Skip to contentSkip to conversation

KVM: x86: Compute the max guest TSC frequency from the current tsc_khz

PATCH Applied Sasha Levin 3 participants 3 messages Oct 24 – Nov 13

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 KVM: x86: Compute the max guest TSC frequency from the current tsc_khz

KVM computes the highest frequency that KVM_SET_TSC_KHZ accepts once, during vendor init, from the tsc_khz of the time. When KVM is built in, that runs before tsc_refine_calibration_work() refines tsc_khz, about a second later. If refinement raises tsc_khz, the cap stays below what the maximum scaling ratio gives for the refined frequency, and KVM_SET_TSC_KHZ rejects frequencies that set_tsc_khz() would accept, as set_tsc_khz() computes the ratio from the current tsc_khz.

nested_tsc_scaling_test runs into this on AMD: its last case asks for the frequency KVM_GET_TSC_KHZ reports times SVM's maximum ratio, and gets -EINVAL on boots where refinement raised tsc_khz. KernelCI sees it intermittently on a DL345 Gen11 (EPYC). On Intel the cap is limited to INT_MAX, so the test does not hit it there.

Like commit 741e511b4208 ("KVM: x86: Don't snapshot "max" TSC if host TSC is constant") did for the default frequency of new VMs, compute the cap from tsc_khz when it is used instead of taking a snapshot at load time. This covers KVM_SET_TSC_KHZ on both the vCPU and the VM file descriptor.

This also drops setting the cap to tsc_khz in the Hyper-V reenlightenment notifier, which left no room for scaling up once the host TSC frequency had changed; commit 0092e4346f49 ("x86/kvm: Support Hyper-V reenlightenment") gives no reason for that limit. The cap now follows the new frequency the same way. The notifier still does not recompute the multipliers of vCPUs that already exist.

Fixes 92a1f12d2598 ("KVM: X86: Implement userspace interface to set virtual_tsc_khz")

---

Reproduced on an AMD DL345 Gen11 booted with tsc_early_khz= 500 kHz below the refined frequency, so that refinement always raises tsc_khz after KVM is initialized. With KernelCI's x86 build of mainline 5b0764a53235, nested_tsc_scaling_test failed 3 out of 3 times with KVM_SET_TSC_KHZ returning -EINVAL. With the same config plus this patch it passed 3 out of 3, and tsc_scaling_sync, tsc_msrs_test, kvm_clock_test and nested_tsc_adjust_test still pass. A nested SVM guest with the same tsc_early_khz trick gives the same result, and with tsc_early_khz above the real frequency (refinement lowers tsc_khz) mainline passes.

This does not change the existing window between the cap check and set_tsc_khz(), which read tsc_khz separately.

 arch/x86/include/asm/kvm_host.h |  2 --
 arch/x86/kvm/x86.c              | 30 +++++++++++++++---------------
 2 files changed, 15 insertions(+), 17 deletions(-)
arch/x86/include/asm/kvm_host.h +0 -2
index 1ee4b127004df..35a74f0eb45e5 100644
@@ -260,8 +260,6 @@ struct kvm_x86_pmu_event_filter;
 struct kvm_caps {
 	/* control of guest tsc rate supported? */
 	bool has_tsc_control;
-	/* maximum supported tsc_khz for guests */
-	u32  max_guest_tsc_khz;
 	/* number of bits of the fractional part of the TSC scaling ratio */
 	u8   tsc_scaling_ratio_frac_bits;
 	/* maximum allowed value of TSC scaling ratio */
arch/x86/kvm/x86.c +15 -15
index 32563a91fba99..9e22072f39c18 100644
@@ -1156,6 +1156,19 @@ static inline u64 __scale_tsc(u64 ratio, u64 tsc)
 	return mul_u64_u64_shr(tsc, ratio, kvm_caps.tsc_scaling_ratio_frac_bits);
 }
 
+/*
+ * Use the current tsc_khz, not the one at load time, as the TSC frequency can
+ * be refined after KVM is initialized.  Make sure the user can only configure
+ * tsc_khz values that fit into a signed integer, otherwise KVM_GET_TSC_KHZ
+ * would return a negative value and confuse userspace.  A min value is not
+ * calculated because it will always be 1 on all machines.
+ */
+static u32 kvm_max_guest_tsc_khz(void)
+{
+	return min(0x7fffffffULL,
+		   __scale_tsc(kvm_caps.max_tsc_scaling_ratio, tsc_khz));
+}
+
 u64 kvm_scale_tsc(u64 tsc, u64 ratio)
 {
 	u64 _tsc = tsc;
@@ -3728,7 +3741,7 @@ long kvm_arch_vcpu_ioctl(struct file *filp,
 		user_tsc_khz = (u32)arg;
 
 		if (kvm_caps.has_tsc_control &&
-		    user_tsc_khz > kvm_caps.max_guest_tsc_khz)
+		    user_tsc_khz > kvm_max_guest_tsc_khz())
 			goto out;
 
