Index: sys/amd64/amd64/mp_machdep.c
===================================================================
--- sys/amd64/amd64/mp_machdep.c	(revision 223435)
+++ sys/amd64/amd64/mp_machdep.c	(working copy)
@@ -604,10 +604,10 @@
 void
 init_secondary(void)
 {
-	cpuset_t tcpuset;
 	struct pcpu *pc;
 	struct nmi_pcpu *np;
 	u_int64_t msr, cr0;
+	u_int cpuid;
 	int cpu, gsel_tss, x;
 	struct region_descriptor ap_gdt;
 
@@ -711,8 +711,9 @@
 	fpuinit();
 
 	/* A quick check from sanity claus */
+	cpuid = PCPU_GET(cpuid);
 	if (PCPU_GET(apic_id) != lapic_id()) {
-		printf("SMP: cpuid = %d\n", PCPU_GET(cpuid));
+		printf("SMP: cpuid = %d\n", cpuid);
 		printf("SMP: actual apic_id = %d\n", lapic_id());
 		printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
 		panic("cpuid mismatch! boom!!");
@@ -734,14 +735,13 @@
 
 	smp_cpus++;
 
-	CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", PCPU_GET(cpuid));
-	printf("SMP: AP CPU #%d Launched!\n", PCPU_GET(cpuid));
-	tcpuset = PCPU_GET(cpumask);
+	CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
+	printf("SMP: AP CPU #%d Launched!\n", cpuid);
 
 	/* Determine if we are a logical CPU. */
 	/* XXX Calculation depends on cpu_logical being a power of 2, e.g. 2 */
 	if (cpu_logical > 1 && PCPU_GET(apic_id) % cpu_logical != 0)
-		CPU_OR(&logical_cpus_mask, &tcpuset);
+		CPU_SET(cpuid, &logical_cpus_mask);
 
 	if (bootverbose)
 		lapic_dump("AP");
@@ -1138,9 +1138,7 @@
 		if (othercpus < 1)
 			return;
 	} else {
-		sched_pin();
-		CPU_NAND(&mask, PCPU_PTR(cpumask));
-		sched_unpin();
+		CPU_CLR(PCPU_GET(cpuid), &mask);
 		if (CPU_EMPTY(&mask))
 			return;
 	}
@@ -1362,7 +1360,7 @@
 int
 ipi_nmi_handler()
 {
-	cpuset_t cpumask;
+	u_int cpuid;
 
 	/*
 	 * As long as there is not a simple way to know about a NMI's
@@ -1370,13 +1368,11 @@
 	 * the global pending bitword an IPI_STOP_HARD has been issued
 	 * and should be handled.
 	 */
-	sched_pin();
-	cpumask = PCPU_GET(cpumask);
-	sched_unpin();
-	if (!CPU_OVERLAP(&ipi_nmi_pending, &cpumask))
+	cpuid = PCPU_GET(cpuid);
+	if (!CPU_ISSET(cpuid, &ipi_nmi_pending))
 		return (1);
 
-	CPU_NAND_ATOMIC(&ipi_nmi_pending, &cpumask);
+	CPU_CLR_ATOMIC(cpuid, &ipi_nmi_pending);
 	cpustop_handler();
 	return (0);
 }
@@ -1388,25 +1384,21 @@
 void
 cpustop_handler(void)
 {
-	cpuset_t cpumask;
 	u_int cpu;
 
-	sched_pin();
 	cpu = PCPU_GET(cpuid);
-	cpumask = PCPU_GET(cpumask);
-	sched_unpin();
 
 	savectx(&stoppcbs[cpu]);
 
 	/* Indicate that we are stopped */
-	CPU_OR_ATOMIC(&stopped_cpus, &cpumask);
+	CPU_SET_ATOMIC(cpu, &stopped_cpus);
 
 	/* Wait for restart */
-	while (!CPU_OVERLAP(&started_cpus, &cpumask))
+	while (!CPU_ISSET(cpu, &started_cpus))
 	    ia32_pause();
 
