diff --git a/sys/amd64/amd64/machdep.c b/sys/amd64/amd64/machdep.c
index 6993dea..39d1757 100644
--- a/sys/amd64/amd64/machdep.c
+++ b/sys/amd64/amd64/machdep.c
@@ -1430,7 +1430,8 @@ hammer_time(u_int64_t modulep, u_int64_t physfree)
 	wrmsr(MSR_GSBASE, (u_int64_t)pc);
 	wrmsr(MSR_KGSBASE, 0);		/* User value while in the kernel */
 
-	pcpu_init(pc, 0, sizeof(struct pcpu));
+	pcpu_init(pc, (void *)(physfree + KERNBASE), 0, sizeof(struct pcpu));
+	physfree += DPCPU_SIZE;
 	PCPU_SET(prvspace, pc);
 	PCPU_SET(curthread, &thread0);
 	PCPU_SET(curpcb, thread0.td_pcb);
diff --git a/sys/amd64/amd64/mp_machdep.c b/sys/amd64/amd64/mp_machdep.c
index a989d10..36736f2 100644
--- a/sys/amd64/amd64/mp_machdep.c
+++ b/sys/amd64/amd64/mp_machdep.c
@@ -93,9 +93,10 @@ static int bootAP;
 /* Free these after use */
 void *bootstacks[MAXCPU];
 
-/* Temporary holder for double fault stack */
+/* Temporary variables for init_secondary()  */
 char *doublefault_stack;
 char *nmi_stack;
+void *dpcpu;
 
 /* Hotwire a 0->4MB V==P mapping */
 extern pt_entry_t *KPTphys;
@@ -589,7 +590,7 @@ init_secondary(void)
 	pc = &__pcpu[cpu];
 
 	/* prime data page for it to use */
-	pcpu_init(pc, cpu, sizeof(struct pcpu));
+	pcpu_init(pc, dpcpu, cpu, sizeof(struct pcpu));
 	pc->pc_apic_id = cpu_apic_ids[cpu];
 	pc->pc_prvspace = pc;
 	pc->pc_curthread = 0;
@@ -885,6 +886,7 @@ start_all_aps(void)
 		bootstacks[cpu] = (void *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
 		doublefault_stack = (char *)kmem_alloc(kernel_map, PAGE_SIZE);
 		nmi_stack = (char *)kmem_alloc(kernel_map, PAGE_SIZE);
+		dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
 
 		bootSTK = (char *)bootstacks[cpu] + KSTACK_PAGES * PAGE_SIZE - 8;
 		bootAP = cpu;
diff --git a/sys/arm/at91/at91_machdep.c b/sys/arm/at91/at91_machdep.c
index af7f97c..3fb1604 100644
--- a/sys/arm/at91/at91_machdep.c
+++ b/sys/arm/at91/at91_machdep.c
@@ -226,10 +226,11 @@ initarm(void *arg, void *arg2)
 	boot_arg2 = arg2;
 	set_cpufuncs();
 	lastaddr = fake_preload_metadata();
-	pcpu_init(pcpup, 0, sizeof(struct pcpu));
+	freemempos = (lastaddr + PAGE_MASK) & ~PAGE_MASK;
+	pcpu_init(pcpup, (void *)freemempos, 0, sizeof(struct pcpu));
+	freemempos += DPCPU_SIZE;
 	PCPU_SET(curthread, &thread0);
 
-	freemempos = (lastaddr + PAGE_MASK) & ~PAGE_MASK;
 	/* Define a macro to simplify memory allocation */
 #define valloc_pages(var, np)                   \
 	alloc_pages((var).pv_va, (np));         \
diff --git a/sys/arm/mv/mv_machdep.c b/sys/arm/mv/mv_machdep.c
index 1c00eb5..5d42b18 100644
--- a/sys/arm/mv/mv_machdep.c
+++ b/sys/arm/mv/mv_machdep.c
@@ -425,7 +425,9 @@ initarm(void *mdp, void *unused __unused)
 	if (platform_pmap_init() != 0)
 		return (NULL);
 
-	pcpu_init(pcpup, 0, sizeof(struct pcpu));
+	freemempos = (lastaddr + PAGE_MASK) & ~PAGE_MASK;
+	pcpu_init(pcpup, (void *)freemempos, 0, sizeof(struct pcpu));
+	freemempos += DPCPU_SIZE;
 	PCPU_SET(curthread, &thread0);
 
 	/* Calculate number of L2 tables needed for mapping vm_page_array */
@@ -442,7 +444,6 @@ initarm(void *mdp, void *unused __unused)
 	l2size = (l2size + 3) & ~3;
 
 #define KERNEL_TEXT_BASE (KERNBASE)
-	freemempos = (lastaddr + PAGE_MASK) & ~PAGE_MASK;
 
