Exemple #1
0
void *
initarm(struct arm_boot_params *abp)
{
#define	next_chunk2(a,b)	(((a) + (b)) &~ ((b)-1))
#define	next_page(a)		next_chunk2(a,PAGE_SIZE)
	struct pv_addr  kernel_l1pt;
	struct pv_addr  dpcpu;
	int loop, i;
	u_int l1pagetable;
	vm_offset_t freemempos;
	vm_offset_t freemem_pt;
	vm_offset_t afterkern;
	vm_offset_t freemem_after;
	vm_offset_t lastaddr;
	uint32_t memsize;

	/* kernel text starts where we were loaded at boot */
#define	KERNEL_TEXT_OFF		(abp->abp_physaddr  - PHYSADDR)
#define	KERNEL_TEXT_BASE	(KERNBASE + KERNEL_TEXT_OFF)
#define	KERNEL_TEXT_PHYS	(PHYSADDR + KERNEL_TEXT_OFF)

	lastaddr = parse_boot_param(abp);
	arm_physmem_kernaddr = abp->abp_physaddr;
	set_cpufuncs();		/* NB: sets cputype */
	pcpu_init(pcpup, 0, sizeof(struct pcpu));
	PCPU_SET(curthread, &thread0);

	if (envmode == 1)
		kern_envp = static_env;
	/* Do basic tuning, hz etc */
      	init_param1();
		
	/*
	 * We allocate memory downwards from where we were loaded
	 * by RedBoot; first the L1 page table, then NUM_KERNEL_PTS
	 * entries in the L2 page table.  Past that we re-align the
	 * allocation boundary so later data structures (stacks, etc)
	 * can be mapped with different attributes (write-back vs
	 * write-through).  Note this leaves a gap for expansion
	 * (or might be repurposed).
	 */
	freemempos = abp->abp_physaddr;

	/* macros to simplify initial memory allocation */
#define alloc_pages(var, np) do {					\
	freemempos -= (np * PAGE_SIZE);					\
	(var) = freemempos;						\
	/* NB: this works because locore maps PA=VA */			\
	memset((char *)(var), 0, ((np) * PAGE_SIZE));			\
} while (0)
#define	valloc_pages(var, np) do {					\
	alloc_pages((var).pv_pa, (np));					\
	(var).pv_va = (var).pv_pa + (KERNVIRTADDR - abp->abp_physaddr);	\
} while (0)

	/* force L1 page table alignment */
	while (((freemempos - L1_TABLE_SIZE) & (L1_TABLE_SIZE - 1)) != 0)
		freemempos -= PAGE_SIZE;
	/* allocate contiguous L1 page table */
	valloc_pages(kernel_l1pt, L1_TABLE_SIZE / PAGE_SIZE);
	/* now allocate L2 page tables; they are linked to L1 below */
	for (loop = 0; loop < NUM_KERNEL_PTS; ++loop) {
		if (!(loop % (PAGE_SIZE / L2_TABLE_SIZE_REAL))) {
			valloc_pages(kernel_pt_table[loop],
			    L2_TABLE_SIZE / PAGE_SIZE);
		} else {
			kernel_pt_table[loop].pv_pa = freemempos +
			    (loop % (PAGE_SIZE / L2_TABLE_SIZE_REAL)) *
			    L2_TABLE_SIZE_REAL;
			kernel_pt_table[loop].pv_va =
			    kernel_pt_table[loop].pv_pa +
				(KERNVIRTADDR - abp->abp_physaddr);
		}
	}
	freemem_pt = freemempos;		/* base of allocated pt's */

	/*
	 * Re-align allocation boundary so we can map the area
	 * write-back instead of write-through for the stacks and
	 * related structures allocated below.
	 */
	freemempos = PHYSADDR + 0x100000;
	/*
	 * Allocate a page for the system page mapped to V0x00000000
	 * This page will just contain the system vectors and can be
	 * shared by all processes.
	 */
	valloc_pages(systempage, 1);

	/* Allocate dynamic per-cpu area. */
	valloc_pages(dpcpu, DPCPU_SIZE / PAGE_SIZE);
	dpcpu_init((void *)dpcpu.pv_va, 0);

	/* Allocate stacks for all modes */
	valloc_pages(irqstack, IRQ_STACK_SIZE);
	valloc_pages(abtstack, ABT_STACK_SIZE);
	valloc_pages(undstack, UND_STACK_SIZE);
	valloc_pages(kernelstack, KSTACK_PAGES);
	alloc_pages(minidataclean.pv_pa, 1);
	valloc_pages(msgbufpv, round_page(msgbufsize) / PAGE_SIZE);

	/*
	 * Now construct the L1 page table.  First map the L2
	 * page tables into the L1 so we can replace L1 mappings
	 * later on if necessary
	 */
	l1pagetable = kernel_l1pt.pv_va;

