Skip to contentSkip to conversation

KVM: x86: Guard against division by zero in adjust_lapic_timer_advance()

PATCH In review Peng Hao 3 participants 6 messages Oct 8 – Oct 15

Subsystems (3)
lore
Share & tools
lore

Peng Hao

author
#0 lore raw

PATCH 1/1 KVM: x86: Guard against division by zero in adjust_lapic_timer_advance()

When TSC calibration fails and userspace never programs a tsc khz, vcpu->arch.virtual_tsc_khz stays zero, but the lapic timer's software-mode expiry path still calls adjust_lapic_timer_advance() unconditionally on every timer expiration (unlike __wait_lapic_expire(), this path is not gated on tsc_scaling_ratio). Both the "too early" and "too late" branches do_div() by virtual_tsc_khz, so a zero khz crashes the host with #DE as soon as the timer advance drifts outside the adjust window.

Bail out before either division.

---
 arch/x86/kvm/lapic.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kvm/lapic.c +3 -0
index 13e32c4915e2..4dcbc66c8e07 100644
@@ -1991,6 +1991,9 @@ static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu,
 	    abs(advance_expire_delta) < LAPIC_TIMER_ADVANCE_ADJUST_MIN)
 		return;
 
+	if (unlikely(!vcpu->arch.virtual_tsc_khz))
+		return;
+
 	/* too early */
 	if (advance_expire_delta < 0) {
 		ns = -advance_expire_delta * 1000000ULL;
Signature · 1 line
2.43.7
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [High] kvm: x86: divide by zero in __wait_lapic_expire() with nested TSC scaling --

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] kvm: x86: divide by zero in __wait_lapic_expire() with nested TSC scaling --

commit 615a870f84e9397c4bfba8db0f187034794e3d9e
Author: Peng Hao <[email protected]>

KVM: x86: Guard against division by zero in adjust_lapic_timer_advance()

When TSC calibration fails, vcpu->arch.virtual_tsc_khz is zero. This patch adds a check to avoid division by zero in adjust_lapic_timer_advance().

Quoted 5 lines

When TSC calibration fails and userspace never programs a tsc khz, vcpu->arch.virtual_tsc_khz stays zero, but the lapic timer's software-mode expiry path still calls adjust_lapic_timer_advance() unconditionally on every timer expiration (unlike __wait_lapic_expire(), this path is not gated on tsc_scaling_ratio).

[ ... ]

arch/x86/kvm/lapic.c static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu, · line 1983

Quoted 11 lines
arch/x86/kvm/lapic.c +3 -0
index 48b019114c196..dc4a669d1d460 100644
@@ -1978,6 +1978,9 @@ static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu,
 	    abs(advance_expire_delta) < LAPIC_TIMER_ADVANCE_ADJUST_MIN)
 		return;
 
+	if (unlikely(!vcpu->arch.virtual_tsc_khz))
+		return;
+

[Severity: High]
Does __wait_lapic_expire() need this same check?

The commit message mentions __wait_lapic_expire() is safely gated by tsc_scaling_ratio, but is that description accurate when nested virtualization is enabled?

An L1 guest can enable nested virtualization and set a custom tsc_multiplier in VMCS12 for its L2 guest. KVM computes vcpu->arch.tsc_scaling_ratio to this non-default value:

arch/x86/kvm/vmx/nested.c:prepare_vmcs02_early() {
    ...
	vcpu->arch.tsc_scaling_ratio = kvm_calc_nested_tsc_multiplier(
			vcpu->arch.l1_tsc_scaling_ratio,
			vmx_get_l2_tsc_multiplier(vcpu));
    ...
}

If host TSC calibration failed and virtual_tsc_khz is 0, when the LAPIC timer expires during L2 execution, __wait_lapic_expire() takes the else branch because tsc_scaling_ratio != default_tsc_scaling_ratio:

arch/x86/kvm/lapic.c:__wait_lapic_expire() {
    ...
	if (vcpu->arch.tsc_scaling_ratio == kvm_caps.default_tsc_scaling_ratio) {
		__delay(min(guest_cycles,
			nsec_to_cycles(vcpu, timer_advance_ns)));
	} else {
		u64 delay_ns = guest_cycles * 1000000ULL;
		do_div(delay_ns, vcpu->arch.virtual_tsc_khz);
		ndelay(min_t(u32, delay_ns, timer_advance_ns));
	}
    ...
}

Can an untrusted L1 guest use this path to reliably crash the host with a divide-by-zero panic?

