Skip to contentSkip to conversation

Fix handling of a fault during NMI unblocked due to IRET

PATCH In review Gleb Natapov 2 participants 6 messages Mar 30, 2009 – Apr 1, 2009

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/5 Fix handling of a fault during NMI unblocked due to IRET

Bit 12 is undefined in any of the following cases:
If the VM exit sets the valid bit in the IDT-vectoring information field. If the VM exit is due to a double fault.

---

 arch/x86/kvm/vmx.c |   17 +++++++++++------
 1 files changed, 11 insertions(+), 6 deletions(-)
arch/x86/kvm/vmx.c +11 -6
index 37ae13d..14e3f48 100644
@@ -3259,36 +3259,41 @@ static void update_tpr_threshold(struct kvm_vcpu *vcpu)
 static void vmx_complete_interrupts(struct vcpu_vmx *vmx)
 {
 	u32 exit_intr_info;
-	u32 idt_vectoring_info;
+	u32 idt_vectoring_info = vmx->idt_vectoring_info;
 	bool unblock_nmi;
 	u8 vector;
 	int type;
 	bool idtv_info_valid;
 	u32 error;
 
+	idtv_info_valid = idt_vectoring_info & VECTORING_INFO_VALID_MASK;
 	exit_intr_info = vmcs_read32(VM_EXIT_INTR_INFO);
 	if (cpu_has_virtual_nmis()) {
 		unblock_nmi = (exit_intr_info & INTR_INFO_UNBLOCK_NMI) != 0;
 		vector = exit_intr_info & INTR_INFO_VECTOR_MASK;
 		/*
-		 * SDM 3: 25.7.1.2
+		 * SDM 3: 27.7.1.2 (September 2008)
 		 * Re-set bit "block by NMI" before VM entry if vmexit caused by
 		 * a guest IRET fault.
+		 * SDM 3: 23.2.2 (September 2008)
+		 * Bit 12 is undefined in any of the following cases:
+		 *  If the VM exit sets the valid bit in the IDT-vectoring
+		 *   information field.
+		 *  If the VM exit is due to a double fault.
 		 */
-		if (unblock_nmi && vector != DF_VECTOR)
+		if ((exit_intr_info & INTR_INFO_VALID_MASK) && unblock_nmi &&
+		    vector != DF_VECTOR && !idtv_info_valid)
 			vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO,
 				      GUEST_INTR_STATE_NMI);
 	} else if (unlikely(vmx->soft_vnmi_blocked))
 		vmx->vnmi_blocked_time +=
 			ktime_to_ns(ktime_sub(ktime_get(), vmx->entry_time));
 
-	idt_vectoring_info = vmx->idt_vectoring_info;
-	idtv_info_valid = idt_vectoring_info & VECTORING_INFO_VALID_MASK;
 	vector = idt_vectoring_info & VECTORING_INFO_VECTOR_MASK;
 	type = idt_vectoring_info & VECTORING_INFO_TYPE_MASK;
 	if (vmx->vcpu.arch.nmi_injected) {
 		/*
-		 * SDM 3: 25.7.1.2
+		 * SDM 3: 27.7.1.2 (September 2008)
 		 * Clear bit "block by NMI" before VM entry if a NMI delivery
 		 * faulted.
 		 */
#1 lore raw

PATCH v2 2/5 Rewrite twisted maze of if() statements with more straightforward switch()

Also fix a bug when NMI could be dropped on exit. Although this should never happen in practice.

---

 arch/x86/kvm/vmx.c |   43 +++++++++++++++++++++++++------------------
 1 files changed, 25 insertions(+), 18 deletions(-)
arch/x86/kvm/vmx.c +25 -18
index 14e3f48..1017544 100644
@@ -3264,7 +3264,6 @@ static void vmx_complete_interrupts(struct vcpu_vmx *vmx)
 	u8 vector;
 	int type;
 	bool idtv_info_valid;
-	u32 error;
 
 	idtv_info_valid = idt_vectoring_info & VECTORING_INFO_VALID_MASK;
 	exit_intr_info = vmcs_read32(VM_EXIT_INTR_INFO);
@@ -3289,34 +3288,42 @@ static void vmx_complete_interrupts(struct vcpu_vmx *vmx)
 		vmx->vnmi_blocked_time +=
 			ktime_to_ns(ktime_sub(ktime_get(), vmx->entry_time));
 