	/* Map the L2 pages tables in the L1 page table */
	pmap_link_l2pt(l1pagetable, ARM_VECTORS_HIGH & ~(0x00100000 - 1),
	    &kernel_pt_table[KERNEL_PT_SYS]);
	pmap_link_l2pt(l1pagetable, IXP425_IO_VBASE,
	    &kernel_pt_table[KERNEL_PT_IO]);
	pmap_link_l2pt(l1pagetable, IXP425_MCU_VBASE,
	    &kernel_pt_table[KERNEL_PT_IO + 1]);
	pmap_link_l2pt(l1pagetable, IXP425_PCI_MEM_VBASE,
	    &kernel_pt_table[KERNEL_PT_IO + 2]);
	pmap_link_l2pt(l1pagetable, KERNBASE,
	    &kernel_pt_table[KERNEL_PT_BEFOREKERN]);
	pmap_map_chunk(l1pagetable, KERNBASE, PHYSADDR, 0x100000,
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	pmap_map_chunk(l1pagetable, KERNBASE + 0x100000, PHYSADDR + 0x100000,
	    0x100000, VM_PROT_READ|VM_PROT_WRITE, PTE_PAGETABLE);
	pmap_map_chunk(l1pagetable, KERNEL_TEXT_BASE, KERNEL_TEXT_PHYS,
	    next_chunk2(((uint32_t)lastaddr) - KERNEL_TEXT_BASE, L1_S_SIZE),
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	freemem_after = next_page((int)lastaddr);
	afterkern = round_page(next_chunk2((vm_offset_t)lastaddr, L1_S_SIZE));
	for (i = 0; i < KERNEL_PT_AFKERNEL_NUM; i++) {
		pmap_link_l2pt(l1pagetable, afterkern + i * 0x00100000,
		    &kernel_pt_table[KERNEL_PT_AFKERNEL + i]);
	}
	pmap_map_entry(l1pagetable, afterkern, minidataclean.pv_pa,
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);


	/* Map the Mini-Data cache clean area. */
	xscale_setup_minidata(l1pagetable, afterkern,
	    minidataclean.pv_pa);

	/* Map the vector page. */
	pmap_map_entry(l1pagetable, ARM_VECTORS_HIGH, systempage.pv_pa,
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	if (cpu_is_ixp43x())
		arm_devmap_bootstrap(l1pagetable, ixp435_devmap);
	else
		arm_devmap_bootstrap(l1pagetable, ixp425_devmap);
	/*
	 * Give the XScale global cache clean code an appropriately
	 * sized chunk of unmapped VA space starting at 0xff000000
	 * (our device mappings end before this address).
	 */
	xscale_cache_clean_addr = 0xff000000U;

	cpu_domains((DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL*2)) | DOMAIN_CLIENT);
	setttb(kernel_l1pt.pv_pa);
	cpu_tlb_flushID();
	cpu_domains(DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL*2));

	/*
	 * Pages were allocated during the secondary bootstrap for the
	 * stacks for different CPU modes.
	 * We must now set the r13 registers in the different CPU modes to
	 * point to these stacks.
	 * Since the ARM stacks use STMFD etc. we must set r13 to the top end
	 * of the stack memory.
	 */
	set_stackptrs(0);

	/*
	 * We must now clean the cache again....
	 * Cleaning may be done by reading new data to displace any
	 * dirty data in the cache. This will have happened in setttb()
	 * but since we are boot strapping the addresses used for the read
	 * may have just been remapped and thus the cache could be out
	 * of sync. A re-clean after the switch will cure this.
	 * After booting there are no gross relocations of the kernel thus
	 * this problem will not occur after initarm().
	 */
	cpu_idcache_wbinv_all();
	cpu_setup();

	/* ready to setup the console (XXX move earlier if possible) */
	cninit();
	/*
	 * Fetch the RAM size from the MCU registers.  The
	 * expansion bus was mapped above so we can now read 'em.
	 */
	if (cpu_is_ixp43x())
		memsize = ixp435_ddram_size();
	else
		memsize = ixp425_sdram_size();

	undefined_init();

	init_proc0(kernelstack.pv_va);

	arm_vector_init(ARM_VECTORS_HIGH, ARM_VEC_ALL);

	pmap_curmaxkvaddr = afterkern + PAGE_SIZE;
	vm_max_kernel_address = 0xe0000000;
	pmap_bootstrap(pmap_curmaxkvaddr, &kernel_l1pt);
	msgbufp = (void*)msgbufpv.pv_va;
	msgbufinit(msgbufp, msgbufsize);
	mutex_init();