Quoted 3 lines
arch/x86/kvm/lapic.c +0 -0
 	/* too early */
 	if (advance_expire_delta < 0) {
 		ns = -advance_expire_delta * 1000000ULL;
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Peng Hao
Quoted 21 lines · Peng Hao

When TSC calibration fails and userspace never programs a tsc khz, vcpu->arch.virtual_tsc_khz stays zero, but the lapic timer's software-mode expiry path still calls adjust_lapic_timer_advance() unconditionally on every timer expiration (unlike __wait_lapic_expire(), this path is not gated on tsc_scaling_ratio). Both the "too early" and "too late" branches do_div() by virtual_tsc_khz, so a zero khz crashes the host with #DE as soon as the timer advance drifts outside the adjust window.

Bail out before either division.

---
 arch/x86/kvm/lapic.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kvm/lapic.c +0 -0
index 13e32c4915e2..4dcbc66c8e07 100644

arch/x86/kvm/lapic.c static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu, · line 1995

Quoted 5 lines
arch/x86/kvm/lapic.c +2 -0
@@ -1991,6 +1991,9 @@ static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu,
 	    abs(advance_expire_delta) < LAPIC_TIMER_ADVANCE_ADJUST_MIN)
 		return;
 
+	if (unlikely(!vcpu->arch.virtual_tsc_khz))
+		return;

Ugh. I'd rather reject KVM_CREATE_VCPU, even though there is a rather surprising amount of code in KVM that does indeed play nice with tsc_khz == 0. I don't see how the guest can possibly function with virtual_tsc_khz==0.

vcpu->arch.virtual_tsc_{shift,mult} will also be left as zero (to avoid #DE there as well):

	/* tsc_khz can be zero if TSC calibration fails */
	if (user_tsc_khz == 0) {
		/* set tsc_scaling_ratio to a safe value */
		kvm_vcpu_write_tsc_multiplier(vcpu, kvm_caps.default_tsc_scaling_ratio);
		return -1;
	}

	/* Compute a scale to convert nanoseconds in TSC cycles */
	kvm_get_time_scale(user_tsc_khz * 1000LL, NSEC_PER_SEC,
			   &vcpu->arch.virtual_tsc_shift,
			   &vcpu->arch.virtual_tsc_mult);

and so nsec_to_cycles() will always return zero due to multiplying by zero. That means things like the APIC timer will always fire immediately:

	apic->lapic_timer.tscdeadline = kvm_read_l1_tsc(apic->vcpu, tscl) +
		nsec_to_cycles(apic->vcpu, deadline);

The first instance of this goes back to 03ba32cae66e ("VMX: x86: handle host TSC calibration failure"), and every "fix" since then has been to avoid #DE.

Quoted 7 lines
arch/x86/kvm/lapic.c +1 -0
+
 	/* too early */
 	if (advance_expire_delta < 0) {
 		ns = -advance_expire_delta * 1000000ULL;
Signature · 1 line
2.43.7
author
#3 lore raw
Quoted 26 lines · Sean Christopherson
Quoted 24 lines · Peng Hao

When TSC calibration fails and userspace never programs a tsc khz, vcpu->arch.virtual_tsc_khz stays zero, but the lapic timer's software-mode expiry path still calls adjust_lapic_timer_advance() unconditionally on every timer expiration (unlike __wait_lapic_expire(), this path is not gated on tsc_scaling_ratio). Both the "too early" and "too late" branches do_div() by virtual_tsc_khz, so a zero khz crashes the host with #DE as soon as the timer advance drifts outside the adjust window.

Bail out before either division.

---
 arch/x86/kvm/lapic.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kvm/lapic.c +0 -0
index 13e32c4915e2..4dcbc66c8e07 100644
@@ -1991,6 +1991,9 @@ static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu,
          abs(advance_expire_delta) < LAPIC_TIMER_ADVANCE_ADJUST_MIN)
              return;
 
Quoted 7 lines

arch/x86/kvm/lapic.c static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu, · line 1995

Quoted 2 lines
arch/x86/kvm/lapic.c +2 -0
+     if (unlikely(!vcpu->arch.virtual_tsc_khz))
+             return;

Ugh. I'd rather reject KVM_CREATE_VCPU, even though there is a rather surprising amount of code in KVM that does indeed play nice with tsc_khz == 0. I don't see how the guest can possibly function with virtual_tsc_khz==0.

Agreed the guest is broken in that state -- virtual_tsc_mult is zero too, so nsec_to_cycles() returns 0 and the deadline timer fires immediately. But I don't think KVM_CREATE_VCPU is where to enforce it, because letting vCPU creation
succeed is what makes the state recoverable:
- kvm_arch_vcpu_create() deliberately ignores kvm_set_tsc_khz()'s -1
- KVM_SET_TSC_KHZ is only reachable once the vCPU exists. With kvm_caps.has_tsc_control and a khz below kvm_caps.max_guest_tsc_khz it recomputes virtual_tsc_{shift,mult} and the guest is fully
functional again.
- max_tsc_khz does the same thing ahead of time, via default_tsc_khz.

