Re: [Qemu-devel] [v3 PATCH] target-mips: implement UserLocal Register

[James Hogan]

On 29/05/14 16:38, Peter Maydell wrote:
> On 29 May 2014 15:29, James Hogan <address@hidden> wrote:
>> On 29/05/14 14:26, Peter Maydell wrote:
>>> On 29 May 2014 14:22, James Hogan <address@hidden> wrote:
>>>> loadvm/migration will write Config3, possibly with a different value if
>>>> the snapshot being loaded is an older one without userlocal support.
>>>
>>> Migration load always happens with a freshly reset system,
>>> so this isn't a problem.
>>
>> Why does that make a difference? The snapshot may have been saved on a
>> slightly different version without that bit set (if that isn't supported
>> then we may as well not bother trying to support loading different
>> snapshot versions).
> 
> A key requirement of migration or vm restore is that the
> machine configuration (including all command line parameters)
> must be exactly identical at both ends. (We don't attempt to
> catch this user error, which typically results in weird stuff
> happening or in a migration failure with an obscure message).
> So you'll never get a migration from a bit-set config to a
> bit-cleared config. Note that this is different from a cross
> version migration, which would be "from QEMU 2.0 with config-X
> to QEMU 2.1 with config-X".

Presumably we want to be able to add the bit to Config3 in existing
machines in target-mips/translate_init.c at some point though to take
advantage of the extra performance (e.g. in v2.1), in which case you
could get a migration between bit-clear and bit-set...

> 
>> Also loadvm can happen at any time, not just after
>> reset.
> 
> The "loadvm" monitor command is implemented as "first reset;
> then load state".

Okay cool. It's fine then as long as there are no old translations lying
around and the Config3 etc are migrated unchanged from the snapshot.

Thanks
James