 	/* Define a macro to simplify memory allocation */
 #define valloc_pages(var, np)                   \
diff --git a/sys/arm/sa11x0/assabet_machdep.c b/sys/arm/sa11x0/assabet_machdep.c
index 71d71da..8a9ee34 100644
--- a/sys/arm/sa11x0/assabet_machdep.c
+++ b/sys/arm/sa11x0/assabet_machdep.c
@@ -222,16 +222,17 @@ initarm(void *arg, void *arg2)
 	set_cpufuncs();
 	lastaddr = fake_preload_metadata();
 	physmem = memsize / PAGE_SIZE;
-	pc = &__pcpu;
-	pcpu_init(pc, 0, sizeof(struct pcpu));
-	PCPU_SET(curthread, &thread0);
-
 	physical_start = (vm_offset_t) KERNBASE;
 	physical_end =  lastaddr;
 	physical_freestart = (((vm_offset_t)physical_end) + PAGE_MASK) & ~PAGE_MASK;
 	md_addr.pv_va = md_addr.pv_pa = MDROOT_ADDR;
 	freemempos = (vm_offset_t)round_page(physical_freestart);
 	memset((void *)freemempos, 0, 256*1024);
+	pc = &__pcpu;
+	pcpu_init(pc, (void *)freemempos, 0, sizeof(struct pcpu));
+	freemempos += DPCPU_SIZE;
+	PCPU_SET(curthread, &thread0);
+
 		/* Define a macro to simplify memory allocation */
 #define	valloc_pages(var, np)			\
 	alloc_pages((var).pv_pa, (np));		\
diff --git a/sys/arm/xscale/i80321/ep80219_machdep.c b/sys/arm/xscale/i80321/ep80219_machdep.c
index 77fad07..3fe2580 100644
--- a/sys/arm/xscale/i80321/ep80219_machdep.c
+++ b/sys/arm/xscale/i80321/ep80219_machdep.c
@@ -197,10 +197,11 @@ initarm(void *arg, void *arg2)
 
 	set_cpufuncs();
 	lastaddr = fake_preload_metadata();
-	pcpu_init(pcpup, 0, sizeof(struct pcpu));
+	freemempos = 0xa0200000;
+	freemempos -= DPCPU_SIZE;
+	pcpu_init(pcpup, (void *)freemempos, 0, sizeof(struct pcpu));
 	PCPU_SET(curthread, &thread0);
 
-	freemempos = 0xa0200000;
 	/* Define a macro to simplify memory allocation */
 #define	valloc_pages(var, np)			\
 	alloc_pages((var).pv_pa, (np));				\
diff --git a/sys/arm/xscale/i80321/iq31244_machdep.c b/sys/arm/xscale/i80321/iq31244_machdep.c
index ba0fd93..065473c 100644
--- a/sys/arm/xscale/i80321/iq31244_machdep.c
+++ b/sys/arm/xscale/i80321/iq31244_machdep.c
@@ -198,10 +198,11 @@ initarm(void *arg, void *arg2)
 
 	set_cpufuncs();
 	lastaddr = fake_preload_metadata();
-	pcpu_init(pcpup, 0, sizeof(struct pcpu));
+	freemempos = 0xa0200000;
+	freemempos -= DPCPU_SIZE;
+	pcpu_init(pcpup, (void *)freemempos, sizeof(struct pcpu));
 	PCPU_SET(curthread, &thread0);
 
-	freemempos = 0xa0200000;
 	/* Define a macro to simplify memory allocation */
 #define	valloc_pages(var, np)			\
 	alloc_pages((var).pv_pa, (np));		\
diff --git a/sys/arm/xscale/i8134x/crb_machdep.c b/sys/arm/xscale/i8134x/crb_machdep.c
index 909a1cd..0220629 100644
--- a/sys/arm/xscale/i8134x/crb_machdep.c
+++ b/sys/arm/xscale/i8134x/crb_machdep.c
@@ -194,10 +194,11 @@ initarm(void *arg, void *arg2)
 
 	set_cpufuncs();
 	lastaddr = fake_preload_metadata();
-	pcpu_init(pcpup, 0, sizeof(struct pcpu));
+	freemempos = 0x00200000;
+	freemempos -= DPCPU_SIZE;
+	pcpu_init(pcpup, (void *)freemempos, 0, sizeof(struct pcpu));
 	PCPU_SET(curthread, &thread0);
 
-	freemempos = 0x00200000;
 	/* Define a macro to simplify memory allocation */
 #define	valloc_pages(var, np)			\
 	alloc_pages((var).pv_pa, (np));		\
diff --git a/sys/arm/xscale/ixp425/avila_machdep.c b/sys/arm/xscale/ixp425/avila_machdep.c
index d199796..fd290c5 100644
--- a/sys/arm/xscale/ixp425/avila_machdep.c
+++ b/sys/arm/xscale/ixp425/avila_machdep.c
@@ -236,7 +236,9 @@ initarm(void *arg, void *arg2)
 
 	set_cpufuncs();		/* NB: sets cputype */
 	lastaddr = fake_preload_metadata();
-	pcpu_init(pcpup, 0, sizeof(struct pcpu));
+	freemempos = KERNPHYSADDR;
+	freemempos -= DPCPU_SIZE;
+	pcpu_init(pcpup, (void *)freemempos, 0, sizeof(struct pcpu));
 	PCPU_SET(curthread, &thread0);
 
