Re: [PATCH v2 02/13] KVM: x86: Protect kvm_hv_msr_[get|set]_crash_data() from Spectre-v1/L1TF attacks

[Christian Borntraeger]

On 12.12.19 18:44, Jim Mattson wrote:
> On Thu, Dec 12, 2019 at 9:31 AM Christian Borntraeger
> <borntraeger@xxxxxxxxxx> wrote:
>>
>>
>>
>> On 11.12.19 21:47, Marios Pomonis wrote:
>>> This fixes Spectre-v1/L1TF vulnerabilities in kvm_hv_msr_get_crash_data()
>>> and kvm_hv_msr_set_crash_data().
>>> These functions contain index computations that use the
>>> (attacker-controlled) MSR number.
>>>
>>> Fixes: commit e7d9513b60e8 ("kvm/x86: added hyper-v crash msrs into kvm hyperv context")
>>>
>>> Signed-off-by: Nick Finco <nifi@xxxxxxxxxx>
>>> Signed-off-by: Marios Pomonis <pomonis@xxxxxxxxxx>
>>> Reviewed-by: Andrew Honig <ahonig@xxxxxxxxxx>
>>> Cc: stable@xxxxxxxxxxxxxxx
>>> ---
>>>  arch/x86/kvm/hyperv.c | 10 ++++++----
>>>  1 file changed, 6 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/arch/x86/kvm/hyperv.c b/arch/x86/kvm/hyperv.c
>>> index 23ff65504d7e..26408434b9bc 100644
>>> --- a/arch/x86/kvm/hyperv.c
>>> +++ b/arch/x86/kvm/hyperv.c
>>> @@ -809,11 +809,12 @@ static int kvm_hv_msr_get_crash_data(struct kvm_vcpu *vcpu,
>>>                                    u32 index, u64 *pdata)
>>>  {
>>>       struct kvm_hv *hv = &vcpu->kvm->arch.hyperv;
>>> +     size_t size = ARRAY_SIZE(hv->hv_crash_param);
>>>
>>> -     if (WARN_ON_ONCE(index >= ARRAY_SIZE(hv->hv_crash_param)))
>>> +     if (WARN_ON_ONCE(index >= size))
>>>               return -EINVAL;
>>
>> The fact that we do a WARN_ON_ONCE here, should actually tell that index is not
>> user controllable. Otherwise this would indicate the possibility to trigger a
>> kernel warning from a malicious user space. So
>> a: we do not need this change
>> or
>> b: we must also fix the WARN_ON_ONCE
> 
> That isn't quite true. The issue is *speculative* execution down this path.
> 
> The call site does constrain the *actual* value of index:
> 
> case HV_X64_MSR_CRASH_P0 ... HV_X64_MSR_CRASH_P4:
>         return kvm_hv_msr_get_crash_data(...);
> 
> However, it is possible to train the branch predictor to go down this
> path using valid indices and subsequently pass what would be an
> invalid index. The CPU will speculatively follow this path and may
> pull interesting data into the cache before it realizes its mistake
> and corrects.

Yes, you are right.