Re: [RFC PATCH v3 1/2] add support for Hyper-V reference time counter

[Peter Lieven <pl@dlhnet.de>]

On 08.01.2014 11:44, Vadim Rozenfeld wrote:
> On Wed, 2014-01-08 at 11:15 +0100, Peter Lieven wrote:
>> On 08.01.2014 10:40, Vadim Rozenfeld wrote:
>>> On Tue, 2014-01-07 at 18:52 +0100, Peter Lieven wrote:
>>>> Am 07.01.2014 10:36, schrieb Vadim Rozenfeld:
>>>>> On Thu, 2014-01-02 at 17:52 +0100, Peter Lieven wrote:
>>>>>> Am 11.12.2013 19:59, schrieb Marcelo Tosatti:
>>>>>>> On Wed, Dec 11, 2013 at 04:53:05PM -0200, Marcelo Tosatti wrote:
>>>>>>>> On Sun, Dec 08, 2013 at 10:33:38PM +1100, Vadim Rozenfeld wrote:
>>>>>>>>> Signed-off: Peter Lieven <pl@dlh.net>
>>>>>>>>> Signed-off: Gleb Natapov <gleb@redhat.com>
>>>>>>>>> Signed-off: Vadim Rozenfeld <vrozenfe@redhat.com>
>>>>>>>>>
>>>>>>>>> v1 -> v2
>>>>>>>>> 1. mark TSC page dirty as suggested by
>>>>>>>>>       Eric Northup <digitaleric@google.com> and Gleb
>>>>>>>>> 2. disable local irq when calling get_kernel_ns,
>>>>>>>>>       as it was done by Peter Lieven <pl@dlhnet.de>
>>>>>>>>> 3. move check for TSC page enable from second patch
>>>>>>>>>       to this one.
>>>>>>>>>
>>>>>>>>> ---
>>>>>>>>>    arch/x86/include/asm/kvm_host.h    |  2 ++
>>>>>>>>>    arch/x86/include/uapi/asm/hyperv.h | 13 +++++++++++++
>>>>>>>>>    arch/x86/kvm/x86.c                 | 39 +++++++++++++++++++++++++++++++++++++-
>>>>>>>>>    include/uapi/linux/kvm.h           |  1 +
>>>>>>>>>    4 files changed, 54 insertions(+), 1 deletion(-)
>>>>>>>>>
>>>>>>>>> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
>>>>>>>>> index ae5d783..2fd0753 100644
>>>>>>>>> --- a/arch/x86/include/asm/kvm_host.h
>>>>>>>>> +++ b/arch/x86/include/asm/kvm_host.h
>>>>>>>>> @@ -605,6 +605,8 @@ struct kvm_arch {
>>>>>>>>>    	/* fields used by HYPER-V emulation */
>>>>>>>>>    	u64 hv_guest_os_id;
>>>>>>>>>    	u64 hv_hypercall;
>>>>>>>>> +	u64 hv_ref_count;
>>>>>>>>> +	u64 hv_tsc_page;
>>>>>>>>>
>>>>>>>>>    	#ifdef CONFIG_KVM_MMU_AUDIT
>>>>>>>>>    	int audit_point;
>>>>>>>>> diff --git a/arch/x86/include/uapi/asm/hyperv.h b/arch/x86/include/uapi/asm/hyperv.h
>>>>>>>>> index b8f1c01..462efe7 100644
>>>>>>>>> --- a/arch/x86/include/uapi/asm/hyperv.h
>>>>>>>>> +++ b/arch/x86/include/uapi/asm/hyperv.h
>>>>>>>>> @@ -28,6 +28,9 @@
>>>>>>>>>    /* Partition Reference Counter (HV_X64_MSR_TIME_REF_COUNT) available*/
>>>>>>>>>    #define HV_X64_MSR_TIME_REF_COUNT_AVAILABLE	(1 << 1)
>>>>>>>>>
>>>>>>>>> +/* A partition's reference time stamp counter (TSC) page */
>>>>>>>>> +#define HV_X64_MSR_REFERENCE_TSC		0x40000021
>>>>>>>>> +
>>>>>>>>>    /*
>>>>>>>>>     * There is a single feature flag that signifies the presence of the MSR
>>>>>>>>>     * that can be used to retrieve both the local APIC Timer frequency as
>>>>>>>>> @@ -198,6 +201,9 @@
>>>>>>>>>    #define HV_X64_MSR_APIC_ASSIST_PAGE_ADDRESS_MASK	\
>>>>>>>>>    		(~((1ull << HV_X64_MSR_APIC_ASSIST_PAGE_ADDRESS_SHIFT) - 1))
>>>>>>>>>
>>>>>>>>> +#define HV_X64_MSR_TSC_REFERENCE_ENABLE		0x00000001
>>>>>>>>> +#define HV_X64_MSR_TSC_REFERENCE_ADDRESS_SHIFT	12
>>>>>>>>> +
>>>>>>>>>    #define HV_PROCESSOR_POWER_STATE_C0		0
>>>>>>>>>    #define HV_PROCESSOR_POWER_STATE_C1		1
>>>>>>>>>    #define HV_PROCESSOR_POWER_STATE_C2		2
>>>>>>>>> @@ -210,4 +216,11 @@
>>>>>>>>>    #define HV_STATUS_INVALID_ALIGNMENT		4
>>>>>>>>>    #define HV_STATUS_INSUFFICIENT_BUFFERS		19
>>>>>>>>>
>>>>>>>>> +typedef struct _HV_REFERENCE_TSC_PAGE {
>>>>>>>>> +	__u32 tsc_sequence;
>>>>>>>>> +	__u32 res1;
>>>>>>>>> +	__u64 tsc_scale;
>>>>>>>>> +	__s64 tsc_offset;
>>>>>>>>> +} HV_REFERENCE_TSC_PAGE, *PHV_REFERENCE_TSC_PAGE;
>>>>>>>>> +
>>>>>>>>>    #endif
>>>>>>>>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>>>>>>>>> index 21ef1ba..5e4e495a 100644