 	/*
@@ -248,7 +250,6 @@ initarm(void *arg, void *arg2)
 	 * write-through).  Note this leaves a gap for expansion
 	 * (or might be repurposed).
 	 */
-	freemempos = KERNPHYSADDR;
 
 	/* macros to simplify initial memory allocation */
 #define alloc_pages(var, np) do {					\
diff --git a/sys/arm/xscale/pxa/pxa_machdep.c b/sys/arm/xscale/pxa/pxa_machdep.c
index f3b137e..da5abc4 100644
--- a/sys/arm/xscale/pxa/pxa_machdep.c
+++ b/sys/arm/xscale/pxa/pxa_machdep.c
@@ -179,10 +179,11 @@ initarm(void *arg, void *arg2)
 	set_cpufuncs();
 
 	lastaddr = fake_preload_metadata();
-	pcpu_init(pcpup, 0, sizeof(struct pcpu));
+	freemempos = 0xa0200000;
+	freemempos -= DPCPU_SIZE;
+	pcpu_init(pcpup, (void *)freemempos, 0, sizeof(struct pcpu));
 	PCPU_SET(curthread, &thread0);
 
-	freemempos = 0xa0200000;
 	/* Define a macro to simplify memory allocation */
 #define	valloc_pages(var, np)			\
 	alloc_pages((var).pv_pa, (np));		\
diff --git a/sys/i386/i386/locore.s b/sys/i386/i386/locore.s
index 94ac670..aaa90e6 100644
--- a/sys/i386/i386/locore.s
+++ b/sys/i386/i386/locore.s
@@ -121,6 +121,12 @@ vm86phystk:	.long	0		/* PA of vm86/bios stack */
 vm86paddr:	.long	0		/* address of vm86 region */
 vm86pa:		.long	0		/* phys addr of vm86 region */
 
+	.globl	__start_set_pcpu, __stop_set_pcpu
+	.globl	dpcpu0_base
+dpcpu0_pa:	.long	0
+dpcpu0_base:	.long	0
+dpcpu0_sz:	.long	0
+
 #ifdef PC98
 	.globl	pc98_system_parameter
 pc98_system_parameter:
@@ -735,6 +741,22 @@ no_kernend:
 	addl	$KERNBASE, %esi
 	movl	%esi, R(proc0kstack)
 
+/* Calculate the size of dpcpu0 */
+	movl	$__stop_set_pcpu,%eax
+	subl	$__start_set_pcpu,%eax
+	addl	$(PAGE_SIZE-1),%eax
+	andl	$~(PAGE_SIZE-1),%eax
+	movl	%eax,R(dpcpu0_sz)
+
+/* Inlined ALLOCPAGES to allocate dpcpu0 */
+	movl	R(physfree), %esi
+	addl	%esi, %eax
+	movl	%eax, R(physfree)
+
+	movl	%esi, R(dpcpu0_pa)
+	addl	$KERNBASE, %esi
+	movl	%esi, R(dpcpu0_base)
+
 	ALLOCPAGES(1)			/* vm86/bios stack */
 	movl	%esi,R(vm86phystk)
 
@@ -793,6 +815,10 @@ no_kernend:
 	movl	$NPGPTD, %ecx
 	fillkptphys($PG_RW)
 
+	movl	R(dpcpu0_pa), %eax
+	movl	R(dpcpu0_sz), %ecx
+	fillkptphys($PG_RW)
+
 /* Map proc0's KSTACK in the physical way ... */
 	movl	R(p0kpa), %eax
 	movl	$(KSTACK_PAGES), %ecx
diff --git a/sys/i386/i386/machdep.c b/sys/i386/i386/machdep.c
index e64bcd2..77ac947 100644
--- a/sys/i386/i386/machdep.c
+++ b/sys/i386/i386/machdep.c
@@ -2369,6 +2369,8 @@ do_next:
 	PT_UPDATES_FLUSH();
 }
 
+extern unsigned long dpcpu0_base;
+
 #ifdef XEN
 #define MTOPSIZE (1<<(14 + PAGE_SHIFT))
 
@@ -2460,7 +2462,8 @@ init386(first)
 		    (unsigned long)Xhypervisor_callback,
 		    GSEL(GCODE_SEL, SEL_KPL), (unsigned long)failsafe_callback);
 #endif
-	pcpu_init(pc, 0, sizeof(struct pcpu));
+	pcpu_init(pc, (void *)(first + KERNBASE), 0, sizeof(struct pcpu));
+	first += DPCPU_SIZE;
 	PCPU_SET(prvspace, pc);
 	PCPU_SET(curthread, &thread0);
 	PCPU_SET(curpcb, thread0.td_pcb);
@@ -2643,7 +2646,7 @@ init386(first)
 	mtx_init(&dt_lock, "descriptor tables", NULL, MTX_SPIN);
 	lgdt(&r_gdt);
 
