KVM: cpuid: set CPUID(EAX=0xd,ECX=1).EBX correctly
authorPaolo Bonzini <pbonzini@redhat.com>
Wed, 3 Dec 2014 13:38:01 +0000 (14:38 +0100)
committerPaolo Bonzini <pbonzini@redhat.com>
Fri, 5 Dec 2014 12:57:17 +0000 (13:57 +0100)
This is the size of the XSAVES area.  This starts providing guest support
for XSAVES (with no support yet for supervisor states, i.e. XSS == 0
always in guests for now).

Wanpeng Li suggested testing XSAVEC as well as XSAVES, since in practice
no real processor exists that only has one of them, and there is no
other way for userspace programs to compute the area of the XSAVEC
save area.  CPUID(EAX=0xd,ECX=1).EBX provides an upper bound.

Suggested-by: Radim Krčmář <rkrcmar@redhat.com>
Reviewed-by: Radim Krčmář <rkrcmar@redhat.com>
Tested-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/cpuid.c

index e24df01ab11808a79011203048ca3f52ce631d49..2f7bc2de991505c17bf7c87605d43ec2cff44f96 100644 (file)
@@ -23,7 +23,7 @@
 #include "mmu.h"
 #include "trace.h"
 
-static u32 xstate_required_size(u64 xstate_bv)
+static u32 xstate_required_size(u64 xstate_bv, bool compacted)
 {
        int feature_bit = 0;
        u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
@@ -31,9 +31,10 @@ static u32 xstate_required_size(u64 xstate_bv)
        xstate_bv &= XSTATE_EXTEND_MASK;
        while (xstate_bv) {
                if (xstate_bv & 0x1) {
-                       u32 eax, ebx, ecx, edx;
+                       u32 eax, ebx, ecx, edx, offset;
                        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
-                       ret = max(ret, eax + ebx);
+                       offset = compacted ? ret : ebx;
+                       ret = max(ret, offset + eax);
                }
 
                xstate_bv >>= 1;
@@ -87,9 +88,13 @@ int kvm_update_cpuid(struct kvm_vcpu *vcpu)
                        (best->eax | ((u64)best->edx << 32)) &
                        kvm_supported_xcr0();
                vcpu->arch.guest_xstate_size = best->ebx =
-                       xstate_required_size(vcpu->arch.xcr0);
+                       xstate_required_size(vcpu->arch.xcr0, false);
        }
 
+       best = kvm_find_cpuid_entry(vcpu, 0xD, 1);
+       if (best && (best->eax & (F(XSAVES) | F(XSAVEC))))
+               best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
+
        /*
         * The existing code assumes virtual address is 48-bit in the canonical
         * address checks; exit if it is ever changed.
@@ -470,9 +475,14 @@ static inline int __do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function,
                                goto out;
 
                        do_cpuid_1_ent(&entry[i], function, idx);
-                       if (idx == 1)
+                       if (idx == 1) {
                                entry[i].eax &= kvm_supported_word10_x86_features;
-                       else if (entry[i].eax == 0 || !(supported & mask))
+                               entry[i].ebx = 0;
+                               if (entry[i].eax & (F(XSAVES)|F(XSAVEC)))
+                                       entry[i].ebx =
+                                               xstate_required_size(supported,
+                                                                    true);
+                       } else if (entry[i].eax == 0 || !(supported & mask))
                                continue;
                        entry[i].flags |=
                               KVM_CPUID_FLAG_SIGNIFCANT_INDEX;