From: David Woodhouse <dwmw@amazon.co.uk>
Date: Thu, 18 Jan 2024 12:57:30 +0000 (+0000)
Subject: KVM: x86: Fix systemic drift in __get_kvmclock() with TSC scaling
X-Git-Url: https://www.infradead.org/git/?a=commitdiff_plain;h=refs%2Fheads%2Fkvmclock2;p=users%2Fdwmw2%2Flinux.git

KVM: x86: Fix systemic drift in __get_kvmclock() with TSC scaling

Converting a host TSC delta directly to nanoseconds results in a systemic
drift compared to the way the guest calculates its kvmclock.

Provide a __kvmclock_from_tsc() function which calculates the kvmclock
just as the guest would, by scaling to the guest TSC first and from that
to nanoseconds.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
---

diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 7bc1daf687417..bf9a36ec64407 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1347,6 +1347,7 @@ struct kvm_arch {
 	u64 last_tsc_nsec;
 	u64 last_tsc_write;
 	u32 last_tsc_khz;
+	u64 last_tsc_scaling_ratio;
 	u64 last_tsc_offset;
 	u64 cur_tsc_nsec;
 	u64 cur_tsc_write;
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 368c3bc400eb1..8e97c47bf1466 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2682,6 +2682,7 @@ static void __kvm_synchronize_tsc(struct kvm_vcpu *vcpu, u64 offset, u64 tsc,
 	kvm->arch.last_tsc_nsec = ns;
 	kvm->arch.last_tsc_write = tsc;
 	kvm->arch.last_tsc_khz = vcpu->arch.virtual_tsc_khz;
+	kvm->arch.last_tsc_scaling_ratio = vcpu->arch.l1_tsc_scaling_ratio;
 	kvm->arch.last_tsc_offset = offset;
 
 	vcpu->arch.last_guest_tsc = tsc;
@@ -3104,11 +3105,53 @@ static unsigned long get_cpu_tsc_khz(void)
 		return __this_cpu_read(cpu_tsc_khz);
 }
 
+static uint64_t __kvmclock_from_tsc(struct kvm *kvm, uint64_t host_tsc)
+{
+	struct pvclock_vcpu_time_info hv_clock;
+	struct kvm_arch *ka = &kvm->arch;
+	uint64_t guest_tsc;
+	uint32_t tsc_khz;
+
+	/*
+	 * Scale the TSC delta (at host TSC frequency) to the guest TSC
+	 * TSC frequency, before scaling that to nanoseconds. Performing
+	 * this two-step calculation in precisely the same way the guest
+	 * does, avoids a systemic drift due to loss of precision in the
+	 * conversions.
+	 *
+	 * Use the frequency of the last-synchronized TSC by preference,
+	 * since the scaling ratio is already calculated for that. Fall
+	 * back to using the default TSC frequency for the KVM (for which
+	 * the ratio must be calculated which takes a little more time),
+	 * and finally just use the host TSC frequency if neither of the
+	 * former were set.
+	 */
+	if ((tsc_khz = ka->last_tsc_khz) && ka->last_tsc_scaling_ratio) {
+		guest_tsc = kvm_scale_tsc(host_tsc - ka->master_cycle_now,
+					  ka->last_tsc_scaling_ratio);
+	} else if ((tsc_khz = READ_ONCE(ka->default_tsc_khz))) {
+		uint64_t ratio = mul_u64_u32_div(1ULL << kvm_caps.tsc_scaling_ratio_frac_bits,
+						 tsc_khz, get_cpu_tsc_khz());
+
+		guest_tsc = kvm_scale_tsc(host_tsc - ka->master_cycle_now,
+					  ratio);
+	} else {
+		guest_tsc = host_tsc - ka->master_cycle_now;
+		tsc_khz = get_cpu_tsc_khz();
+	}
+
+	hv_clock.tsc_timestamp = 0;
+	hv_clock.system_time = ka->master_kernel_ns + ka->kvmclock_offset;
+	kvm_get_time_scale(NSEC_PER_SEC, tsc_khz * 1000LL,
+			   &hv_clock.tsc_shift,
+			   &hv_clock.tsc_to_system_mul);
+	return __pvclock_read_cycles(&hv_clock, guest_tsc);
+
+}
 /* Called within read_seqcount_begin/retry for kvm->pvclock_sc.  */
 static void __get_kvmclock(struct kvm *kvm, struct kvm_clock_data *data)
 {
 	struct kvm_arch *ka = &kvm->arch;
-	struct pvclock_vcpu_time_info hv_clock;
 
 	/* both __this_cpu_read() and rdtsc() should be on the same cpu */
 	get_cpu();
@@ -3127,12 +3170,7 @@ static void __get_kvmclock(struct kvm *kvm, struct kvm_clock_data *data)
 		data->host_tsc = rdtsc();
 
 		data->flags |= KVM_CLOCK_TSC_STABLE;
-		hv_clock.tsc_timestamp = ka->master_cycle_now;
-		hv_clock.system_time = ka->master_kernel_ns + ka->kvmclock_offset;
-		kvm_get_time_scale(NSEC_PER_SEC, get_cpu_tsc_khz() * 1000LL,
-				   &hv_clock.tsc_shift,
-				   &hv_clock.tsc_to_system_mul);
-		data->clock = __pvclock_read_cycles(&hv_clock, data->host_tsc);
+		data->clock = __kvmclock_from_tsc(kvm, data->host_tsc);
 	} else {
 		data->clock = get_kvmclock_base_ns() + ka->kvmclock_offset;
 	}