>>>>>>>>> --- a/arch/x86/kvm/x86.c
>>>>>>>>> +++ b/arch/x86/kvm/x86.c
>>>>>>>>> @@ -840,7 +840,7 @@ EXPORT_SYMBOL_GPL(kvm_rdpmc);
>>>>>>>>>    static u32 msrs_to_save[] = {
>>>>>>>>>    	MSR_KVM_SYSTEM_TIME, MSR_KVM_WALL_CLOCK,
>>>>>>>>>    	MSR_KVM_SYSTEM_TIME_NEW, MSR_KVM_WALL_CLOCK_NEW,
>>>>>>>>> -	HV_X64_MSR_GUEST_OS_ID, HV_X64_MSR_HYPERCALL,
>>>>>>>>> +	HV_X64_MSR_GUEST_OS_ID, HV_X64_MSR_HYPERCALL, HV_X64_MSR_TIME_REF_COUNT,
>>>>>>>>>    	HV_X64_MSR_APIC_ASSIST_PAGE, MSR_KVM_ASYNC_PF_EN, MSR_KVM_STEAL_TIME,
>>>>>>>>>    	MSR_KVM_PV_EOI_EN,
>>>>>>>>>    	MSR_IA32_SYSENTER_CS, MSR_IA32_SYSENTER_ESP, MSR_IA32_SYSENTER_EIP,
>>>>>>>>> @@ -1826,6 +1826,8 @@ static bool kvm_hv_msr_partition_wide(u32 msr)
>>>>>>>>>    	switch (msr) {
>>>>>>>>>    	case HV_X64_MSR_GUEST_OS_ID:
>>>>>>>>>    	case HV_X64_MSR_HYPERCALL:
>>>>>>>>> +	case HV_X64_MSR_REFERENCE_TSC:
>>>>>>>>> +	case HV_X64_MSR_TIME_REF_COUNT:
>>>>>>>>>    		r = true;
>>>>>>>>>    		break;
>>>>>>>>>    	}
>>>>>>>>> @@ -1865,6 +1867,29 @@ static int set_msr_hyperv_pw(struct kvm_vcpu *vcpu, u32 msr, u64 data)
>>>>>>>>>    		if (__copy_to_user((void __user *)addr, instructions, 4))
>>>>>>>>>    			return 1;
>>>>>>>>>    		kvm->arch.hv_hypercall = data;
>>>>>>>>> +		local_irq_disable();
>>>>>>>>> +		kvm->arch.hv_ref_count = get_kernel_ns() + kvm->arch.kvmclock_offset;
>>>>>>>>> +		local_irq_enable();
>>>>>>>>
>>>>>>>> Where does the docs say that HV_X64_MSR_HYPERCALL is the where the clock
>>>>>>>> starts counting?
>>>>>>>>
>>>>>>>> No need to store kvmclock_offset in hv_ref_count? (moreover
>>>>>>>> the name is weird, better name would be "hv_ref_start_time".
>>>>>>>
>>>>>>> Just add kvmclock_offset when reading the values (otherwise you have a
>>>>>>> "stale copy" of kvmclock_offset in hv_ref_count).
>>>>>>>
>>>>>>
>>>>>> After some experiments I think we do no need kvm->arch.hv_ref_count at all.
>>>>>>
>>>>>> I was debugging some weird clockjump issues and I think the problem is that after live migration
>>>>>> kvm->arch.hv_ref_count is initialized to 0. Depending on the uptime of the vServer when the
>>>>>> hypercall was set up this can lead to series jumps.
>>>>>>
>>>>>> So I would suggest to completely drop kvm->arch.hv_ref_count.
>>>>>>
>>>>>> And use simply this in get_msr_hyperv_pw().
>>>>>>
>>>>>>           case HV_X64_MSR_TIME_REF_COUNT: {
>>>>>>                   data = div_u64(get_kernel_ns() + kvm->arch.kvmclock_offset, 100);
>>>>>>                   break;
>>>>>>           }
>>>>>>
>>>>>
>>>>> Agreed. It should work as long as we rely on kvmclock_offset.
>>>>> Vadim.
>>>>
>>>> I think we can rely on kvmclock_offset. Had you had a chance to do further testing
>>>> during the weekend?
>>>
>>> Yes, and still testing.
>>>
>>> There is still some inconsistency in the value returned by
>>> QueryPerformanceCounter during migration.
>>
>> With my proposal?
>>
>
> Right.
> Please see below QTC vs. NTP time stamp
> before and after migration:
>
> C:\timedrift\objchk_win7_x86\i386>timedrift.exe
> quantum = 15
> QPC     NTP     Delta
> 1311    1307    4
> 2621    2618    3
> 3932    3931    1
> .......
> .......
> 38018   38014   4
> Unable to wait for NTP reply from the SNTP server, GetLastError returns
> 7279088
> 142210  40113   102097
> 143458  41357   102101
> .......
> .......
> 156500  54615   101885
>
> NTP time changed from  38014 to 40113 ((40113 - 38014) / 100 = 21 mSec)
> while QPC jumped from  38018 to 142210 ((142210 - 38014) / 100 = 1042
> mSec)

How long was the downtime when you migrated the vServer?

I have done similar tests and have not seen this...
Can you show the relevant parts of your patch. Remember I use the REFERENCE_COUTNER not iTSC.

Peter