-	CPU_NAND_ATOMIC(&started_cpus, &cpumask);
-	CPU_NAND_ATOMIC(&stopped_cpus, &cpumask);
+	CPU_CLR_ATOMIC(cpu, &started_cpus);
+	CPU_CLR_ATOMIC(cpu, &stopped_cpus);
 
 	if (cpu == 0 && cpustop_restartfunc != NULL) {
 		cpustop_restartfunc();
@@ -1421,19 +1413,17 @@
 void
 cpususpend_handler(void)
 {
-	cpuset_t cpumask;
 	register_t cr3, rf;
 	u_int cpu;
 
 	cpu = PCPU_GET(cpuid);
-	cpumask = PCPU_GET(cpumask);
 
 	rf = intr_disable();
 	cr3 = rcr3();
 
 	if (savectx(susppcbs[cpu])) {
 		wbinvd();
-		CPU_OR_ATOMIC(&stopped_cpus, &cpumask);
+		CPU_SET_ATOMIC(cpu, &stopped_cpus);
 	} else {
 		pmap_init_pat();
 		PCPU_SET(switchtime, 0);
@@ -1441,11 +1431,11 @@
 	}
 
 	/* Wait for resume */
-	while (!CPU_OVERLAP(&started_cpus, &cpumask))
+	while (!CPU_ISSET(cpu, &started_cpus))
 		ia32_pause();
 
-	CPU_NAND_ATOMIC(&started_cpus, &cpumask);
-	CPU_NAND_ATOMIC(&stopped_cpus, &cpumask);
+	CPU_CLR_ATOMIC(cpu, &started_cpus);
+	CPU_CLR_ATOMIC(cpu, &stopped_cpus);
 
 	/* Restore CR3 and enable interrupts */
 	load_cr3(cr3);
Index: sys/amd64/amd64/pmap.c
===================================================================
--- sys/amd64/amd64/pmap.c	(revision 223435)
+++ sys/amd64/amd64/pmap.c	(working copy)
@@ -925,17 +925,18 @@
 void
 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
 {
-	cpuset_t cpumask, other_cpus;
+	cpuset_t other_cpus;
+	u_int cpuid;
 
 	sched_pin();
 	if (pmap == kernel_pmap || !CPU_CMP(&pmap->pm_active, &all_cpus)) {
 		invlpg(va);
 		smp_invlpg(va);
 	} else {
-		cpumask = PCPU_GET(cpumask);
+		cpuid = PCPU_GET(cpuid);
 		other_cpus = all_cpus;
-		CPU_CLR(PCPU_GET(cpuid), &other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		CPU_CLR(cpuid, &other_cpus);
+		if (CPU_ISSET(cpuid, &pmap->pm_active))
 			invlpg(va);
 		CPU_AND(&other_cpus, &pmap->pm_active);
 		if (!CPU_EMPTY(&other_cpus))
@@ -947,8 +948,9 @@
 void
 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
 {
-	cpuset_t cpumask, other_cpus;
+	cpuset_t other_cpus;
 	vm_offset_t addr;
+	u_int cpuid;
 
 	sched_pin();
 	if (pmap == kernel_pmap || !CPU_CMP(&pmap->pm_active, &all_cpus)) {
@@ -956,10 +958,10 @@
 			invlpg(addr);
 		smp_invlpg_range(sva, eva);
 	} else {
-		cpumask = PCPU_GET(cpumask);
+		cpuid = PCPU_GET(cpuid);
 		other_cpus = all_cpus;
-		CPU_CLR(PCPU_GET(cpuid), &other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		CPU_CLR(cpuid, &other_cpus);
+		if (CPU_ISSET(cpuid, &pmap->pm_active))
 			for (addr = sva; addr < eva; addr += PAGE_SIZE)
 				invlpg(addr);
 		CPU_AND(&other_cpus, &pmap->pm_active);
@@ -972,17 +974,18 @@
 void
 pmap_invalidate_all(pmap_t pmap)
 {
-	cpuset_t cpumask, other_cpus;
+	cpuset_t other_cpus;
+	u_int cpuid;
 
