[PATCH v2] KVM: SVM: Set/clear SRSO's BP_SPEC_REDUCE on 0 <=> 1 VM count transitions

[Sean Christopherson <seanjc@google.com>]

Set the magic BP_SPEC_REDUCE bit to mitigate SRSO when running VMs if and
only if KVM has at least one active VM.  Leaving the bit set at all times
unfortunately degrades performance by a wee bit more than expected.

Use a dedicated spinlock and counter instead of hooking virtualization
enablement, as changing the behavior of kvm.enable_virt_at_load based on
SRSO_BP_SPEC_REDUCE is painful, and has its own drawbacks, e.g. could
result in performance issues for flows that are sensitive to VM creation
latency.

Defer setting BP_SPEC_REDUCE until VMRUN is imminent to avoid impacting
performance on CPUs that aren't running VMs, e.g. if a setup is using
housekeeping CPUs.  Setting BP_SPEC_REDUCE in task context, i.e. without
blasting IPIs to all CPUs, also helps avoid serializing 1<=>N transitions
without incurring a gross amount of complexity (see the Link for details
on how ugly coordinating via IPIs gets).

Link: https://lore.kernel.org/all/aBOnzNCngyS_pQIW@google.com
Fixes: 8442df2b49ed ("x86/bugs: KVM: Add support for SRSO_MSR_FIX")
Reported-by: Michael Larabel <Michael@michaellarabel.com>
Closes: https://www.phoronix.com/review/linux-615-amd-regression
Cc: Borislav Petkov <bp@alien8.de>
Signed-off-by: Sean Christopherson <seanjc@google.com>
---

v2: Defer setting BP_SPEC_REDUCE until VMRUN is imminent, which in turn
    allows for eliding the lock on 0<=>1 transitions as there is no race
    with CPUs doing VMRUN before receiving the IPI to set the bit, and
    having multiple tasks take the lock during svm_srso_vm_init() is a-ok.

v1: https://lore.kernel.org/all/20250502223456.887618-1-seanjc@google.com

 arch/x86/kvm/svm/svm.c | 71 ++++++++++++++++++++++++++++++++++++++----
 arch/x86/kvm/svm/svm.h |  2 ++
 2 files changed, 67 insertions(+), 6 deletions(-)

diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index cc1c721ba067..15f7a0703c16 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -607,9 +607,6 @@ static void svm_disable_virtualization_cpu(void)
 	kvm_cpu_svm_disable();
 
 	amd_pmu_disable_virt();
-
-	if (cpu_feature_enabled(X86_FEATURE_SRSO_BP_SPEC_REDUCE))
-		msr_clear_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_BP_SPEC_REDUCE_BIT);
 }
 
 static int svm_enable_virtualization_cpu(void)
@@ -687,9 +684,6 @@ static int svm_enable_virtualization_cpu(void)
 		rdmsr(MSR_TSC_AUX, sev_es_host_save_area(sd)->tsc_aux, msr_hi);
 	}
 
-	if (cpu_feature_enabled(X86_FEATURE_SRSO_BP_SPEC_REDUCE))
-		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_BP_SPEC_REDUCE_BIT);
-
 	return 0;
 }
 
@@ -1518,6 +1512,63 @@ static void svm_vcpu_free(struct kvm_vcpu *vcpu)
 	__free_pages(virt_to_page(svm->msrpm), get_order(MSRPM_SIZE));
 }
 
+#ifdef CONFIG_CPU_MITIGATIONS
+static DEFINE_SPINLOCK(srso_lock);
+static atomic_t srso_nr_vms;
+
+static void svm_srso_clear_bp_spec_reduce(void *ign)
+{
+	struct svm_cpu_data *sd = this_cpu_ptr(&svm_data);
+
+	if (!sd->bp_spec_reduce_set)
+		return;
+
+	msr_clear_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_BP_SPEC_REDUCE_BIT);
+	sd->bp_spec_reduce_set = false;
+}
+
+static void svm_srso_vm_destroy(void)
+{
+	if (!cpu_feature_enabled(X86_FEATURE_SRSO_BP_SPEC_REDUCE))
+		return;
+
+	if (atomic_dec_return(&srso_nr_vms))
+		return;
+
+	guard(spinlock)(&srso_lock);
+
+	/*
+	 * Verify a new VM didn't come along, acquire the lock, and increment
+	 * the count before this task acquired the lock.
+	 */
+	if (atomic_read(&srso_nr_vms))
+		return;
+
+	on_each_cpu(svm_srso_clear_bp_spec_reduce, NULL, 1);
+}
+
+static void svm_srso_vm_init(void)
+{
+	if (!cpu_feature_enabled(X86_FEATURE_SRSO_BP_SPEC_REDUCE))
+		return;
+
+	/*
+	 * Acquire the lock on 0 => 1 transitions to ensure a potential 1 => 0
+	 * transition, i.e. destroying the last VM, is fully complete, e.g. so
+	 * that a delayed IPI doesn't clear BP_SPEC_REDUCE after a vCPU runs.
+	 */
+	if (atomic_inc_not_zero(&srso_nr_vms))
+		return;
+
+	guard(spinlock)(&srso_lock);
+
+	atomic_inc(&srso_nr_vms);
+}
+#else
+static void svm_srso_vm_init(void) { }
+static void svm_srso_vm_destroy(void) { }
+#endif
+
 static void svm_prepare_switch_to_guest(struct kvm_vcpu *vcpu)
 {
 	struct vcpu_svm *svm = to_svm(vcpu);
@@ -1550,6 +1601,11 @@ static void svm_prepare_switch_to_guest(struct kvm_vcpu *vcpu)
 	    (!boot_cpu_has(X86_FEATURE_V_TSC_AUX) || !sev_es_guest(vcpu->kvm)))
 		kvm_set_user_return_msr(tsc_aux_uret_slot, svm->tsc_aux, -1ull);
 
+	if (cpu_feature_enabled(X86_FEATURE_SRSO_BP_SPEC_REDUCE) &&
+	    !sd->bp_spec_reduce_set) {
+		sd->bp_spec_reduce_set = true;
+		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_BP_SPEC_REDUCE_BIT);
+	}
 	svm->guest_state_loaded = true;
 }
 
@@ -5036,6 +5092,8 @@ static void svm_vm_destroy(struct kvm *kvm)
 {
 	avic_vm_destroy(kvm);
 	sev_vm_destroy(kvm);
+
+	svm_srso_vm_destroy();
 }
 
 static int svm_vm_init(struct kvm *kvm)
@@ -5061,6 +5119,7 @@ static int svm_vm_init(struct kvm *kvm)
 			return ret;
 	}
 
+	svm_srso_vm_init();
 	return 0;
 }
 
diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h
index d4490eaed55d..f16b068c4228 100644
--- a/arch/x86/kvm/svm/svm.h
+++ b/arch/x86/kvm/svm/svm.h
@@ -335,6 +335,8 @@ struct svm_cpu_data {
 	u32 next_asid;
 	u32 min_asid;
 
+	bool bp_spec_reduce_set;
+
 	struct vmcb *save_area;
 	unsigned long save_area_pa;
 

base-commit: 45eb29140e68ffe8e93a5471006858a018480a45
-- 
2.49.0.967.g6a0df3ecc3-goog