Re: x-blockdev-reopen & block-dirty-bitmaps

[Kevin Wolf]

Am 14.02.2020 um 19:54 hat John Snow geschrieben:
> Hi, what work remains to make this a stable interface, is it known?
> 
> We're having a problem with bitmaps where in some cases libvirt wants to
> commit an image back down to a base image but -- for various reasons --
> the bitmap was left enabled in the backing image, so it would accrue new
> writes during the commit.
> 
> Normally, when creating a snapshot using blockdev-snapshot, the backing
> file becomes RO and all of the bitmaps become RO too.
> 
> The goal is to be able to disable (or enable) bitmaps from a backing
> file before (or atomically just before) a commit operation to allow
> libvirt greater control on snapshot commit.
> 
> Now, in my own testing, we can reopen a backing file just fine, delete
> or disable a bitmap and be done with it -- but the interface isn't
> stable, so libvirt will be reluctant to use such tricks.
> 
> Probably a loaded question, but:
> 
> - What's needed to make the interface stable?
> - Are there known problem points?
> - Any suggestions for workarounds in the meantime?

I think I've asked this before, but I don't remember the answer...

What would be the problem with letting the enable/disable command
temporarily reopen the backing file read-write, like the commit job [1]
does?

Kevin

[1] I mean, I know that this code is wrong strictly speaking because we
    really should be counting read-write users [2] rather than blindly
    making the node read-only at the end of the operation - but somehow
    it seems to work in practice for commit jobs.

[2] Counting really means just looking at parent BdrvChild links that
    have WRITE permissions. I guess doing it right would mean getting
    rid of BlockDriverState.read_only (which is redundant) and using
    only permissions.