Re: [PATCH v3 5/9] KVM: vmx/pmu: Add MSR_ARCH_LBR_DEPTH emulation for Arch LBR

[Xu, Like]

On 2021/3/5 0:12, Sean Christopherson wrote:
> On Thu, Mar 04, 2021, Xu, Like wrote:
> > Hi Sean,
> >
> > Thanks for your detailed review on the patch set.
> >
> > On 2021/3/4 0:58, Sean Christopherson wrote:
> > > On Wed, Mar 03, 2021, Like Xu wrote:
> > > > @@ -348,10 +352,26 @@ static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
> > > >    	return true;
> > > >    }
> > > > +/*
> > > > + * Check if the requested depth values is supported
> > > > + * based on the bits [0:7] of the guest cpuid.1c.eax.
> > > > + */
> > > > +static bool arch_lbr_depth_is_valid(struct kvm_vcpu *vcpu, u64 depth)
> > > > +{
> > > > +	struct kvm_cpuid_entry2 *best;
> > > > +
> > > > +	best = kvm_find_cpuid_entry(vcpu, 0x1c, 0);
> > > > +	if (best && depth && !(depth % 8))
> > > This is still wrong, it fails to weed out depth > 64.
> > How come ? The testcases depth = {65, 127, 128} get #GP as expected.
> @depth is a u64, throw in a number that is a multiple of 8 and >= 520, and the
> "(1ULL << (depth / 8 - 1))" will trigger undefined behavior due to shifting
> beyond the capacity of a ULL / u64.

Extra:

when we say "undefined behavior" if shifting beyond the capacity of a ULL,
do you mean that the actual behavior depends on the machine, architecture 
or compiler?

> Adding the "< 64" check would also allow dropping the " & 0xff" since the check
> would ensure the shift doesn't go beyond bit 7.  I'm not sure the cleverness is
> worth shaving a cycle, though.

Finally how about:

    if (best && depth && (depth < 65) && !(depth & 7))
        return best->eax & BIT_ULL(depth / 8 - 1);

    return false;

Do you see the room for optimization ？

> > > Not that this is a hot path, but it's probably worth double checking that the
> > > compiler generates simple code for "depth % 8", e.g. it can be "depth & 7)".
> > Emm, the "%" operation is quite normal over kernel code.
> So is "&" :-)  I was just pointing out that the compiler should optimize this,
> and it did.
>
> > if (best && depth && !(depth % 8))
> >     10659:       48 85 c0                test   rax,rax
> >     1065c:       74 c7                   je     10625 <intel_pmu_set_msr+0x65>
> >     1065e:       4d 85 e4                test   r12,r12
> >     10661:       74 c2                   je     10625 <intel_pmu_set_msr+0x65>
> >     10663:       41 f6 c4 07             test   r12b,0x7
> >     10667:       75 bc                   jne    10625 <intel_pmu_set_msr+0x65>
> >
> > It looks like the compiler does the right thing.
> > Do you see the room for optimization ？
> >
> > > > +		return (best->eax & 0xff) & (1ULL << (depth / 8 - 1));
> Actually, looking at this again, I would explicitly use BIT() instead of 1ULL
> (or BIT_ULL), since the shift must be 7 or less.
>
> > > > +
> > > > +	return false;
> > > > +}
> > > > +