-	pcpu_init(pc, 0, sizeof(struct pcpu));
+	pcpu_init(pc, (void *)dpcpu0_base, 0, sizeof(struct pcpu));
 	PCPU_SET(prvspace, pc);
 	PCPU_SET(curthread, &thread0);
 	PCPU_SET(curpcb, thread0.td_pcb);
diff --git a/sys/i386/i386/mp_machdep.c b/sys/i386/i386/mp_machdep.c
index c233f25..3fbccc0 100644
--- a/sys/i386/i386/mp_machdep.c
+++ b/sys/i386/i386/mp_machdep.c
@@ -143,6 +143,7 @@ static int bootAP;
 
 /* Free these after use */
 void *bootstacks[MAXCPU];
+static void *dpcpu;
 
 /* Hotwire a 0->4MB V==P mapping */
 extern pt_entry_t *KPTphys;
@@ -609,7 +610,7 @@ init_secondary(void)
 	pc = &__pcpu[myid];
 
 	/* prime data page for it to use */
-	pcpu_init(pc, myid, sizeof(struct pcpu));
+	pcpu_init(pc, dpcpu, myid, sizeof(struct pcpu));
 	pc->pc_apic_id = cpu_apic_ids[myid];
 	pc->pc_prvspace = pc;
 	pc->pc_curthread = 0;
@@ -897,8 +898,9 @@ start_all_aps(void)
 		apic_id = cpu_apic_ids[cpu];
 
 		/* allocate and set up a boot stack data page */
-		bootstacks[cpu] = (char *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
-
+		bootstacks[cpu] =
+		    (char *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
+		dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
 		/* setup a vector to our boot code */
 		*((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
 		*((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
diff --git a/sys/i386/i386/pmap.c b/sys/i386/i386/pmap.c
index 5d7dc97..2162b65 100644
--- a/sys/i386/i386/pmap.c
+++ b/sys/i386/i386/pmap.c
@@ -414,6 +414,8 @@ pmap_bootstrap(vm_paddr_t firstaddr)
 		mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF);
 		SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1)
 		SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1)
+		*(sysmaps->CMAP1) = 0;
+		*(sysmaps->CMAP2) = 0;
 	}
 	SYSMAP(caddr_t, CMAP1, CADDR1, 1)
 	SYSMAP(caddr_t, CMAP3, CADDR3, 1)
diff --git a/sys/i386/xen/mp_machdep.c b/sys/i386/xen/mp_machdep.c
index f16d25f..04468b9 100644
--- a/sys/i386/xen/mp_machdep.c
+++ b/sys/i386/xen/mp_machdep.c
@@ -743,7 +743,8 @@ start_all_aps(void)
 
 		/* Get per-cpu data */
 		pc = &__pcpu[bootAP];
-		pcpu_init(pc, bootAP, sizeof(struct pcpu));
+		pcpu_init(pc, (void *)kmem_alloc(kernel_map, DPCPU_SIZE),
+		    bootAP, sizeof(struct pcpu));
 		pc->pc_apic_id = cpu_apic_ids[bootAP];
 		pc->pc_prvspace = pc;
 		pc->pc_curthread = 0;
diff --git a/sys/ia64/ia64/machdep.c b/sys/ia64/ia64/machdep.c
index dda1a08..8bc260c 100644
--- a/sys/ia64/ia64/machdep.c
+++ b/sys/ia64/ia64/machdep.c
@@ -647,13 +647,26 @@ ia64_init(void)
 		bootverbose = 1;
 
 	/*
+	 * Find the beginning and end of the kernel.
+	 */
+	kernstart = trunc_page(kernel_text);
+#ifdef DDB
+	ksym_start = bootinfo.bi_symtab;
+	ksym_end = bootinfo.bi_esymtab;
+	kernend = (vm_offset_t)round_page(ksym_end);
+#else
+	kernend = (vm_offset_t)round_page(_end);
+#endif
+
+	/*
 	 * Setup the PCPU data for the bootstrap processor. It is needed
 	 * by printf(). Also, since printf() has critical sections, we
 	 * need to initialize at least pc_curthread.
 	 */
 	pcpup = &pcpu0;
 	ia64_set_k4((u_int64_t)pcpup);
-	pcpu_init(pcpup, 0, sizeof(pcpu0));
+	pcpu_init(pcpup, (void *)kernend, 0, sizeof(pcpu0));
+	kernend += DPCPU_SIZE;
 	PCPU_SET(curthread, &thread0);
 
 	/*
@@ -682,18 +695,6 @@ ia64_init(void)
 	ia64_sal_init();
 	calculate_frequencies();
 