+	vmx->vcpu.arch.nmi_injected = false;
+	kvm_clear_exception_queue(&vmx->vcpu);
+	kvm_clear_interrupt_queue(&vmx->vcpu);
+
+	if (!idtv_info_valid)
+		return;
+
 	vector = idt_vectoring_info & VECTORING_INFO_VECTOR_MASK;
 	type = idt_vectoring_info & VECTORING_INFO_TYPE_MASK;
-	if (vmx->vcpu.arch.nmi_injected) {
+	
+	switch(type) {
+	case INTR_TYPE_NMI_INTR:
+		vmx->vcpu.arch.nmi_injected = true;
 		/*
 		 * SDM 3: 27.7.1.2 (September 2008)
-		 * Clear bit "block by NMI" before VM entry if a NMI delivery
-		 * faulted.
+		 * Clear bit "block by NMI" before VM entry if a NMI
+		 * delivery faulted.
 		 */
-		if (idtv_info_valid && type == INTR_TYPE_NMI_INTR)
-			vmcs_clear_bits(GUEST_INTERRUPTIBILITY_INFO,
-					GUEST_INTR_STATE_NMI);
-		else
-			vmx->vcpu.arch.nmi_injected = false;
-	}
-	kvm_clear_exception_queue(&vmx->vcpu);
-	if (idtv_info_valid && (type == INTR_TYPE_HARD_EXCEPTION ||
-				type == INTR_TYPE_SOFT_EXCEPTION)) {
+		vmcs_clear_bits(GUEST_INTERRUPTIBILITY_INFO,
+				GUEST_INTR_STATE_NMI);
+		break;
+	case INTR_TYPE_HARD_EXCEPTION:
+	case INTR_TYPE_SOFT_EXCEPTION:
 		if (idt_vectoring_info & VECTORING_INFO_DELIVER_CODE_MASK) {
-			error = vmcs_read32(IDT_VECTORING_ERROR_CODE);
-			kvm_queue_exception_e(&vmx->vcpu, vector, error);
+			u32 err = vmcs_read32(IDT_VECTORING_ERROR_CODE);
+			kvm_queue_exception_e(&vmx->vcpu, vector, err);
 		} else
 			kvm_queue_exception(&vmx->vcpu, vector);
 		vmx->idt_vectoring_info = 0;
-	}
-	kvm_clear_interrupt_queue(&vmx->vcpu);
-	if (idtv_info_valid && type == INTR_TYPE_EXT_INTR) {
+		break;
+	case INTR_TYPE_EXT_INTR:
 		kvm_queue_interrupt(&vmx->vcpu, vector);
 		vmx->idt_vectoring_info = 0;
+		break;
+	default:
+		break;
 	}
 }
 
#2 lore raw

PATCH v2 3/5 Do not zero idt_vectoring_info in vmx_complete_interrupts().

In reply to Gleb Natapov

We will need it later in task_switch().
Code in handle_exception() is dead. is_external_interrupt(vect_info) will always be false since idt_vectoring_info is zeroed in vmx_complete_interrupts().

---

 arch/x86/kvm/vmx.c |    7 -------
 1 files changed, 0 insertions(+), 7 deletions(-)
arch/x86/kvm/vmx.c +0 -7
index 1017544..0da7a9e 100644
@@ -2613,11 +2613,6 @@ static int handle_exception(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
 		printk(KERN_ERR "%s: unexpected, vectoring info 0x%x "
 		       "intr info 0x%x\n", __func__, vect_info, intr_info);
 
-	if (!irqchip_in_kernel(vcpu->kvm) && is_external_interrupt(vect_info)) {
-		int irq = vect_info & VECTORING_INFO_VECTOR_MASK;
-		kvm_push_irq(vcpu, irq);
-	}
-
 	if ((intr_info & INTR_INFO_INTR_TYPE_MASK) == INTR_TYPE_NMI_INTR)
 		return 1;  /* already handled by vmx_vcpu_run() */
 
