Re: [Qemu-devel] [PATCH 4/8] i386/kvm: implement 'hv-all' pass-through mode

[Vitaly Kuznetsov]

Roman Kagan <address@hidden> writes:

> On Fri, Mar 29, 2019 at 03:18:28PM +0100, Vitaly Kuznetsov wrote:
>> In many case we just want to give Windows guests all currently supported
>> Hyper-V enlightenments and that's where this new mode may come handy. We
>> pass through what was returned by KVM_GET_SUPPORTED_HV_CPUID.
>
> The only one out of those "many cases" I can think of is when you've
> developed a new hyperv feature in the kernel and you want to test it
> with a version of QEMU that's not aware of it.  Are there any others?
>

I can recall the following case I had: benchmark Windows guest
performance with different kernels like try to get the best number. As
these kernels were supporting different set of hv-* enlightenments I had
to do non-trivial work to figure out what's supported and adjust QEMU
command line accordingly.

Would've been much easier with 'hv-all'

>> 
>> hv_cpuid_check_and_set() is modified to also set cpu->hyperv_* flags as
>> we may want to check them later (and we actually do for hv_runtime,
>> hv_synic,...).
>> 
>> 'hv-all' is a development only feature, a migration blocker is added to
>> prevent issues while migrating between hosts with different feature sets.
>> 
>> Signed-off-by: Vitaly Kuznetsov <address@hidden>
>> ---
>>  docs/hyperv.txt   |  10 ++++
>>  target/i386/cpu.c |   1 +
>>  target/i386/cpu.h |   1 +
>>  target/i386/kvm.c | 148 +++++++++++++++++++++++++++++++++++++---------
>>  4 files changed, 132 insertions(+), 28 deletions(-)
>> 
>> diff --git a/docs/hyperv.txt b/docs/hyperv.txt
>> index 397f2517b8..d1299aba81 100644
>> --- a/docs/hyperv.txt
>> +++ b/docs/hyperv.txt
>> @@ -174,6 +174,16 @@ without the feature to find out if enabling it is 
>> beneficial.
>>  Requires: hv-vapic
>>  
>>  
>> +4. Development features
>> +========================
>> +In some cases (e.g. during development) it may make sense to use QEMU in
>> +'pass-through' mode and give Windows guests all enlightenments currently
>> +supported by KVM. This pass-through mode is enabled by "hv-all" CPU flag.
>> +Note: enabling this flag effectively prevents migration as supported 
>> features
>> +may differ between target and destination.
>
> I find 'hv-passthrough' a more adequate name for this.

Sure, will adjust.

>
>> +Note: "hv-all" doesn't include 'hv-evmcs', it needs to be enabled 
>> explicitly.
>
> This is extremely confusing, when some features are more equal than
> others.  I think it'd make more sense instead to support filtering out
> some features, like in "hv-passthrough,hv-evmcs=off".

hv-evmcs is probably the only enlightenment which is not an obvious
'win': when enabled, some features (e.g. posted interrupts) are getting
disabled. But as 'hv-all' is now a developer-only feature I see no
problem with enabling evmcs too.

-- 
Vitaly