Re: [PATCH 2/2] migration: dirty-bitmap: Allow control of bitmap persistence on destination

[Peter Krempa <pkrempa@redhat.com>]

On Wed, Feb 03, 2021 at 14:27:49 +0100, Peter Krempa wrote:
> On Wed, Feb 03, 2021 at 16:23:21 +0300, Vladimir Sementsov-Ogievskiy wrote:
> > 03.02.2021 16:00, Peter Krempa wrote:
> > > Bitmap's source persistence is transported over the migration stream and
> > > the destination mirrors it. In some cases the destination might want to
> > > persist bitmaps which are not persistent on the source (e.g. the result
> > > of merge of bitmaps from a number of layers on the source when migrating
> > > into a squashed image)
> > 
> > Why not make merge target on source be persistent itself? Then it will be persistent on migration destination.
> 
> Because they are temporary on the source. I don't want to make it
> persistent in case of a failure so that it doesn't get written to the
> disk e.g. in case of VM shutdown.

To be a bit more specific, I don't want the bitmaps to stay in the
image, that means that I'd have to also delete them on the source after
a successfull migration before qemu is terminated, which might not even
be possible since source deactivates storage after migration.

So making them persistent on source is impossible.

> 
> > 
> > > but currently it would need to create another set
> > > of persistent bitmaps and merge them.
> > > 
> > > This adds 'dest-persistent' optional property to
> > > 'BitmapMigrationBitmapAlias' which when present overrides the bitmap
> > > presence state from the source.
> > 
> > It's seems simpler to make a separate qmp command block-dirty-bitmap-make-persistent.. Didn't you consider this way?
> 
> I'm not sure how the internals work entirely. In my case it's way
> simpler to do this setup when generating the mapping which I need to do
> anyways rather than calling separate commands.

Similarly here, after a successful migration I'd have to go and make all
the bitmaps persistent, which is an extra step, and also a vector for
possible failures after migration which also doesn't seem appealing.