@@ -3316,11 +3311,9 @@ static void vmx_complete_interrupts(struct vcpu_vmx *vmx)
 			kvm_queue_exception_e(&vmx->vcpu, vector, err);
 		} else
 			kvm_queue_exception(&vmx->vcpu, vector);
-		vmx->idt_vectoring_info = 0;
 		break;
 	case INTR_TYPE_EXT_INTR:
 		kvm_queue_interrupt(&vmx->vcpu, vector);
-		vmx->idt_vectoring_info = 0;
 		break;
 	default:
 		break;
#3 lore raw

PATCH v2 4/5 Fix task switch back link handling.

In reply to Gleb Natapov

Back link is written to a wrong TSS now.

---

 arch/x86/kvm/x86.c |   40 ++++++++++++++++++++++++++++++++--------
 1 files changed, 32 insertions(+), 8 deletions(-)
arch/x86/kvm/x86.c +32 -8
index ae4918c..f14c622 100644
@@ -3697,7 +3697,6 @@ static void save_state_to_tss32(struct kvm_vcpu *vcpu,
 	tss->fs = get_segment_selector(vcpu, VCPU_SREG_FS);
 	tss->gs = get_segment_selector(vcpu, VCPU_SREG_GS);
 	tss->ldt_selector = get_segment_selector(vcpu, VCPU_SREG_LDTR);
-	tss->prev_task_link = get_segment_selector(vcpu, VCPU_SREG_TR);
 }
 
 static int load_state_from_tss32(struct kvm_vcpu *vcpu,
@@ -3794,8 +3793,8 @@ static int load_state_from_tss16(struct kvm_vcpu *vcpu,
 }
 
 static int kvm_task_switch_16(struct kvm_vcpu *vcpu, u16 tss_selector,
-		       u32 old_tss_base,
-		       struct desc_struct *nseg_desc)
+			      u16 old_tss_sel, u32 old_tss_base,
+			      struct desc_struct *nseg_desc)
 {
 	struct tss_segment_16 tss_segment_16;
 	int ret = 0;
@@ -3814,6 +3813,16 @@ static int kvm_task_switch_16(struct kvm_vcpu *vcpu, u16 tss_selector,
 			   &tss_segment_16, sizeof tss_segment_16))
 		goto out;
 
+	if (old_tss_sel != 0xffff) {
+		tss_segment_16.prev_task_link = old_tss_sel;
+
+		if (kvm_write_guest(vcpu->kvm,
+				    get_tss_base_addr(vcpu, nseg_desc),
+				    &tss_segment_16.prev_task_link,
+				    sizeof tss_segment_16.prev_task_link))
+			goto out;
+	}
+
 	if (load_state_from_tss16(vcpu, &tss_segment_16))
 		goto out;
 
@@ -3823,7 +3832,7 @@ out:
 }
 
 static int kvm_task_switch_32(struct kvm_vcpu *vcpu, u16 tss_selector,
-		       u32 old_tss_base,
+		       u16 old_tss_sel, u32 old_tss_base,
 		       struct desc_struct *nseg_desc)
 {
 	struct tss_segment_32 tss_segment_32;
@@ -3843,6 +3852,16 @@ static int kvm_task_switch_32(struct kvm_vcpu *vcpu, u16 tss_selector,
 			   &tss_segment_32, sizeof tss_segment_32))
 		goto out;
 
+	if (old_tss_sel != 0xffff) {
+		tss_segment_32.prev_task_link = old_tss_sel;
+
+		if (kvm_write_guest(vcpu->kvm,
+				    get_tss_base_addr(vcpu, nseg_desc),
+				    &tss_segment_32.prev_task_link,
+				    sizeof tss_segment_32.prev_task_link))
+			goto out;
+	}
+
 	if (load_state_from_tss32(vcpu, &tss_segment_32))
 		goto out;
 
@@ -3898,12 +3917,17 @@ int kvm_task_switch(struct kvm_vcpu *vcpu, u16 tss_selector, int reason)
 
 	kvm_x86_ops->skip_emulated_instruction(vcpu);
 
