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

[Vadim Rozenfeld <vrozenfe@redhat.com>]

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.

> It seems that get_kernel_ns() + kvm->arch.kvmclock_offset is exactly the vServer uptime
> in nanoseconds which starts counting at 0.
> 
> Peter
> 
>