Re: [Qemu-devel] [PATCH v4] target-i386: Register QOM properties for feature flags

[Eduardo Habkost]

On Tue, Apr 14, 2015 at 01:49:19PM +0200, Igor Mammedov wrote:
> On Fri, 10 Apr 2015 17:02:29 -0300
> Eduardo Habkost <address@hidden> wrote:
> 
> > This uses the feature name arrays to register QOM properties for feature
> > flags. This simply adds properties that can be configured using -global,
> > but doesn't change x86_cpu_parse_featurestr() to use them yet.
> > 
> > Signed-off-by: Eduardo Habkost <address@hidden>
> > ---
> > Changes v1 -> v2:
> > * Use "cpuid-" prefix instead of "feat-"
> > * Register release function for property
> > * Convert '_' to '-' on feature name before registering property
> > * Add dev->realized check to property setter
> > 
> > Changes v2 -> v3:
> > * Register alias properties for feature name aliases
> > * Patch is based on x86 tree, at:
> >   https://github.com/ehabkost/qemu.git x86
> > 
> > Changes v3 -> v4:
> > * Remove "cpuid-" prefix altogether
> > ---
> >  target-i386/cpu.c | 115 
> > ++++++++++++++++++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 115 insertions(+)
> > 
> > diff --git a/target-i386/cpu.c b/target-i386/cpu.c
> > index e657f10..1d97656 100644
> > --- a/target-i386/cpu.c
> > +++ b/target-i386/cpu.c
> > @@ -2848,12 +2848,120 @@ out:
> >      }
> >  }
> >  
> > +typedef struct FeatureProperty {
> > +    FeatureWord word;
> > +    uint32_t mask;
> > +} FeatureProperty;
> > +
> > +static void x86_cpu_get_feature_prop(Object *obj,
> > +                                     struct Visitor *v,
> > +                                     void *opaque,
> > +                                     const char *name,
> > +                                     Error **errp)
> > +{
> > +    X86CPU *cpu = X86_CPU(obj);
> > +    CPUX86State *env = &cpu->env;
> > +    FeatureProperty *fp = opaque;
> > +    bool value = (env->features[fp->word] & fp->mask) == fp->mask;
> maybe s/env->features[fp->word]/*fp->feature_word_ptr/
> it would consolidate possible out of bound error place to 1
> instead of several as it's now.
> Also it makes it easier to read for me :).

That makes sense, and would make it easier to convert this code to
generic QOM code later.

> 
> > +    visit_type_bool(v, &value, name, errp);
> > +}
> > +
> > +static void x86_cpu_set_feature_prop(Object *obj,
> > +                                     struct Visitor *v,
> > +                                     void *opaque,
> > +                                     const char *name,
> > +                                     Error **errp)
> > +{
> > +    X86CPU *cpu = X86_CPU(obj);
> > +    DeviceState *dev = DEVICE(obj);
> > +    CPUX86State *env = &cpu->env;
> > +    FeatureProperty *fp = opaque;
> > +    bool value;
> > +
> > +    if (dev->realized) {
> > +        qdev_prop_set_after_realize(dev, name, errp);
> > +        return;
> > +    }
> > +
> > +    visit_type_bool(v, &value, name, errp);
> should check error here and return in case of it's set,
> + local_err + error_propagate()

Will do it. Thanks for catching it!

> 
> > +    if (value) {
> > +        env->features[fp->word] |= fp->mask;
> > +    } else {
> > +        env->features[fp->word] &= ~fp->mask;
> > +    }
> > +}
> > +
> > +static void x86_cpu_release_feature_prop(Object *obj, const char *name,
> > +                                         void *opaque)
> > +{
> > +    FeatureProperty *prop = opaque;
> > +    g_free(prop);
> > +}
> > +
> > +/* Register a boolean feature-bits property.
> > + * If mask has multiple bits, all must be set for the property to return 
> > true.
> > + * The same property name can be registered multiple times to make it 
> > affect
> > + * multiple bits in the same FeatureWord.
> > + */
> > +static void x86_cpu_register_feature_prop(X86CPU *cpu,
> > +                                          const char *prop_name,
> > +                                          FeatureWord w,
> > +                                          uint32_t mask)
> isn't used as mask by caller, s/mask/bit/ ???

