[pve-devel] [RFC v2 kernel] backport patch to fix TSC scaling for SVM

[Fiona Ebner]

The following issue reported on the community forum [0] is likely
fixed by this.

In my case, loading a VM snapshot that originally was taken on an
Intel CPU on my AMD-based host often caused problems in other VMs. In
particular, it often led to CPU stalls, and sometimes clock jumps far
into the future. With this backport applied, everything seems to run
smoothly even after loading the "bad" snapshot 10 times.

The backport from upstream commit 11d39e8cc43e ("KVM: SVM: fix tsc
scaling cache logic consisted of dropping the parts for nested TSC
scaling, which is not yet present in our kernel, renaming the constant
for the default ratio, and some context changes.

[0] https://forum.proxmox.com/threads/112756/

Signed-off-by: Fiona Ebner <f.ebner at proxmox.com>
---

Changes from v1:
    * Alternative approach backporting the fix of the bad commit.

 ...-KVM-SVM-fix-tsc-scaling-cache-logic.patch | 113 ++++++++++++++++++
 1 file changed, 113 insertions(+)
 create mode 100644 patches/kernel/0032-KVM-SVM-fix-tsc-scaling-cache-logic.patch

diff --git a/patches/kernel/0032-KVM-SVM-fix-tsc-scaling-cache-logic.patch b/patches/kernel/0032-KVM-SVM-fix-tsc-scaling-cache-logic.patch
new file mode 100644
index 0000000..15016ea
--- /dev/null
+++ b/patches/kernel/0032-KVM-SVM-fix-tsc-scaling-cache-logic.patch
@@ -0,0 +1,113 @@
+From 955c2b4d64593baecc87c64d1354cd5786975bd9 Mon Sep 17 00:00:00 2001
+From: Maxim Levitsky <mlevitsk at redhat.com>
+Date: Mon, 6 Jun 2022 21:11:49 +0300
+Subject: [PATCH] KVM: SVM: fix tsc scaling cache logic
+
+SVM uses a per-cpu variable to cache the current value of the
+tsc scaling multiplier msr on each cpu.
+
+Commit 1ab9287add5e2
+("KVM: X86: Add vendor callbacks for writing the TSC multiplier")
+broke this caching logic.
+
+Refactor the code so that all TSC scaling multiplier writes go through
+a single function which checks and updates the cache.
+
+This fixes the following scenario:
+
+1. A CPU runs a guest with some tsc scaling ratio.
+
+2. New guest with different tsc scaling ratio starts on this CPU
+   and terminates almost immediately.
+
+   This ensures that the short running guest had set the tsc scaling ratio just
+   once when it was set via KVM_SET_TSC_KHZ. Due to the bug,
+   the per-cpu cache is not updated.
+
+3. The original guest continues to run, it doesn't restore the msr
+   value back to its own value, because the cache matches,
+   and thus continues to run with a wrong tsc scaling ratio.
+
+Fixes: 1ab9287add5e2 ("KVM: X86: Add vendor callbacks for writing the TSC multiplier")
+Signed-off-by: Maxim Levitsky <mlevitsk at redhat.com>
+Message-Id: <20220606181149.103072-1-mlevitsk at redhat.com>
+Cc: stable at vger.kernel.org
+Signed-off-by: Paolo Bonzini <pbonzini at redhat.com>
+[FE: backport, mainly dropped parts for the not yet present
+     5228eb96a487 ("KVM: x86: nSVM: implement nested TSC scaling")]
+Signed-off-by: Fiona Ebner <f.ebner at proxmox.com>
+---
+ arch/x86/kvm/svm/svm.c | 30 +++++++++++++++++++-----------
+ 1 file changed, 19 insertions(+), 11 deletions(-)
+
+diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
+index 938b9b24f0ee..ea87d06bdaba 100644
+--- a/arch/x86/kvm/svm/svm.c
++++ b/arch/x86/kvm/svm/svm.c
+@@ -468,11 +468,24 @@ static int has_svm(void)
+ 	return 1;
+ }
+ 
++void __svm_write_tsc_multiplier(u64 multiplier)
++{
++	preempt_disable();
++
++	if (multiplier == __this_cpu_read(current_tsc_ratio))
++		goto out;
++
++	wrmsrl(MSR_AMD64_TSC_RATIO, multiplier);
++	__this_cpu_write(current_tsc_ratio, multiplier);
++out:
++	preempt_enable();
++}
++
+ static void svm_hardware_disable(void)
+ {
+ 	/* Make sure we clean up behind us */
+ 	if (static_cpu_has(X86_FEATURE_TSCRATEMSR))
+-		wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
++		__svm_write_tsc_multiplier(TSC_RATIO_DEFAULT);
+ 
+ 	cpu_svm_disable();
+ 
+@@ -514,8 +527,7 @@ static int svm_hardware_enable(void)
+ 	wrmsrl(MSR_VM_HSAVE_PA, __sme_page_pa(sd->save_area));
+ 
+ 	if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
+-		wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
+-		__this_cpu_write(current_tsc_ratio, TSC_RATIO_DEFAULT);
++		__svm_write_tsc_multiplier(TSC_RATIO_DEFAULT);
+ 	}
+ 
+ 
+@@ -1128,9 +1140,10 @@ static void svm_write_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
+ 
+ static void svm_write_tsc_multiplier(struct kvm_vcpu *vcpu, u64 multiplier)
+ {
+-	wrmsrl(MSR_AMD64_TSC_RATIO, multiplier);
++	__svm_write_tsc_multiplier(multiplier);
+ }
+ 
++
+ /* Evaluate instruction intercepts that depend on guest CPUID features. */
+ static void svm_recalc_instruction_intercepts(struct kvm_vcpu *vcpu,
+ 					      struct vcpu_svm *svm)
+@@ -1453,13 +1466,8 @@ static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
+ 		vmsave(__sme_page_pa(sd->save_area));
+ 	}
+ 
+-	if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
+-		u64 tsc_ratio = vcpu->arch.tsc_scaling_ratio;
+-		if (tsc_ratio != __this_cpu_read(current_tsc_ratio)) {
+-			__this_cpu_write(current_tsc_ratio, tsc_ratio);
+-			wrmsrl(MSR_AMD64_TSC_RATIO, tsc_ratio);
+-		}
+-	}
++	if (static_cpu_has(X86_FEATURE_TSCRATEMSR))
++		__svm_write_tsc_multiplier(vcpu->arch.tsc_scaling_ratio);
+ 
+ 	if (likely(tsc_aux_uret_slot >= 0))
+ 		kvm_set_user_return_msr(tsc_aux_uret_slot, svm->tsc_aux, -1ull);
+-- 
+2.30.2
+
-- 
2.30.2