	/*
	 * Add the physical ram we have available.
	 *
	 * Exclude the kernel, and all the things we allocated which immediately
	 * follow the kernel, from the VM allocation pool but not from crash
	 * dumps.  virtual_avail is a global variable which tracks the kva we've
	 * "allocated" while setting up pmaps.
	 *
	 * Prepare the list of physical memory available to the vm subsystem.
	 */
	arm_physmem_hardware_region(PHYSADDR, memsize);
	arm_physmem_exclude_region(freemem_pt, KERNPHYSADDR -
	    freemem_pt, EXFLAG_NOALLOC);
	arm_physmem_exclude_region(freemempos, KERNPHYSADDR - 0x100000 -
	    freemempos, EXFLAG_NOALLOC);
	arm_physmem_exclude_region(abp->abp_physaddr, 
	    virtual_avail - KERNVIRTADDR, EXFLAG_NOALLOC);
	arm_physmem_init_kernel_globals();

	init_param2(physmem);
	kdb_init();

	/* use static kernel environment if so configured */
	if (envmode == 1)
		kern_envp = static_env;

	return ((void *)(kernelstack.pv_va + USPACE_SVC_STACK_TOP -
	    sizeof(struct pcb)));
#undef next_page
#undef next_chunk2
}
Exemple #2
0
void *
initarm(void *arg, void *arg2)
{
#define	next_chunk2(a,b)	(((a) + (b)) &~ ((b)-1))
#define	next_page(a)		next_chunk2(a,PAGE_SIZE)
	struct pv_addr  kernel_l1pt;
	struct pv_addr  dpcpu;
	int loop, i;
	u_int l1pagetable;
	vm_offset_t freemempos;
	vm_offset_t freemem_pt;
	vm_offset_t afterkern;
	vm_offset_t freemem_after;
	vm_offset_t lastaddr;
	uint32_t memsize;

	set_cpufuncs();		/* NB: sets cputype */
	lastaddr = fake_preload_metadata();
	pcpu_init(pcpup, 0, sizeof(struct pcpu));
	PCPU_SET(curthread, &thread0);

	/* Do basic tuning, hz etc */
      	init_param1();
		
	/*
	 * We allocate memory downwards from where we were loaded
	 * by RedBoot; first the L1 page table, then NUM_KERNEL_PTS
	 * entries in the L2 page table.  Past that we re-align the
	 * allocation boundary so later data structures (stacks, etc)
	 * can be mapped with different attributes (write-back vs
	 * 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 {					\
	freemempos -= (np * PAGE_SIZE);					\
	(var) = freemempos;						\
	/* NB: this works because locore maps PA=VA */			\
	memset((char *)(var), 0, ((np) * PAGE_SIZE));			\
} while (0)
#define	valloc_pages(var, np) do {					\
	alloc_pages((var).pv_pa, (np));					\
	(var).pv_va = (var).pv_pa + (KERNVIRTADDR - KERNPHYSADDR);	\
} while (0)

	/* force L1 page table alignment */
	while (((freemempos - L1_TABLE_SIZE) & (L1_TABLE_SIZE - 1)) != 0)
		freemempos -= PAGE_SIZE;
	/* allocate contiguous L1 page table */
	valloc_pages(kernel_l1pt, L1_TABLE_SIZE / PAGE_SIZE);
	/* now allocate L2 page tables; they are linked to L1 below */
	for (loop = 0; loop < NUM_KERNEL_PTS; ++loop) {
		if (!(loop % (PAGE_SIZE / L2_TABLE_SIZE_REAL))) {
			valloc_pages(kernel_pt_table[loop],
			    L2_TABLE_SIZE / PAGE_SIZE);
		} else {
			kernel_pt_table[loop].pv_pa = freemempos +
			    (loop % (PAGE_SIZE / L2_TABLE_SIZE_REAL)) *
			    L2_TABLE_SIZE_REAL;
			kernel_pt_table[loop].pv_va = 
			    kernel_pt_table[loop].pv_pa +
				(KERNVIRTADDR - KERNPHYSADDR);
		}
	}
	freemem_pt = freemempos;		/* base of allocated pt's */

	/*
	 * Re-align allocation boundary so we can map the area
	 * write-back instead of write-through for the stacks and
	 * related structures allocated below.
	 */
	freemempos = PHYSADDR + 0x100000;
	/*
	 * Allocate a page for the system page mapped to V0x00000000
	 * This page will just contain the system vectors and can be
	 * shared by all processes.
	 */
	valloc_pages(systempage, 1);

	/* Allocate dynamic per-cpu area. */
	valloc_pages(dpcpu, DPCPU_SIZE / PAGE_SIZE);
	dpcpu_init((void *)dpcpu.pv_va, 0);

