On 2/9/2021 1:24 AM, Sean Christopherson wrote:
On Mon, Feb 08, 2021, Dave Hansen wrote:
On 2/8/21 8:16 AM, Jing Liu wrote:
-#define XSTATE_COMPACTION_ENABLED (1ULL << 63)
-
static void fill_xsave(u8 *dest, struct kvm_vcpu *vcpu)
{
struct xregs_state *xsave = &vcpu->arch.guest_fpu->state.xsave;
@@ -4494,7 +4492,8 @@ static void load_xsave(struct kvm_vcpu *vcpu, u8 *src)
/* Set XSTATE_BV and possibly XCOMP_BV. */
xsave->header.xfeatures = xstate_bv;
if (boot_cpu_has(X86_FEATURE_XSAVES))
- xsave->header.xcomp_bv = host_xcr0 | XSTATE_COMPACTION_ENABLED;
+ xsave->header.xcomp_bv = XCOMP_BV_COMPACTED_FORMAT |
+ xfeatures_mask_all;
This is wrong, xfeatures_mask_all also tracks supervisor states.
When looking at SDM Vol2 XSAVES instruction Operation part, it says as
follows,
RFBM ← (XCR0 OR IA32_XSS) AND EDX:EAX;
COMPMASK ← RFBM OR 80000000_00000000H;
...
XCOMP_BV field in XSAVE header ← COMPMASK;
So it seems xcomp_bv also tracks supervisor states?
BRs,
Jing
Are 'host_xcr0' and 'xfeatures_mask_all' really interchangeable? If so,
shouldn't we just remove 'host_xcr0' everywhere?
I think so? But use xfeatures_mask_user().
In theory, host_xss can also be replaced with the _supervisor() and _dynamic()
variants. That code needs a good hard look at the _dynamic() features, which is
currently just architectural LBRs. E.g. I wouldn't be surprised if KVM currently
fails to save/restore arch LBRs due to the bit not being set in host_xss.
