Re: [PATCH RFC v1 2/2] i386: reorder kvm_put_sregs2() and kvm_put_nested_state() when vCPU is reset

[Vitaly Kuznetsov]

Maxim Levitsky <mlevitsk@redhat.com> writes:

> On Wed, 2022-08-10 at 16:00 +0200, Vitaly Kuznetsov wrote:
>> Setting nested state upon migration needs to happen after kvm_put_sregs2()
>> to e.g. have EFER.SVME set. This, however, doesn't work for vCPU reset:
>> when vCPU is in VMX root operation, certain CR bits are locked and
>> kvm_put_sregs2() may fail. As nested state is fully cleaned up upon
>> vCPU reset (kvm_arch_reset_vcpu() -> kvm_init_nested_state()), calling
>> kvm_put_nested_state() before kvm_put_sregs2() is OK, this will ensure
>> that vCPU is *not* in VMX root opertaion.
>> 
>> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
>> ---
>>  target/i386/kvm/kvm.c | 20 ++++++++++++++++++--
>>  1 file changed, 18 insertions(+), 2 deletions(-)
>> 
>> diff --git a/target/i386/kvm/kvm.c b/target/i386/kvm/kvm.c
>> index 4f8dacc1d4b5..73e3880fa57b 100644
>> --- a/target/i386/kvm/kvm.c
>> +++ b/target/i386/kvm/kvm.c
>> @@ -4529,18 +4529,34 @@ int kvm_arch_put_registers(CPUState *cpu, int level)
>>  
>>      assert(cpu_is_stopped(cpu) || qemu_cpu_is_self(cpu));
>>  
>> -    /* must be before kvm_put_nested_state so that EFER.SVME is set */
>> +    /*
>> +     * When resetting a vCPU, make sure to reset nested state first to
>> +     * e.g clear VMXON state and unlock certain CR4 bits.
>> +     */
>> +    if (level == KVM_PUT_RESET_STATE) {
>> +        ret = kvm_put_nested_state(x86_cpu);
>> +        if (ret < 0) {
>> +            return ret;
>> +        }
>
> I should have mentioned this, I actually already debugged the same issue while
> trying to reproduce the smm int window bug.
> 100% my fault.
>
> I also share the same feeling that this might be yet another 'whack a mole' 
> and
> break somewhere else, but overall it does make sense.

This certainly *is* a 'whack a mole' and I'm sure there are other cases
when one of calls in kvm_arch_put_registers() fails. We need to work on
what's missing so we can expose kvm_vcpu_reset() to VMMs.

>
>
> Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
>

Thanks!

-- 
Vitaly