KVM: TDX: restore host xsave state when exit from the guest TD
authorIsaku Yamahata <isaku.yamahata@intel.com>
Wed, 29 Jan 2025 09:58:56 +0000 (11:58 +0200)
committerPaolo Bonzini <pbonzini@redhat.com>
Fri, 14 Mar 2025 18:20:54 +0000 (14:20 -0400)
On exiting from the guest TD, xsave state is clobbered; restore it.
Do not use kvm_load_host_xsave_state(), as it relies on vcpu->arch
to find out whether other KVM_RUN code has loaded guest state into
XCR0/PKRU/XSS or not.  In the case of TDX, the exit values are known
independent of the guest CR0 and CR4, and in fact the latter are not
available.

Signed-off-by: Isaku Yamahata <isaku.yamahata@intel.com>
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Message-ID: <20250129095902.16391-8-adrian.hunter@intel.com>
[Rewrite to not use kvm_load_host_xsave_state. - Paolo]
Reviewed-by: Xiayao Li <xiaoyao.li@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/vmx/tdx.c

index eac5fa7..9a08a82 100644 (file)
@@ -2,6 +2,7 @@
 #include <linux/cleanup.h>
 #include <linux/cpu.h>
 #include <asm/cpufeature.h>
+#include <asm/fpu/xcr.h>
 #include <linux/misc_cgroup.h>
 #include <linux/mmu_context.h>
 #include <asm/tdx.h>
@@ -740,6 +741,30 @@ static noinstr void tdx_vcpu_enter_exit(struct kvm_vcpu *vcpu)
                                 BIT_ULL(VCPU_REGS_R14) | \
                                 BIT_ULL(VCPU_REGS_R15))
 
+static void tdx_load_host_xsave_state(struct kvm_vcpu *vcpu)
+{
+       struct kvm_tdx *kvm_tdx = to_kvm_tdx(vcpu->kvm);
+
+       /*
+        * All TDX hosts support PKRU; but even if they didn't,
+        * vcpu->arch.host_pkru would be 0 and the wrpkru would be
+        * skipped.
+        */
+       if (vcpu->arch.host_pkru != 0)
+               wrpkru(vcpu->arch.host_pkru);
+
+       if (kvm_host.xcr0 != (kvm_tdx->xfam & kvm_caps.supported_xcr0))
+               xsetbv(XCR_XFEATURE_ENABLED_MASK, kvm_host.xcr0);
+
+       /*
+        * Likewise, even if a TDX hosts didn't support XSS both arms of
+        * the comparison would be 0 and the wrmsrl would be skipped.
+        */
+       if (kvm_host.xss != (kvm_tdx->xfam & kvm_caps.supported_xss))
+               wrmsrl(MSR_IA32_XSS, kvm_host.xss);
+}
+EXPORT_SYMBOL_GPL(kvm_load_host_xsave_state);
+
 fastpath_t tdx_vcpu_run(struct kvm_vcpu *vcpu, bool force_immediate_exit)
 {
        /*
@@ -756,6 +781,8 @@ fastpath_t tdx_vcpu_run(struct kvm_vcpu *vcpu, bool force_immediate_exit)
 
        tdx_vcpu_enter_exit(vcpu);
 
+       tdx_load_host_xsave_state(vcpu);
+
        vcpu->arch.regs_avail &= TDX_REGS_AVAIL_SET;
 
        trace_kvm_exit(vcpu, KVM_ISA_VMX);
@@ -2332,6 +2359,11 @@ int __init tdx_bringup(void)
                goto success_disable_tdx;
        }
 
+       if (!cpu_feature_enabled(X86_FEATURE_OSXSAVE)) {
+               pr_err("tdx: OSXSAVE is required for TDX\n");
+               goto success_disable_tdx;
+       }
+
        if (!cpu_feature_enabled(X86_FEATURE_MOVDIR64B)) {
                pr_err("tdx: MOVDIR64B is required for TDX\n");
                goto success_disable_tdx;