+	/* set back link to prev task only if NT bit is set in eflags
+	   note that old_tss_sel is not used afetr this point */
+	if (reason != TASK_SWITCH_CALL && reason != TASK_SWITCH_GATE)
+		old_tss_sel = 0xffff;
+
 	if (nseg_desc.type & 8)
-		ret = kvm_task_switch_32(vcpu, tss_selector, old_tss_base,
-					 &nseg_desc);
+		ret = kvm_task_switch_32(vcpu, tss_selector, old_tss_sel,
+					 old_tss_base, &nseg_desc);
 	else
-		ret = kvm_task_switch_16(vcpu, tss_selector, old_tss_base,
-					 &nseg_desc);
+		ret = kvm_task_switch_16(vcpu, tss_selector, old_tss_sel,
+					 old_tss_base, &nseg_desc);
 
 	if (reason == TASK_SWITCH_CALL || reason == TASK_SWITCH_GATE) {
 		u32 eflags = kvm_x86_ops->get_rflags(vcpu);
#4 lore raw

PATCH v2 5/5 Fix unneeded instruction skipping during task switching.

In reply to Gleb Natapov

There is no need to skip instruction if the reason for a task switch is a task gate in IDT and access to it is caused by an external even. The problem is currently solved only for VMX since there is no reliable way to skip an instruction in SVM. We should emulate it instead.

---

 arch/x86/include/asm/svm.h |    1 +
 arch/x86/kvm/svm.c         |   25 ++++++++++++++++++-------
 arch/x86/kvm/vmx.c         |   40 +++++++++++++++++++++++++++++-----------
 arch/x86/kvm/x86.c         |    5 ++++-
 4 files changed, 52 insertions(+), 19 deletions(-)
arch/x86/include/asm/svm.h +1 -0
index 82ada75..85574b7 100644
@@ -225,6 +225,7 @@ struct __attribute__ ((__packed__)) vmcb {
 #define SVM_EVTINJ_VALID_ERR (1 << 11)
 
 #define SVM_EXITINTINFO_VEC_MASK SVM_EVTINJ_VEC_MASK
+#define SVM_EXITINTINFO_TYPE_MASK SVM_EVTINJ_TYPE_MASK
 
 #define	SVM_EXITINTINFO_TYPE_INTR SVM_EVTINJ_TYPE_INTR
 #define	SVM_EXITINTINFO_TYPE_NMI SVM_EVTINJ_TYPE_NMI
arch/x86/kvm/svm.c +18 -7
index 1fcbc17..3ffb695 100644
@@ -1823,17 +1823,28 @@ static int task_switch_interception(struct vcpu_svm *svm,
 				    struct kvm_run *kvm_run)
 {
 	u16 tss_selector;
+	int reason;
+	int int_type = svm->vmcb->control.exit_int_info &
+		SVM_EXITINTINFO_TYPE_MASK;
 
 	tss_selector = (u16)svm->vmcb->control.exit_info_1;
+
 	if (svm->vmcb->control.exit_info_2 &
 	    (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
-		return kvm_task_switch(&svm->vcpu, tss_selector,
-				       TASK_SWITCH_IRET);
-	if (svm->vmcb->control.exit_info_2 &
-	    (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
-		return kvm_task_switch(&svm->vcpu, tss_selector,
-				       TASK_SWITCH_JMP);
-	return kvm_task_switch(&svm->vcpu, tss_selector, TASK_SWITCH_CALL);
+		reason = TASK_SWITCH_IRET;
+	else if (svm->vmcb->control.exit_info_2 &
+		 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
+		reason = TASK_SWITCH_JMP;
+	else if (svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID)
+		reason = TASK_SWITCH_GATE;
+	else
+		reason = TASK_SWITCH_CALL;
+
+
+	if (reason != TASK_SWITCH_GATE || int_type == SVM_EXITINTINFO_TYPE_SOFT)
+		skip_emulated_instruction(&svm->vcpu);
+
+	return kvm_task_switch(&svm->vcpu, tss_selector, reason);
 }
 
