diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
index 19128d994ee9..27811f382734 100644
--- a/arch/arm64/kernel/perf_event.c
+++ b/arch/arm64/kernel/perf_event.c
@@ -372,6 +372,73 @@ static inline bool armv8pmu_event_is_chained(struct perf_event *event)
 #define	ARMV8_IDX_TO_COUNTER(x)	\
 	(((x) - ARMV8_IDX_COUNTER0) & ARMV8_PMU_COUNTER_MASK)
 
+/*
+ * This code is really good
+ */
+
+#define PMEVN_CASE(n, case_macro) \
+	case n: case_macro(n); break
+
+#define PMEVN_SWITCH(x, case_macro)				\
+	do {							\
+		switch (x) {					\
+		PMEVN_CASE(0,  case_macro);			\
+		PMEVN_CASE(1,  case_macro);			\
+		PMEVN_CASE(2,  case_macro);			\
+		PMEVN_CASE(3,  case_macro);			\
+		PMEVN_CASE(4,  case_macro);			\
+		PMEVN_CASE(5,  case_macro);			\
+		PMEVN_CASE(6,  case_macro);			\
+		PMEVN_CASE(7,  case_macro);			\
+		PMEVN_CASE(8,  case_macro);			\
+		PMEVN_CASE(9,  case_macro);			\
+		PMEVN_CASE(10, case_macro);			\
+		PMEVN_CASE(11, case_macro);			\
+		PMEVN_CASE(12, case_macro);			\
+		PMEVN_CASE(13, case_macro);			\
+		PMEVN_CASE(14, case_macro);			\
+		PMEVN_CASE(15, case_macro);			\
+		PMEVN_CASE(16, case_macro);			\
+		PMEVN_CASE(17, case_macro);			\
+		PMEVN_CASE(18, case_macro);			\
+		PMEVN_CASE(19, case_macro);			\
+		PMEVN_CASE(20, case_macro);			\
+		PMEVN_CASE(21, case_macro);			\
+		PMEVN_CASE(22, case_macro);			\
+		PMEVN_CASE(23, case_macro);			\
+		PMEVN_CASE(24, case_macro);			\
+		PMEVN_CASE(25, case_macro);			\
+		PMEVN_CASE(26, case_macro);			\
+		PMEVN_CASE(27, case_macro);			\
+		PMEVN_CASE(28, case_macro);			\
+		PMEVN_CASE(29, case_macro);			\
+		PMEVN_CASE(30, case_macro);			\
+		default: WARN(1, "Invalid PMEV* index\n");	\
+		}						\
+	} while (0)
+
+#define RETURN_READ_PMEVCNTRN(n) \
+	return read_sysreg(pmevcntr##n##_el0)
+static unsigned long read_pmevcntrn(int n)
+{
+	PMEVN_SWITCH(n, RETURN_READ_PMEVCNTRN);
+	return 0;
+}
+
+#define WRITE_PMEVCNTRN(n) \
+	write_sysreg(val, pmevcntr##n##_el0)
+static void write_pmevcntrn(int n, unsigned long val)
+{
+	PMEVN_SWITCH(n, WRITE_PMEVCNTRN);
+}
+
+#define WRITE_PMEVTYPERN(n) \
+	write_sysreg(val, pmevtyper##n##_el0)
+static void write_pmevtypern(int n, unsigned long val)
+{
+	PMEVN_SWITCH(n, WRITE_PMEVTYPERN);
+}
+
 static inline u32 armv8pmu_pmcr_read(void)
 {
 	return read_sysreg(pmcr_el0);
@@ -400,17 +467,11 @@ static inline int armv8pmu_counter_has_overflowed(u32 pmnc, int idx)
 	return pmnc & BIT(ARMV8_IDX_TO_COUNTER(idx));
 }
 
-static inline void armv8pmu_select_counter(int idx)
-{
-	u32 counter = ARMV8_IDX_TO_COUNTER(idx);
-	write_sysreg(counter, pmselr_el0);
-	isb();
-}
-
 static inline u32 armv8pmu_read_evcntr(int idx)
 {
-	armv8pmu_select_counter(idx);
-	return read_sysreg(pmxevcntr_el0);
+	u32 counter = ARMV8_IDX_TO_COUNTER(idx);
+
+	return read_pmevcntrn(counter);
 }
 
 static inline u64 armv8pmu_read_hw_counter(struct perf_event *event)
@@ -444,8 +505,9 @@ static u64 armv8pmu_read_counter(struct perf_event *event)
 
 static inline void armv8pmu_write_evcntr(int idx, u32 value)
 {
-	armv8pmu_select_counter(idx);
-	write_sysreg(value, pmxevcntr_el0);
+	u32 counter = ARMV8_IDX_TO_COUNTER(idx);
+
+	write_pmevcntrn(counter, value);
 }
 
 static inline void armv8pmu_write_hw_counter(struct perf_event *event,
@@ -486,9 +548,10 @@ static void armv8pmu_write_counter(struct perf_event *event, u64 value)
 
 static inline void armv8pmu_write_evtype(int idx, u32 val)
 {
-	armv8pmu_select_counter(idx);
+	u32 counter = ARMV8_IDX_TO_COUNTER(idx);
+
 	val &= ARMV8_PMU_EVTYPE_MASK;
-	write_sysreg(val, pmxevtyper_el0);
+	write_pmevtypern(counter, val);
 }
 
 static inline void armv8pmu_write_event_type(struct perf_event *event)
@@ -508,13 +571,21 @@ static inline void armv8pmu_write_event_type(struct perf_event *event)
 		armv8pmu_write_evtype(idx - 1, hwc->config_base);
 		armv8pmu_write_evtype(idx, chain_evt);
 	} else {
-		armv8pmu_write_evtype(idx, hwc->config_base);
+		if (idx == ARMV8_IDX_CYCLE_COUNTER)
+			write_sysreg(hwc->config_base, pmccfiltr_el0);
+		else
+			armv8pmu_write_evtype(idx, hwc->config_base);
 	}
 }
 
 static inline int armv8pmu_enable_counter(int idx)
 {
 	u32 counter = ARMV8_IDX_TO_COUNTER(idx);
+	/*
+         * Make sure event configuration register writes are visible before we
+         * enable the counter.
+         * */
+        isb();
 	write_sysreg(BIT(counter), pmcntenset_el0);
 	return idx;
 }
@@ -542,6 +613,11 @@ static inline int armv8pmu_disable_counter(int idx)
 {
 	u32 counter = ARMV8_IDX_TO_COUNTER(idx);
 	write_sysreg(BIT(counter), pmcntenclr_el0);
+	/*
+         * Make sure the effects of disabling the counter are visible before we
+         * start configuring the event.
+         */
+        isb();
 	return idx;
 }
 
@@ -610,15 +686,10 @@ static inline u32 armv8pmu_getreset_flags(void)
 
 static void armv8pmu_enable_event(struct perf_event *event)
 {
-	unsigned long flags;
-	struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
-	struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
-
 	/*
 	 * Enable counter and interrupt, and set the counter to count
 	 * the event that we're interested in.
 	 */
-	raw_spin_lock_irqsave(&events->pmu_lock, flags);
 
 	/*
 	 * Disable counter
@@ -626,7 +697,7 @@ static void armv8pmu_enable_event(struct perf_event *event)
 	armv8pmu_disable_event_counter(event);
 
 	/*
-	 * Set event (if destined for PMNx counters).
+	 * Set event.
 	 */
 	armv8pmu_write_event_type(event);
 
@@ -639,21 +710,10 @@ static void armv8pmu_enable_event(struct perf_event *event)
 	 * Enable counter
 	 */
 	armv8pmu_enable_event_counter(event);
-
-	raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
 }
 
 static void armv8pmu_disable_event(struct perf_event *event)
 {
-	unsigned long flags;
-	struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
-	struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
-
-	/*
-	 * Disable counter and interrupt
-	 */
-	raw_spin_lock_irqsave(&events->pmu_lock, flags);
-
 	/*
 	 * Disable counter
 	 */
@@ -663,30 +723,18 @@ static void armv8pmu_disable_event(struct perf_event *event)
 	 * Disable interrupt for this counter
 	 */
 	armv8pmu_disable_event_irq(event);
-
-	raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
 }
 
 static void armv8pmu_start(struct arm_pmu *cpu_pmu)
 {
-	unsigned long flags;
-	struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
-
-	raw_spin_lock_irqsave(&events->pmu_lock, flags);
 	/* Enable all counters */
 	armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_E);
-	raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
 }
 