	/* Allocate stacks for all modes */
	valloc_pages(irqstack, IRQ_STACK_SIZE);
	valloc_pages(abtstack, ABT_STACK_SIZE);
	valloc_pages(undstack, UND_STACK_SIZE);
	valloc_pages(kernelstack, KSTACK_PAGES);
	alloc_pages(minidataclean.pv_pa, 1);
	valloc_pages(msgbufpv, round_page(msgbufsize) / PAGE_SIZE);
#ifdef ARM_USE_SMALL_ALLOC
	freemempos -= PAGE_SIZE;
	freemem_pt = trunc_page(freemem_pt);
	freemem_after = freemempos - ((freemem_pt - (PHYSADDR + 0x100000)) /
	    PAGE_SIZE) * sizeof(struct arm_small_page);
	arm_add_smallalloc_pages(
	    (void *)(freemem_after + (KERNVIRTADDR - KERNPHYSADDR)),
	    (void *)0xc0100000,
	    freemem_pt - (PHYSADDR + 0x100000), 1);
	freemem_after -= ((freemem_after - (PHYSADDR + 0x1000)) / PAGE_SIZE) *
	    sizeof(struct arm_small_page);
	arm_add_smallalloc_pages(
	    (void *)(freemem_after + (KERNVIRTADDR - KERNPHYSADDR)),
	    (void *)0xc0001000,
	    trunc_page(freemem_after) - (PHYSADDR + 0x1000), 0);
	freemempos = trunc_page(freemem_after);
	freemempos -= PAGE_SIZE;
#endif

	/*
	 * Now construct the L1 page table.  First map the L2
	 * page tables into the L1 so we can replace L1 mappings
	 * later on if necessary
	 */
	l1pagetable = kernel_l1pt.pv_va;

	/* Map the L2 pages tables in the L1 page table */
	pmap_link_l2pt(l1pagetable, ARM_VECTORS_HIGH & ~(0x00100000 - 1),
	    &kernel_pt_table[KERNEL_PT_SYS]);
	pmap_link_l2pt(l1pagetable, IXP425_IO_VBASE,
	    &kernel_pt_table[KERNEL_PT_IO]);
	pmap_link_l2pt(l1pagetable, IXP425_MCU_VBASE,
	    &kernel_pt_table[KERNEL_PT_IO + 1]);
	pmap_link_l2pt(l1pagetable, IXP425_PCI_MEM_VBASE,
	    &kernel_pt_table[KERNEL_PT_IO + 2]);
	pmap_link_l2pt(l1pagetable, KERNBASE,
	    &kernel_pt_table[KERNEL_PT_BEFOREKERN]);
	pmap_map_chunk(l1pagetable, KERNBASE, PHYSADDR, 0x100000,
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	pmap_map_chunk(l1pagetable, KERNBASE + 0x100000, PHYSADDR + 0x100000,
	    0x100000, VM_PROT_READ|VM_PROT_WRITE, PTE_PAGETABLE);
	pmap_map_chunk(l1pagetable, KERNEL_TEXT_BASE, KERNEL_TEXT_PHYS,
	    next_chunk2(((uint32_t)lastaddr) - KERNEL_TEXT_BASE, L1_S_SIZE),
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	freemem_after = next_page((int)lastaddr);
	afterkern = round_page(next_chunk2((vm_offset_t)lastaddr, L1_S_SIZE));
	for (i = 0; i < KERNEL_PT_AFKERNEL_NUM; i++) {
		pmap_link_l2pt(l1pagetable, afterkern + i * 0x00100000,
		    &kernel_pt_table[KERNEL_PT_AFKERNEL + i]);
	}
	pmap_map_entry(l1pagetable, afterkern, minidataclean.pv_pa, 
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);

#ifdef ARM_USE_SMALL_ALLOC
	if ((freemem_after + 2 * PAGE_SIZE) <= afterkern) {
		arm_add_smallalloc_pages((void *)(freemem_after),
		    (void*)(freemem_after + PAGE_SIZE),
		    afterkern - (freemem_after + PAGE_SIZE), 0);
		    
	}
#endif

	/* Map the Mini-Data cache clean area. */
	xscale_setup_minidata(l1pagetable, afterkern,
	    minidataclean.pv_pa);

	/* Map the vector page. */
	pmap_map_entry(l1pagetable, ARM_VECTORS_HIGH, systempage.pv_pa,
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	if (cpu_is_ixp43x())
		pmap_devmap_bootstrap(l1pagetable, ixp435_devmap);
	else
		pmap_devmap_bootstrap(l1pagetable, ixp425_devmap);
	/*
	 * Give the XScale global cache clean code an appropriately
	 * sized chunk of unmapped VA space starting at 0xff000000
	 * (our device mappings end before this address).
	 */
	xscale_cache_clean_addr = 0xff000000U;

	cpu_domains((DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL*2)) | DOMAIN_CLIENT);
	setttb(kernel_l1pt.pv_pa);
	cpu_tlb_flushID();
	cpu_domains(DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL*2));