-	/*
-	 * Find the beginning and end of the kernel.
-	 */
-	kernstart = trunc_page(kernel_text);
-#ifdef DDB
-	ksym_start = bootinfo.bi_symtab;
-	ksym_end = bootinfo.bi_esymtab;
-	kernend = (vm_offset_t)round_page(ksym_end);
-#else
-	kernend = (vm_offset_t)round_page(_end);
-#endif
-
 	/* But if the bootstrap tells us otherwise, believe it! */
 	if (bootinfo.bi_kernend)
 		kernend = round_page(bootinfo.bi_kernend);
diff --git a/sys/ia64/ia64/mp_machdep.c b/sys/ia64/ia64/mp_machdep.c
index 6dba43a..f61d5c8 100644
--- a/sys/ia64/ia64/mp_machdep.c
+++ b/sys/ia64/ia64/mp_machdep.c
@@ -207,6 +207,7 @@ cpu_mp_add(u_int acpiid, u_int apicid, u_int apiceid)
 {
 	struct pcpu *pc;
 	u_int64_t lid;
+	void *dpcpu;
 
 	/* Ignore any processor numbers outside our range */
 	if (acpiid > mp_maxid)
@@ -224,7 +225,8 @@ cpu_mp_add(u_int acpiid, u_int apicid, u_int apiceid)
 
 	if (acpiid != 0) {
 		pc = (struct pcpu *)malloc(sizeof(*pc), M_SMP, M_WAITOK);
-		pcpu_init(pc, acpiid, sizeof(*pc));
+		dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
+		pcpu_init(pc, dpcpu, acpiid, sizeof(*pc));
 	} else
 		pc = pcpup;
 
diff --git a/sys/kern/subr_pcpu.c b/sys/kern/subr_pcpu.c
index 74082c2..c0f7d24 100644
--- a/sys/kern/subr_pcpu.c
+++ b/sys/kern/subr_pcpu.c
@@ -56,6 +56,9 @@ __FBSDID("$FreeBSD$");
 #include <sys/smp.h>
 #include <ddb/ddb.h>
 
+DPCPU_DEFINE(int, id);
+
+uintptr_t dpcpu_ptrs[MAXCPU];
 struct pcpu *cpuid_to_pcpu[MAXCPU];
 struct cpuhead cpuhead = SLIST_HEAD_INITIALIZER(cpuhead);
 
@@ -63,7 +66,7 @@ struct cpuhead cpuhead = SLIST_HEAD_INITIALIZER(cpuhead);
  * Initialize the MI portions of a struct pcpu.
  */
 void
-pcpu_init(struct pcpu *pcpu, int cpuid, size_t size)
+pcpu_init(struct pcpu *pcpu, void *dpcpu, int cpuid, size_t size)
 {
 
 	bzero(pcpu, size);
@@ -71,6 +74,7 @@ pcpu_init(struct pcpu *pcpu, int cpuid, size_t size)
 	    ("pcpu_init: invalid cpuid %d", cpuid));
 	pcpu->pc_cpuid = cpuid;
 	pcpu->pc_cpumask = 1 << cpuid;
+	pcpu->pc_dynamic = (uintptr_t)dpcpu;
 	cpuid_to_pcpu[cpuid] = pcpu;
 	SLIST_INSERT_HEAD(&cpuhead, pcpu, pc_allcpu);
 	cpu_pcpu_init(pcpu, cpuid, size);
@@ -80,6 +84,17 @@ pcpu_init(struct pcpu *pcpu, int cpuid, size_t size)
 	snprintf(pcpu->pc_name, sizeof(pcpu->pc_name), "CPU %d", cpuid);
 #endif
 
+	/*
+	 * Initialize defaults from our linker section.
+	 */
+	memcpy(dpcpu, (void *)DPCPU_START, DPCPU_SIZE);
+
+	/*
+	 * Place it in the global pcpu array.
+	 */
+	dpcpu_ptrs[pcpu->pc_cpuid] = pcpu->pc_dynamic;
+
+	DPCPU_ID_SET(pcpu->pc_cpuid, id, pcpu->pc_cpuid);
 }
 
 /*
@@ -91,6 +106,7 @@ pcpu_destroy(struct pcpu *pcpu)
 
 	SLIST_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu);
 	cpuid_to_pcpu[pcpu->pc_cpuid] = NULL;
+	dpcpu_ptrs[pcpu->pc_cpuid] = 0;
 }
 
 /*
@@ -110,7 +126,9 @@ show_pcpu(struct pcpu *pc)
 {
 	struct thread *td;
 
-	db_printf("cpuid        = %d\n", pc->pc_cpuid);
+	db_printf("cpuid        = %d (%d)\n",
+	    pc->pc_cpuid, DPCPU_ID_GET(pc->pc_cpuid, id));
+	db_printf("curcpuid	= %d\n", DPCPU_GET(id));
 	db_printf("curthread    = ");
 	td = pc->pc_curthread;
 	if (td != NULL)
diff --git a/sys/mips/mips/mp_machdep.c b/sys/mips/mips/mp_machdep.c
index 3f7d731..6380145 100644
--- a/sys/mips/mips/mp_machdep.c
+++ b/sys/mips/mips/mp_machdep.c
@@ -224,12 +224,14 @@ static int
 smp_start_secondary(int cpuid)
 {
 	struct pcpu *pcpu;
+	void *dpcpu;
 	int i;
 
 	if (bootverbose)
 		printf("smp_start_secondary: starting cpu %d\n", cpuid);
 
