11:59 a.m.
(CCing libvir-list and Jiri Denemark for libvirt-related discussion
about -cpu host/none, and live-migration safety expectations)
On Wed, Apr 01, 2015 at 06:31:23PM +0200, Michael Mueller wrote:
On Wed, 1 Apr 2015 10:01:13 -0300
Eduardo Habkost <ehabkost(a)redhat.com> wrote:
> On Tue, Mar 31, 2015 at 10:09:09PM +0200, Michael Mueller wrote:
> > On Tue, 31 Mar 2015 15:35:26 -0300
> > Eduardo Habkost <ehabkost(a)redhat.com> wrote:
> >
> > > On Mon, Mar 30, 2015 at 04:28:24PM +0200, Michael Mueller wrote:
> > > > This patch implements a new QMP request named
'query-cpu-model'.
> > > > It returns the cpu model of cpu 0 and its backing accelerator.
> > > >
> > > > request:
> > > > {"execute" : "query-cpu-model" }
> > > >
> > > > answer:
> > > > {"return" : {"name": "2827-ga2",
"accel": "kvm" }}
> > > >
> > > > Alias names are resolved to their respective machine type and GA
names
> > > > already during cpu instantiation. Thus, also a cpu model like
'host'
> > > > which is implemented as alias will return its normalized cpu model
name.
> > > >
> > > > Furthermore the patch implements the following function:
> > > >
> > > > - s390_cpu_models_used(), returns true if S390 cpu models are in use
> > > >
> > > > Signed-off-by: Michael Mueller <mimu(a)linux.vnet.ibm.com>
> > > > ---
> > > [...]
> > > > +static inline char *strdup_s390_cpu_name(S390CPUClass *cc)
> > > > +{
> > > > + return g_strdup_printf("%04x-ga%u", cc->proc.type,
cc->mach.ga);
> > > > +}
> > >
> > > How exactly is this information going to be used by clients? If getting
> > > the correct type and ga values is important for them, maybe you could
> > > add them as integer fields, instead of requiring clients to parse the
> > > CPU model name?
> >
> > The consumer don't need to parse the name, it is just important for them to
have
> > distinctive names that correlate with the names returned by
query-cpu-definitions.
> > Once the name of an active guest is known, e.g. ("2827-ga2",
"kvm") a potential
> > migration target can be verified, i.e. its query-cpu-definitions answer for
"kvm"
> > has to contain "2827-ga2" with the attribute runnable set to true.
With that mechanism
> > also the largest common denominator can be calculated. That model will be used
then.
>
> Understood. So the point is to really have a name that can be found at
> query-cpu-definitions. Makes sense.
>
> (BTW, if you reused strdup_s390_cpu_name() inside
> s390_cpu_compare_class_name() too, you would automatically ensure that
> query-cpus, query-cpu-definitions and s390_cpu_class_by_name() will
> always agree with each other).
I have to verify but it seems to make sense from reading... I will do that if it works.
:-)
>
> >
> > I also changed the above mentioned routine to map the cpu model none case:
> >
> > static inline char *strdup_s390_cpu_name(S390CPUClass *cc)
> > {
> > if (cpuid(cc->proc)) {
> > return g_strdup_printf("%04x-ga%u", cc->proc.type,
cc->mach.ga);
> > } else {
> > return g_strdup("none");
> > }
> > }
>
> What about:
>
> static const char *s390_cpu_name(S390CPUClass *cc)
> {
> return cc->model_name;
> }
>
> And then you can just set cc->model_name=_name inside S390_PROC_DEF (and
> set it to "none" inside s390_cpu_class_init()).
>
Wouldn't that store redundant information... but it would at least shift the work
into the
initialization phase and do the format just once per model.
To be honest, calculating the CPU model name on the fly with
strdup_s390_cpu_name() like you did above wouldn't be a problem to me.
I just wanted to avoid having the same CPU model name logic (and special
cases like "none") duplicated inside strdup_s390_cpu_name(),
S390_PROC_DEF, s390_cpu_class_by_name(), and maybe other places. Maybe
this duplication can be eliminated by simply reusing
strdup_s390_cpu_name() inside s390_cpu_compare_class_name().
> I wonder if this class->model_name conversion could be made generic
> inside the CPU class. We already have a CPU::class_by_name() method, so
> it makes sense to have the opposite function too.
>
> (But I wouldn't mind making this s390-specific first, and converted
> later to generic code if appropriate).
ok
>
> >
> > This implicitly will fail a comparison for cpu model ("none",
"kvm") as that will
> > never be part of the query-cpu-definitions answer.
>
> I am not sure I follow. If ("none", "kvm") is never in the list,
is
> "-cpu none -machine accel=kvm" always an invalid use case?
Not directly invalid as "-cpu none" will be the same as omitting the -cpu
option.
KVM will setup the vcpu properties withou any QEMU control to whatever the hosting
machine and the kvm kernel code offers. That will allow to run QEMU against a KVM
version that is not aware of the s390 cpu model ioctls.
It looks like we have conflicting expectations about
query-cpu-definitions, it seems: on the one hand, if "-cpu none" is
valid I believe it should appear on the query-cpu-definitions return
value; on the other hand, it is not (always?) migration-safe, so just
comparing the source query-cpus data with the target
query-cpu-definitions data wouldn't be enough to ensure live-migration
safety.
On x86, we have a similar problem with "-cpu host", that changes
depending on the host hardware and host kernel. We solve this problem by
making libvirt code aware of the set of valid CPU models, and libvirt
has special cases for "-cpu host".
If you don't want to encode that knowledge in libvirt or other
management software for s390, it looks like you need something like a
"stable-abi-safe" field on CpuDefinitionInfo?
>
> (I don't understand completely the meaning of "-cpu none" yet. How
does
> the CPU look like for the guest in this case? Is it possible to
> live-migrate when using -cpu none?)
And yes, that does not make sense in a migration context. The answer on query-cpu-model
(or query-cpus) will be ("none", "kvm") and that will never match a
runnable model.
The guest cpu will be derived from the hosting system and the kvm kernel as it is
currently
without the cpu model interface.
I hope I made it better to understand now...
Yes, thanks!
--
Eduardo