	/*
	 * Pages were allocated during the secondary bootstrap for the
	 * stacks for different CPU modes.
	 * We must now set the r13 registers in the different CPU modes to
	 * point to these stacks.
	 * Since the ARM stacks use STMFD etc. we must set r13 to the top end
	 * of the stack memory.
	 */
	set_stackptr(PSR_IRQ32_MODE, irqstack.pv_va + IRQ_STACK_SIZE*PAGE_SIZE);
	set_stackptr(PSR_ABT32_MODE, abtstack.pv_va + ABT_STACK_SIZE*PAGE_SIZE);
	set_stackptr(PSR_UND32_MODE, undstack.pv_va + UND_STACK_SIZE*PAGE_SIZE);

	/*
	 * We must now clean the cache again....
	 * Cleaning may be done by reading new data to displace any
	 * dirty data in the cache. This will have happened in setttb()
	 * but since we are boot strapping the addresses used for the read
	 * may have just been remapped and thus the cache could be out
	 * of sync. A re-clean after the switch will cure this.
	 * After booting there are no gross relocations of the kernel thus
	 * this problem will not occur after initarm().
	 */
	cpu_idcache_wbinv_all();
	/* ready to setup the console (XXX move earlier if possible) */
	cninit();
	/*
	 * Fetch the RAM size from the MCU registers.  The
	 * expansion bus was mapped above so we can now read 'em.
	 */
	if (cpu_is_ixp43x())
		memsize = ixp435_ddram_size();
	else
		memsize = ixp425_sdram_size();
	physmem = memsize / PAGE_SIZE;

	/* Set stack for exception handlers */

	data_abort_handler_address = (u_int)data_abort_handler;
	prefetch_abort_handler_address = (u_int)prefetch_abort_handler;
	undefined_handler_address = (u_int)undefinedinstruction_bounce;
	undefined_init();

	proc_linkup0(&proc0, &thread0);
	thread0.td_kstack = kernelstack.pv_va;
	thread0.td_pcb = (struct pcb *)
		(thread0.td_kstack + KSTACK_PAGES * PAGE_SIZE) - 1;
	thread0.td_pcb->pcb_flags = 0;
	thread0.td_frame = &proc0_tf;
	pcpup->pc_curpcb = thread0.td_pcb;

	arm_vector_init(ARM_VECTORS_HIGH, ARM_VEC_ALL);

	pmap_curmaxkvaddr = afterkern + PAGE_SIZE;
	dump_avail[0] = PHYSADDR;
	dump_avail[1] = PHYSADDR + memsize;
	dump_avail[2] = 0;
	dump_avail[3] = 0;

	pmap_bootstrap(pmap_curmaxkvaddr, 0xd0000000, &kernel_l1pt);
	msgbufp = (void*)msgbufpv.pv_va;
	msgbufinit(msgbufp, msgbufsize);
	mutex_init();

	i = 0;
#ifdef ARM_USE_SMALL_ALLOC
	phys_avail[i++] = PHYSADDR;
	phys_avail[i++] = PHYSADDR + PAGE_SIZE; 	/*
					 *XXX: Gross hack to get our
					 * pages in the vm_page_array.
					 */
#endif
	phys_avail[i++] = round_page(virtual_avail - KERNBASE + PHYSADDR);
	phys_avail[i++] = trunc_page(PHYSADDR + memsize - 1);
	phys_avail[i++] = 0;
	phys_avail[i] = 0;

	init_param2(physmem);
	kdb_init();

	/* use static kernel environment if so configured */
	if (envmode == 1)
		kern_envp = static_env;

	return ((void *)(kernelstack.pv_va + USPACE_SVC_STACK_TOP -
	    sizeof(struct pcb)));
#undef next_page
#undef next_chunk2
}
/*
 * u_int initarm(...)
 *
 * Initial entry point on startup. This gets called before main() is
 * entered.
 * It should be responsible for setting up everything that must be
 * in place when main is called.
 * This includes
 *   Taking a copy of the boot configuration structure.
 *   Initialising the physical console so characters can be printed.
 *   Setting up page tables for the kernel
 *   Relocating the kernel to the bottom of physical memory
 */
u_int
initarm(void *arg)
{
	extern vaddr_t xscale_cache_clean_addr;
#ifdef DIAGNOSTIC
	extern vsize_t xscale_minidata_clean_size;
#endif
	int loop;
	int loop1;
	u_int kerneldatasize;
	u_int l1pagetable;
	u_int freemempos;

	/*
	 * Since we map v0xf0000000 == p0xc8000000, it's possible for
	 * us to initialize the console now.
	 */
	consinit();

#ifdef VERBOSE_INIT_ARM
	/* Talk to the user */
	printf("\nNetBSD/evbarm (Intel IXDP425) booting ...\n");
#endif

	/*
	 * Heads up ... Setup the CPU / MMU / TLB functions
	 */
	if (set_cpufuncs())
		panic("cpu not recognized!");