-	pcpu_init(&__pcpu[cpuid], cpuid, sizeof(struct pcpu));
+	dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
+	pcpu_init(&__pcpu[cpuid], dpcpu, cpuid, sizeof(struct pcpu));
 
 	if (bootverbose)
 		printf("smp_start_secondary: cpu %d started\n", cpuid);
diff --git a/sys/pc98/pc98/machdep.c b/sys/pc98/pc98/machdep.c
index 0025a8a..2720256 100644
--- a/sys/pc98/pc98/machdep.c
+++ b/sys/pc98/pc98/machdep.c
@@ -2009,7 +2009,9 @@ init386(first)
 	mtx_init(&dt_lock, "descriptor tables", NULL, MTX_SPIN);
 	lgdt(&r_gdt);
 
-	pcpu_init(pc, 0, sizeof(struct pcpu));
+	pcpu_init(pc, (void *)(first + KERNBASE), 0, sizeof(struct pcpu));
+	first += DPCPU_SIZE;
+
 	PCPU_SET(prvspace, pc);
 	PCPU_SET(curthread, &thread0);
 	PCPU_SET(curpcb, thread0.td_pcb);
diff --git a/sys/powerpc/aim/machdep.c b/sys/powerpc/aim/machdep.c
index 5d2665a..3dfa8fc 100644
--- a/sys/powerpc/aim/machdep.c
+++ b/sys/powerpc/aim/machdep.c
@@ -278,7 +278,8 @@ powerpc_init(u_int startkernel, u_int endkernel, u_int basekernel, void *mdp)
 			ksym_end = MD_FETCH(kmdp, MODINFOMD_ESYM, uintptr_t);
 #endif
 		}
-	}
+	} else
+		end = endkernel;
 
 	/*
 	 * Init params/tunables that can be overridden by the loader
@@ -295,7 +296,8 @@ powerpc_init(u_int startkernel, u_int endkernel, u_int basekernel, void *mdp)
 	 * Set up per-cpu data.
 	 */
 	pc = __pcpu;
-	pcpu_init(pc, 0, sizeof(struct pcpu));
+	pcpu_init(pc, (void *)end, 0, sizeof(struct pcpu));
+	end += DPCPU_SIZE;
 	pc->pc_curthread = &thread0;
 	pc->pc_cpuid = 0;
 
@@ -488,7 +490,7 @@ powerpc_init(u_int startkernel, u_int endkernel, u_int basekernel, void *mdp)
 	else
 		pmap_mmu_install(MMU_TYPE_OEA, BUS_PROBE_GENERIC);
 
-	pmap_bootstrap(startkernel, endkernel);
+	pmap_bootstrap(startkernel, end);
 	mtmsr(mfmsr() | PSL_IR|PSL_DR|PSL_ME|PSL_RI);
 	isync();
 
diff --git a/sys/powerpc/booke/machdep.c b/sys/powerpc/booke/machdep.c
index 8787098..45a3deb 100644
--- a/sys/powerpc/booke/machdep.c
+++ b/sys/powerpc/booke/machdep.c
@@ -390,7 +390,8 @@ e500_init(u_int32_t startkernel, u_int32_t endkernel, void *mdp)
 
 	/* Set up per-cpu data and store the pointer in SPR general 0. */
 	pc = &__pcpu[0];
-	pcpu_init(pc, 0, sizeof(struct pcpu));
+	pcpu_init(pc, (void *)end, 0, sizeof(struct pcpu));
+	end += DPCPU_SIZE;
 	pc->pc_curthread = &thread0;
 	__asm __volatile("mtsprg 0, %0" :: "r"(pc));
 
diff --git a/sys/powerpc/powerpc/mp_machdep.c b/sys/powerpc/powerpc/mp_machdep.c
index 986259f..1428ae5 100644
--- a/sys/powerpc/powerpc/mp_machdep.c
+++ b/sys/powerpc/powerpc/mp_machdep.c
@@ -146,8 +146,11 @@ cpu_mp_start(void)
 			goto next;
 		}
 		if (cpu.cr_cpuid != bsp.cr_cpuid) {
+			void *dpcpu;
+
 			pc = &__pcpu[cpu.cr_cpuid];
-			pcpu_init(pc, cpu.cr_cpuid, sizeof(*pc));
+			dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
+			pcpu_init(pc, dpcpu, cpu.cr_cpuid, sizeof(*pc));
 		} else {
 			pc = pcpup;
 			pc->pc_cpuid = bsp.cr_cpuid;
diff --git a/sys/sparc64/sparc64/machdep.c b/sys/sparc64/sparc64/machdep.c
index 164a8b3..52f0ebc 100644
--- a/sys/sparc64/sparc64/machdep.c
+++ b/sys/sparc64/sparc64/machdep.c
@@ -301,6 +301,15 @@ sparc64_init(caddr_t mdp, u_long o1, u_long o2, u_long o3, ofw_vec_t *vec)
 		}
 	}
 