There will be an use case for mask containing multiple bits, later.  My
plan is to remove the duplicate "kvmclock" alias from kvm_feature_name,
and call this manually:

x86_cpu_register_feature_prop(cpu, "kvmclock", FEAT_KVM,
                              (1 << KVM_FEATURE_CLOCKSOURCE) |
                              (1 << KVM_FEATURE_CLOCKSOURCE2));

I didn't do that yet because I need the existing
x86_cpu_parse_featurestr() code to keep working until it is converted to
use object_property_set().

> 
> > +{
> > +    FeatureProperty *fp;
> > +    ObjectProperty *op;
> white line  here?

Will do it.

> 
> > +    op = object_property_find(OBJECT(cpu), prop_name, NULL);
> > +    if (op) {
> > +        fp = op->opaque;
> > +        assert(fp->word == w);
> > +        fp->mask |= mask;

^^^ This is the block of code that will be removed once I add the manual
"kvmclock" registration call I mentioned above.

> > +    } else {
> > +        fp = g_new0(FeatureProperty, 1);
> > +        fp->word = w;
> > +        fp->mask = mask;
> > +        object_property_add(OBJECT(cpu), prop_name, "bool",
> > +                            x86_cpu_get_feature_prop,
> > +                            x86_cpu_set_feature_prop,
> > +                            x86_cpu_release_feature_prop, fp, 
> > &error_abort);
> > +    }
> > +}
> > +
> > +static void x86_cpu_register_feature_bit_props(X86CPU *cpu,
> this adds 1 property and possibly aliases, _props() is confusing here.

Alias properties are still properties like any other, aren't they? The
function is still responsible for registering multiple properties. Is
the "_props()" suffix really that confusing?


> I'd rename it to x86_cpu_add_feature_bit_prop() and inline
> above x86_cpu_register_feature_prop() since it's not going to be reused

I prefer to keep the single-property function separated, as it may
become a generic bitmap property registration function inside generic
QOM code later. With your feature_word_ptr suggestion, it would be even
more generic and non-x86-specific.

(To be honest, I would prefer to keep the single-property function
registration code clearly separated even if it was never going to be
reused anywhere. 20-line functions are already too long for my taste.)


> 
> > +                                               FeatureWord w,
> > +                                               int bit)
> > +{
> > +    Object *obj = OBJECT(cpu);
> > +    int i;
> > +    char **names;
> > +    FeatureWordInfo *fi = &feature_word_info[w];
> > +
> > +    if (!fi->feat_names) {
> > +        return;
> > +    }
> > +    if (!fi->feat_names[bit]) {
> > +        return;
> > +    }
> > +
> > +    names = g_strsplit(fi->feat_names[bit], "|", 0);
> > +
> > +    feat2prop(names[0]);
> > +    x86_cpu_register_feature_prop(cpu, names[0], w, (1UL << bit));
> > +
> > +    for (i = 1; names[i]; i++) {
> > +        feat2prop(names[i]);
> > +        object_property_add_alias(obj, names[i], obj, names[0], 
> > &error_abort);
> > +    }
> > +
> > +    g_strfreev(names);
> > +}
> > +
> >  static void x86_cpu_initfn(Object *obj)
> >  {
> >      CPUState *cs = CPU(obj);
> >      X86CPU *cpu = X86_CPU(obj);
> >      X86CPUClass *xcc = X86_CPU_GET_CLASS(obj);
> >      CPUX86State *env = &cpu->env;
> > +    FeatureWord w;
> >      static int inited;
> >  
> >      cs->env_ptr = env;
> > @@ -2894,6 +3002,13 @@ static void x86_cpu_initfn(Object *obj)
> >      cpu->apic_id = -1;
> >  #endif
> >  
> > +    for (w = 0; w < FEATURE_WORDS; w++) {
> > +        int bit;
> white line  here?

Will change it. Thanks!

> 
> > +        for (bit = 0; bit < 32; bit++) {
> > +            x86_cpu_register_feature_bit_props(cpu, w, bit);
> > +        }
> > +    }
> > +
> >      x86_cpu_load_def(cpu, xcc->cpu_def, &error_abort);
> >  
> >      /* init various static tables used in TCG mode */
> 

-- 
Eduardo