 static void armv8pmu_stop(struct arm_pmu *cpu_pmu)
 {
-	unsigned long flags;
-	struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
-
-	raw_spin_lock_irqsave(&events->pmu_lock, flags);
 	/* Disable all counters */
 	armv8pmu_pmcr_write(armv8pmu_pmcr_read() & ~ARMV8_PMU_PMCR_E);
-	raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
 }
 
 static irqreturn_t armv8pmu_handle_irq(struct arm_pmu *cpu_pmu)
@@ -739,20 +787,16 @@ static irqreturn_t armv8pmu_handle_irq(struct arm_pmu *cpu_pmu)
 		if (!armpmu_event_set_period(event))
 			continue;
 
+		/*
+		 * Perf event overflow will queue the processing of the event as
+		 * an irq_work which will be taken care of in the handling of
+		 * IPI_IRQ_WORK.
+		 */
 		if (perf_event_overflow(event, &data, regs))
 			cpu_pmu->disable(event);
 	}
 	armv8pmu_start(cpu_pmu);
 
-	/*
-	 * Handle the pending perf events.
-	 *
-	 * Note: this call *must* be run with interrupts disabled. For
-	 * platforms that can have the PMU interrupts raised as an NMI, this
-	 * will not work.
-	 */
-	irq_work_run();
-
 	return IRQ_HANDLED;
 }
 
diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c
index b377872a8f9d..7fd11ef5cb8a 100644
--- a/drivers/perf/arm_pmu.c
+++ b/drivers/perf/arm_pmu.c
@@ -26,8 +26,84 @@
 
 #include <asm/irq_regs.h>
 
+static int armpmu_count_irq_users(const int irq);
+
+struct pmu_irq_ops {
+	void (*enable_pmuirq)(unsigned int irq);
+	void (*disable_pmuirq)(unsigned int irq);
+	void (*free_pmuirq)(unsigned int irq, int cpu, void __percpu *devid);
+};
+
+static void armpmu_free_pmuirq(unsigned int irq, int cpu, void __percpu *devid)
+{
+	free_irq(irq, per_cpu_ptr(devid, cpu));
+}
+
+static const struct pmu_irq_ops pmuirq_ops = {
+	.enable_pmuirq = enable_irq,
+	.disable_pmuirq = disable_irq_nosync,
+	.free_pmuirq = armpmu_free_pmuirq
+};
+
+static void armpmu_free_pmunmi(unsigned int irq, int cpu, void __percpu *devid)
+{
+	free_nmi(irq, per_cpu_ptr(devid, cpu));
+}
+
+static const struct pmu_irq_ops pmunmi_ops = {
+	.enable_pmuirq = enable_nmi,
+	.disable_pmuirq = disable_nmi_nosync,
+	.free_pmuirq = armpmu_free_pmunmi
+};
+
+static void armpmu_enable_percpu_pmuirq(unsigned int irq)
+{
+	enable_percpu_irq(irq, IRQ_TYPE_NONE);
+}
+
+static void armpmu_free_percpu_pmuirq(unsigned int irq, int cpu,
+				   void __percpu *devid)
+{
+	if (armpmu_count_irq_users(irq) == 1)
+		free_percpu_irq(irq, devid);
+}
+
+static const struct pmu_irq_ops percpu_pmuirq_ops = {
+	.enable_pmuirq = armpmu_enable_percpu_pmuirq,
+	.disable_pmuirq = disable_percpu_irq,
+	.free_pmuirq = armpmu_free_percpu_pmuirq
+};
+
+static void armpmu_enable_percpu_pmunmi(unsigned int irq)
+{
+	if (!prepare_percpu_nmi(irq))
+		enable_percpu_nmi(irq, IRQ_TYPE_NONE);
+}
+
+static void armpmu_disable_percpu_pmunmi(unsigned int irq)
+{
+	disable_percpu_nmi(irq);
+	teardown_percpu_nmi(irq);
+}
+
+static void armpmu_free_percpu_pmunmi(unsigned int irq, int cpu,
+				      void __percpu *devid)
+{
+	if (armpmu_count_irq_users(irq) == 1)
+		free_percpu_nmi(irq, devid);
+}
+
+static const struct pmu_irq_ops percpu_pmunmi_ops = {
+	.enable_pmuirq = armpmu_enable_percpu_pmunmi,
+	.disable_pmuirq = armpmu_disable_percpu_pmunmi,
+	.free_pmuirq = armpmu_free_percpu_pmunmi
+};
+
 static DEFINE_PER_CPU(struct arm_pmu *, cpu_armpmu);
 static DEFINE_PER_CPU(int, cpu_irq);
+static DEFINE_PER_CPU(const struct pmu_irq_ops *, cpu_irq_ops);
+
+static bool has_nmi;
 
 static inline u64 arm_pmu_event_max_period(struct perf_event *event)
 {
@@ -542,6 +618,23 @@ static int armpmu_count_irq_users(const int irq)
 	return count;
 }
 