 		if (user_tsc_khz == 0)
@@ -4695,7 +4708,7 @@ int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
 		user_tsc_khz = (u32)arg;
 
 		if (kvm_caps.has_tsc_control &&
-		    user_tsc_khz > kvm_caps.max_guest_tsc_khz)
+		    user_tsc_khz > kvm_max_guest_tsc_khz())
 			goto out;
 
 		if (user_tsc_khz == 0)
@@ -6749,7 +6762,6 @@ static void kvm_hyperv_tsc_notifier(void)
 		for_each_present_cpu(cpu)
 			per_cpu(cpu_tsc_khz, cpu) = tsc_khz;
 	}
-	kvm_caps.max_guest_tsc_khz = tsc_khz;
 
 	list_for_each_entry(kvm, &vm_list, vm_list) {
 		__kvm_start_pvclock_update(kvm);
@@ -7188,18 +7200,6 @@ int kvm_x86_vendor_init(struct kvm_x86_init_ops *ops)
 	if (!tdp_enabled)
 		kvm_caps.supported_quirks &= ~KVM_X86_QUIRK_IGNORE_GUEST_PAT;
 
-	if (kvm_caps.has_tsc_control) {
-		/*
-		 * Make sure the user can only configure tsc_khz values that
-		 * fit into a signed integer, otherwise KVM_GET_TSC_KHZ would
-		 * return a negative value and confuse userspace.
-		 * A min value is not calculated because it will always
-		 * be 1 on all machines.
-		 */
-		u64 max = min(0x7fffffffULL,
-			      __scale_tsc(kvm_caps.max_tsc_scaling_ratio, tsc_khz));
-		kvm_caps.max_guest_tsc_khz = max;
-	}
 	kvm_caps.default_tsc_scaling_ratio = 1ULL << kvm_caps.tsc_scaling_ratio_frac_bits;
 	kvm_init_msr_lists();
 	return 0;
Signature · 1 line
2.53.0
Reviewed-by Tested-by on v1 1/1
#1 lore raw
Quoted 26 lines · Sasha Levin

KVM computes the highest frequency that KVM_SET_TSC_KHZ accepts once, during vendor init, from the tsc_khz of the time. When KVM is built in, that runs before tsc_refine_calibration_work() refines tsc_khz, about a second later. If refinement raises tsc_khz, the cap stays below what the maximum scaling ratio gives for the refined frequency, and KVM_SET_TSC_KHZ rejects frequencies that set_tsc_khz() would accept, as set_tsc_khz() computes the ratio from the current tsc_khz.

nested_tsc_scaling_test runs into this on AMD: its last case asks for the frequency KVM_GET_TSC_KHZ reports times SVM's maximum ratio, and gets -EINVAL on boots where refinement raised tsc_khz. KernelCI sees it intermittently on a DL345 Gen11 (EPYC). On Intel the cap is limited to INT_MAX, so the test does not hit it there.

Like commit 741e511b4208 ("KVM: x86: Don't snapshot "max" TSC if host TSC is constant") did for the default frequency of new VMs, compute the cap from tsc_khz when it is used instead of taking a snapshot at load time. This covers KVM_SET_TSC_KHZ on both the vCPU and the VM file descriptor.

This also drops setting the cap to tsc_khz in the Hyper-V reenlightenment notifier, which left no room for scaling up once the host TSC frequency had changed; commit 0092e4346f49 ("x86/kvm: Support Hyper-V reenlightenment") gives no reason for that limit. The cap now follows the new frequency the same way. The notifier still does not recompute the multipliers of vCPUs that already exist.

Quoted 4 lines
Fixes 92a1f12d2598 ("KVM: X86: Implement userspace interface to set virtual_tsc_khz")

---

Tested on an AMD EPYC 9996 (Venice) with KVM built in (CONFIG_KVM_AMD=y), booting with tsc_early_khz set to the refined tsc_khz minus 500.

The Hyper-V notifier path is only reachable when KVM runs as a Hyper-V guest; I haven't tested that path.

Quoted 99 lines

Reproduced on an AMD DL345 Gen11 booted with tsc_early_khz= 500 kHz below the refined frequency, so that refinement always raises tsc_khz after KVM is initialized. With KernelCI's x86 build of mainline 5b0764a53235, nested_tsc_scaling_test failed 3 out of 3 times with KVM_SET_TSC_KHZ returning -EINVAL. With the same config plus this patch it passed 3 out of 3, and tsc_scaling_sync, tsc_msrs_test, kvm_clock_test and nested_tsc_adjust_test still pass. A nested SVM guest with the same tsc_early_khz trick gives the same result, and with tsc_early_khz above the real frequency (refinement lowers tsc_khz) mainline passes.

