Re: [PATCH v3 1/2] KVM: LAPIC: Optimize timer latency consider world switch time

From: Radim KrÄmÃÅ
Date: Wed Jun 12 2019 - 15:27:27 EST


2019-06-12 08:14-0700, Sean Christopherson:
> On Wed, Jun 12, 2019 at 05:40:18PM +0800, Wanpeng Li wrote:
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > @@ -145,6 +145,12 @@ module_param(tsc_tolerance_ppm, uint, S_IRUGO | S_IWUSR);
> > static int __read_mostly lapic_timer_advance_ns = -1;
> > module_param(lapic_timer_advance_ns, int, S_IRUGO | S_IWUSR);
> >
> > +/*
> > + * lapic timer vmentry advance (tscdeadline mode only) in nanoseconds.
> > + */
> > +u32 __read_mostly vmentry_advance_ns = 300;
>
> Enabling this by default makes me nervous, e.g. nothing guarantees that
> future versions of KVM and/or CPUs will continue to have 300ns of overhead
> between wait_lapic_expire() and VM-Enter.
>
> If we want it enabled by default so that it gets tested, the default
> value should be extremely conservative, e.g. set the default to a small
> percentage (25%?) of the latency of VM-Enter itself on modern CPUs,
> VM-Enter latency being the min between VMLAUNCH and VMLOAD+VMRUN+VMSAVE.

I share the sentiment. We definitely must not enter the guest before
the deadline has expired and CPUs are approaching 5 GHz (in turbo), so
300 ns would be too much even today.

I wrote a simple testcase for rough timing and there are 267 cycles
(111 ns @ 2.4 GHz) between doing rdtsc() right after
kvm_wait_lapic_expire() [1] and doing rdtsc() in the guest as soon as
possible (see the attached kvm-unit-test).

That is on a Haswell, where vmexit.flat reports 2120 cycles for a
vmcall. This would linearly (likely incorrect method in this case)
translate to 230 cycles on a machine with 1800 cycles for a vmcall,
which is less than 50 ns @ 5 GHz.

I wouldn't go above 25 ns for a hard-coded default.

(We could also do a similar measurement when initializing KVM and have a
dynamic default, but I'm thinking it's going to be way too much code
for the benefit.)

---
1: This is how the TSC is read in KVM.

diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index da24f1858acc..a7251ac0109b 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -6449,6 +6449,8 @@ static void vmx_vcpu_run(struct kvm_vcpu *vcpu)
vcpu->arch.apic->lapic_timer.timer_advance_ns)
kvm_wait_lapic_expire(vcpu);

+ vcpu->last_seen_tsc = kvm_read_l1_tsc(vcpu, rdtsc());
+
/*
* If this vCPU has touched SPEC_CTRL, restore the guest's value if
* it's non-zero. Since vmentry is serialising on affected CPUs, there
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 6200d5a51f13..5e0ce8ca31e7 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -7201,6 +7201,9 @@ int kvm_emulate_hypercall(struct kvm_vcpu *vcpu)
case KVM_HC_SEND_IPI:
ret = kvm_pv_send_ipi(vcpu->kvm, a0, a1, a2, a3, op_64_bit);
break;
+ case KVM_HC_LAST_SEEN_TSC:
+ ret = vcpu->last_seen_tsc;
+ break;
default:
ret = -KVM_ENOSYS;
break;
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index abafddb9fe2c..7f70fe7a28b1 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -323,6 +323,8 @@ struct kvm_vcpu {
bool preempted;
struct kvm_vcpu_arch arch;
struct dentry *debugfs_dentry;
+
+ u64 last_seen_tsc;
};

static inline int kvm_vcpu_exiting_guest_mode(struct kvm_vcpu *vcpu)
diff --git a/include/uapi/linux/kvm_para.h b/include/uapi/linux/kvm_para.h
index 6c0ce49931e5..dfbc6e9ad7a1 100644
--- a/include/uapi/linux/kvm_para.h
+++ b/include/uapi/linux/kvm_para.h
@@ -28,6 +28,7 @@
#define KVM_HC_MIPS_CONSOLE_OUTPUT 8
#define KVM_HC_CLOCK_PAIRING 9
#define KVM_HC_SEND_IPI 10
+#define KVM_HC_LAST_SEEN_TSC 11

/*
* hypercalls use architecture specific