+	/*
+	 * Sanity check the kernel end, which is important.
+	 */
+	if (end == 0) {
+		printf("sparc64_init: warning, kernel end not specified.\n"
+		    "Attempting to continue anyway.\n");
+		end = (vm_offset_t)_end;
+	}
+
 	init_param1();
 
 	/*
@@ -315,7 +324,8 @@ sparc64_init(caddr_t mdp, u_long o1, u_long o2, u_long o3, ofw_vec_t *vec)
 	 * or it'll zero it out from under us.
 	 */
 	pc = (struct pcpu *)(pcpu0 + (PCPU_PAGES * PAGE_SIZE)) - 1;
-	pcpu_init(pc, 0, sizeof(struct pcpu));
+	pcpu_init(pc, (void *)end, 0, sizeof(struct pcpu));
+	end += DPCPU_SIZE;
 	pc->pc_addr = (vm_offset_t)pcpu0;
 	pc->pc_mid = UPA_CR_GET_MID(ldxa(0, ASI_UPA_CONFIG_REG));
 	pc->pc_tlb_ctx = TLB_CTX_USER_MIN;
@@ -374,15 +384,6 @@ sparc64_init(caddr_t mdp, u_long o1, u_long o2, u_long o3, ofw_vec_t *vec)
 	}
 
 	/*
-	 * Sanity check the kernel end, which is important.
-	 */
-	if (end == 0) {
-		printf("sparc64_init: warning, kernel end not specified.\n"
-		    "Attempting to continue anyway.\n");
-		end = (vm_offset_t)_end;
-	}
-
-	/*
 	 * Determine the TLB slot maxima, which are expected to be
 	 * equal across all CPUs.
 	 * NB: for Cheetah-class CPUs, these properties only refer
diff --git a/sys/sparc64/sparc64/mp_machdep.c b/sys/sparc64/sparc64/mp_machdep.c
index ee4ffe7..d211c1e 100644
--- a/sys/sparc64/sparc64/mp_machdep.c
+++ b/sys/sparc64/sparc64/mp_machdep.c
@@ -289,7 +289,8 @@ cpu_mp_start(void)
 
 		va = kmem_alloc(kernel_map, PCPU_PAGES * PAGE_SIZE);
 		pc = (struct pcpu *)(va + (PCPU_PAGES * PAGE_SIZE)) - 1;
-		pcpu_init(pc, cpuid, sizeof(*pc));
+		pcpu_init(pc, (void *)kmem_alloc(kernel_map, DPCPU_SIZE),
+		    cpuid, sizeof(*pc));
 		pc->pc_addr = va;
 		pc->pc_clock = clock;
 		pc->pc_mid = mid;
diff --git a/sys/sun4v/sun4v/machdep.c b/sys/sun4v/sun4v/machdep.c
index 7975040..48accdc 100644
--- a/sys/sun4v/sun4v/machdep.c
+++ b/sys/sun4v/sun4v/machdep.c
@@ -339,6 +339,15 @@ sparc64_init(caddr_t mdp, u_long o1, u_long o2, u_long o3, ofw_vec_t *vec)
 		}
 	}
 
+	/*
+	 * Sanity check the kernel end, which is important.
+	 */
+	if (end == 0) {
+		printf("sparc64_init: warning, kernel end not specified.\n"
+		       "Attempting to continue anyway.\n");
+		end = (vm_offset_t)_end;
+	}
+
         if (boothowto & RB_VERBOSE)
                 bootverbose = 1;
 
@@ -385,7 +394,8 @@ sparc64_init(caddr_t mdp, u_long o1, u_long o2, u_long o3, ofw_vec_t *vec)
 	 * it'll zero it out from under us.
 	 */
 	pc = (struct pcpu *)(pcpu0 + (PCPU_PAGES * PAGE_SIZE)) - 1;
-	pcpu_init(pc, 0, sizeof(struct pcpu));
+	pcpu_init(pc, (void *)end, 0, sizeof(struct pcpu));
+	end += DPCPU_SIZE;
 	pc->pc_curthread = &thread0;
 	pc->pc_addr = (vm_offset_t)pcpu0;
 
@@ -405,15 +415,6 @@ sparc64_init(caddr_t mdp, u_long o1, u_long o2, u_long o3, ofw_vec_t *vec)
 		panic("sparc64_init");
 	}
 