	/* XXX overwrite bootconfig to hardcoded values */
	bootconfig.dramblocks = 1;
	bootconfig.dram[0].address = 0x10000000;
	bootconfig.dram[0].pages = ixp425_sdram_size() / PAGE_SIZE;

	kerneldatasize = (u_int32_t)&end - (u_int32_t)KERNEL_TEXT_BASE;

#ifdef VERBOSE_INIT_ARM
        printf("kernsize=0x%x\n", kerneldatasize);
#endif
        kerneldatasize = ((kerneldatasize - 1) & ~(PAGE_SIZE * 4 - 1)) + PAGE_SIZE * 8;

	/*
	 * Set up the variables that define the availablilty of
	 * physical memory.  For now, we're going to set
	 * physical_freestart to 0x10200000 (where the kernel
	 * was loaded), and allocate the memory we need downwards.
	 * If we get too close to the L1 table that we set up, we
	 * will panic.  We will update physical_freestart and
	 * physical_freeend later to reflect what pmap_bootstrap()
	 * wants to see.
	 *
	 * XXX pmap_bootstrap() needs an enema.
	 */
	physical_start = bootconfig.dram[0].address;
	physical_end = physical_start + (bootconfig.dram[0].pages * PAGE_SIZE);

	physical_freestart = physical_start
                + (KERNEL_TEXT_BASE - KERNEL_BASE) + kerneldatasize;
        physical_freeend = physical_end;

	physmem = (physical_end - physical_start) / PAGE_SIZE;

	/* Tell the user about the memory */
#ifdef VERBOSE_INIT_ARM
	printf("physmemory: %d pages at 0x%08lx -> 0x%08lx\n", physmem,
	    physical_start, physical_end - 1);

	printf("Allocating page tables\n");
#endif
	free_pages = (physical_freeend - physical_freestart) / PAGE_SIZE;

	freemempos = 0x10000000;

#ifdef VERBOSE_INIT_ARM
        printf("physical_start = 0x%08lx, physical_end = 0x%08lx\n",
                physical_start, physical_end);
#endif

	/* Define a macro to simplify memory allocation */
#define	valloc_pages(var, np)				\
	alloc_pages((var).pv_pa, (np));			\
	(var).pv_va = KERNEL_BASE + (var).pv_pa - physical_start;

#if 0
#define alloc_pages(var, np)				\
	physical_freeend -= ((np) * PAGE_SIZE);		\
	if (physical_freeend < physical_freestart)	\
		panic("initarm: out of memory");	\
	(var) = physical_freeend;			\
	free_pages -= (np);				\
	memset((char *)(var), 0, ((np) * PAGE_SIZE));
#else
#define alloc_pages(var, np)				\
        (var) = freemempos;                             \
        memset((char *)(var), 0, ((np) * PAGE_SIZE));   \
        freemempos += (np) * PAGE_SIZE;
#endif

	loop1 = 0;
	for (loop = 0; loop <= NUM_KERNEL_PTS; ++loop) {
		/* Are we 16KB aligned for an L1 ? */
		if (((physical_freeend - L1_TABLE_SIZE) & (L1_TABLE_SIZE - 1)) == 0
		    && kernel_l1pt.pv_pa == 0) {
			valloc_pages(kernel_l1pt, L1_TABLE_SIZE / PAGE_SIZE);
		} else {
			valloc_pages(kernel_pt_table[loop1],
			    L2_TABLE_SIZE / PAGE_SIZE);
			++loop1;
		}
	}

	/* This should never be able to happen but better confirm that. */
	if (!kernel_l1pt.pv_pa || (kernel_l1pt.pv_pa & (L1_TABLE_SIZE-1)) != 0)
		panic("initarm: Failed to align the kernel page directory");

	/*
	 * Allocate a page for the system page.
	 * This page will just contain the system vectors and can be
	 * shared by all processes.
	 */
	alloc_pages(systempage.pv_pa, 1);

	/* Allocate stacks for all modes */
	valloc_pages(irqstack, IRQ_STACK_SIZE);
	valloc_pages(abtstack, ABT_STACK_SIZE);
	valloc_pages(undstack, UND_STACK_SIZE);
	valloc_pages(kernelstack, UPAGES);

	/* Allocate enough pages for cleaning the Mini-Data cache. */
	KASSERT(xscale_minidata_clean_size <= PAGE_SIZE);
	valloc_pages(minidataclean, 1);

#ifdef VERBOSE_INIT_ARM
	printf("IRQ stack: p0x%08lx v0x%08lx\n", irqstack.pv_pa,
	    irqstack.pv_va); 
	printf("ABT stack: p0x%08lx v0x%08lx\n", abtstack.pv_pa,
	    abtstack.pv_va); 
	printf("UND stack: p0x%08lx v0x%08lx\n", undstack.pv_pa,
	    undstack.pv_va); 
	printf("SVC stack: p0x%08lx v0x%08lx\n", kernelstack.pv_pa,
	    kernelstack.pv_va); 
#endif

