Re: [RFC PATCH 2/3] x86/cpufeatures: Combine word 11 and 12 into new scattered features word 11

[Sean Christopherson <sean.j.christopherson@intel.com>]

On Fri, Jun 14, 2019 at 04:21:39PM +0200, Borislav Petkov wrote:
> On Fri, Jun 14, 2019 at 07:14:24AM -0700, Sean Christopherson wrote:
> > This is wrong.  KVM isn't complaining about shuffling the order of feature
> > words, it's complaining that code is trying to do a reverse CPUID lookup
> > to a feature that isn't in the reverse_cpuid table.   Filtering out
> > checks dynamically is just hiding bugs.
> 
> No no, reverse_cpuid is hardcoding our feature leafs. This is wrong as
> we want to be able to change those. And reverse_cpuid[] should be able
> to handle that.
> 
> KVM is complaining because he removed one leaf. He adds it later in
> patch 3 as a Linux-defined leaf.

Yes, because removing that leaf breaks 'enum cpuid_leafs'.  Patch 3/3
"fixes" it by re-inserting a leaf, which causes 'enum cpuid_leafs' to
align with the CPU features.

For example, this assertion also fails:

diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index 5b0e9d869ce5..c273b99702d0 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -823,6 +823,7 @@ void get_cpu_cap(struct cpuinfo_x86 *c)
                c->x86_capability[CPUID_7_0_EBX] = ebx;
                c->x86_capability[CPUID_7_ECX] = ecx;
                c->x86_capability[CPUID_7_EDX] = edx;
+               BUILD_BUG_ON(CPUID_7_EDX != X86_FEATURE_ARCH_CAPABILITIES/32);
        }
 
        /* Extended state features: level 0x0000000d */

In function ‘x86_feature_cpuid’,
    inlined from ‘guest_cpuid_get_register’ at arch/x86/kvm/cpuid.h:71:25,
    inlined from ‘guest_cpuid_has’ at arch/x86/kvm/cpuid.h:100:6,
    inlined from ‘kvm_get_msr_common’ at arch/x86/kvm/x86.c:2824:8:
include/linux/compiler.h:345:38: error: call to ‘__compiletime_assert_62’ declared with attribute error: BUILD_BUG_ON failed: x86_leaf >= ARRAY_SIZE(reverse_cpuid)
  _compiletime_assert(condition, msg, __compiletime_assert_, __LINE__)


But this assertion passes because its word is 10, i.e. below the 11/12
words that are getting mucked with.

diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index 5b0e9d869ce5..aada9d2fa4df 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -830,6 +830,7 @@ void get_cpu_cap(struct cpuinfo_x86 *c)
                cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
 
                c->x86_capability[CPUID_D_1_EAX] = eax;
+               BUILD_BUG_ON(CPUID_D_1_EAX != X86_FEATURE_XSAVES/32);
        }
 
        /* AMD-defined flags: level 0x80000001 */


> All that doesn't matter for KVM - if KVM wants to do reverse lookup,
> then it should handle Linux-defined leafs just fine.

KVM can't handle Linux-defined leafs without extra tricks, which is why
I removed get_scattered_cpuid_leaf() or whatever it was called.