On 22/11/17 11:31, Wanpeng Li wrote:
> 2017-11-22 17:07 GMT+08:00 Liran Alon <[email protected]>:
>>
>>
>> On 22/11/17 10:45, Liran Alon wrote:
>>>
>>>
>>>
>>> On 22/11/17 09:56, Wanpeng Li wrote:
>>>>
>>>> From: Wanpeng Li <[email protected]>
>>>>
>>>> Reported by syzkaller:
>>>>
>>>> ------------[ cut here ]------------
>>>> WARNING: CPU: 5 PID: 2939 at arch/x86/kvm/vmx.c:3844
>>>> free_loaded_vmcs+0x77/0x80 [kvm_intel]
>>>> CPU: 5 PID: 2939 Comm: repro Not tainted 4.14.0+ #26
>>>> RIP: 0010:free_loaded_vmcs+0x77/0x80 [kvm_intel]
>>>> Call Trace:
>>>> vmx_free_vcpu+0xda/0x130 [kvm_intel]
>>>> kvm_arch_destroy_vm+0x192/0x290 [kvm]
>>>> kvm_put_kvm+0x262/0x560 [kvm]
>>>> kvm_vm_release+0x2c/0x30 [kvm]
>>>> __fput+0x190/0x370
>>>> task_work_run+0xa1/0xd0
>>>> do_exit+0x4d2/0x13e0
>>>> do_group_exit+0x89/0x140
>>>> get_signal+0x318/0xb80
>>>> do_signal+0x8c/0xb40
>>>> exit_to_usermode_loop+0xe4/0x140
>>>> syscall_return_slowpath+0x206/0x230
>>>> entry_SYSCALL_64_fastpath+0x98/0x9a
>>>>
>>>> The syzkaller testcase will execute VMXON/VMLAUCH instructions, so the
>>>> vmx->nested stuff is populated, it will also issue KVM_SMI ioctl.
>>>> However,
>>>> the testcase is just a simple c program and not be lauched by something
>>>> like seabios which implements smi_handler. Commit 05cade71cf (KVM: nSVM:
>>>> fix SMI injection in guest mode) gets out of guest mode and set
>>>> nested.vmxon
>>>> to false for the duration of SMM according to SDM 34.14.1 "leave VMX
>>>> operation" upon entering SMM. We can't alloc/free the vmx->nested stuff
>>>> each time when entering/exiting SMM since it will induce more
>>>> overhead. So
>>>> the function vmx_pre_enter_smm() marks nested.vmxon false even if
>>>> vmx->nested
>>>> stuff is still populated. What it expected is em_rsm() can mark
>>>> nested.vmxon
>>>> to be true again. However, the smi_handler/rsm will not execute since
>>>> there
>>>> is no something like seabios in this scenario. The function free_nested()
>>>> fails to free the vmx->nested stuff since the vmx->nested.vmxon is false
>>>> which results in the above warning.
>>>>
>>>> This patch fixes it by also considering the no SMI handler case, luckily
>>>> vmx->nested.smm.vmxon is marked according to the value of
>>>> vmx->nested.vmxon
>>>> in vmx_pre_enter_smm(), we can take advantage of it and free vmx->nested
>>>> stuff when L1 goes down.
>>>>
>>>> Reported-by: Dmitry Vyukov <[email protected]>
>>>> Cc: Paolo Bonzini <[email protected]>
>>>> Cc: Radim Krčmář <[email protected]>
>>>> Cc: Dmitry Vyukov <[email protected]>
>>>> Fixes: 05cade71cf (KVM: nSVM: fix SMI injection in guest mode)
>>>> Signed-off-by: Wanpeng Li <[email protected]>
>>>> ---
>>>> arch/x86/kvm/vmx.c | 2 +-
>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
>>>> index dccc0f7..ed22425 100644
>>>> --- a/arch/x86/kvm/vmx.c
>>>> +++ b/arch/x86/kvm/vmx.c
>>>> @@ -7372,7 +7372,7 @@ static inline void nested_release_vmcs12(struct
>>>> vcpu_vmx *vmx)
>>>> */
>>>> static void free_nested(struct vcpu_vmx *vmx)
>>>> {
>>>> - if (!vmx->nested.vmxon)
>>>> + if (!vmx->nested.vmxon && !vmx->nested.smm.vmxon)
>>>> return;
>>>>
>>>> vmx->nested.vmxon = false;
>>>>
>>> Funny bug. Great analysis.
>>> Reviewed-by: Liran Alon <[email protected]>
>>
>> Actually, I would add one more thing to patch:
>> I think we should also set "vmx->nested.smm.vmxon = false;" after
>> "vmx->nested.vmxon = false;" to correctlyhandle the case VMXOFF is executed
>> from SMI handler. Otherwise, when SMI handler executes RSM, we will reach
>> vmx_pre_leave_smm() which will set again "vmx->nested.vmxon = true;" which I
>> think shouldn't happen.
>
> I didn't see a real scenario for this.
Actually I later saw that handle_vmoff() calls
nested_vmx_check_permission() which indeed won't allow to continue
executing if running from SMI because vmx->nested.vmxon=false; and
therefore this will raise a #UD. So you are right. :)
>
> Regards,
> Wanpeng Li
>
From 1584758165742760813@xxx Wed Nov 22 09:32:40 +0000 2017
X-GM-THRID: 1584752194833300147
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread