With retirement of cpumask_t and usage of cpuset_t for representing a

mask of CPUs, pc_other_cpus and pc_cpumask become highly inefficient.

Remove them and replace their usage with custom pc_cpuid magic (as,
atm, pc_cpumask can be easilly represented by (1 << pc_cpuid) and
pc_other_cpus by (all_cpus & ~(1 << pc_cpuid))).

This change is not targeted for MFC because of struct pcpu members
removal and dependency by cpumask_t retirement.

MD review by:	marcel, marius, alc
Tested by:	pluknet
MD testing by:	marcel, marius, gonzo, andreast

5 files changed, 125 insertions, 152 deletions

diff --git a/sys/i386/i386/mp_machdep.c b/sys/i386/i386/mp_machdep.c
index 78c90c0..0900901 100644
--- a/sys/i386/i386/mp_machdep.c
+++ b/sys/i386/i386/mp_machdep.c
@@ -658,12 +658,11 @@ cpu_mp_announce(void)
 void
 init_secondary(void)
 {
-	cpuset_t tcpuset, tallcpus;
 	struct pcpu *pc;
 	vm_offset_t addr;
 	int	gsel_tss;
 	int	x, myid;
-	u_int	cr0;
+	u_int	cpuid, cr0;
 
 	/* bootAP is set in start_ap() to our ID. */
 	myid = bootAP;
@@ -758,8 +757,9 @@ init_secondary(void)
 #endif
 
 	/* 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!!");
@@ -781,19 +781,13 @@ init_secondary(void)
 
 	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);
-
-	/* Build our map of 'other' CPUs. */
-	tallcpus = all_cpus;
-	CPU_NAND(&tallcpus, &tcpuset);
-	PCPU_SET(other_cpus, tallcpus);
+		CPU_SET(cpuid, &logical_cpus_mask);
 
 	if (bootverbose)
 		lapic_dump("AP");
@@ -934,7 +928,6 @@ assign_cpu_ids(void)
 static int
 start_all_aps(void)
 {
-	cpuset_t tallcpus;
 #ifndef PC98
 	u_char mpbiosreason;
 #endif
@@ -997,11 +990,6 @@ start_all_aps(void)
 		CPU_SET(cpu, &all_cpus);	/* record AP in CPU map */
 	}
 
-	/* build our map of 'other' CPUs */
-	tallcpus = all_cpus;
-	CPU_NAND(&tallcpus, PCPU_PTR(cpumask));
-	PCPU_SET(other_cpus, tallcpus);
-
 	/* restore the warmstart vector */
 	*(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
 
@@ -1253,9 +1241,7 @@ smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, vm_offset_t addr1, vm_of
 		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;
 	}
@@ -1452,11 +1438,12 @@ ipi_cpu(int cpu, u_int ipi)
 void
 ipi_all_but_self(u_int ipi)
 {
+	cpuset_t other_cpus;
 
-	sched_pin();
+	other_cpus = all_cpus;
+	CPU_CLR(PCPU_GET(cpuid), &other_cpus);
 	if (IPI_IS_BITMAPED(ipi)) {
-		ipi_selected(PCPU_GET(other_cpus), ipi);
-		sched_unpin();
+		ipi_selected(other_cpus, ipi);
 		return;
 	}
 
@@ -1466,8 +1453,7 @@ ipi_all_but_self(u_int ipi)
 	 * Set the mask of receiving CPUs for this purpose.
 	 */
 	if (ipi == IPI_STOP_HARD)
-		CPU_OR_ATOMIC(&ipi_nmi_pending, PCPU_PTR(other_cpus));
-	sched_unpin();
+		CPU_OR_ATOMIC(&ipi_nmi_pending, &other_cpus);
 
