Re: [PATCH v3 7/7] x86/kvm: use Enlightened VMCS when running on Hyper-V

[Paolo Bonzini]

On 15/03/2018 10:56, Vitaly Kuznetsov wrote:
>> +	EVMCS1_FIELD(VM_EXIT_MSR_STORE_ADDR, vm_exit_msr_store_addr,
>> +		     HV_VMX_ENLIGHTENED_CLEAN_FIELD_ALL),
>> +	EVMCS1_FIELD(VM_EXIT_MSR_LOAD_ADDR, vm_exit_msr_load_addr,
>> +		     HV_VMX_ENLIGHTENED_CLEAN_FIELD_ALL),
>> +	EVMCS1_FIELD(VM_ENTRY_MSR_LOAD_ADDR, vm_entry_msr_load_addr,
>> +		     HV_VMX_ENLIGHTENED_CLEAN_FIELD_ALL),
>> +	EVMCS1_FIELD(VM_EXIT_MSR_STORE_COUNT, vm_exit_msr_store_count,
>> +		     HV_VMX_ENLIGHTENED_CLEAN_FIELD_ALL),
>> +	EVMCS1_FIELD(VM_EXIT_MSR_LOAD_COUNT, vm_exit_msr_load_count,
>> +		     HV_VMX_ENLIGHTENED_CLEAN_FIELD_ALL),
>> +	EVMCS1_FIELD(VM_ENTRY_MSR_LOAD_COUNT, vm_entry_msr_load_count,
>> +		     HV_VMX_ENLIGHTENED_CLEAN_FIELD_ALL),

Hmm, actually these six are used.  I guess
HV_VMX_ENLIGHTENED_CLEAN_FIELD_ALL is the best we can do, apart from
asking Microsoft to fix the spec.


>>> +{
>>> +	*pin_based_exec_ctrl &= ~PIN_BASED_POSTED_INTR;> +	*cpu_based_2nd_exec_ctrl &= ~SECONDARY_EXEC_VIRTUAL_INTR_DELIVERY;
>>> +	*cpu_based_2nd_exec_ctrl &= ~SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES;
>>> +	*cpu_based_2nd_exec_ctrl &= ~SECONDARY_EXEC_APIC_REGISTER_VIRT;
>>> +	*cpu_based_2nd_exec_ctrl &= ~SECONDARY_EXEC_ENABLE_PML;
>>> +	*cpu_based_2nd_exec_ctrl &= ~SECONDARY_EXEC_ENABLE_VMFUNC;
>>> +	*cpu_based_2nd_exec_ctrl &= ~SECONDARY_EXEC_SHADOW_VMCS;
>>> +	*cpu_based_2nd_exec_ctrl &= ~SECONDARY_EXEC_TSC_SCALING;
>>> +	*cpu_based_2nd_exec_ctrl &= ~SECONDARY_EXEC_PAUSE_LOOP_EXITING;
>>> +	*pin_based_exec_ctrl &= ~PIN_BASED_VMX_PREEMPTION_TIMER;
>> How can these be set?
>>
> They can not if Hyper-V behaves but Radim didn't want to trust it -- so
> the suggestion was to forcefully disable unsupported controls.

Yeah, it's good to have, especially if placed before we start using the
values that are read.

>> This is wrong, we're reading the VMCS so the values must already be
>> sanitized (and if not, that's the bug and we want dump_vmcs to print the
>> "wrong" values).
> 
> The problem is that we vmcs_read these fields later in the function and
> this will now WARN(). Initally, there was no WARN() for non-existent
> fields so this could work (we would just print zeroes for unsupported
> fields). Maybe, additional WARN_ON() is not a big deal here.

If you WARN(), isn't it because the secondary_exec_control had a bad
value to begin with?  As you say, the controls should never be set.

Thanks,

Paolo