Re: [Xen-devel] [PATCH v23 12/15] x86/VPMU: Handle PMU interrupts for PV(H) guests

[Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>]

On 06/05/2015 12:09 PM, Jan Beulich wrote:
> > @@ -201,27 +202,56 @@ static inline void context_load(struct vcpu *v)
> >       }
> >   }
> >   
> > -static void amd_vpmu_load(struct vcpu *v)
> > +static int amd_vpmu_load(struct vcpu *v, bool_t from_guest)
> >   {
> >       struct vpmu_struct *vpmu = vcpu_vpmu(v);
> > -    struct xen_pmu_amd_ctxt *ctxt = vpmu->context;
> > -    uint64_t *ctrl_regs = vpmu_reg_pointer(ctxt, ctrls);
> > +    struct xen_pmu_amd_ctxt *ctxt;
> > +    uint64_t *ctrl_regs;
> > +    unsigned int i;
> >   
> >       vpmu_reset(vpmu, VPMU_FROZEN);
> >   
> > -    if ( vpmu_is_set(vpmu, VPMU_CONTEXT_LOADED) )
> > +    if ( !from_guest && vpmu_is_set(vpmu, VPMU_CONTEXT_LOADED) )
> >       {
> > -        unsigned int i;
> > +        ctxt = vpmu->context;
> > +        ctrl_regs = vpmu_reg_pointer(ctxt, ctrls);
> >   
> >           for ( i = 0; i < num_counters; i++ )
> >               wrmsrl(ctrls[i], ctrl_regs[i]);
> >   
> > -        return;
> > +        return 0;
> > +    }
> > +
> > +    if ( from_guest )
> > +    {
> > +        unsigned int num_enabled = 0;
> > +
> > +        ASSERT(!is_hvm_vcpu(v));
> > +
> > +        ctxt = &vpmu->xenpmu_data->pmu.c.amd;
> > +        ctrl_regs = vpmu_reg_pointer(ctxt, ctrls);
> > +        for ( i = 0; i < num_counters; i++ )
> > +        {
> > +            if ( (ctrl_regs[i] & CTRL_RSVD_MASK) != ctrl_rsvd[i] )
> > +                return -EINVAL;
> > +
> > +            if ( is_pmu_enabled(ctrl_regs[i]) )
> > +                num_enabled++;
> > +        }
> > +
> > +        if ( num_enabled )
> > +            vpmu_set(vpmu, VPMU_RUNNING);
> > +        else
> > +            vpmu_reset(vpmu, VPMU_RUNNING);
> > +
> > +        memcpy(vpmu->context, &vpmu->xenpmu_data->pmu.c.amd, ctxt_sz);
> For one, why don't you use ctxt here? And then I think I made very
> clear that all shared data must be read exactly once during any
> hypervisor operation. Yet afaict you do two reads (one in the
> verification loop above, and another in the memcpy()).

Yes, and this is what is done on Intel side as well. I will need to copy 
first then and verifiy the copied context.
-boris

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel