Hi Zide, thanks for the review.
"Chen, Zide" <[email protected]> writes:
On 6/12/2026 2:29 PM, Colton Lewis wrote:
Apply dynamic guest counter reservations by checking if the requested
guest mask collides with any events the host has scheduled and calling
pmu_perf_resched_update() with a hook that updates the mask of
available counters in between schedule out and schedule in.
Signed-off-by: Colton Lewis <[email protected]>
---
arch/arm64/kvm/pmu-direct.c | 69 +++++++++++++++++++++++++++++++++++-
include/linux/perf/arm_pmu.h | 1 +
2 files changed, 69 insertions(+), 1 deletion(-)
diff --git a/arch/arm64/kvm/pmu-direct.c b/arch/arm64/kvm/pmu-direct.c
index 49f1feb5d280c..044f011c9c84b 100644
--- a/arch/arm64/kvm/pmu-direct.c
+++ b/arch/arm64/kvm/pmu-direct.c
@@ -87,6 +87,73 @@ u64 kvm_pmu_direct_pmcr_read(struct kvm_vcpu *vcpu)
ARMV8_PMU_PMCR_N);
}
+/* Callback to update counter mask between perf scheduling */
+static void kvm_pmu_update_mask(struct pmu *pmu, void *data)
+{
+ struct arm_pmu *arm_pmu = to_arm_pmu(pmu);
+ unsigned long *new_mask = data;
+
+ bitmap_copy(arm_pmu->cntr_mask, new_mask, ARMPMU_MAX_HWEVENTS);
+}
+
+/**
+ * kvm_pmu_set_guest_counters() - Handle dynamic counter reservations
+ * @cpu_pmu: struct arm_pmu to potentially modify
+ * @guest_mask: new guest mask for the pmu
+ *
+ * Check if guest counters will interfere with current host events and
+ * call into perf_pmu_resched_update if a reschedule is required.
+ */
+static void kvm_pmu_set_guest_counters(struct arm_pmu *cpu_pmu, u64
guest_mask)
+{
+ struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
+ DECLARE_BITMAP(guest_bitmap, ARMPMU_MAX_HWEVENTS);
+ DECLARE_BITMAP(new_mask, ARMPMU_MAX_HWEVENTS);
+ bool need_resched = false;
+
+ bitmap_from_arr64(guest_bitmap, &guest_mask, ARMPMU_MAX_HWEVENTS);
+ bitmap_copy(new_mask, cpu_pmu->hw_cntr_impl, ARMPMU_MAX_HWEVENTS);
+
+ if (guest_mask) {
+ /* Subtract guest counters from available host mask */
+ bitmap_andnot(new_mask, new_mask, guest_bitmap,
ARMPMU_MAX_HWEVENTS);
+
+ /* Did we collide with an active host event? */
+ if (bitmap_intersects(cpuc->used_mask, guest_bitmap,
ARMPMU_MAX_HWEVENTS)) {
+ int idx;
+
+ need_resched = true;
+ cpuc->host_squeezed = true;
+
+ /* Look for pinned events that are about to be
preempted */
+ for_each_set_bit(idx, guest_bitmap,
ARMPMU_MAX_HWEVENTS) {
+ if (test_bit(idx, cpuc->used_mask) && cpuc->events[idx]
&&
+ cpuc->events[idx]->attr.pinned) {
+ pr_warn_once("perf: Pinned host event squeezed out by KVM guest
PMU partition\n");
+ break;
+ }
+ }
+ }
+ } else {
+ /*
+ * Restoring to hw_cntr_impl.
+ * Only resched if we previously squeezed an event.
+ */
+ if (cpuc->host_squeezed) {
+ need_resched = true;
+ cpuc->host_squeezed = false;
+ }
+ }
+
+ if (need_resched) {
+ /* Collision: run full perf reschedule */
+ perf_pmu_resched_update(&cpu_pmu->pmu, kvm_pmu_update_mask,
new_mask);
+ } else {
+ /* Host was never using guest counters anyway */
+ bitmap_copy(cpu_pmu->cntr_mask, new_mask, ARMPMU_MAX_HWEVENTS);
+ }
+}
+
/**
* kvm_pmu_host_counter_mask() - Compute bitmask of host-reserved
counters
* @pmu: Pointer to arm_pmu struct
@@ -209,6 +276,7 @@ void kvm_pmu_load(struct kvm_vcpu *vcpu)
pmu = vcpu->kvm->arch.arm_pmu;
guest_counters = kvm_pmu_guest_counter_mask(pmu);
+ kvm_pmu_set_guest_counters(pmu, guest_counters);
kvm_pmu_apply_event_filter(vcpu);
for_each_set_bit(i, &guest_counters, ARMPMU_MAX_HWEVENTS) {
@@ -317,7 +385,6 @@ void kvm_pmu_put(struct kvm_vcpu *vcpu)
/* Stop guest counters and disable interrupts in hardware. */
write_sysreg(mask, pmcntenclr_el0);
write_sysreg(mask, pmintenclr_el1);
-
kvm_pmu_set_guest_counters(pmu, 0);
I must have missed something.
Here, vcpu->kvm->arch.arm_pmu->cntr_mask is restored to cntr_mask_impl.
Is this same struct shared with the perf driver
(to_arm_pmu(event->pmu)), and shared across other vCPUs as well?
If so, if another pCPU still has a PMU-partition-enabled vcpu loaded
with guest-owned counters (MDCR_EL2.HPMN is set), could a host event be
scheduled onto one of those guest-owned counters, conflicting with the
guest?
You're correct that is shared more widely than appropriate and could
cause such a conflict.
cntr_mask should be moved to the per-CPU struct pmu_hw_events inside
struct arm_pmu.
preempt_enable();
}