 	CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
 	lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
@@ -1476,7 +1462,7 @@ ipi_all_but_self(u_int ipi)
 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
@@ -1484,13 +1470,11 @@ ipi_nmi_handler()
 	 * 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);
 }
@@ -1502,25 +1486,21 @@ ipi_nmi_handler()
 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();
diff --git a/sys/i386/i386/pmap.c b/sys/i386/i386/pmap.c
index e63cddd..002e529 100644
--- a/sys/i386/i386/pmap.c
+++ b/sys/i386/i386/pmap.c
@@ -932,16 +932,18 @@ pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde)
 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);
-		other_cpus = PCPU_GET(other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		cpuid = PCPU_GET(cpuid);
+		other_cpus = all_cpus;
+		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))
@@ -953,8 +955,9 @@ pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
 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)) {
@@ -962,9 +965,10 @@ pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
 			invlpg(addr);
 		smp_invlpg_range(sva, eva);
 	} else {
-		cpumask = PCPU_GET(cpumask);
-		other_cpus = PCPU_GET(other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		cpuid = PCPU_GET(cpuid);
+		other_cpus = all_cpus;
+		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);
@@ -977,16 +981,18 @@ pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
 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);
-		other_cpus = PCPU_GET(other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		cpuid = PCPU_GET(cpuid);
+		other_cpus = all_cpus;
+		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))
@@ -1006,11 +1012,11 @@ pmap_invalidate_cache(void)
 }
 
 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
@@ -1020,9 +1026,7 @@ pmap_update_pde_kernel(void *arg)
 	pd_entry_t *pde;
 	pmap_t pmap;
 
-	sched_pin();
-	if (!CPU_CMP(&act->store, PCPU_PTR(cpumask))) {
-		sched_unpin();
+	if (act->store == PCPU_GET(cpuid)) {
 
 		/*
 		 * Elsewhere, this operation requires allpmaps_lock for
@@ -1033,8 +1037,7 @@ pmap_update_pde_kernel(void *arg)
 			pde = pmap_pde(pmap, act->va);
 			pde_store(pde, act->newpde);
 		}
-	} else
-		sched_unpin();
+	}
 }
 
 static void
@@ -1042,12 +1045,8 @@ pmap_update_pde_user(void *arg)
 {
 	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
@@ -1055,12 +1054,8 @@ pmap_update_pde_teardown(void *arg)
 {
 	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();
 }
 
 /*
@@ -1075,23 +1070,25 @@ static void
 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);
-	other_cpus = PCPU_GET(other_cpus);
+	cpuid = PCPU_GET(cpuid);
+	other_cpus = all_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 == kernel_pmap ?
 		    pmap_update_pde_kernel : pmap_update_pde_user,
 		    pmap_update_pde_teardown, &act);
@@ -1100,7 +1097,7 @@ pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
 			pmap_kenter_pde(va, newpde);
 		else
 			pde_store(pde, newpde);
-		if (CPU_OVERLAP(&active, &cpumask))
+		if (CPU_ISSET(cpuid, &active))
 			pmap_update_pde_invalidate(va, newpde);
 	}
 	sched_unpin();
@@ -1925,12 +1922,12 @@ pmap_lazyfix_action(void)
 }
 
 static void
-pmap_lazyfix_self(cpuset_t mymask)
+pmap_lazyfix_self(u_int cpuid)
 {
 
 	if (rcr3() == lazyptd)
 		load_cr3(PCPU_GET(curpcb)->pcb_cr3);
-	CPU_NAND_ATOMIC(lazymask, &mymask);
+	CPU_CLR_ATOMIC(cpuid, lazymask);
 }
 
 
@@ -1938,7 +1935,7 @@ static void
 pmap_lazyfix(pmap_t pmap)
 {
 	cpuset_t mymask, mask;
-	u_int spins;
+	u_int cpuid, spins;
 	int lsb;
 
 	mask = pmap->pm_active;
@@ -1956,10 +1953,13 @@ pmap_lazyfix(pmap_t pmap)
 #else
 		lazyptd = vtophys(pmap->pm_pdir);
 #endif
-		mymask = PCPU_GET(cpumask);
+		cpuid = PCPU_GET(cpuid);
+
+		/* Use a cpuset just for having an easy check. */
+		CPU_SETOF(cpuid, &mymask);
 		if (!CPU_CMP(&mask, &mymask)) {
 			lazymask = &pmap->pm_active;
-			pmap_lazyfix_self(mymask);
+			pmap_lazyfix_self(cpuid);
 		} else {
 			atomic_store_rel_int((u_int *)&lazymask,
 			    (u_int)&pmap->pm_active);
@@ -5098,17 +5098,19 @@ void
 pmap_activate(struct thread *td)
 {
 	pmap_t	pmap, oldpmap;
+	u_int	cpuid;
 	u_int32_t  cr3;
 
 	critical_enter();
 	pmap = vmspace_pmap(td->td_proc->p_vmspace);
 	oldpmap = PCPU_GET(curpmap);
+	cpuid = PCPU_GET(cpuid);
 #if defined(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
 #ifdef PAE
 	cr3 = vtophys(pmap->pm_pdpt);
diff --git a/sys/i386/i386/vm_machdep.c b/sys/i386/i386/vm_machdep.c
index a084e09..0a7bc21 100644
--- a/sys/i386/i386/vm_machdep.c
+++ b/sys/i386/i386/vm_machdep.c
@@ -602,8 +602,8 @@ cpu_reset()
 	u_int cnt;
 
 	if (smp_active) {
-		sched_pin();
-		map = PCPU_GET(other_cpus);
+		map = all_cpus;
+		CPU_CLR(PCPU_GET(cpuid), &map);
 		CPU_NAND(&map, &stopped_cpus);
 		if (!CPU_EMPTY(&map)) {
 			printf("cpu_reset: Stopping other CPUs\n");
@@ -612,7 +612,6 @@ cpu_reset()
 
 		if (PCPU_GET(cpuid) != 0) {
 			cpu_reset_proxyid = PCPU_GET(cpuid);
-			sched_unpin();
 			cpustop_restartfunc = cpu_reset_proxy;
 			cpu_reset_proxy_active = 0;
 			printf("cpu_reset: Restarting BSP\n");
@@ -632,8 +631,7 @@ cpu_reset()
 
 			while (1);
 			/* NOTREACHED */
-		} else
-			sched_unpin();
+		}
 
 		DELAY(1000000);
 	}
@@ -802,7 +800,8 @@ sf_buf_alloc(struct vm_page *m, int flags)
 	struct sf_head *hash_list;
 	struct sf_buf *sf;
 #ifdef SMP
-	cpuset_t cpumask, other_cpus;
+	cpuset_t other_cpus;
+	u_int cpuid;
 #endif
 	int error;
 
@@ -877,13 +876,14 @@ sf_buf_alloc(struct vm_page *m, int flags)
 		CPU_ZERO(&sf->cpumask);
 shootdown:
 	sched_pin();
-	cpumask = PCPU_GET(cpumask);
-	if (!CPU_OVERLAP(&cpumask, &sf->cpumask)) {
-		CPU_OR(&sf->cpumask, &cpumask);
+	cpuid = PCPU_GET(cpuid);
+	if (!CPU_ISSET(cpuid, &sf->cpumask)) {
+		CPU_SET(cpuid, &sf->cpumask);
 		invlpg(sf->kva);
 	}
 	if ((flags & SFB_CPUPRIVATE) == 0) {
-		other_cpus = PCPU_GET(other_cpus);
+		other_cpus = all_cpus;
+		CPU_CLR(cpuid, &other_cpus);
 		CPU_NAND(&other_cpus, &sf->cpumask);
 		if (!CPU_EMPTY(&other_cpus)) {
 			CPU_OR(&sf->cpumask, &other_cpus);
diff --git a/sys/i386/xen/mp_machdep.c b/sys/i386/xen/mp_machdep.c
index 2d05596..a0ef1e8 100644
--- a/sys/i386/xen/mp_machdep.c
+++ b/sys/i386/xen/mp_machdep.c
@@ -523,8 +523,8 @@ xen_smp_intr_init_cpus(void *unused)
 void
 init_secondary(void)
 {
-	cpuset_t tcpuset, tallcpus;
 	vm_offset_t addr;
+	u_int	cpuid;
 	int	gsel_tss;
 	
