it to fix bug in current Rhel's qemu? How much effort on libvirt part is needed?


-- 
Best regards,
Vladimir