diff --git a/sys/amd64/amd64/mp_machdep.c b/sys/amd64/amd64/mp_machdep.c
--- a/sys/amd64/amd64/mp_machdep.c
+++ b/sys/amd64/amd64/mp_machdep.c
@@ -215,6 +215,8 @@
 	/* See comment in pmap_bootstrap(). */
 	pc->pc_pcid_next = PMAP_PCID_KERN + 2;
 	pc->pc_pcid_gen = 1;
+	pc->pc_kpmap_store.pm_pcid = PMAP_PCID_KERN;
+	pc->pc_kpmap_store.pm_gen = 1;
 
 	pc->pc_smp_tlb_gen = 1;
 
@@ -288,29 +290,33 @@
 	init_secondary_tail();
 }
 
-/*******************************************************************
- * local functions and data
- */
-
-#ifdef NUMA
 static void
-mp_realloc_pcpu(int cpuid, int domain)
+amd64_mp_alloc_pcpu(void)
 {
 	vm_page_t m;
-	vm_offset_t oa, na;
-
-	oa = (vm_offset_t)&__pcpu[cpuid];
-	if (vm_phys_domain(pmap_kextract(oa)) == domain)
-		return;
-	m = vm_page_alloc_noobj_domain(domain, 0);
-	if (m == NULL)
-		return;
-	na = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
-	pagecopy((void *)oa, (void *)na);
-	pmap_qenter((vm_offset_t)&__pcpu[cpuid], &m, 1);
-	/* XXX old pcpu page leaked. */
-}
+	int cpu;
+
+	/* Allocate pcpu areas to the correct domain. */
+	for (cpu = 1; cpu < mp_ncpus; cpu++) {
+#ifdef NUMA
+		m = NULL;
+		if (vm_ndomains > 1) {
+			m = vm_page_alloc_noobj_domain(
+			    acpi_pxm_get_cpu_locality(cpu_apic_ids[cpu]), 0);
+		}
+		if (m == NULL)
+#else
+			m = vm_page_alloc_noobj(0);
 #endif
+		if (m == NULL)
+			panic("cannot alloc pcpu page for cpu %d", cpu);
+		pmap_qenter((vm_offset_t)&__pcpu[cpu], &m, 1);
+	}
+}
+
+/*******************************************************************
+ * local functions and data
+ */
 
 /*
  * start each AP in our list
@@ -328,6 +334,7 @@
 	int apic_id, cpu, domain, i;
 	u_char mpbiosreason;
 
+	amd64_mp_alloc_pcpu();
 	mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
 
 	MPASS(bootMP_size <= PAGE_SIZE);
@@ -401,16 +408,6 @@
 	outb(CMOS_REG, BIOS_RESET);
 	outb(CMOS_DATA, BIOS_WARM);	/* 'warm-start' */
 
-	/* Relocate pcpu areas to the correct domain. */
-#ifdef NUMA
-	if (vm_ndomains > 1)
-		for (cpu = 1; cpu < mp_ncpus; cpu++) {
-			apic_id = cpu_apic_ids[cpu];
-			domain = acpi_pxm_get_cpu_locality(apic_id);
-			mp_realloc_pcpu(cpu, domain);
-		}
-#endif
-
 	/* start each AP */
 	domain = 0;
 	for (cpu = 1; cpu < mp_ncpus; cpu++) {
diff --git a/sys/amd64/amd64/pmap.c b/sys/amd64/amd64/pmap.c
--- a/sys/amd64/amd64/pmap.c
+++ b/sys/amd64/amd64/pmap.c
@@ -1902,7 +1902,7 @@
 	vm_offset_t va;
 	pt_entry_t *pte, *pcpu_pte;
 	struct region_descriptor r_gdt;
-	uint64_t cr4, pcpu_phys;
+	uint64_t cr4, pcpu0_phys;
 	u_long res;
 	int i;
 
@@ -1917,7 +1917,7 @@
 	 */
 	create_pagetables(firstaddr);
 
-	pcpu_phys = allocpages(firstaddr, MAXCPU);
+	pcpu0_phys = allocpages(firstaddr, 1);
 
 	/*
 	 * Add a physical memory segment (vm_phys_seg) corresponding to the
@@ -1995,10 +1995,15 @@
 	SYSMAP(struct pcpu *, pcpu_pte, __pcpu, MAXCPU);
 	virtual_avail = va;
 
-	for (i = 0; i < MAXCPU; i++) {
-		pcpu_pte[i] = (pcpu_phys + ptoa(i)) | X86_PG_V | X86_PG_RW |
-		    pg_g | pg_nx | X86_PG_M | X86_PG_A;
-	}
+	/*
+	 * Map the BSP PCPU now, rest PCPUs are mapped by
+	 * amd64_mp_alloc_pcpu()/start_all_aps() when we know the
+	 * number of CPUs and NUMA affinity.
+	 */
+	pcpu_pte[0] = pcpu0_phys | X86_PG_V | X86_PG_RW | pg_g | pg_nx |
+	    X86_PG_M | X86_PG_A;
+	for (i = 1; i < MAXCPU; i++)
+		pcpu_pte[i] = 0;
 
 	/*
 	 * Re-initialize PCPU area for BSP after switching.
@@ -2036,13 +2041,9 @@
 	if (pmap_pcid_enabled) {
 		kernel_pmap->pm_pcidp = (void *)(uintptr_t)
 		    offsetof(struct pcpu, pc_kpmap_store);
-		for (i = 0; i < MAXCPU; i++) {
-			struct pmap_pcid *pcidp;
 
-			pcidp = zpcpu_get_cpu(kernel_pmap->pm_pcidp, i);
-			pcidp->pm_pcid = PMAP_PCID_KERN;
-			pcidp->pm_gen = 1;
-		}
+		__pcpu[0].pc_kpmap_store.pm_pcid = PMAP_PCID_KERN;
+		__pcpu[0].pc_kpmap_store.pm_gen = 1;
 
 		/*
 		 * PMAP_PCID_KERN + 1 is used for initialization of