 	
@@ -601,23 +601,18 @@ init_secondary(void)
 #endif
 	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);
+	cpuid = PCPU_GET(cpuid);
+	CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
+	printf("SMP: AP CPU #%d Launched!\n", cpuid);
 
 	/* Determine if we are a logical CPU. */
 	if (logical_cpus > 1 && PCPU_GET(apic_id) % logical_cpus != 0)
-		CPU_OR(&logical_cpus_mask, &tcpuset);
+		CPU_SET(cpuid, &logical_cpus_mask);
 	
 	/* Determine if we are a hyperthread. */
 	if (hyperthreading_cpus > 1 &&
 	    PCPU_GET(apic_id) % hyperthreading_cpus != 0)
-		CPU_OR(&hyperthreading_cpus_mask, &tcpuset);
-
-	/* Build our map of 'other' CPUs. */
-	tallcpus = all_cpus;
-	CPU_NAND(&tallcpus, &tcpuset);
-	PCPU_SET(other_cpus, tallcpus);
+		CPU_SET(cpuid, &hyperthreading_cpus_mask);
 #if 0
 	if (bootverbose)
 		lapic_dump("AP");
@@ -731,7 +726,6 @@ assign_cpu_ids(void)
 int
 start_all_aps(void)
 {
-	cpuset_t tallcpus;
 	int x,apic_id, cpu;
 	struct pcpu *pc;
 	
@@ -789,11 +783,6 @@ start_all_aps(void)
 	}
 	
 
-	/* build our map of 'other' CPUs */
-	tallcpus = all_cpus;
-	CPU_NAND(&tallcpus, PCPU_PTR(cpumask));
-	PCPU_SET(other_cpus, tallcpus);
-
 	pmap_invalidate_range(kernel_pmap, 0, NKPT * NBPDR - 1);
 	
 	/* number of APs actually started */
@@ -1031,9 +1020,7 @@ smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, vm_offset_t addr1, vm_of
 		if (othercpus < 1)
 			return;
 	} else {
-		critical_enter();
-		CPU_NAND(&mask, PCPU_PTR(cpumask));
-		critical_exit();
+		CPU_CLR(PCPU_GET(cpuid), &mask);
 		if (CPU_EMPTY(&mask))
 			return;
 	}
@@ -1184,9 +1171,8 @@ ipi_all_but_self(u_int ipi)
 	 * of help in order to understand what is the source.
 	 * Set the mask of receiving CPUs for this purpose.
 	 */
-	sched_pin();
-	other_cpus = PCPU_GET(other_cpus);
-	sched_unpin();
+	other_cpus = all_cpus;
+	CPU_CLR(PCPU_GET(cpuid), &other_cpus);
 	if (ipi == IPI_STOP_HARD)
 		CPU_OR_ATOMIC(&ipi_nmi_pending, &other_cpus);
 
@@ -1197,7 +1183,7 @@ ipi_all_but_self(u_int ipi)
 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
@@ -1205,13 +1191,11 @@ ipi_nmi_handler()
 	 * 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);
 }
@@ -1223,25 +1207,21 @@ ipi_nmi_handler()
 void
 cpustop_handler(void)
 {
-	cpuset_t cpumask;
 	int cpu;
 
-	sched_pin();
-	cpumask = PCPU_GET(cpumask);
 	cpu = PCPU_GET(cpuid);
-	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();
diff --git a/sys/i386/xen/pmap.c b/sys/i386/xen/pmap.c
index 87bc0a0..b5e71da 100644
--- a/sys/i386/xen/pmap.c
+++ b/sys/i386/xen/pmap.c
@@ -802,7 +802,8 @@ pmap_cache_bits(int mode, boolean_t is_pde)
 void
 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
 {
-	cpuset_t cpumask, other_cpus;
+	cpuset_t other_cpus;
+	u_int cpuid;
 