+static const struct pmu_irq_ops *armpmu_find_irq_ops(int irq)
+{
+	const struct pmu_irq_ops *ops = NULL;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		if (per_cpu(cpu_irq, cpu) != irq)
+			continue;
+
+		ops = per_cpu(cpu_irq_ops, cpu);
+		if (ops)
+			break;
+	}
+
+	return ops;
+}
+
 void armpmu_free_irq(int irq, int cpu)
 {
 	if (per_cpu(cpu_irq, cpu) == 0)
@@ -549,18 +642,18 @@ void armpmu_free_irq(int irq, int cpu)
 	if (WARN_ON(irq != per_cpu(cpu_irq, cpu)))
 		return;
 
-	if (!irq_is_percpu_devid(irq))
-		free_irq(irq, per_cpu_ptr(&cpu_armpmu, cpu));
-	else if (armpmu_count_irq_users(irq) == 1)
-		free_percpu_irq(irq, &cpu_armpmu);
+	per_cpu(cpu_irq_ops, cpu)->free_pmuirq(irq, cpu, &cpu_armpmu);
 
 	per_cpu(cpu_irq, cpu) = 0;
+	per_cpu(cpu_irq_ops, cpu) = NULL;
 }
 
 int armpmu_request_irq(int irq, int cpu)
 {
 	int err = 0;
 	const irq_handler_t handler = armpmu_dispatch_irq;
+	const struct pmu_irq_ops *irq_ops;
+
 	if (!irq)
 		return 0;
 
@@ -580,17 +673,44 @@ int armpmu_request_irq(int irq, int cpu)
 			    IRQF_NO_THREAD;
 
 		irq_set_status_flags(irq, IRQ_NOAUTOEN);
-		err = request_irq(irq, handler, irq_flags, "arm-pmu",
+
+		err = request_nmi(irq, handler, irq_flags, "arm-pmu",
 				  per_cpu_ptr(&cpu_armpmu, cpu));
+
+		/* If cannot get an NMI, get a normal interrupt */
+		if (err) {
+			err = request_irq(irq, handler, irq_flags, "arm-pmu",
+					  per_cpu_ptr(&cpu_armpmu, cpu));
+			irq_ops = &pmuirq_ops;
+		} else {
+			has_nmi = true;
+			irq_ops = &pmunmi_ops;
+		}
 	} else if (armpmu_count_irq_users(irq) == 0) {
-		err = request_percpu_irq(irq, handler, "arm-pmu",
-					 &cpu_armpmu);
+		err = request_percpu_nmi(irq, handler, "arm-pmu", &cpu_armpmu);
+
+		/* If cannot get an NMI, get a normal interrupt */
+		if (err) {
+			err = request_percpu_irq(irq, handler, "arm-pmu",
+						 &cpu_armpmu);
+			irq_ops = &percpu_pmuirq_ops;
+		} else {
+			has_nmi= true;
+			irq_ops = &percpu_pmunmi_ops;
+		}
+	} else {
+		/* Per cpudevid irq was already requested by another CPU */
+		irq_ops = armpmu_find_irq_ops(irq);
+
+		if (WARN_ON(!irq_ops))
+			err = -EINVAL;
 	}
 
 	if (err)
 		goto err_out;
 
 	per_cpu(cpu_irq, cpu) = irq;
+	per_cpu(cpu_irq_ops, cpu) = irq_ops;
 	return 0;
 
 err_out:
@@ -623,12 +743,8 @@ static int arm_perf_starting_cpu(unsigned int cpu, struct hlist_node *node)
 	per_cpu(cpu_armpmu, cpu) = pmu;
 
