[Qemu-devel] cpu_set vs device_add (was Re: [PATCH 03/10] target-i386: move hyperv_* static globals to CPUState)

[Eduardo Habkost]

On Tue, Feb 26, 2013 at 08:03:16PM +0100, Igor Mammedov wrote:
[...]
> > > > > > > > Won't -cpu/parse_featurestr() simply set global properties? In 
> > > > > > > > this
> > > > > > > > case, the common case would be to call "device_add XXX" with no
> > > > > > > > extra options at all, so there's no option to be excluded and no
> > > > > > > > special case to care about.
> > > > > > > That is if global properties will made to 1.5  which I highly 
> > > > > > > doubt
> > > > > > > taking in account how fast patches are reviewed and accepted.
> > > > > > > Otherwise release would be broken. 
> > > > > > 
> > > > > > IMO it _has_ to make 1.5 and is a requirement to make device_add 
> > > > > > usable
> > > > > > for CPU hotplug. Otherwise we would have to change the behavior of 
> > > > > > -cpu
> > > > > > + device_add in an incompatible way.
> > > > > if all -cpu features are converted to static properties, we do not 
> > > > > have to
> > > > > have global properties working. In absence of 'global properties', 
> > > > > user
> > > > > will have to use the same properties at device_add that was specified 
> > > > > at
> > > > > -cpu. And introduction of global properties won't regress it, it will
> > > > > just allow to use device_add without features specified on -cpu
> > > > 
> > > > Strictly, we do not have to, but changing -cpu to set global properties
> > > > only later would change the behavior of "-cpu XXX,foo=1,bar=2" followed
> > > > by "device_add XXX" in an incompatible way. So if our long-term plan is
> > > Could you explain how ^^^ it will be incompatible, pls?
> > 
> > Suppose that "foo" defaults to 0, and we run: "-cpu XXX,foo=1", followed
> > by "device_add XXX".
> > 
> > Without globals/defaults set by -cpu, the above will create a new CPU
> > with foo=0.
> > 
> > With globals/defaults set by -cpu, the above will create a new CPU with
> > foo=1.
> > 
> > If I recall correctly, we agreed that the latter is the behavior we
> > wanted (because it is simpler for users, matches the fact that "-cpu"
> > already affects multiple CPU devices [it already affects all the CPUs
> > created on startup], and is the most common use-case [creating CPUs that
> > look basically the same]).
> Yes, that is the goal. I wouldn't say incompatible if user will start
> QEMU with "-cpu XXX,foo=1" and then use "device_add XXX,foo=1". That's a
> strict minimum that would work for hot-plug. Plain "device_add XXX" is an
> invalid in this case since it won't produce the same CPU.
> So later on top of "-cpu XXX,foo=1" + "device_add XXX,foo=1" we add up
> conversion of -cpu to global properties it shouldn't break anything, only add
> new option to create the same CPU usin "device_add XXX", users will still be
> able to use "device_add XXX,foo=1" if desired.
> 
> I hope that -cpu => global properties will make it in 1.5, but it is not
> must have show-stopper for hot-plug if it misses it.
> 

OK, so without it we would only support a very strict subset of
device_add command (specifically: only if the device_add arguments match
exactly what was given to -cpu).

This would work if carefully and clearly documented, true. But if we are
going to have such a severe limitation, why not use "cpu_set" in 1.5
instead? cpu_set is not perfect, not the long-term solution we want, and
it is also very limited, but at least it is already supported by libvirt
and won't let users shoot their own foot.

-- 
Eduardo