	/*
	 * XXX Defer this to later so that we can reclaim the memory
	 * XXX used by the RedBoot page tables.
	 */
	alloc_pages(msgbufphys, round_page(MSGBUFSIZE) / PAGE_SIZE);

	/*
	 * Ok we have allocated physical pages for the primary kernel
	 * page tables
	 */

#ifdef VERBOSE_INIT_ARM
	printf("Creating L1 page table at 0x%08lx\n", kernel_l1pt.pv_pa);
#endif

	/*
	 * Now we start construction of the L1 page table
	 * We start by mapping the L2 page tables into the L1.
	 * This means that we can replace L1 mappings later on if necessary
	 */
	l1pagetable = kernel_l1pt.pv_pa;

	/* Map the L2 pages tables in the L1 page table */
	pmap_link_l2pt(l1pagetable, ARM_VECTORS_HIGH & ~(0x00400000 - 1),
	    &kernel_pt_table[KERNEL_PT_SYS]);
	for (loop = 0; loop < KERNEL_PT_KERNEL_NUM; loop++)
		pmap_link_l2pt(l1pagetable, KERNEL_BASE + loop * 0x00400000,
		    &kernel_pt_table[KERNEL_PT_KERNEL + loop]);
	for (loop = 0; loop < KERNEL_PT_VMDATA_NUM; loop++)
		pmap_link_l2pt(l1pagetable, KERNEL_VM_BASE + loop * 0x00400000,
		    &kernel_pt_table[KERNEL_PT_VMDATA + loop]);

	/* update the top of the kernel VM */
	pmap_curmaxkvaddr =
	    KERNEL_VM_BASE + (KERNEL_PT_VMDATA_NUM * 0x00400000);

	pmap_link_l2pt(l1pagetable, IXP425_IO_VBASE,
	    &kernel_pt_table[KERNEL_PT_IO]);

#ifdef VERBOSE_INIT_ARM
	printf("Mapping kernel\n");
#endif

	/* Now we fill in the L2 pagetable for the kernel static code/data */
	{
		extern char etext[], _end[];
		size_t textsize = (uintptr_t) etext - KERNEL_TEXT_BASE;
		size_t totalsize = (uintptr_t) _end - KERNEL_TEXT_BASE;
		u_int logical;

		textsize = (textsize + PGOFSET) & ~PGOFSET;
		totalsize = (totalsize + PGOFSET) & ~PGOFSET;
		
		logical = 0x00200000;	/* offset of kernel in RAM */

		logical += pmap_map_chunk(l1pagetable, KERNEL_BASE + logical,
		    physical_start + logical, textsize,
		    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
		logical += pmap_map_chunk(l1pagetable, KERNEL_BASE + logical,
		    physical_start + logical, totalsize - textsize,
		    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	}

#ifdef VERBOSE_INIT_ARM
	printf("Constructing L2 page tables\n");
#endif

	/* Map the stack pages */
	pmap_map_chunk(l1pagetable, irqstack.pv_va, irqstack.pv_pa,
	    IRQ_STACK_SIZE * PAGE_SIZE, VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	pmap_map_chunk(l1pagetable, abtstack.pv_va, abtstack.pv_pa,
	    ABT_STACK_SIZE * PAGE_SIZE, VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	pmap_map_chunk(l1pagetable, undstack.pv_va, undstack.pv_pa,
	    UND_STACK_SIZE * PAGE_SIZE, VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);
	pmap_map_chunk(l1pagetable, kernelstack.pv_va, kernelstack.pv_pa,
	    UPAGES * PAGE_SIZE, VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);

	pmap_map_chunk(l1pagetable, kernel_l1pt.pv_va, kernel_l1pt.pv_pa,
	    L1_TABLE_SIZE, VM_PROT_READ|VM_PROT_WRITE, PTE_PAGETABLE);

	for (loop = 0; loop < NUM_KERNEL_PTS; ++loop) {
		pmap_map_chunk(l1pagetable, kernel_pt_table[loop].pv_va,
		    kernel_pt_table[loop].pv_pa, L2_TABLE_SIZE,
		    VM_PROT_READ|VM_PROT_WRITE, PTE_PAGETABLE);
	}

	/* Map the Mini-Data cache clean area. */
	xscale_setup_minidata(l1pagetable, minidataclean.pv_va,
	    minidataclean.pv_pa);

	/* Map the vector page. */
	pmap_map_entry(l1pagetable, ARM_VECTORS_HIGH, systempage.pv_pa,
	    VM_PROT_READ|VM_PROT_WRITE, PTE_CACHE);

        /*
         * Map the IXP425 registers
         */
	pmap_devmap_bootstrap(l1pagetable, ixp425_devmap);

