Re: [Qemu-devel] [PATCH v7 42/42] Inhibit ballooning during postcopy

[Dr. David Alan Gilbert]

* Amit Shah (address@hidden) wrote:
> On (Tue) 16 Jun 2015 [11:26:55], Dr. David Alan Gilbert (git) wrote:
> > From: "Dr. David Alan Gilbert" <address@hidden>
> > 
> > Postcopy detects accesses to pages that haven't been transferred yet
> > using userfaultfd, and it causes exceptions on pages that are 'not
> > present'.
> > Ballooning also causes pages to be marked as 'not present' when the
> > guest inflates the balloon.
> > Potentially a balloon could be inflated to discard pages that are
> > currently inflight during postcopy and that may be arriving at about
> > the same time.
> > 
> > To avoid this confusion, disable ballooning during postcopy.
> > 
> > When disabled we drop balloon requests from the guest.  Since ballooning
> > is generally initiated by the host, the management system should avoid
> > initiating any balloon instructions to the guest during migration,
> > although it's not possible to know how long it would take a guest to
> > process a request made prior to the start of migration.
> > 
> > Queueing the requests until after migration would be nice, but is
> > non-trivial, since the set of inflate/deflate requests have to
> > be compared with the state of the page to know what the final
> > outcome is allowed to be.
> 
> I didn't track the previous discussion, but there were plans to have
> guest-initiated balloon requests for cases where the guest wants to
> co-operate with hosts and return any free mem available We don't
> currently have guests that do this, but we also don't want to have a
> dependency between the host and guest -- they should be independent.
> 
> This approach here seems the simplest possible, short of maintaining
> another bitmap for the duration of postcopy which indicates
> guest-freed memory pages which postcopy should not populate, after
> receiving them at the dest (this sounds better to me than queuing up
> guest requests).
> 
> The downside here is that the guest offered some memory back, and we
> don't use it.  The guest also doesn't use it -- so it's a double loss,
> of sorts.
> 
> Thoughts?  I don't have a problem with this current approach, but if
> we could get something better, that'll be good too.

It needs something like that bitmap, but it would take quite a bit
of care to manage the interaction between:
    a) The guest emitting balloon notifications
    b) Pages being received from the source
    c) Destination use of that page

  we also have to think what to do with a page that's been ballooned
after reception of the source page; the madvise(dontneed) that's used
normally would cause userfault to fire again, and we can't allow that.
(We could make it the same as receiving a zero page).   But then we would
also have to cope with  the source sending us a page after the destination
has ballooned it and make sure to discard that (I suspect there are further
ordering examples that have to also be considered).

Dave
> 
>               Amit
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK