Re: [Qemu-devel] [PATCH v5 20/45] Modify savevm handlers for postcopy

[David Gibson]

On Wed, Mar 25, 2015 at 04:40:11PM +0000, Dr. David Alan Gilbert wrote:
> * Dr. David Alan Gilbert (address@hidden) wrote:
> > * David Gibson (address@hidden) wrote:
> > > On Tue, Mar 24, 2015 at 08:04:14PM +0000, Dr. David Alan Gilbert wrote:
> > > > * David Gibson (address@hidden) wrote:
> > > > > On Fri, Mar 20, 2015 at 12:37:59PM +0000, Dr. David Alan Gilbert 
> > > > > wrote:
> > > > > > * David Gibson (address@hidden) wrote:
> > > > > > > On Fri, Mar 13, 2015 at 10:19:54AM +0000, Dr. David Alan Gilbert 
> > > > > > > wrote:
> > > > > > > > * David Gibson (address@hidden) wrote:
> > > > > > > > > On Wed, Feb 25, 2015 at 04:51:43PM +0000, Dr. David Alan 
> > > > > > > > > Gilbert (git) wrote:
> > > > > > > > > > From: "Dr. David Alan Gilbert" <address@hidden>
> > > > > > > > > > 
> > > > > > > > > > Modify save_live_pending to return separate postcopiable and
> > > > > > > > > > non-postcopiable counts.
> > > > > > > > > > 
> > > > > > > > > > Add 'can_postcopy' to allow a device to state if it can 
> > > > > > > > > > postcopy
> > > > > > > > > 
> > > > > > > > > What's the purpose of the can_postcopy callback?  There are 
> > > > > > > > > no callers
> > > > > > > > > in this patch - is it still necessary with the change to
> > > > > > > > > save_live_pending?
> > > > > > > > 
> > > > > > > > The patch 'qemu_savevm_state_complete: Postcopy changes' uses
> > > > > > > > it in qemu_savevm_state_postcopy_complete and 
> > > > > > > > qemu_savevm_state_complete
> > > > > > > > to decide which devices must be completed at that point.
> > > > > > > 
> > > > > > > Couldn't they check for non-zero postcopiable state from
> > > > > > > save_live_pending instead?
> > > > > > 
> > > > > > That would be a bit weird.
> > > > > > 
> > > > > > At the moment for each device we call the:
> > > > > >        save_live_setup method (from qemu_savevm_state_begin)
> > > > > > 
> > > > > >    0...multiple times we call:
> > > > > >        save_live_pending
> > > > > >        save_live_iterate
> > > > > > 
> > > > > >    and then we always call
> > > > > >        save_live_complete
> > > > > > 
> > > > > > 
> > > > > > To my mind we have to call save_live_complete for any device
> > > > > > that we've called save_live_setup on (maybe it allocated something
> > > > > > in _setup that it clears up in _complete).
> > > > > > 
> > > > > > save_live_pending could perfectly well return 0 remaining at the 
> > > > > > end of
> > > > > > the migrate for our device, and thus if we used that then we 
> > > > > > wouldn't
> > > > > > call save_live_complete.
> > > > > 
> > > > > Um.. I don't follow.  I was suggesting that at the precopy->postcopy
> > > > > transition point you call save_live_complete for everything that
> > > > > reports 0 post-copiable state.
> > > > > 
> > > > > 
> > > > > Then again, a different approach would be to split the
> > > > > save_live_complete hook into (possibly NULL) "complete precopy" and
> > > > > "complete postcopy" hooks.  The core would ensure that every chunk of
> > > > > state has both completion hooks called (unless NULL).  That might also
> > > > > address my concerns about the no longer entirely accurate
> > > > > save_live_complete function name.
> > > > 
> > > > OK, that one I prefer.  Are you OK with:
> > > >     qemu_savevm_state_complete_precopy
> > > >        calls -> save_live_complete_precopy
> > > > 
> > > >     qemu_savevm_state_complete_postcopy
> > > >        calls -> save_live_complete_postcopy
> > > > 
> > > > ?
> > > 
> > > Sounds ok to me.  Fwiw, I was thinking that both the complete_precopy
> > > and complete_postcopy hooks should always be called.  For a
> > > non-postcopy migration, the postcopy hooks would just be called
> > > immediately after the precopy hooks.
> > 
> > OK, I've made the change as described in my last mail; but I haven't called
> > the complete_postcopy hook in the precopy case.  If it was as simple as 
> > making
> > all devices use one or the other then it would work, however there are
> > existing (precopy) assumptions about ordering of device state on the wire 
> > that
> > I want to be careful not to alter; for example RAM must come first is the 
> > one
> > I know.
> 
> Actually, I spoke too soon; testing this found a bad breakage.
> 
> the functions in savevm.c add the per-section headers, and then call the 
> _complete
> methods on the devices.  Those _complete methods can't elect to do nothing, 
> because
> a header has already been planted.

Hrm.. couldn't you move the test for presence of the hook earlier so
you don't sent the header if the hook is NULL?

> I've ended up with something between the two;  we still have a 
> complete_precopy and
> complete_postcopy method on the devices; if the complete_postcopy method 
> exists and
> we're in postcopy mode, the complete_precopy method isn't called at all.
> A device could decide to do something different in complete_postcopy from 
> complete_precopy
> but it must do something to complete the section.
> Effectively the presence of the complete_postcopy is now doing what
> can_postcopy() used to do.

Hmm.. but it means there's no per-device hook for the precopy to
postcopy transition point.  I'm not sure if that might matter.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

pgpZsaHHKDcxI.pgp
Description: PGP signature