	/*
	 * Give the XScale global cache clean code an appropriately
	 * sized chunk of unmapped VA space starting at 0xff000000
	 * (our device mappings end before this address).
	 */
	xscale_cache_clean_addr = 0xff000000U;

	/*
	 * Now we have the real page tables in place so we can switch to them.
	 * Once this is done we will be running with the REAL kernel page
	 * tables.
	 */

	/*
	 * Update the physical_freestart/physical_freeend/free_pages
	 * variables.
	 */
	{
		extern char _end[];

		physical_freestart = physical_start +
		    (((((uintptr_t) _end) + PGOFSET) & ~PGOFSET) -
		     KERNEL_BASE);
		physical_freeend = physical_end;
		free_pages =
		    (physical_freeend - physical_freestart) / PAGE_SIZE;
	}

	/* Switch tables */
#ifdef VERBOSE_INIT_ARM
	printf("freestart = 0x%08lx, free_pages = %d (0x%x)\n",
	       physical_freestart, free_pages, free_pages);
	printf("switching to new L1 page table  @%#lx...", kernel_l1pt.pv_pa);
#endif
	cpu_domains((DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL*2)) | DOMAIN_CLIENT);
	setttb(kernel_l1pt.pv_pa);
	cpu_tlb_flushID();
	cpu_domains(DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL*2));

	/*
	 * Moved from cpu_startup() as data_abort_handler() references
	 * this during uvm init
	 */
	proc0paddr = (struct user *)kernelstack.pv_va;
	lwp0.l_addr = proc0paddr;

#ifdef VERBOSE_INIT_ARM
	printf("bootstrap done.\n");
#endif

	arm32_vector_init(ARM_VECTORS_HIGH, ARM_VEC_ALL);

	/*
	 * Pages were allocated during the secondary bootstrap for the
	 * stacks for different CPU modes.
	 * We must now set the r13 registers in the different CPU modes to
	 * point to these stacks.
	 * Since the ARM stacks use STMFD etc. we must set r13 to the top end
	 * of the stack memory.
	 */
#ifdef VERBOSE_INIT_ARM
	printf("init subsystems: stacks ");
#endif

	set_stackptr(PSR_IRQ32_MODE,
	    irqstack.pv_va + IRQ_STACK_SIZE * PAGE_SIZE);
	set_stackptr(PSR_ABT32_MODE,
	    abtstack.pv_va + ABT_STACK_SIZE * PAGE_SIZE);
	set_stackptr(PSR_UND32_MODE,
	    undstack.pv_va + UND_STACK_SIZE * PAGE_SIZE);

	/*
	 * Well we should set a data abort handler.
	 * Once things get going this will change as we will need a proper
	 * handler.
	 * Until then we will use a handler that just panics but tells us
	 * why.
	 * Initialisation of the vectors will just panic on a data abort.
	 * This just fills in a slightly better one.
	 */
#ifdef VERBOSE_INIT_ARM
	printf("vectors ");
#endif
	data_abort_handler_address = (u_int)data_abort_handler;
	prefetch_abort_handler_address = (u_int)prefetch_abort_handler;
	undefined_handler_address = (u_int)undefinedinstruction_bounce;

	/* Initialise the undefined instruction handlers */
#ifdef VERBOSE_INIT_ARM
	printf("undefined ");
#endif
	undefined_init();

	/* Load memory into UVM. */
#ifdef VERBOSE_INIT_ARM
	printf("page ");
#endif
	uvm_setpagesize();	/* initialize PAGE_SIZE-dependent variables */
	uvm_page_physload(atop(physical_freestart), atop(physical_freeend),
	    atop(physical_freestart), atop(physical_freeend),
	    VM_FREELIST_DEFAULT);

	/* Boot strap pmap telling it where the kernel page table is */
#ifdef VERBOSE_INIT_ARM
	printf("pmap ");
#endif
	pmap_bootstrap(KERNEL_VM_BASE, KERNEL_VM_BASE + KERNEL_VM_SIZE);

	/* Setup the IRQ system */
#ifdef VERBOSE_INIT_ARM
	printf("irq ");
#endif
	ixp425_intr_init();
#ifdef VERBOSE_INIT_ARM
	printf("\nAll initialize done!\nNow Starting NetBSD, Hear we go!\n");
#endif

#ifdef BOOTHOWTO
	boothowto = BOOTHOWTO;
#endif

#if NKSYMS || defined(DDB) || defined(LKM)
	/* Firmware doesn't load symbols. */
	ksyms_init(0, NULL, NULL);
#endif

#ifdef DDB
	db_machine_init();
	if (boothowto & RB_KDB)
		Debugger();
#endif

	/* We return the new stack pointer address */
	return(kernelstack.pv_va + USPACE_SVC_STACK_TOP);
}