Re: [RFC PATCH 0/7] migration/multifd: Introduce storage slots

[Peter Xu <peterx@redhat.com>]

On Sun, Jun 23, 2024 at 10:25:05PM +0200, Maciej S. Szmigiero wrote:
> > I appreciated a lot you worked out VFIO on top of multifd, because IMHO
> > that's really the right direction.  However even with that, I don't think
> > the whole design is yet fully settled, not to mention the details. And that
> > implies it may miss 9.1.
> 
> I appreciate your work and review Peter - it would be nice if we could
> at least make some progress on this subject for 9.1.

Let's try and see, I didn't mean it won't hit 9.1 at all, it just feels
still challenging, but who knows!  I just don't want to give you that
feeling then right before softfreeze I start to say "it's not ready".

I think one reason it'll be more challenge is also since Fabiano will be
missing for weeks, and since you look like to agree on his RFC as a start,
it means it might be good idea we wait for his back and see how that goes
from there even if we can reach some consensus; it'll just be challenging
already.

I also left my (slightly lengthy) comment on the high level design of that
series here (I know you'll see that, but just to keep a record of this
discussion and link them together):

https://lore.kernel.org/r/ZniFH14DT6ycjbrL@x1n

Let's discuss there, let me know if something I just made it wrong, and
also if you're targeting landing part of the series you can try to
prioritize / provision what can already be landed and easier.

I actually have a wild idea where I don't know whether "switchover phase"
hooks are even needed.  I mean, currently save_live_iterate() allows
"returning 1 means all data ready".  Logically in switchover phase the
migration core could simply call save_live_iterate() until it returns 1.
Then switchover hooks do not need to exist at all.  I didn't check the
details, but if that works that'll simplify all register_savevm_live()
users, and also for VFIO it may mean iteration phase is more important too,
as when it's resolved it naturally works with switchover phase.  You can
ignore this idea because it can be too wild and definitely not helpful on
landing things fast, just in case it may bring some thoughts.

Thanks,

-- 
Peter Xu