 	sched_pin();
 	if (pmap == kernel_pmap || !CPU_CMP(&pmap->pm_active, &all_cpus)) {
 		invltlb();
 		smp_invltlb();
 	} else {
-		cpumask = PCPU_GET(cpumask);
+		cpuid = PCPU_GET(cpuid);
 		other_cpus = all_cpus;
-		CPU_CLR(PCPU_GET(cpuid), &other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		CPU_CLR(cpuid, &other_cpus);
+		if (CPU_ISSET(cpuid, &pmap->pm_active))
 			invltlb();
 		CPU_AND(&other_cpus, &pmap->pm_active);
 		if (!CPU_EMPTY(&other_cpus))
@@ -1002,11 +1005,11 @@
 }
 
 struct pde_action {
-	cpuset_t store;		/* processor that updates the PDE */
 	cpuset_t invalidate;	/* processors that invalidate their TLB */
 	vm_offset_t va;
 	pd_entry_t *pde;
 	pd_entry_t newpde;
+	u_int store;		/* processor that updates the PDE */
 };
 
 static void
@@ -1014,12 +1017,8 @@
 {
 	struct pde_action *act = arg;
 
-	sched_pin();
-	if (!CPU_CMP(&act->store, PCPU_PTR(cpumask))) {
-		sched_unpin();
+	if (act->store == PCPU_GET(cpuid))
 		pde_store(act->pde, act->newpde);
-	} else
-		sched_unpin();
 }
 
 static void
@@ -1027,12 +1026,8 @@
 {
 	struct pde_action *act = arg;
 
-	sched_pin();
-	if (CPU_OVERLAP(&act->invalidate, PCPU_PTR(cpumask))) {
-		sched_unpin();
+	if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate))
 		pmap_update_pde_invalidate(act->va, act->newpde);
-	} else
-		sched_unpin();
 }
 
 /*
@@ -1047,29 +1042,30 @@
 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
 {
 	struct pde_action act;
-	cpuset_t active, cpumask, other_cpus;
+	cpuset_t active, other_cpus;
+	u_int cpuid;
 
 	sched_pin();
-	cpumask = PCPU_GET(cpumask);
+	cpuid = PCPU_GET(cpuid);
 	other_cpus = all_cpus;
-	CPU_CLR(PCPU_GET(cpuid), &other_cpus);
+	CPU_CLR(cpuid, &other_cpus);
 	if (pmap == kernel_pmap)
 		active = all_cpus;
 	else
 		active = pmap->pm_active;
 	if (CPU_OVERLAP(&active, &other_cpus)) { 
-		act.store = cpumask;
+		act.store = cpuid;
 		act.invalidate = active;
 		act.va = va;
 		act.pde = pde;
 		act.newpde = newpde;
-		CPU_OR(&cpumask, &active);
-		smp_rendezvous_cpus(cpumask,
+		CPU_SET(cpuid, &active);
+		smp_rendezvous_cpus(active,
 		    smp_no_rendevous_barrier, pmap_update_pde_action,
 		    pmap_update_pde_teardown, &act);
 	} else {
 		pde_store(pde, newpde);
-		if (CPU_OVERLAP(&active, &cpumask))
+		if (CPU_ISSET(cpuid, &active))
 			pmap_update_pde_invalidate(va, newpde);
 	}
 	sched_unpin();
@@ -5100,16 +5096,18 @@
 {
 	pmap_t	pmap, oldpmap;
 	u_int64_t  cr3;
+	u_int cpuid;
 
 	critical_enter();
 	pmap = vmspace_pmap(td->td_proc->p_vmspace);
 	oldpmap = PCPU_GET(curpmap);
+	cpuid = PCPU_GET(cpuid);
 #ifdef SMP
-	CPU_NAND_ATOMIC(&oldpmap->pm_active, PCPU_PTR(cpumask));
-	CPU_OR_ATOMIC(&pmap->pm_active, PCPU_PTR(cpumask));
+	CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active);
+	CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
 #else
-	CPU_NAND(&oldpmap->pm_active, PCPU_PTR(cpumask));
-	CPU_OR(&pmap->pm_active, PCPU_PTR(cpumask));
+	CPU_CLR(cpuid, &oldpmap->pm_active);
+	CPU_SET(cpuid, &pmap->pm_active);
 #endif
 	cr3 = DMAP_TO_PHYS((vm_offset_t)pmap->pm_pml4);
 	td->td_pcb->pcb_cr3 = cr3;