Rejecting KVM_CREATE_VCPU removes both, the guard is the kind of small,
obviously-correct fix that is.

Quoted 31 lines

vcpu->arch.virtual_tsc_{shift,mult} will also be left as zero (to avoid #DE there as well):

        /* tsc_khz can be zero if TSC calibration fails */
        if (user_tsc_khz == 0) {
                /* set tsc_scaling_ratio to a safe value */
                kvm_vcpu_write_tsc_multiplier(vcpu, kvm_caps.default_tsc_scaling_ratio);
                return -1;
        }

        /* Compute a scale to convert nanoseconds in TSC cycles */
        kvm_get_time_scale(user_tsc_khz * 1000LL, NSEC_PER_SEC,
                           &vcpu->arch.virtual_tsc_shift,
                           &vcpu->arch.virtual_tsc_mult);

and so nsec_to_cycles() will always return zero due to multiplying by zero. That means things like the APIC timer will always fire immediately:

        apic->lapic_timer.tscdeadline = kvm_read_l1_tsc(apic->vcpu, tscl) +
                nsec_to_cycles(apic->vcpu, deadline);

The first instance of this goes back to 03ba32cae66e ("VMX: x86: handle host TSC calibration failure"), and every "fix" since then has been to avoid #DE.

Quoted 7 lines
arch/x86/kvm/lapic.c +1 -1
+
      /* too early */
      if (advance_expire_delta < 0) {
              ns = -advance_expire_delta * 1000000ULL;
--

2.43.7

#4 lore raw
Quoted 34 lines · Hao Peng
Quoted 26 lines · Sean Christopherson
Quoted 24 lines · Peng Hao

When TSC calibration fails and userspace never programs a tsc khz, vcpu->arch.virtual_tsc_khz stays zero, but the lapic timer's software-mode expiry path still calls adjust_lapic_timer_advance() unconditionally on every timer expiration (unlike __wait_lapic_expire(), this path is not gated on tsc_scaling_ratio). Both the "too early" and "too late" branches do_div() by virtual_tsc_khz, so a zero khz crashes the host with #DE as soon as the timer advance drifts outside the adjust window.

Bail out before either division.

---
 arch/x86/kvm/lapic.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kvm/lapic.c +0 -0
index 13e32c4915e2..4dcbc66c8e07 100644
@@ -1991,6 +1991,9 @@ static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu,
          abs(advance_expire_delta) < LAPIC_TIMER_ADVANCE_ADJUST_MIN)
              return;
 
Quoted 7 lines

arch/x86/kvm/lapic.c static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu, · line 1995

Quoted 2 lines
arch/x86/kvm/lapic.c +2 -0
+     if (unlikely(!vcpu->arch.virtual_tsc_khz))
+             return;

Ugh. I'd rather reject KVM_CREATE_VCPU, even though there is a rather surprising amount of code in KVM that does indeed play nice with tsc_khz == 0. I don't see how the guest can possibly function with virtual_tsc_khz==0.

Quoted 10 lines

Agreed the guest is broken in that state -- virtual_tsc_mult is zero too, so nsec_to_cycles() returns 0 and the deadline timer fires immediately. But I don't think KVM_CREATE_VCPU is where to enforce it, because letting vCPU creation
succeed is what makes the state recoverable:
- kvm_arch_vcpu_create() deliberately ignores kvm_set_tsc_khz()'s -1
- KVM_SET_TSC_KHZ is only reachable once the vCPU exists. With kvm_caps.has_tsc_control and a khz below kvm_caps.max_guest_tsc_khz it recomputes virtual_tsc_{shift,mult} and the guest is fully
functional again.
- max_tsc_khz does the same thing ahead of time, via default_tsc_khz.

I don't see how it's recoverable. kvm_set_tsc_khz() will compute thresh_{lo,hi}=0, and so use_scaling will always be true. Which means this:

	ratio = mul_u64_u32_div(1ULL << kvm_caps.tsc_scaling_ratio_frac_bits,
				user_tsc_khz, tsc_khz);

will again hit a #DE, no?

