[Qemu-devel] Snapshots not bound to an architecture?

[Qemu-devel] Snapshots not bound to an architecture?

[Blue Swirl]

Hi,

While testing the savevm/loadvm functions, I noticed that it's
possible to attempt to load a snapshot made for entirely different
architecture. There are a lot of warnings, of course.

Could we prevent this somehow? Or is there a use case for this, for
example loading a snapshot made on i386 to an x86_64 emulator?

Re: [Qemu-devel] Snapshots not bound to an architecture?

[Anthony Liguori]

Blue Swirl wrote:
> Hi,
>
> While testing the savevm/loadvm functions, I noticed that it's
> possible to attempt to load a snapshot made for entirely different
> architecture. There are a lot of warnings, of course.
>
> Could we prevent this somehow? Or is there a use case for this, for
> example loading a snapshot made on i386 to an x86_64 emulator?
>   

I think we should introduce a machine section for save/restore that 
included that information.  It should also be versioned in such a way 
that it could be incremented whenever a new piece of hardware is added 
to the default machine type.

Regards,

Anthony Liguori

Re: [Qemu-devel] Snapshots not bound to an architecture?

[Paul Brook]

On Tuesday 07 October 2008, Anthony Liguori wrote:
> Blue Swirl wrote:
> > Hi,
> >
> > While testing the savevm/loadvm functions, I noticed that it's
> > possible to attempt to load a snapshot made for entirely different
> > architecture. There are a lot of warnings, of course.
> >
> > Could we prevent this somehow? Or is there a use case for this, for
> > example loading a snapshot made on i386 to an x86_64 emulator?
>
> I think we should introduce a machine section for save/restore that
> included that information.  It should also be versioned in such a way
> that it could be incremented whenever a new piece of hardware is added
> to the default machine type.

Using a single version number to determine the "base" machine is IMHO a bad 
idea. The base peripherals should be identified (and mismatches detected) 
that same way as any other peripherals.

The simplest way to avoid loading the wrong type of machine is to give the 
cpus different names (e.g. cpu_i386/cpu_amd64) instead of just calling 
them "cpu".

Paul

Re: [Qemu-devel] Snapshots not bound to an architecture?

[Anthony Liguori]

Paul Brook wrote:
> On Tuesday 07 October 2008, Anthony Liguori wrote:
>   
>> Blue Swirl wrote:
>>     
>>> Hi,
>>>
>>> While testing the savevm/loadvm functions, I noticed that it's
>>> possible to attempt to load a snapshot made for entirely different
>>> architecture. There are a lot of warnings, of course.
>>>
>>> Could we prevent this somehow? Or is there a use case for this, for
>>> example loading a snapshot made on i386 to an x86_64 emulator?
>>>       
>> I think we should introduce a machine section for save/restore that
>> included that information.  It should also be versioned in such a way
>> that it could be incremented whenever a new piece of hardware is added
>> to the default machine type.
>>     
>
> Using a single version number to determine the "base" machine is IMHO a bad 
> idea.

I should have said, that I expected this to be an intermediate step.

>  The base peripherals should be identified (and mismatches detected) 
> that same way as any other peripherals.
>   

Yes.  This is particularly important in the context of hot plug because 
the bus/dev/fn of the plugged device may be different than if the device 
were just added directly.

> The simplest way to avoid loading the wrong type of machine is to give the 
> cpus different names (e.g. cpu_i386/cpu_amd64) instead of just calling 
> them "cpu".
>   

Yes, that works too, but it breaks backwards compatibility in a way 
that's hard to reconcile.  Bumping the version of "cpu" and including a 
machine type identifier would also work and would be easier to do in a 
backwards compatible way.

Regards,

Anthony Liguori

> Paul
>   

Re: [Qemu-devel] Snapshots not bound to an architecture?

[Blue Swirl]

On 10/7/08, Anthony Liguori <anthony@codemonkey.ws> wrote:
> Paul Brook wrote:
>
> > On Tuesday 07 October 2008, Anthony Liguori wrote:
> >
> >
> > > Blue Swirl wrote:
> > >
> > >
> > > > Hi,
> > > >
> > > > While testing the savevm/loadvm functions, I noticed that it's
> > > > possible to attempt to load a snapshot made for entirely different
> > > > architecture. There are a lot of warnings, of course.
> > > >
> > > > Could we prevent this somehow? Or is there a use case for this, for
> > > > example loading a snapshot made on i386 to an x86_64 emulator?
> > > >
> > > >
> > > I think we should introduce a machine section for save/restore that
> > > included that information.  It should also be versioned in such a way
> > > that it could be incremented whenever a new piece of hardware is added
> > > to the default machine type.
> > >
> > >
> >
> > Using a single version number to determine the "base" machine is IMHO a
> bad idea.
> >
>
>  I should have said, that I expected this to be an intermediate step.
>
>
> >  The base peripherals should be identified (and mismatches detected) that
> same way as any other peripherals.
> >
> >
>
>  Yes.  This is particularly important in the context of hot plug because the
> bus/dev/fn of the plugged device may be different than if the device were
> just added directly.
>
>
> > The simplest way to avoid loading the wrong type of machine is to give the
> cpus different names (e.g. cpu_i386/cpu_amd64) instead of just calling them
> "cpu".
> >
> >
>
>  Yes, that works too, but it breaks backwards compatibility in a way that's
> hard to reconcile.  Bumping the version of "cpu" and including a machine
> type identifier would also work and would be easier to do in a backwards
> compatible way.

There could be a compatibility check in the loadvm code just for "cpu"
and "cpu_common". The check would be removed when no target wants to
handle those versions anymore.