Index: sys/arm64/arm64/machdep.c
===================================================================
--- sys/arm64/arm64/machdep.c
+++ sys/arm64/arm64/machdep.c
@@ -105,7 +105,8 @@
 
 enum arm64_bus arm64_bus_method = ARM64_BUS_NONE;
 
-struct pcpu __pcpu[MAXCPU];
+struct pcpu pcpu0;
+struct pcpu *__pcpu[MAXCPU];
 
 static struct trapframe proc0_tf;
 
@@ -925,7 +926,10 @@
 static void
 init_proc0(vm_offset_t kstack)
 {
-	struct pcpu *pcpup = &__pcpu[0];
+	struct pcpu *pcpup;
+
+	pcpup = __pcpu[0];
+	MPASS(pcpup != NULL);
 
 	proc_linkup0(&proc0, &thread0);
 	thread0.td_kstack = kstack;
@@ -1357,7 +1361,7 @@
 	physmem_exclude_region(0x80000000, 0x10000, EXFLAG_NOALLOC);
 
 	/* Set the pcpu data, this is needed by pmap_bootstrap */
-	pcpup = &__pcpu[0];
+	pcpup = __pcpu[0] = &pcpu0;
 	pcpu_init(pcpup, 0, sizeof(struct pcpu));
 
 	/*
Index: sys/arm64/arm64/mp_machdep.c
===================================================================
--- sys/arm64/arm64/mp_machdep.c
+++ sys/arm64/arm64/mp_machdep.c
@@ -40,6 +40,7 @@
 #include <sys/bus.h>
 #include <sys/cpu.h>
 #include <sys/csan.h>
+#include <sys/domainset.h>
 #include <sys/kernel.h>
 #include <sys/ktr.h>
 #include <sys/malloc.h>
@@ -152,7 +153,7 @@
 is_boot_cpu(uint64_t target_cpu)
 {
 
-	return (__pcpu[0].pc_mpidr == (target_cpu & CPU_AFF_MASK));
+	return (__pcpu[0]->pc_mpidr == (target_cpu & CPU_AFF_MASK));
 }
 
 static void
@@ -216,15 +217,17 @@
 	 * they can pass random value in it.
 	 */
 	mpidr = READ_SPECIALREG(mpidr_el1) & CPU_AFF_MASK;
-	if  (cpu >= MAXCPU || __pcpu[cpu].pc_mpidr != mpidr) {
+	if (cpu >= MAXCPU || __pcpu[cpu] == NULL ||
+	    __pcpu[cpu]->pc_mpidr != mpidr) {
 		for (cpu = 0; cpu < mp_maxid; cpu++)
-			if (__pcpu[cpu].pc_mpidr == mpidr)
+			if (__pcpu[cpu] != NULL &&
+			    __pcpu[cpu]->pc_mpidr == mpidr)
 				break;
 		if ( cpu >= MAXCPU)
 			panic("MPIDR for this CPU is not in pcpu table");
 	}
 
-	pcpup = &__pcpu[cpu];
+	pcpup = __pcpu[cpu];
 	/*
 	 * Set the pcpu pointer with a backup in tpidr_el1 to be
 	 * loaded when entering the kernel from userland.
@@ -485,7 +488,7 @@
  * do nothing. Returns true if the CPU is present and running.
  */
 static bool
-start_cpu(u_int cpuid, uint64_t target_cpu)
+start_cpu(u_int cpuid, uint64_t target_cpu, int domain)
 {
 	struct pcpu *pcpup;
 	vm_paddr_t pa;
@@ -502,14 +505,17 @@
 
 	KASSERT(cpuid < MAXCPU, ("Too many CPUs"));
 
-	pcpup = &__pcpu[cpuid];
+	pcpup = __pcpu[cpuid] = (void *)kmem_malloc_domainset(
+	    DOMAINSET_PREF(domain), sizeof(*pcpup), M_WAITOK | M_ZERO);;
 	pcpu_init(pcpup, cpuid, sizeof(struct pcpu));
 	pcpup->pc_mpidr = target_cpu & CPU_AFF_MASK;
 