Even if KVM somehow makes it past that, kvm_guest_time_update() will put the vCPU into an infinite request loop as get_cpu_tsc_khz() will always return 0 for modern CPUs with a constant TSC:

static unsigned long get_cpu_tsc_khz(void)

  {
	if (cpu_feature_enabled(X86_FEATURE_CONSTANT_TSC))
		return tsc_khz;    <======================== always zero
	else
		return __this_cpu_read(cpu_tsc_khz);

}

And so this code will re-request KVM_REQ_CLOCK_UPDATE and the vCPU will hang.

	tgt_tsc_hz = (u64)get_cpu_tsc_khz() * HZ_PER_KHZ;
	if (unlikely(tgt_tsc_hz == 0)) {
		preempt_enable();
		kvm_make_request(KVM_REQ_CLOCK_UPDATE, v);
		return 1;
	}
author
#5 lore raw
Quoted 50 lines · Sean Christopherson
Quoted 34 lines · Hao Peng
Quoted 26 lines · Sean Christopherson
Quoted 24 lines · Peng Hao

When TSC calibration fails and userspace never programs a tsc khz, vcpu->arch.virtual_tsc_khz stays zero, but the lapic timer's software-mode expiry path still calls adjust_lapic_timer_advance() unconditionally on every timer expiration (unlike __wait_lapic_expire(), this path is not gated on tsc_scaling_ratio). Both the "too early" and "too late" branches do_div() by virtual_tsc_khz, so a zero khz crashes the host with #DE as soon as the timer advance drifts outside the adjust window.

Bail out before either division.

---
 arch/x86/kvm/lapic.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kvm/lapic.c +0 -0
index 13e32c4915e2..4dcbc66c8e07 100644
@@ -1991,6 +1991,9 @@ static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu,
          abs(advance_expire_delta) < LAPIC_TIMER_ADVANCE_ADJUST_MIN)
              return;
 
Quoted 7 lines

arch/x86/kvm/lapic.c static inline void adjust_lapic_timer_advance(struct kvm_vcpu *vcpu, · line 1995

Quoted 2 lines
arch/x86/kvm/lapic.c +2 -0
+     if (unlikely(!vcpu->arch.virtual_tsc_khz))
+             return;

Ugh. I'd rather reject KVM_CREATE_VCPU, even though there is a rather surprising amount of code in KVM that does indeed play nice with tsc_khz == 0. I don't see how the guest can possibly function with virtual_tsc_khz==0.

Quoted 10 lines

Agreed the guest is broken in that state -- virtual_tsc_mult is zero too, so nsec_to_cycles() returns 0 and the deadline timer fires immediately. But I don't think KVM_CREATE_VCPU is where to enforce it, because letting vCPU creation
succeed is what makes the state recoverable:
- kvm_arch_vcpu_create() deliberately ignores kvm_set_tsc_khz()'s -1
- KVM_SET_TSC_KHZ is only reachable once the vCPU exists. With kvm_caps.has_tsc_control and a khz below kvm_caps.max_guest_tsc_khz it recomputes virtual_tsc_{shift,mult} and the guest is fully
functional again.
- max_tsc_khz does the same thing ahead of time, via default_tsc_khz.

I don't see how it's recoverable. kvm_set_tsc_khz() will compute thresh_{lo,hi}=0, and so use_scaling will always be true. Which means this:

Quoted 9 lines
        ratio = mul_u64_u32_div(1ULL << kvm_caps.tsc_scaling_ratio_frac_bits,
                                user_tsc_khz, tsc_khz);

will again hit a #DE, no?

Even if KVM somehow makes it past that, kvm_guest_time_update() will put the vCPU into an infinite request loop as get_cpu_tsc_khz() will always return 0 for modern CPUs with a constant TSC:

You're right, and it's worse than I argued. I went back and checked:
- adjust_tsc_khz() gives thresh_lo == thresh_hi == 0, so use_scaling
is always true, and
set_tsc_khz() lands on mul_u64_u32_div(..., user_tsc_khz, tsc_khz)
with tsc_khz == 0.
So KVM_SET_TSC_KHZ itself #DEs -- the "recoverable via KVM_SET_TSC_KHZ"
argument I made is wrong; retract it.
- get_cpu_tsc_khz() (x86.c:1652) returns tsc_khz for
CONSTANT_TSC, and mark_tsc_unstable()
doesn't clear that CPUID bit, so kvm_guest_time_update() returns 1 forever. Since vcpu_run() only
breaks on r <= 0 , the vCPU thread spins in the host kernel rather than even failing the guest.

Worse for my approach: that first one is a third division by zero in
the same state, reachable from
a plain ioctl and not covered by either of my patches. "Add a guard at each use site" is a losing
strategy here.

Quoted 17 lines
  static unsigned long get_cpu_tsc_khz(void)
  {
        if (cpu_feature_enabled(X86_FEATURE_CONSTANT_TSC))
                return tsc_khz;    <======================== always zero
        else
                return __this_cpu_read(cpu_tsc_khz);
  }

And so this code will re-request KVM_REQ_CLOCK_UPDATE and the vCPU will hang.

        tgt_tsc_hz = (u64)get_cpu_tsc_khz() * HZ_PER_KHZ;
        if (unlikely(tgt_tsc_hz == 0)) {
                preempt_enable();
                kvm_make_request(KVM_REQ_CLOCK_UPDATE, v);
                return 1;
        }