Re: [PATCH 07/14] hw/i386/vmport: Add support for CMD_GETBIOSUUID

[Michael S. Tsirkin]

On Tue, Mar 10, 2020 at 04:24:45PM +0200, Liran Alon wrote:
> 
> On 10/03/2020 13:22, Michael S. Tsirkin wrote:
> > On Tue, Mar 10, 2020 at 01:13:21PM +0200, Liran Alon wrote:
> > > On 10/03/2020 11:34, Michael S. Tsirkin wrote:
> > > > On Tue, Mar 10, 2020 at 01:54:04AM +0200, Liran Alon wrote:
> > > > > This is VMware documented functionallity that some guests rely on.
> > > > > Returns the BIOS UUID of the current virtual machine.
> > > > > 
> > > > > Reviewed-by: Nikita Leshenko <address@hidden>
> > > > > Signed-off-by: Liran Alon <address@hidden>
> > > > So this at least seems guest-visible.
> > > > 
> > > > So I suspect you need to add properties to
> > > > disable this for old machine types, to avoid
> > > > breaking compatibility with live-migration.
> > > It is indeed guest visible.
> > > In theory, you are right that for every guest-visible change, we should 
> > > make
> > > sure to expose it to only new machine-types.
> > > 
> > > However, in this case, I feel it just unnecessary over-complicates the 
> > > code.
> > > I don't see how a guest which previously failed to use this command, will
> > > fail because after Live-Migration it could succeed.
> > The reverse can happen, start guest on a new qemu, command seems to
> > work, then we migrate and it fails.
> > 
> > And I guess this applies to the version right?
> > 
> > > If you insist, I will add such functionality. In that case, do you think a
> > > single flag will suffice for the addition of all new commands
> > > (i.e. "commands-version" that it's number specifies set of commands to
> > > expose), or you want to have a per-command flag?
> > > 
> > > -Liran
> > Can be a single flag but I'd just do it a boolean that enables a group
> > of commands. E.g. "commands-v2".
> > 
> Re-thinking about this...
> 
> QEMU VMPort interface was quite broken already (See first patch in series
> "hw/i386/vmport: Propagate IOPort read to vCPU EAX register").
> The introduction of that fix already changes the result of all existing
> commands from guest perspective which relied on return-value from
> vmport_ioport_read().
> E.g. CMD_GETVERSION and CMD_GETRAMSIZE.
> 
> In theory, we should have also made that bug-fix be tied to machine-type. To
> similarly avoid the issue of migrating a VM from a working VMPort command
> implementation to a non-working one.
> i.e. In case of migrating from new QEMU to old QEMU. Do we wish to create a
> property-flag for that fix as-well?

Yes, I meants that too. Just include everything in the same property.

> Or can we just drop all the machine-type
> flags alltogether (Including the suggested "commands-v2")
> and declare this the first actually working VMPort implementation?
> 
> -Liran

It's hard to be sure no one used this, and the overhead isn't big.  But
that would be a maintainer call. In any case you need to call this out
explicitly in the commit log, and I guess block migration for old
machine types.