 	CTR2(KTR_PMAP, "pmap_invalidate_page: pmap=%p va=0x%x",
 	    pmap, va);
@@ -812,9 +813,10 @@ pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
 		invlpg(va);
 		smp_invlpg(va);
 	} else {
-		cpumask = PCPU_GET(cpumask);
-		other_cpus = PCPU_GET(other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		cpuid = PCPU_GET(cpuid);
+		other_cpus = all_cpus;
+		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))
@@ -827,8 +829,9 @@ pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
 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;
 
 	CTR3(KTR_PMAP, "pmap_invalidate_page: pmap=%p eva=0x%x sva=0x%x",
 	    pmap, sva, eva);
@@ -839,9 +842,10 @@ pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
 			invlpg(addr);
 		smp_invlpg_range(sva, eva);
 	} else {
-		cpumask = PCPU_GET(cpumask);
-		other_cpus = PCPU_GET(other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		cpuid = PCPU_GET(cpuid);
+		other_cpus = all_cpus;
+		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);
@@ -855,7 +859,8 @@ pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
 void
 pmap_invalidate_all(pmap_t pmap)
 {
-	cpuset_t cpumask, other_cpus;
+	cpuset_t other_cpus;
+	u_int cpuid;
 
 	CTR1(KTR_PMAP, "pmap_invalidate_page: pmap=%p", pmap);
 
@@ -864,9 +869,10 @@ pmap_invalidate_all(pmap_t pmap)
 		invltlb();
 		smp_invltlb();
 	} else {
-		cpumask = PCPU_GET(cpumask);
-		other_cpus = PCPU_GET(other_cpus);
-		if (CPU_OVERLAP(&pmap->pm_active, &cpumask))
+		cpuid = PCPU_GET(cpuid);
+		other_cpus = all_cpus;
+		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))
@@ -1708,12 +1714,12 @@ pmap_lazyfix_action(void)
 }
 
 static void
-pmap_lazyfix_self(cpuset_t mymask)
+pmap_lazyfix_self(u_int cpuid)
 {
 
 	if (rcr3() == lazyptd)
 		load_cr3(PCPU_GET(curpcb)->pcb_cr3);
-	CPU_NAND_ATOMIC(lazymask, &mymask);
+	CPU_CLR_ATOMIC(cpuid, lazymask);
 }
 
 
@@ -1721,7 +1727,7 @@ static void
 pmap_lazyfix(pmap_t pmap)
 {
 	cpuset_t mymask, mask;
-	u_int spins;
+	u_int cpuid, spins;
 	int lsb;
 
 	mask = pmap->pm_active;
@@ -1739,10 +1745,13 @@ pmap_lazyfix(pmap_t pmap)
 #else
 		lazyptd = vtophys(pmap->pm_pdir);
 #endif
-		mymask = PCPU_GET(cpumask);
+		cpuid = PCPU_GET(cpuid);
+
+		/* Use a cpuset just for having an easy check. */
+		CPU_SETOF(cpuid, &mymask);
 		if (!CPU_CMP(&mask, &mymask)) {
 			lazymask = &pmap->pm_active;
-			pmap_lazyfix_self(mymask);
+			pmap_lazyfix_self(cpuid);
 		} else {
 			atomic_store_rel_int((u_int *)&lazymask,
 			    (u_int)&pmap->pm_active);
@@ -4126,17 +4135,19 @@ void
 pmap_activate(struct thread *td)
 {
 	pmap_t	pmap, oldpmap;
+	u_int	cpuid;
 	u_int32_t  cr3;
 
 	critical_enter();
 	pmap = vmspace_pmap(td->td_proc->p_vmspace);
 	oldpmap = PCPU_GET(curpmap);
+	cpuid = PCPU_GET(cpuid);
 #if defined(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
 #ifdef PAE
 	cr3 = vtophys(pmap->pm_pdpt);