Re: [Qemu-block] ping Re: [PATCH v7 03/16] migration: split common postcopy out of ram postcopy

[Kevin Wolf]

Am 26.09.2017 um 12:21 hat Dr. David Alan Gilbert geschrieben:
> * Kevin Wolf (address@hidden) wrote:
> > Am 25.09.2017 um 17:27 hat Dr. David Alan Gilbert geschrieben:
> > > > > Whatever you think the preferred way to set up postcopy migration is: 
> > > > > If
> > > > > something worked before this patch and doesn't after it, that's a
> > > > > regression and breaks backwards compatibility.
> > > > > 
> > > > > If we were talking about a graceful failure, maybe we could discuss
> > > > > whether carefully and deliberately breaking compatibility could be
> > > > > justified in this specific case. But the breakage is neither mentioned
> > > > > in the commit message nor is it graceful, so I can only call it a bug.
> > > > > 
> > > > > Kevin
> > > > 
> > > > It's of course my fault, I don't mean "it's wrong test, so it's not my
> > > > problem") And I've already sent a patch.
> > > 
> > > Why does this fail so badly, asserts etc - I was hoping for something
> > > a bit more obvious from the migration code.
> > > 
> > > postcopy did originally work without the destination having the flag on
> > > but setting the flag on the destination was always good practice because
> > > it detected whether the host support was there early on.
> > 
> > So what does this mean for 2.11? Do you think it is acceptable breaking
> > cases where the flag isn't set on the destination?
> 
> I think so, because we've always recommended setting it on the
> destination for the early detection.

Okay, I'll include the test case patch in my pull request today then.

> > If so, just changing the test case is enough. But if not, I'd rather
> > keep the test case as it is and fix only the migration code.
> 
> I'd take the test case fix, but I also want to dig why it fails so
> badly; it would be nice just to have a clean failure telling you
> that postcopy was expected.

Yes, that would be nice.

Kevin