 static int cpuid_interception(struct vcpu_svm *svm, struct kvm_run *kvm_run)
arch/x86/kvm/vmx.c +29 -11
index 0da7a9e..01db958 100644
@@ -3025,22 +3025,40 @@ static int handle_task_switch(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
 	struct vcpu_vmx *vmx = to_vmx(vcpu);
 	unsigned long exit_qualification;
 	u16 tss_selector;
-	int reason;
+	int reason, type, idt_v;
+
+	idt_v = (vmx->idt_vectoring_info & VECTORING_INFO_VALID_MASK);
+	type = (vmx->idt_vectoring_info & VECTORING_INFO_TYPE_MASK);
 
 	exit_qualification = vmcs_readl(EXIT_QUALIFICATION);
 
 	reason = (u32)exit_qualification >> 30;
-	if (reason == TASK_SWITCH_GATE && vmx->vcpu.arch.nmi_injected &&
-	    (vmx->idt_vectoring_info & VECTORING_INFO_VALID_MASK) &&
-	    (vmx->idt_vectoring_info & VECTORING_INFO_TYPE_MASK)
-	    == INTR_TYPE_NMI_INTR) {
-		vcpu->arch.nmi_injected = false;
-		if (cpu_has_virtual_nmis())
-			vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO,
-				      GUEST_INTR_STATE_NMI);
+	if (reason == TASK_SWITCH_GATE && idt_v) {
+		switch (type) {
+		case INTR_TYPE_NMI_INTR:
+			vcpu->arch.nmi_injected = false;
+			if (cpu_has_virtual_nmis())
+				vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO,
+					      GUEST_INTR_STATE_NMI);
+			break;
+		case INTR_TYPE_EXT_INTR:
+			kvm_clear_interrupt_queue(vcpu);
+			break;
+		case INTR_TYPE_HARD_EXCEPTION:
+		case INTR_TYPE_SOFT_EXCEPTION:
+			kvm_clear_exception_queue(vcpu);
+			break;
+		default:
+			break;
+		}
 	}
 	tss_selector = exit_qualification;
 
+	if (!idt_v || (type != INTR_TYPE_HARD_EXCEPTION &&
+		       type != INTR_TYPE_EXT_INTR &&
+		       type != INTR_TYPE_NMI_INTR))
+		skip_emulated_instruction(vcpu);
+
 	if (!kvm_task_switch(vcpu, tss_selector, reason))
 		return 0;
 
@@ -3292,8 +3310,8 @@ static void vmx_complete_interrupts(struct vcpu_vmx *vmx)
 
 	vector = idt_vectoring_info & VECTORING_INFO_VECTOR_MASK;
 	type = idt_vectoring_info & VECTORING_INFO_TYPE_MASK;
-	
-	switch(type) {
+
+	switch (type) {
 	case INTR_TYPE_NMI_INTR:
 		vmx->vcpu.arch.nmi_injected = true;
 		/*
arch/x86/kvm/x86.c +4 -1
index f14c622..9415a03 100644
@@ -3915,7 +3915,10 @@ int kvm_task_switch(struct kvm_vcpu *vcpu, u16 tss_selector, int reason)
 		kvm_x86_ops->set_rflags(vcpu, eflags & ~X86_EFLAGS_NT);
 	}
 
-	kvm_x86_ops->skip_emulated_instruction(vcpu);
+	/* set back link to prev task only if NT bit is set in eflags
+	   note that old_tss_sel is not used afetr this point */
+	if (reason != TASK_SWITCH_CALL && reason != TASK_SWITCH_GATE)
+		old_tss_sel = 0xffff;
 
 	/* set back link to prev task only if NT bit is set in eflags
 	   note that old_tss_sel is not used afetr this point */
#5 lore raw
In reply to Gleb Natapov
Quoted 4 lines · Gleb Natapov

Bit 12 is undefined in any of the following cases:
If the VM exit sets the valid bit in the IDT-vectoring information field. If the VM exit is due to a double fault.

Applied the entire series, thanks.

Signature · 2 lines
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.