-	dpcpu[cpuid - 1] = (void *)kmem_malloc(DPCPU_SIZE, M_WAITOK | M_ZERO);
+	dpcpu[cpuid - 1] = (void *)kmem_malloc_domainset(
+	    DOMAINSET_PREF(domain), DPCPU_SIZE, M_WAITOK | M_ZERO);
 	dpcpu_init(dpcpu[cpuid - 1], cpuid);
 
-	bootstacks[cpuid] = (void *)kmem_malloc(PAGE_SIZE, M_WAITOK | M_ZERO);
+	bootstacks[cpuid] = (void *)kmem_malloc_domainset(
+	    DOMAINSET_PREF(domain), PAGE_SIZE, M_WAITOK | M_ZERO);
 
 	naps = atomic_load_int(&aps_started);
 	bootstack = (char *)bootstacks[cpuid] + PAGE_SIZE;
@@ -552,6 +558,7 @@
 	ACPI_MADT_GENERIC_INTERRUPT *intr;
 	u_int *cpuid;
 	u_int id;
+	int domain;
 
 	switch(entry->Type) {
 	case ACPI_MADT_TYPE_GENERIC_INTERRUPT:
@@ -563,8 +570,14 @@
 		else
 			id = *cpuid;
 
-		if (start_cpu(id, intr->ArmMpidr)) {
-			__pcpu[id].pc_acpi_id = intr->Uid;
+#ifdef NUMA
+		domain = acpi_pxm_get_cpu_locality(*cpuid);
+#else
+		domain = 0;
+#endif
+		if (start_cpu(id, intr->ArmMpidr, domain)) {
+			MPASS(__pcpu[id] != NULL);
+			__pcpu[id]->pc_acpi_id = intr->Uid;
 			/*
 			 * Don't increment for the boot CPU, its CPU ID is
 			 * reserved.
@@ -627,7 +640,7 @@
 	else
 		cpuid = fdt_cpuid;
 
-	if (!start_cpu(cpuid, target_cpu))
+	if (!start_cpu(cpuid, target_cpu, 0))
 		return (FALSE);
 
 	/*
@@ -640,7 +653,7 @@
 	if (vm_ndomains == 1 ||
 	    OF_getencprop(node, "numa-node-id", &domain, sizeof(domain)) <= 0)
 		domain = 0;
-	__pcpu[cpuid].pc_domain = domain;
+	__pcpu[cpuid]->pc_domain = domain;
 	if (domain < MAXMEMDOM)
 		CPU_SET(cpuid, &cpuset_domain[domain]);
 	return (TRUE);
@@ -671,7 +684,7 @@
 
 	/* CPU 0 is always boot CPU. */
 	CPU_SET(0, &all_cpus);
-	__pcpu[0].pc_mpidr = READ_SPECIALREG(mpidr_el1) & CPU_AFF_MASK;
+	__pcpu[0]->pc_mpidr = READ_SPECIALREG(mpidr_el1) & CPU_AFF_MASK;
 
 	switch(arm64_bus_method) {
 #ifdef DEV_ACPI
Index: sys/arm64/include/counter.h
===================================================================
--- sys/arm64/include/counter.h
+++ sys/arm64/include/counter.h
@@ -32,7 +32,7 @@
 #include <sys/pcpu.h>
 #include <machine/atomic.h>
 
-#define	EARLY_COUNTER	&__pcpu[0].pc_early_dummy_counter
+#define	EARLY_COUNTER	&pcpu0.pc_early_dummy_counter
 
 #define	counter_enter()	do {} while (0)
 #define	counter_exit()	do {} while (0)
Index: sys/arm64/include/pcpu_aux.h
===================================================================
--- sys/arm64/include/pcpu_aux.h
+++ sys/arm64/include/pcpu_aux.h
@@ -47,6 +47,7 @@
  */
 _Static_assert(PAGE_SIZE % sizeof(struct pcpu) == 0, "fix pcpu size");
 
-extern struct pcpu __pcpu[];
+extern struct pcpu pcpu0;
+extern struct pcpu *__pcpu[];
 
 #endif	/* _MACHINE_PCPU_AUX_H_ */