[PATCH RFC] KVM: nSVM: Fix L1 state corruption upon return from SMM

[Vitaly Kuznetsov]

VMCB split commit 4995a3685f1b ("KVM: SVM: Use a separate vmcb for the
nested L2 guest") broke return from SMM when we entered there from guest
(L2) mode. Gen2 WS2016/Hyper-V is known to do this on boot. The problem
manifests itself like this:

  kvm_exit:             reason EXIT_RSM rip 0x7ffbb280 info 0 0
  kvm_emulate_insn:     0:7ffbb280: 0f aa
  kvm_smm_transition:   vcpu 0: leaving SMM, smbase 0x7ffb3000
  kvm_nested_vmrun:     rip: 0x000000007ffbb280 vmcb: 0x0000000008224000
    nrip: 0xffffffffffbbe119 int_ctl: 0x01020000 event_inj: 0x00000000
    npt: on
  kvm_nested_intercepts: cr_read: 0000 cr_write: 0010 excp: 40060002
    intercepts: fd44bfeb 0000217f 00000000
  kvm_entry:            vcpu 0, rip 0xffffffffffbbe119
  kvm_exit:             reason EXIT_NPF rip 0xffffffffffbbe119 info
    200000006 1ab000
  kvm_nested_vmexit:    vcpu 0 reason npf rip 0xffffffffffbbe119 info1
    0x0000000200000006 info2 0x00000000001ab000 intr_info 0x00000000
    error_code 0x00000000
  kvm_page_fault:       address 1ab000 error_code 6
  kvm_nested_vmexit_inject: reason EXIT_NPF info1 200000006 info2 1ab000
    int_info 0 int_info_err 0
  kvm_entry:            vcpu 0, rip 0x7ffbb280
  kvm_exit:             reason EXIT_EXCP_GP rip 0x7ffbb280 info 0 0
  kvm_emulate_insn:     0:7ffbb280: 0f aa
  kvm_inj_exception:    #GP (0x0)

Note: return to L2 succeeded but upon first exit to L1 its RIP points to
'RSM' instruction but we're not in SMM.

The problem appears to be that VMCB01 gets irreversibly destroyed during
SMM execution. Previously, we used to have 'hsave' VMCB where regular
(pre-SMM) L1's state was saved upon nested_svm_vmexit() but now we just
switch to VMCB01 from VMCB02.

Pre-split (working) flow looked like:
- SMM is triggered during L2's execution
- L2's state is pushed to SMRAM
- nested_svm_vmexit() restores L1's state from 'hsave'
- SMM -> RSM
- enter_svm_guest_mode() switches to L2 but keeps 'hsave' intact so we have
  pre-SMM (and pre L2 VMRUN) L1's state there
- L2's state is restored from SMRAM
- upon first exit L1's state is restored from L1.

This was always broken with regards to svm_get_nested_state()/
svm_set_nested_state(): 'hsave' was never a part of what's being
save and restored so migration happening during SMM triggered from L2 would
never restore L1's state correctly.

Post-split flow (broken) looks like:
- SMM is triggered during L2's execution
- L2's state is pushed to SMRAM
- nested_svm_vmexit() switches to VMCB01 from VMCB02
- SMM -> RSM
- enter_svm_guest_mode() switches from VMCB01 to VMCB02 but pre-SMM VMCB01
  is already lost.
- L2's state is restored from SMRAM
- upon first exit L1's state is restored from VMCB01 but it is corrupted
 (reflects the state during 'RSM' execution).

VMX doesn't have this problem because unlike VMCB, VMCS keeps both guest
and host state so when we switch back to VMCS02 L1's state is intact there.

To resolve the issue we need to save L1's state somewhere. We could've
created a third VMCB for SMM but that would require us to modify saved
state format. L1's architectural HSAVE area (pointed by MSR_VM_HSAVE_PA)
seems appropriate: L0 is free to save any (or none) of L1's state there.
Currently, KVM does 'none'.

Note, for nested state migration to succeed, both source and destination
hypervisors must have the fix. We, however, don't need to create a new
flag indicating the fact that HSAVE area is now populated as migration
during SMM triggered from L2 was always broken.

Fixes: 4995a3685f1b ("KVM: SVM: Use a separate vmcb for the nested L2 guest")
Signed-off-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>
---
- RFC: I'm not 100% sure my 'smart' idea to use currently-unused HSAVE area
is that smart. Also, we don't even seem to check that L1 set it up upon
nested VMRUN so hypervisors which don't do that may remain broken. A very
much needed selftest is also missing.
---
 arch/x86/kvm/svm/svm.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)

diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 12c06ea28f5c..d110bfe0e208 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -4286,6 +4286,7 @@ static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
 static int svm_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
 {
 	struct vcpu_svm *svm = to_svm(vcpu);
+	struct kvm_host_map map_save;
 	int ret;
 
 	if (is_guest_mode(vcpu)) {
@@ -4301,6 +4302,13 @@ static int svm_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
 		ret = nested_svm_vmexit(svm);
 		if (ret)
 			return ret;
+
+		/* Save L1 state to L1 HSAVE area as vmcb01 will be used in SMM */
+		if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr),
+				 &map_save) == -EINVAL)
+			return 1;
+		memcpy(map_save.hva, &svm->vmcb01.ptr->save, sizeof(svm->vmcb01.ptr->save));
+		kvm_vcpu_unmap(vcpu, &map_save, true);
 	}
 	return 0;
 }
@@ -4308,7 +4316,7 @@ static int svm_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
 static int svm_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
 {
 	struct vcpu_svm *svm = to_svm(vcpu);
-	struct kvm_host_map map;
+	struct kvm_host_map map, map_save;
 	int ret = 0;
 
 	if (guest_cpuid_has(vcpu, X86_FEATURE_LM)) {
@@ -4332,6 +4340,13 @@ static int svm_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
 
 			ret = enter_svm_guest_mode(vcpu, vmcb12_gpa, map.hva);
 			kvm_vcpu_unmap(vcpu, &map, true);
+
+			/* Restore L1 state from L1 HSAVE area as vmcb01 was used in SMM */
+			if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr),
+					 &map_save) == -EINVAL)
+				return 1;
+			memcpy(&svm->vmcb01.ptr->save, map_save.hva, sizeof(svm->vmcb01.ptr->save));
+			kvm_vcpu_unmap(vcpu, &map_save, true);
 		}
 	}
 
-- 
2.31.1