This does not change the existing window between the cap check and set_tsc_khz(), which read tsc_khz separately.

 arch/x86/include/asm/kvm_host.h |  2 --
 arch/x86/kvm/x86.c              | 30 +++++++++++++++---------------
 2 files changed, 15 insertions(+), 17 deletions(-)
arch/x86/include/asm/kvm_host.h +0 -2
index 1ee4b127004df..35a74f0eb45e5 100644
@@ -260,8 +260,6 @@ struct kvm_x86_pmu_event_filter;
 struct kvm_caps {
 	/* control of guest tsc rate supported? */
 	bool has_tsc_control;
-	/* maximum supported tsc_khz for guests */
-	u32  max_guest_tsc_khz;
 	/* number of bits of the fractional part of the TSC scaling ratio */
 	u8   tsc_scaling_ratio_frac_bits;
 	/* maximum allowed value of TSC scaling ratio */
arch/x86/kvm/x86.c +15 -15
index 32563a91fba99..9e22072f39c18 100644
@@ -1156,6 +1156,19 @@ static inline u64 __scale_tsc(u64 ratio, u64 tsc)
 	return mul_u64_u64_shr(tsc, ratio, kvm_caps.tsc_scaling_ratio_frac_bits);
 }
 
+/*
+ * Use the current tsc_khz, not the one at load time, as the TSC frequency can
+ * be refined after KVM is initialized.  Make sure the user can only configure
+ * tsc_khz values that fit into a signed integer, otherwise KVM_GET_TSC_KHZ
+ * would return a negative value and confuse userspace.  A min value is not
+ * calculated because it will always be 1 on all machines.
+ */
+static u32 kvm_max_guest_tsc_khz(void)
+{
+	return min(0x7fffffffULL,
+		   __scale_tsc(kvm_caps.max_tsc_scaling_ratio, tsc_khz));
+}
+
 u64 kvm_scale_tsc(u64 tsc, u64 ratio)
 {
 	u64 _tsc = tsc;
@@ -3728,7 +3741,7 @@ long kvm_arch_vcpu_ioctl(struct file *filp,
 		user_tsc_khz = (u32)arg;
 
 		if (kvm_caps.has_tsc_control &&
-		    user_tsc_khz > kvm_caps.max_guest_tsc_khz)
+		    user_tsc_khz > kvm_max_guest_tsc_khz())
 			goto out;
 
 		if (user_tsc_khz == 0)
@@ -4695,7 +4708,7 @@ int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
 		user_tsc_khz = (u32)arg;
 
 		if (kvm_caps.has_tsc_control &&
-		    user_tsc_khz > kvm_caps.max_guest_tsc_khz)
+		    user_tsc_khz > kvm_max_guest_tsc_khz())
 			goto out;
 
 		if (user_tsc_khz == 0)
@@ -6749,7 +6762,6 @@ static void kvm_hyperv_tsc_notifier(void)
 		for_each_present_cpu(cpu)
 			per_cpu(cpu_tsc_khz, cpu) = tsc_khz;
 	}
-	kvm_caps.max_guest_tsc_khz = tsc_khz;
 
 	list_for_each_entry(kvm, &vm_list, vm_list) {
 		__kvm_start_pvclock_update(kvm);
@@ -7188,18 +7200,6 @@ int kvm_x86_vendor_init(struct kvm_x86_init_ops *ops)
 	if (!tdp_enabled)
 		kvm_caps.supported_quirks &= ~KVM_X86_QUIRK_IGNORE_GUEST_PAT;
 
-	if (kvm_caps.has_tsc_control) {
-		/*
-		 * Make sure the user can only configure tsc_khz values that
-		 * fit into a signed integer, otherwise KVM_GET_TSC_KHZ would
-		 * return a negative value and confuse userspace.
-		 * A min value is not calculated because it will always
-		 * be 1 on all machines.
-		 */
-		u64 max = min(0x7fffffffULL,
-			      __scale_tsc(kvm_caps.max_tsc_scaling_ratio, tsc_khz));
-		kvm_caps.max_guest_tsc_khz = max;
-	}
 	kvm_caps.default_tsc_scaling_ratio = 1ULL << kvm_caps.tsc_scaling_ratio_frac_bits;
 	kvm_init_msr_lists();
 	return 0;
maintainer Applied
#2 lore raw
In reply to Sasha Levin
Quoted 9 lines · Sasha Levin

KVM computes the highest frequency that KVM_SET_TSC_KHZ accepts once, during vendor init, from the tsc_khz of the time. When KVM is built in, that runs before tsc_refine_calibration_work() refines tsc_khz, about a second later. If refinement raises tsc_khz, the cap stays below what the maximum scaling ratio gives for the refined frequency, and KVM_SET_TSC_KHZ rejects frequencies that set_tsc_khz() would accept, as set_tsc_khz() computes the ratio from the current tsc_khz.

[...]

Applied to kvm-x86 misc, thanks!

[1/1] KVM: x86: Compute the max guest TSC frequency from the current tsc_khz

      https://github.com/kvm-x86/linux/commit/87ea335ed2bc

--
https://github.com/kvm-x86/linux/tree/next