2021-09-03 07:57:34

by Jiasheng Jiang

Subject: [PATCH 4/4] KVM: X86: Potential 'index out of range' bug

The kvm_get_vcpu() will call for the array_index_nospec()
with the value of atomic_read(&(v->kvm)->online_vcpus) as size,
and the value of constant '0' as index.
If the size is also '0', it will be unreasonabe
that the index is no less than the size.

Signed-off-by: Jiang Jiasheng <[email protected]>
---
arch/x86/kvm/x86.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index e0f4a46..c59013c 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2871,7 +2871,7 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
offsetof(struct compat_vcpu_info, time));
if (vcpu->xen.vcpu_time_info_set)
kvm_setup_pvclock_page(v, &vcpu->xen.vcpu_time_info_cache, 0);
- if (v == kvm_get_vcpu(v->kvm, 0))
+ if (atomic_read(&(v->kvm)->online_vcpus) > 0 && v == kvm_get_vcpu(v->kvm, 0))
kvm_hv_setup_tsc_page(v->kvm, &vcpu->hv_clock);
return 0;
}
--
2.7.4

2021-09-03 12:09:34

by Vitaly Kuznetsov

Subject: Re: [PATCH 4/4] KVM: X86: Potential 'index out of range' bug

Jiang Jiasheng <[email protected]> writes:

> The kvm_get_vcpu() will call for the array_index_nospec()
> with the value of atomic_read(&(v->kvm)->online_vcpus) as size,
> and the value of constant '0' as index.
> If the size is also '0', it will be unreasonabe
> that the index is no less than the size.
>

Can this really happen?

'online_vcpus' is never decreased, it is increased with every
kvm_vm_ioctl_create_vcpu() call when a new vCPU is created and is set to
0 when all vCPUs are destroyed (kvm_free_vcpus()).

kvm_guest_time_update() takes a vcpu as a parameter, this means that at
least 1 vCPU is currently present so 'online_vcpus' just can't be zero.

> Signed-off-by: Jiang Jiasheng <[email protected]>
> ---
> arch/x86/kvm/x86.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index e0f4a46..c59013c 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -2871,7 +2871,7 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
> offsetof(struct compat_vcpu_info, time));
> if (vcpu->xen.vcpu_time_info_set)
> kvm_setup_pvclock_page(v, &vcpu->xen.vcpu_time_info_cache, 0);
> - if (v == kvm_get_vcpu(v->kvm, 0))
> + if (atomic_read(&(v->kvm)->online_vcpus) > 0 && v == kvm_get_vcpu(v->kvm, 0))
> kvm_hv_setup_tsc_page(v->kvm, &vcpu->hv_clock);
> return 0;
> }

--
Vitaly

2021-09-03 16:18:04

by Sean Christopherson

Subject: Re: [PATCH 4/4] KVM: X86: Potential 'index out of range' bug

On Fri, Sep 03, 2021, Vitaly Kuznetsov wrote:
> Jiang Jiasheng <[email protected]> writes:
>
> > The kvm_get_vcpu() will call for the array_index_nospec()
> > with the value of atomic_read(&(v->kvm)->online_vcpus) as size,
> > and the value of constant '0' as index.
> > If the size is also '0', it will be unreasonabe
> > that the index is no less than the size.
> >
>
> Can this really happen?
>
> 'online_vcpus' is never decreased, it is increased with every
> kvm_vm_ioctl_create_vcpu() call when a new vCPU is created and is set to
> 0 when all vCPUs are destroyed (kvm_free_vcpus()).
>
> kvm_guest_time_update() takes a vcpu as a parameter, this means that at
> least 1 vCPU is currently present so 'online_vcpus' just can't be zero.

Agreed, but doing kvm_get_vcpu() is ugly and overkill.

