Re: [Qemu-devel] [PATCH 3/4] ppc: add CPU access_type into the migration stream

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 2660 bytes --]

On Mon, Sep 11, 2017 at 05:52:10PM +0100, Mark Cave-Ayland wrote:
> On 11/09/17 11:57, David Gibson wrote:
> 
> > On Sun, Sep 10, 2017 at 03:37:34PM +0100, Mark Cave-Ayland wrote:
> >> This is referenced in cpu_ppc_handle_mmu_fault() and so should be included
> >> in the migration stream.
> > 
> > That is not, on its own, sufficient reason.
> > 
> >> Note: the vmstate_ppc version number has already been bumped by the previous
> >> patch in this series.
> >>
> >> Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
> > 
> > As with 2/4 it breaks backwards migration.
> > 
> > But more, I really disklike the idea of migrating this.  It's internal
> > state for one, and it's also essentially transitory state.  Can we
> > avoid putting it in the otherwise persistent structure at all? Can we
> > derive the state from elsewhere?  Can we prevent migration from
> > occurring in the small windows where this data is live?
> 
> >From what I can see references to access_type are scattered throughout
> mmu_helper.c although I'm not necessarily familiar enough with PPC to
> know whether this is something that can be derived elsewhere instead.
> And once again it was something that was removed by a90db15.

Right, but the migration code prior to a90db15 was a complete mess.
It definitely included a number of things it didn't need to and
shouldn't as well as being missing other things that were needed.
It's not a good model.  And although it might have more-or-less worked
for certain machines like the ones you're reviving here, it was never
properly tested

> When pausing a VM, does execution stop at the end of the current TB
> rather than immediately? If so, perhaps someone could confirm that
> guarantee is good enough for access_type?

I'm pretty sure it has to; we'd have to come up out of an individual
TB in order to get to the main loop where we check the "please pause"
flag.  I'm not sure if that helps us here though - I *think* access
type is about carrying information from the point where we trigger an
exception to the point where we actually start processing the
exception.

This code is really ugly and I've never understood it well :(. It's
always seemed bogus to me that we have an essentially global variable
to carry information over that small gap, though.  Unfortunately it's
unlikely that I'd be able to dive into this and work out if it's
really needed any time soon.

-- 
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

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]