Re: [PATCH v4 0/8] KVM: x86: hyperv: PV TLB flush for Windows guests

[Vitaly Kuznetsov]

Wanpeng Li <kernellwp@xxxxxxxxx> writes:

> 2018-05-18 20:42 GMT+08:00 Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>:
>> Wanpeng Li <kernellwp@xxxxxxxxx> writes:
>>
>>> 2018-05-18 19:19 GMT+08:00 Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>:
>>>> Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> writes:
>>>>
>>>>> Wanpeng Li <kernellwp@xxxxxxxxx> writes:
>>>>>
>>>>>> Hi Vitaly,
>>>>>> 2018-05-16 23:21 GMT+08:00 Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>:
>>>>>>> Changes since v3 [Radim Krcmar]:
>>>>>>> - PATCH2 fixing 'HV_GENERIC_SET_SPARCE_4K' typo added.
>>>>>>> - PATCH5 introducing kvm_make_vcpus_request_mask() API added.
>>>>>>> - Fix undefined behavior for hv->vp_index >= 64.
>>>>>>> - Merge kvm_hv_flush_tlb() and kvm_hv_flush_tlb_ex()
>>>>>>> - For -ex case preload all banks with a single kvm_read_guest().
>>>>>>>
>>>>>>> Description:
>>>>>>>
>>>>>>> This is both a new feature and a bugfix.
>>>>>>>
>>>>>>> Bugfix description:
>>>>>>>
>>>>>>> It was found that Windows 2016 guests on KVM crash when they have > 64
>>>>>>> vCPUs, non-flat topology (>1 core/thread per socket; in case it has >64
>>>>>>> sockets Windows just ignores vCPUs above 64) and Hyper-V enlightenments
>>>>>>
>>>>>> We try the below command line, the Windows 2016 guest successfully to
>>>>>> login and there are 80 vCPUs can be observed in the guest w/o the
>>>>>> patchset, why you mentioned the crash and ignore?
>>>>>>
>>>>>> /usr/local/bin/qemu-system-x86_64 -machine  pc-i440fx-rhel7.3.0  -m
>>>>>> 8192 -smp 80,sockets=2,cores=40,threads=1  -device
>>>>>> ide-drive,bus=ide.0,drive=test -drive
>>>>>> id=test,if=none,file=/instanceimage/359b18ab-05bb-460d-9b53-89505bca68ed/359b18ab-05bb-460d-9b53-89505bca68ed_vda_1.qcow2
>>>>>>  -net nic,model=virtio -net user -monitor stdio -usb -usbdevice tablet
>>>>>> --enable-kvm --cpu host -vnc 0.0.0.0:2
>>>>>
>>>>> Crash happens when you manifest yourself as Hyper-V, you can do this by
>>>>> adding any 'hv-*' feature (e.g. try '-cpu host,hv_vpindex').
>>>>
>>>> Oh, and the 'ignore' happens when you pass more than 64 sockets
>>>> (somthing like "-smp 128,sockets=128,cores=1,threads=1") -- and this
>>>> happens regardless of Hyper-V enlightenments. But I guess it's just
>>>> because Windows doesn't support more than 64 sockets.
>>>
>>> Is there an option in the guest to avoid to check pvtlb support in hyperv?
>>>
>>
>> You mean to tell Windows to not use PV TLB flush when it's available? I
>> have no idea. My guess would be that it's left up to the hypervisor: if
>> the feature is available Windows will use it.
>
> I mean a way to work around Windows guest crash since there is no PV
> TLB flush enabled in product environment currently.
>

Unfortunately I don't know of such option, all real Hyper-V servers
support it so I think nodoby ever tested how Windows behaves without
it. I did and oops, it crashes...

-- 
  Vitaly