Re: [PATCH] KVM: LAPIC: Tune lapic_timer_advance_ns automatically

[Liran Alon]

> On 8 Oct 2018, at 13:59, Wanpeng Li <kernellwp@xxxxxxxxx> wrote:
> 
> On Mon, 8 Oct 2018 at 05:02, Liran Alon <liran.alon@xxxxxxxxxx> wrote:
>> 
>> 
>> 
>>> On 28 Sep 2018, at 9:12, Wanpeng Li <kernellwp@xxxxxxxxx> wrote:
>>> 
>>> From: Wanpeng Li <wanpengli@xxxxxxxxxxx>
>>> 
>>> In cloud environment, lapic_timer_advance_ns is needed to be tuned for every CPU
>>> generations, and every host kernel versions(the kvm-unit-tests/tscdeadline_latency.flat
>>> is 5700 cycles for upstream kernel and 9600 cycles for our 3.10 product kernel,
>>> both preemption_timer=N, Skylake server).
>>> 
>>> This patch adds the capability to automatically tune lapic_timer_advance_ns
>>> step by step, the initial value is 1000ns as d0659d946be05 (KVM: x86: add
>>> option to advance tscdeadline hrtimer expiration) recommended, it will be
>>> reduced when it is too early, and increased when it is too late. The guest_tsc
>>> and tsc_deadline are hard to equal, so we assume we are done when the delta
>>> is within a small scope e.g. 100 cycles. This patch reduces latency
>>> (kvm-unit-tests/tscdeadline_latency, busy waits, preemption_timer enabled)
>>> from ~2600 cyles to ~1200 cyles on our Skylake server.
>>> 
>>> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
>>> Cc: Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>
>>> Signed-off-by: Wanpeng Li <wanpengli@xxxxxxxxxxx>
>>> ---
>>> arch/x86/kvm/lapic.c | 7 +++++++
>>> arch/x86/kvm/x86.c   | 2 +-
>>> 2 files changed, 8 insertions(+), 1 deletion(-)
>>> 
>>> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
>>> index fbb0e6d..b756f12 100644
>>> --- a/arch/x86/kvm/lapic.c
>>> +++ b/arch/x86/kvm/lapic.c
>>> @@ -70,6 +70,8 @@
>>> #define APIC_BROADCAST                        0xFF
>>> #define X2APIC_BROADCAST              0xFFFFFFFFul
>>> 
>>> +static bool __read_mostly lapic_timer_advance_adjust_done = false;
>>> +
>>> static inline int apic_test_vector(int vec, void *bitmap)
>>> {
>>>      return test_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
>>> @@ -1492,6 +1494,11 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
>>>      if (guest_tsc < tsc_deadline)
>>>              __delay(min(tsc_deadline - guest_tsc,
>>>                      nsec_to_cycles(vcpu, lapic_timer_advance_ns)));
>>> +     if (!lapic_timer_advance_adjust_done) {
>>> +             lapic_timer_advance_ns += (s64)(guest_tsc - tsc_deadline) / 8;
>> 
>> I donât understand how this â/ 8â converts between guest TSC units to host nanoseconds.
> 
> Oh, I miss it. In addition, /8 here I mean adjust
> lapic_timer_advance_ns step by step. I can observe big fluctuated

If thatâs the case, I would also put the â8â as a #define to make it more clear of itâs purpose.

> value between early and late when running real guest os like linux
> instead of kvm-unit-tests. After more testing, I saw
> lapic_timer_advance_ns can be overflow since the delta between
> guest_tsc and tsc_deadline is too huge.
> 
>> 
>> I think that instead you should do something like:
>> s64 ns = (s64)(guest_tsc - tsc_deadline) * 1000000ULL;
>> do_div(ns, vcpu->arch.virtual_tsc_khz);
>> lapic_timer_advance_ns += ns;
>> 
>>> +             if (abs(guest_tsc - tsc_deadline) < 100)
>> 
>> I would put this â100â hard-coded value as some â#defineâ to make code more clear.
> 
> How about something like below:
> 
> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> index fbb0e6d..354eb13c 100644
> --- a/arch/x86/kvm/lapic.c
> +++ b/arch/x86/kvm/lapic.c
> @@ -70,6 +70,9 @@
> #define APIC_BROADCAST            0xFF
> #define X2APIC_BROADCAST        0xFFFFFFFFul
> 
> +static bool __read_mostly lapic_timer_advance_adjust_done = false;
> +#define LAPIC_TIMER_ADVANCE_ADJUST_DONE 100
> +
> static inline int apic_test_vector(int vec, void *bitmap)
> {
>     return test_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
> @@ -1472,7 +1475,7 @@ static bool lapic_timer_int_injected(struct
> kvm_vcpu *vcpu)
> void wait_lapic_expire(struct kvm_vcpu *vcpu)
> {
>     struct kvm_lapic *apic = vcpu->arch.apic;
> -    u64 guest_tsc, tsc_deadline;
> +    u64 guest_tsc, tsc_deadline, ns;
> 
>     if (!lapic_in_kernel(vcpu))
>         return;
> @@ -1492,6 +1495,19 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
>     if (guest_tsc < tsc_deadline)
>         __delay(min(tsc_deadline - guest_tsc,
>             nsec_to_cycles(vcpu, lapic_timer_advance_ns)));
> +    if (!lapic_timer_advance_adjust_done) {
> +        if (guest_tsc < tsc_deadline) {
> +            ns = (tsc_deadline - guest_tsc) * 1000000ULL;
> +            do_div(ns, vcpu->arch.virtual_tsc_khz);
> +            lapic_timer_advance_ns -= min((unsigned int)ns,
> lapic_timer_advance_ns / 8);
> +        } else {
> +            ns = (guest_tsc - tsc_deadline) * 1000000ULL;
> +            do_div(ns, vcpu->arch.virtual_tsc_khz);
> +            lapic_timer_advance_ns += min((unsigned int)ns,
> lapic_timer_advance_ns / 8);
> +        }
> +        if (ns < LAPIC_TIMER_ADVANCE_ADJUST_DONE)

Didnât you meant to compare here that abs(guest_tsc - tsc_deadline) < LAPIC_TIMER_ADVANCE_ADJUST_DONE?
This is also a good example on why I would also rename LAPIC_TIMER_ADVANCE_ADJUST_DONE to something
which indicates it represents a number in guest TSC units.

> +            lapic_timer_advance_adjust_done = true;
> +    }
> }
> 
> static void start_sw_tscdeadline(struct kvm_lapic *apic)
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index ca71773..1f3f955 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -136,7 +136,7 @@ static u32 __read_mostly tsc_tolerance_ppm = 250;
> module_param(tsc_tolerance_ppm, uint, S_IRUGO | S_IWUSR);
> 
> /* lapic timer advance (tscdeadline mode only) in nanoseconds */
> -unsigned int __read_mostly lapic_timer_advance_ns = 0;
> +unsigned int __read_mostly lapic_timer_advance_ns = 1000;
> module_param(lapic_timer_advance_ns, uint, S_IRUGO | S_IWUSR);
> EXPORT_SYMBOL_GPL(lapic_timer_advance_ns);