 	irq = armpmu_get_cpu_irq(pmu, cpu);
-	if (irq) {
-		if (irq_is_percpu_devid(irq))
-			enable_percpu_irq(irq, IRQ_TYPE_NONE);
-		else
-			enable_irq(irq);
-	}
+	if (irq)
+		per_cpu(cpu_irq_ops, cpu)->enable_pmuirq(irq);
 
 	return 0;
 }
@@ -642,12 +758,8 @@ static int arm_perf_teardown_cpu(unsigned int cpu, struct hlist_node *node)
 		return 0;
 
 	irq = armpmu_get_cpu_irq(pmu, cpu);
-	if (irq) {
-		if (irq_is_percpu_devid(irq))
-			disable_percpu_irq(irq);
-		else
-			disable_irq_nosync(irq);
-	}
+	if (irq)
+		per_cpu(cpu_irq_ops, cpu)->disable_pmuirq(irq);
 
 	per_cpu(cpu_armpmu, cpu) = NULL;
 
@@ -868,8 +980,9 @@ int armpmu_register(struct arm_pmu *pmu)
 	if (!__oprofile_cpu_pmu)
 		__oprofile_cpu_pmu = pmu;
 
-	pr_info("enabled with %s PMU driver, %d counters available\n",
-		pmu->name, pmu->num_events);
+	pr_info("enabled with %s PMU driver, %d counters available%s\n",
+		pmu->name, pmu->num_events,
+		has_nmi ? ", using NMIs" : "");
 
 	return 0;
 
diff --git a/include/kvm/arm_pmu.h b/include/kvm/arm_pmu.h
index 6db030439e29..dbf4f08d42e5 100644
--- a/include/kvm/arm_pmu.h
+++ b/include/kvm/arm_pmu.h
@@ -27,6 +27,7 @@ struct kvm_pmu {
 	bool ready;
 	bool created;
 	bool irq_level;
+	struct irq_work overflow_work;
 };
 
 #define kvm_arm_pmu_v3_ready(v)		((v)->arch.pmu.ready)
diff --git a/virt/kvm/arm/pmu.c b/virt/kvm/arm/pmu.c
index 4c08fd009768..d522a1000497 100644
--- a/virt/kvm/arm/pmu.c
+++ b/virt/kvm/arm/pmu.c
@@ -259,6 +259,7 @@ void kvm_pmu_vcpu_destroy(struct kvm_vcpu *vcpu)
 
 	for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++)
 		kvm_pmu_release_perf_event(&pmu->pmc[i]);
+	irq_work_sync(&vcpu->arch.pmu.overflow_work);
 }
 
 u64 kvm_pmu_valid_counter_mask(struct kvm_vcpu *vcpu)
@@ -435,6 +436,22 @@ void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu)
 	kvm_pmu_update_state(vcpu);
 }
 
+/**
+ * When perf interrupt is an NMI, we cannot safely notify the vcpu corresponding
+ * to the event.
+ * This is why we need a callback to do it once outside of the NMI context.
+ */
+static void kvm_pmu_perf_overflow_notify_vcpu(struct irq_work *work)
+{
+	struct kvm_vcpu *vcpu;
+	struct kvm_pmu *pmu;
+
+	pmu = container_of(work, struct kvm_pmu, overflow_work);
+	vcpu = kvm_pmc_to_vcpu(pmu->pmc);
+
+	kvm_vcpu_kick(vcpu);
+}
+
 /**
  * When the perf event overflows, set the overflow status and inform the vcpu.
  */
@@ -467,7 +484,11 @@ static void kvm_pmu_perf_overflow(struct perf_event *perf_event,
 
 	if (kvm_pmu_overflow_status(vcpu)) {
 		kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
-		kvm_vcpu_kick(vcpu);
+
+		if (!in_nmi())
+			kvm_vcpu_kick(vcpu);
+		else
+			irq_work_queue(&vcpu->arch.pmu.overflow_work);
 	}
 
 	cpu_pmu->pmu.start(perf_event, PERF_EF_RELOAD);
@@ -760,6 +781,9 @@ static int kvm_arm_pmu_v3_init(struct kvm_vcpu *vcpu)
 			return ret;
 	}
 
+	init_irq_work(&vcpu->arch.pmu.overflow_work,
+		      kvm_pmu_perf_overflow_notify_vcpu);
+
 	vcpu->arch.pmu.created = true;
 	return 0;
 }