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 86539c1686fa..cc1cb9a401cd 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2969,7 +2969,7 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
offsetof(struct compat_vcpu_info, time));
if (vcpu->xen.vcpu_time_info_set)
kvm_setup_pvclock_page(v, &vcpu->xen.vcpu_time_info_cache, 0);
- if (v == kvm_get_vcpu(v->kvm, 0))
+ if (!kvm_vcpu_get_idx(v))
kvm_hv_setup_tsc_page(v->kvm, &vcpu->hv_clock);
return 0;
}

2021-09-06 11:20:22

by Vitaly Kuznetsov

Subject: Re: [PATCH 4/4] KVM: X86: Potential 'index out of range' bug

Sean Christopherson <[email protected]> writes:

> On Fri, Sep 03, 2021, Vitaly Kuznetsov wrote:
>> Jiang Jiasheng <[email protected]> writes:
>>
>> > The kvm_get_vcpu() will call for the array_index_nospec()
>> > with the value of atomic_read(&(v->kvm)->online_vcpus) as size,
>> > and the value of constant '0' as index.
>> > If the size is also '0', it will be unreasonabe
>> > that the index is no less than the size.
>> >
>>
>> Can this really happen?
>>
>> 'online_vcpus' is never decreased, it is increased with every
>> kvm_vm_ioctl_create_vcpu() call when a new vCPU is created and is set to
>> 0 when all vCPUs are destroyed (kvm_free_vcpus()).
>>
>> kvm_guest_time_update() takes a vcpu as a parameter, this means that at
>> least 1 vCPU is currently present so 'online_vcpus' just can't be zero.
>
> Agreed, but doing kvm_get_vcpu() is ugly and overkill.
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 86539c1686fa..cc1cb9a401cd 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -2969,7 +2969,7 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
> offsetof(struct compat_vcpu_info, time));
> if (vcpu->xen.vcpu_time_info_set)
> kvm_setup_pvclock_page(v, &vcpu->xen.vcpu_time_info_cache, 0);
> - if (v == kvm_get_vcpu(v->kvm, 0))
> + if (!kvm_vcpu_get_idx(v))

Do we really need to keep kvm_vcpu_get_idx() around though? It has only
3 users, all in arch/x86/kvm/hyperv.[ch], and the inline simpy returns
'vcpu->vcpu_idx'.

> kvm_hv_setup_tsc_page(v->kvm, &vcpu->hv_clock);
> return 0;
> }
>

--
Vitaly

2021-09-08 18:44:54

by Sean Christopherson

Subject: Re: [PATCH 4/4] KVM: X86: Potential 'index out of range' bug

On Mon, Sep 06, 2021, Vitaly Kuznetsov wrote:
> Sean Christopherson <[email protected]> writes:
>
> > On Fri, Sep 03, 2021, Vitaly Kuznetsov wrote:
> >> Jiang Jiasheng <[email protected]> writes:
> >>
> >> > The kvm_get_vcpu() will call for the array_index_nospec()
> >> > with the value of atomic_read(&(v->kvm)->online_vcpus) as size,
> >> > and the value of constant '0' as index.
> >> > If the size is also '0', it will be unreasonabe
> >> > that the index is no less than the size.
> >> >
> >>
> >> Can this really happen?
> >>
> >> 'online_vcpus' is never decreased, it is increased with every
> >> kvm_vm_ioctl_create_vcpu() call when a new vCPU is created and is set to
> >> 0 when all vCPUs are destroyed (kvm_free_vcpus()).
> >>
> >> kvm_guest_time_update() takes a vcpu as a parameter, this means that at
> >> least 1 vCPU is currently present so 'online_vcpus' just can't be zero.
> >
> > Agreed, but doing kvm_get_vcpu() is ugly and overkill.
> >
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 86539c1686fa..cc1cb9a401cd 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -2969,7 +2969,7 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
> > offsetof(struct compat_vcpu_info, time));
> > if (vcpu->xen.vcpu_time_info_set)
> > kvm_setup_pvclock_page(v, &vcpu->xen.vcpu_time_info_cache, 0);
> > - if (v == kvm_get_vcpu(v->kvm, 0))
> > + if (!kvm_vcpu_get_idx(v))
>
> Do we really need to keep kvm_vcpu_get_idx() around though? It has only
> 3 users, all in arch/x86/kvm/hyperv.[ch], and the inline simpy returns
> 'vcpu->vcpu_idx'.

Nope, looks like it's a holdover from before the introduction of vcpu_idx. I'll
send a small series to jettison the wrapper and make the above change.

Thanks!