The PEBS event counters scheduled by host may different to the counters
required by guest. The host counter index will be leaked into the guest
PEBS record and the guest driver will be confused by the counter indexes
in the "Applicable Counters" field of the PEBS records and ignore them.
Before the guest PEBS overflow PMI is injected into the guest through
global status, KVM needs to rewrite the "Applicable Counters" field with
the right enabled guest pebs counter idx(s) in the guest PEBS records.
Co-developed-by: Luwei Kang <luwei.kang@xxxxxxxxx>
Signed-off-by: Luwei Kang <luwei.kang@xxxxxxxxx>
Signed-off-by: Like Xu <like.xu@xxxxxxxxxxxxxxx>
---
arch/x86/include/asm/kvm_host.h | 1 +
arch/x86/kvm/pmu.c | 1 +
arch/x86/kvm/vmx/pmu_intel.c | 85 +++++++++++++++++++++++++++++++--
3 files changed, 82 insertions(+), 5 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 77b529b8c16a..cdc3c6efdd8e 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -457,6 +457,7 @@ struct kvm_pmu {
bool counter_cross_mapped;
bool need_rewrite_ds_pebs_interrupt_threshold;
+ bool need_rewrite_pebs_records;
/*
* The gate to release perf_events not marked in
diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
index 7c8e3ca5b7ad..64dce19644e3 100644
--- a/arch/x86/kvm/pmu.c
+++ b/arch/x86/kvm/pmu.c
@@ -77,6 +77,7 @@ static void kvm_perf_overflow_intr(struct perf_event *perf_event,
if (!test_and_set_bit(pmc->idx, pmu->reprogram_pmi)) {
if (perf_event->attr.precise_ip) {
+ pmu->need_rewrite_pebs_records = pmu->counter_cross_mapped;
/* Indicate PEBS overflow PMI to guest. */
__set_bit(62, (unsigned long *)&pmu->global_status);
} else
diff --git a/arch/x86/kvm/vmx/pmu_intel.c b/arch/x86/kvm/vmx/pmu_intel.c
index 346b1104e674..d58d04ee13a5 100644
--- a/arch/x86/kvm/vmx/pmu_intel.c
+++ b/arch/x86/kvm/vmx/pmu_intel.c
@@ -551,22 +551,97 @@ static int rewrite_ds_pebs_interrupt_threshold(struct kvm_vcpu *vcpu)
return ret;
}
+static int rewrite_ds_pebs_records(struct kvm_vcpu *vcpu)
+{
+ struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
+ struct kvm_pmc *pmc = NULL;
+ struct debug_store *ds = NULL;
+ gpa_t gpa;
+ u64 pebs_buffer_base, offset, status, new_status, format_size;
+ int srcu_idx, bit, ret = 0;
+
+ if (!pmu->counter_cross_mapped)
+ return ret;
+
+ ds = kmalloc(sizeof(struct debug_store), GFP_KERNEL);
+ if (!ds)
+ return -ENOMEM;
+
+ ret = -EFAULT;
+ srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
+ gpa = kvm_mmu_gva_to_gpa_system(vcpu, pmu->ds_area, NULL);
+ if (gpa == UNMAPPED_GVA)
+ goto out;
+
+ if (kvm_read_guest(vcpu->kvm, gpa, ds, sizeof(struct debug_store)))
+ goto out;
+
+ if (ds->pebs_index <= ds->pebs_buffer_base)
+ goto out;
+
+ pebs_buffer_base = ds->pebs_buffer_base;
+ offset = offsetof(struct pebs_basic, applicable_counters);
+
+ do {
+ ret = -EFAULT;
+ gpa = kvm_mmu_gva_to_gpa_system(vcpu, pebs_buffer_base + offset, NULL);
+ if (gpa == UNMAPPED_GVA)
+ goto out;
+ if (kvm_read_guest(vcpu->kvm, gpa, &status, sizeof(u64)))
+ goto out;
+
+ new_status = 0ull;
+ for_each_set_bit(bit, (unsigned long *)&pmu->pebs_enable, X86_PMC_IDX_MAX) {
+ pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, bit);
+
+ if (!pmc || !pmc->perf_event)
+ continue;
+
+ if (test_bit(pmc->perf_event->hw.idx, (unsigned long *)&status))
+ new_status |= BIT_ULL(pmc->idx);
+ }
+ if (kvm_write_guest(vcpu->kvm, gpa, &new_status, sizeof(u64)))
+ goto out;
+
+ gpa = kvm_mmu_gva_to_gpa_system(vcpu, pebs_buffer_base, NULL);
+ if (gpa == UNMAPPED_GVA)
+ goto out;
+ if (kvm_read_guest(vcpu->kvm, gpa, &format_size, sizeof(u64)))
+ goto out;
+ ret = 0;
+
+ pebs_buffer_base = pebs_buffer_base + (format_size >> 48);
+ } while (pebs_buffer_base < ds->pebs_index);
+
+out:
+ srcu_read_unlock(&vcpu->kvm->srcu, srcu_idx);
+ kfree(ds);
+ return ret;
+}
+
void intel_pmu_handle_event(struct kvm_vcpu *vcpu)
{
struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
- int ret;
+ int ret1, ret2;
+
+ if (pmu->need_rewrite_pebs_records) {
+ pmu->need_rewrite_pebs_records = false;
+ ret1 = rewrite_ds_pebs_records(vcpu);
+ }
if (!(pmu->global_ctrl & pmu->pebs_enable))
- return;
+ goto out;
if (pmu->counter_cross_mapped && pmu->need_rewrite_ds_pebs_interrupt_threshold) {
- ret = rewrite_ds_pebs_interrupt_threshold(vcpu);
pmu->need_rewrite_ds_pebs_interrupt_threshold = false;
+ ret2 = rewrite_ds_pebs_interrupt_threshold(vcpu);
}
- if (ret == -ENOMEM)
+out:
+
+ if (ret1 == -ENOMEM || ret2 == -ENOMEM)
pr_debug_ratelimited("%s: Fail to emulate guest PEBS due to OOM.", __func__);
- else if (ret == -EFAULT)
+ else if (ret1 == -EFAULT || ret2 == -EFAULT)
pr_debug_ratelimited("%s: Fail to emulate guest PEBS due to GPA fault.", __func__);
}
--
2.21.3