-	/*
-	 * Sanity check the kernel end, which is important.
-	 */
-	if (end == 0) {
-		printf("sparc64_init: warning, kernel end not specified.\n"
-		       "Attempting to continue anyway.\n");
-		end = (vm_offset_t)_end;
-	}
-
 	cpu_block_copy = bcopy;
 	cpu_block_zero = bzero;
 
diff --git a/sys/sun4v/sun4v/mp_machdep.c b/sys/sun4v/sun4v/mp_machdep.c
index 5f4f632..cd7b4d3 100644
--- a/sys/sun4v/sun4v/mp_machdep.c
+++ b/sys/sun4v/sun4v/mp_machdep.c
@@ -323,7 +323,8 @@ cpu_mp_start(void)
 #endif
 		va = kmem_alloc(kernel_map, PCPU_PAGES * PAGE_SIZE);
 		pc = (struct pcpu *)(va + (PCPU_PAGES * PAGE_SIZE)) - 1;
-		pcpu_init(pc, cpuid, sizeof(*pc));
+		pcpu_init(pc, (void *)kmem_alloc(kernel_map, DPCPU_SIZE),
+		    cpuid, sizeof(*pc));
 		pc->pc_addr = va;
 
 		all_cpus |= 1 << cpuid;
diff --git a/sys/sys/pcpu.h b/sys/sys/pcpu.h
index dc26c1a..ab9c4ef 100644
--- a/sys/sys/pcpu.h
+++ b/sys/sys/pcpu.h
@@ -45,6 +45,61 @@
 struct pcb;
 struct thread;
 
+/*
+ * Define a set for pcpu data.
+ * 
+ * We don't use SET_DECLARE because it defines the set as 'a' when we
+ * want 'aw'.  GCC considers uninitialized data in a seperate section
+ * writable and there is no generic zero initializer that works for
+ * structs and scalars.
+ */
+extern uintptr_t *__start_set_pcpu;
+extern uintptr_t *__stop_set_pcpu;
+__asm__(".section set_pcpu, \"aw\", @progbits");
+__asm__(".previous");
+
+/*
+ * Array of dynamic pcpu base values.  Indexed by id.
+ */
+extern uintptr_t dpcpu_ptrs[];
+
+/*
+ * Convenience defines.
+ */
+#define	DPCPU_NAME(n)		pcpu_entry_##n
+#define	DPCPU_START		(uintptr_t)&__start_set_pcpu
+#define	DPCPU_STOP		(uintptr_t)&__stop_set_pcpu
+#define	DPCPU_SIZE		roundup2(DPCPU_STOP - DPCPU_START, PAGE_SIZE)
+#define	DPCPU_OFFSET(n)		((uintptr_t)&DPCPU_NAME(n) - DPCPU_START)
+
+/*
+ * Declaration and definition.
+ */
+#define	DPCPU_DECLARE(t, n)	extern t DPCPU_NAME(n)
+#define	DPCPU_DEFINE(t, n)	t DPCPU_NAME(n) __section("set_pcpu") __used
+
+/*
+ * Accessors with a given base.
+ */
+#define	_DPCPU_PTR(b, n)						\
+    (__typeof(DPCPU_NAME(n))*)((b) + DPCPU_OFFSET(n))
+#define	_DPCPU_GET(b, n)	(*_DPCPU_PTR(b, n))
+#define	_DPCPU_SET(b, n, v)	(*_DPCPU_PTR(b, n) = v)
+
+/*
+ * Accessors for the current cpu.
+ */
+#define	DPCPU_PTR(n)		_DPCPU_PTR(PCPU_GET(dynamic), n)
+#define	DPCPU_GET(n)		(*DPCPU_PTR(n))
+#define	DPCPU_SET(n, v)		(*DPCPU_PTR(n) = v)
+
+/*
+ * Accessors for remote cpus.
+ */
+#define	DPCPU_ID_PTR(i, n)	_DPCPU_PTR(dpcpu_ptrs[(i)], n)
+#define	DPCPU_ID_GET(i, n)	(*DPCPU_ID_PTR(i, n))
+#define	DPCPU_ID_SET(i, n, v)	(*DPCPU_ID_PTR(i, n) = v)
+
 /* 
  * XXXUPS remove as soon as we have per cpu variable
  * linker sets and  can define rm_queue in _rm_lock.h
@@ -93,6 +148,11 @@ struct pcpu {
 	struct rm_queue  pc_rm_queue; 
 
 	/*
+	 * Dynamic per-cpu data area.
+	 */
+	uintptr_t	pc_dynamic;
+
+	/*
 	 * Keep MD fields last, so that CPU-specific variations on a
 	 * single architecture don't result in offset variations of
 	 * the machine-independent fields of the pcpu. Even though
@@ -132,7 +192,7 @@ void	db_show_mdpcpu(struct pcpu *pcpu);
 
 void	pcpu_destroy(struct pcpu *pcpu);
 struct	pcpu *pcpu_find(u_int cpuid);
-void	pcpu_init(struct pcpu *pcpu, int cpuid, size_t size);
+void	pcpu_init(struct pcpu *pcpu, void *dpcpu, int cpuid, size_t size);
 
 #endif	/* _KERNEL */