Wanpeng Li <kernellwp@xxxxxxxxx> wrote:
> 2017-07-20 6:53 GMT+08:00 Nadav Amit <nadav.amit@xxxxxxxxx>:
>> Wanpeng Li <kernellwp@xxxxxxxxx> wrote:
>>
>>> 2017-07-20 0:25 GMT+08:00 Nadav Amit <nadav.amit@xxxxxxxxx>:
>>>> Radim Krčmář <rkrcmar@xxxxxxxxxx> wrote:
>>>>
>>>>> 2017-07-19 08:14-0700, Nadav Amit:
>>>>>> Radim Krčmář <rkrcmar@xxxxxxxxxx> wrote:
>>>>>>> @@ -2363,6 +2368,8 @@ static unsigned long vmx_get_rflags(struct kvm_vcpu *vcpu)
>>>>>>>
>>>>>>> static void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
>>>>>>> {
>>>>>>> + unsigned long old_rflags = to_vmx(vcpu)->rflags;
>>>>>>
>>>>>> It assumes rflags was decached from the VMCS before. Probably it is true, but…
>>>>>
>>>>> Right, it's better to use accessors everywhere, thanks.
>>>>> The line should read:
>>>>>
>>>>> + unsigned long old_rflags = vmx_get_rflags(vcpu);
>>>>>
>>>>> ---8<---
>>>>> This can be reproduced by EPT=1, unrestricted_guest=N, emulate_invalid_state=Y
>>>>> or EPT=0, the trace of kvm-unit-tests/taskswitch2.flat is like below, it
>>>>> tries to emulate invalid guest state task-switch:
>>>>>
>>>>> kvm_exit: reason TASK_SWITCH rip 0x0 info 40000058 0
>>>>> kvm_emulate_insn: 42000:0:0f 0b (0x2)
>>>>> kvm_emulate_insn: 42000:0:0f 0b (0x2) failed
>>>>> kvm_inj_exception: #UD (0x0)
>>>>> kvm_entry: vcpu 0
>>>>> kvm_exit: reason TASK_SWITCH rip 0x0 info 40000058 0
>>>>> kvm_emulate_insn: 42000:0:0f 0b (0x2)
>>>>> kvm_emulate_insn: 42000:0:0f 0b (0x2) failed
>>>>> kvm_inj_exception: #UD (0x0)
>>>>>
>>>>> It appears that the task-switch emulation updates rflags (and vm86 flag)
>>>>> only after the segments are loaded, causing vmx->emulation_required to
>>>>> be set, when in fact invalid guest state emulation is not needed.
>>>>>
>>>>> This patch fixes it by updating vmx->emulation_required after the rflags
>>>>> (and vm86 flag) is updated.
>>>>>
>>>>> Suggested-by: Nadav Amit <nadav.amit@xxxxxxxxx>
>>>>> Signed-off-by: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
>>>>> [Wanpeng wrote the commit message with initial patch and Radim moved the
>>>>> update to vmx_set_rflags and added Paolo's suggestion for the check.]
>>>>> Signed-off-by: Radim Krčmář <rkrcmar@xxxxxxxxxx>
>>>>> ---
>>>>> arch/x86/kvm/vmx.c | 15 ++++++++++-----
>>>>> 1 file changed, 10 insertions(+), 5 deletions(-)
>>>>>
>>>>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
>>>>> index 84e62acf2dd8..a776aea0043a 100644
>>>>> --- a/arch/x86/kvm/vmx.c
>>>>> +++ b/arch/x86/kvm/vmx.c
>>>>> @@ -2326,6 +2326,11 @@ static void vmx_vcpu_put(struct kvm_vcpu *vcpu)
>>>>> __vmx_load_host_state(to_vmx(vcpu));
>>>>> }
>>>>>
>>>>> +static bool emulation_required(struct kvm_vcpu *vcpu)
>>>>> +{
>>>>> + return emulate_invalid_guest_state && !guest_state_valid(vcpu);
>>>>> +}
>>>>> +
>>>>> static void vmx_decache_cr0_guest_bits(struct kvm_vcpu *vcpu);
>>>>>
>>>>> /*
>>>>> @@ -2363,6 +2368,8 @@ static unsigned long vmx_get_rflags(struct kvm_vcpu *vcpu)
>>>>>
>>>>> static void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
>>>>> {
>>>>> + unsigned long old_rflags = vmx_get_rflags(vcpu);
>>>>> +
>>>>> __set_bit(VCPU_EXREG_RFLAGS, (ulong *)&vcpu->arch.regs_avail);
>>>>> to_vmx(vcpu)->rflags = rflags;
>>>>> if (to_vmx(vcpu)->rmode.vm86_active) {
>>>>> @@ -2370,6 +2377,9 @@ static void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
>>>>> rflags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
>>>>> }
>>>>> vmcs_writel(GUEST_RFLAGS, rflags);
>>>>> +
>>>>> + if ((old_rflags ^ rflags) & X86_EFLAGS_VM)
>>>>> + to_vmx(vcpu)->emulation_required = emulation_required(vcpu);
>>>>
>>>> Sorry for not pointing it before, but here you compare the old_rflags with
>>>> the new rflags but after you already “massaged” it. So the value you compare
>>>> with is not what the guest “sees”.
>>>
>>> So you mean we should use unsigned long old_rflags =
>>> vmcs_readl(GUEST_RFLAGS); right?
>>
>> No. The problem is not with old_rflags now, but with rflags. If vm86_active,
>> then rflags is changed and you don’t compare the guest-visible rflags
>> anymore.
>
> Ah, I see. So we should compare the old_flags with the
> rmode->save_rflags(guest-visible rflags) instead of the rflags (shadow
> rflags), right?
Not exactly, since rmode->save_rflags are invalid if !vm86_active. Instead,
I think you should have a save_rflags variable on the stack that would hold
the rflags before “massaging” and use it instead.
