On Tue, Mar 28, 2017 at 09:31:05AM -0300, Eduardo Habkost wrote:
> On Tue, Mar 28, 2017 at 01:26:04PM +0200, Alexander Graf wrote:
> > On 03/28/2017 02:41 AM, Eduardo Habkost wrote:
> > > On Tue, Mar 28, 2017 at 12:19:37AM +0200, Alexander Graf wrote:
> > > > KVM has a feature bitmap of CPUID bits that it knows works for guests.
> > > > QEMU removes bits that are not part of that bitmap automatically on VM
> > > > start.
> > > >
> > > > However, some times we just don't list features in that list because
> > > > they don't make sense for normal scenarios, but may be useful in
> > > > specific,
> > > > targeted workloads.
> > > >
> > > > For that purpose, add a new =force option to all CPUID feature flags in
> > > > the CPU property. With that we can override the accel filtering and give
> > > > users full control over the CPUID feature bits exposed into guests.
> > > >
> > > > Signed-off-by: Alexander Graf <ag...@suse.de>
> > > > ---
> > > > target/i386/cpu.c | 30 ++++++++++++++++++++++++++----
> > > > target/i386/cpu.h | 3 +++
> > > > 2 files changed, 29 insertions(+), 4 deletions(-)
> > > >
> > > > diff --git a/target/i386/cpu.c b/target/i386/cpu.c
> > > > index 7aa7622..5a22f9a 100644
> > > > --- a/target/i386/cpu.c
> > > > +++ b/target/i386/cpu.c
> > > > @@ -2229,7 +2229,7 @@ void x86_cpu_list(FILE *f, fprintf_function
> > > > cpu_fprintf)
> > > > g_slist_foreach(list, x86_cpu_list_entry, &s);
> > > > g_slist_free(list);
> > > > - (*cpu_fprintf)(f, "\nRecognized CPUID flags:\n");
> > > > + (*cpu_fprintf)(f, "\nRecognized CPUID flags (=on|=off|=force):\n");
> > > > for (i = 0; i < ARRAY_SIZE(feature_word_info); i++) {
> > > > FeatureWordInfo *fw = &feature_word_info[i];
> > > > @@ -3460,6 +3460,7 @@ static int x86_cpu_filter_features(X86CPU *cpu)
> > > > x86_cpu_get_supported_feature_word(w, false);
> > > > uint32_t requested_features = env->features[w];
> > > > env->features[w] &= host_feat;
> > > > + env->features[w] |= cpu->forced_features[w];
> > > > cpu->filtered_features[w] = requested_features &
> > > > ~env->features[w];
> > > > if (cpu->filtered_features[w]) {
> > > > rv = 1;
> > > > @@ -3693,6 +3694,7 @@ static void x86_cpu_unrealizefn(DeviceState *dev,
> > > > Error **errp)
> > > > typedef struct BitProperty {
> > > > uint32_t *ptr;
> > > > + uint32_t *force_ptr;
> > > > uint32_t mask;
> > > > } BitProperty;
> > > Please take a look at:
> > > Subject: [PATCH for-2.9 v2 1/2] i386: Replace uint32_t* with
> > > FeatureWord on feature getter/setter
> > >
> > > I plan to include that series in 2.9, and it would make the
> > > force_ptr field unnecessary.
> > >
> > > > @@ -3701,7 +3703,15 @@ static void x86_cpu_get_bit_prop(Object *obj,
> > > > Visitor *v, const char *name,
> > > > {
> > > > BitProperty *fp = opaque;
> > > > bool value = (*fp->ptr & fp->mask) == fp->mask;
> > > > - visit_type_bool(v, name, &value, errp);
> > > > + bool forced = (*fp->force_ptr & fp->mask) == fp->mask;
> > > > + char str[] = "force";
> > > > + char *strval = str;
> > > > +
> > > > + if (!forced) {
> > > > + strcpy(str, value ? "on" : "off");
> > > > + }
> > > > +
> > > > + visit_type_str(v, name, &strval, errp);
> > > You can define an enum type in qapi-schema.json, and use
> > > visit_type_<YourEnumType>(). You can grep for
> > > visit_type_OnOffAuto to find examples.
> > >
> > > (But I suggest naming the enum something like
> > > "X86CPUFeatureSetting" instead of "OnOffForce", because we will
> > > probably add other enum values in the future).
> > >
> > > However: we need to find a way to do this and not break
> > > compatibility with "feat=yes|true|no|false", that's supported by
> > > StringInputVisitor (which is used by object_property_parse()).
> > > Maybe fallback to visit_type_bool() in case
> > > visit_type_<YourEnumType>() fails?
> >
> > Putting it into a special enum sounds much more fragile than the current
> > solution to me. We need to bool fallback either way, so I fail to see any
> > benefit from having the enum.
>
> I don't see why the enum would be more fragile. With the QAPI
> enum, we:
> * Have a meaningful value for the QOM property 'type' field,
> and have some hope to make type information for QOM properties
> really useful one day;
> * Have the possible values for the property well-documented in
> the QAPI schema;
> * Have the string<->enum translation code automatically generated
> for us;
> * Can easily add other values later (I have been planning to
> support "feat=host" so "-cpu host/max" aren't special cases in
> the code.
Well, we have one problem:
* Breaking compatibility with management code that expects the
feature getter to return boolean values, not enums/strings.
This affects users of query-cpu-model-expansion, in particular.
We could still make the setter accept an OnOffForce enum, but we
would have to keep returning a boolean from the property getter.
We could create a set of "force-FEAT=on|off" boolean properties,
and make the existing property setter accept "FEAT=force" as
input just to make the command-line easier to use. But I'm not
sure if making the list of X86CPU properties even larger is worth
it.
This means Alex' patch might be a reasonable solution, except for
the property getter that was changed to return a string. I will
make specific comments about the code as reply to the patch
itself.
--
Eduardo
