xref: /linux/arch/arm/mm/mmu.c (revision c537b994505099b7197e7d3125b942ecbcc51eb6)
1 /*
2  *  linux/arch/arm/mm/mmu.c
3  *
4  *  Copyright (C) 1995-2005 Russell King
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  */
10 #include <linux/module.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/init.h>
14 #include <linux/bootmem.h>
15 #include <linux/mman.h>
16 #include <linux/nodemask.h>
17 
18 #include <asm/mach-types.h>
19 #include <asm/setup.h>
20 #include <asm/sizes.h>
21 #include <asm/tlb.h>
22 
23 #include <asm/mach/arch.h>
24 #include <asm/mach/map.h>
25 
26 #include "mm.h"
27 
28 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
29 
30 extern void _stext, _etext, __data_start, _end;
31 extern pgd_t swapper_pg_dir[PTRS_PER_PGD];
32 
33 /*
34  * empty_zero_page is a special page that is used for
35  * zero-initialized data and COW.
36  */
37 struct page *empty_zero_page;
38 
39 /*
40  * The pmd table for the upper-most set of pages.
41  */
42 pmd_t *top_pmd;
43 
44 #define CPOLICY_UNCACHED	0
45 #define CPOLICY_BUFFERED	1
46 #define CPOLICY_WRITETHROUGH	2
47 #define CPOLICY_WRITEBACK	3
48 #define CPOLICY_WRITEALLOC	4
49 
50 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
51 static unsigned int ecc_mask __initdata = 0;
52 pgprot_t pgprot_user;
53 pgprot_t pgprot_kernel;
54 
55 EXPORT_SYMBOL(pgprot_user);
56 EXPORT_SYMBOL(pgprot_kernel);
57 
58 struct cachepolicy {
59 	const char	policy[16];
60 	unsigned int	cr_mask;
61 	unsigned int	pmd;
62 	unsigned int	pte;
63 };
64 
65 static struct cachepolicy cache_policies[] __initdata = {
66 	{
67 		.policy		= "uncached",
68 		.cr_mask	= CR_W|CR_C,
69 		.pmd		= PMD_SECT_UNCACHED,
70 		.pte		= 0,
71 	}, {
72 		.policy		= "buffered",
73 		.cr_mask	= CR_C,
74 		.pmd		= PMD_SECT_BUFFERED,
75 		.pte		= PTE_BUFFERABLE,
76 	}, {
77 		.policy		= "writethrough",
78 		.cr_mask	= 0,
79 		.pmd		= PMD_SECT_WT,
80 		.pte		= PTE_CACHEABLE,
81 	}, {
82 		.policy		= "writeback",
83 		.cr_mask	= 0,
84 		.pmd		= PMD_SECT_WB,
85 		.pte		= PTE_BUFFERABLE|PTE_CACHEABLE,
86 	}, {
87 		.policy		= "writealloc",
88 		.cr_mask	= 0,
89 		.pmd		= PMD_SECT_WBWA,
90 		.pte		= PTE_BUFFERABLE|PTE_CACHEABLE,
91 	}
92 };
93 
94 /*
95  * These are useful for identifing cache coherency
96  * problems by allowing the cache or the cache and
97  * writebuffer to be turned off.  (Note: the write
98  * buffer should not be on and the cache off).
99  */
100 static void __init early_cachepolicy(char **p)
101 {
102 	int i;
103 
104 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
105 		int len = strlen(cache_policies[i].policy);
106 
107 		if (memcmp(*p, cache_policies[i].policy, len) == 0) {
108 			cachepolicy = i;
109 			cr_alignment &= ~cache_policies[i].cr_mask;
110 			cr_no_alignment &= ~cache_policies[i].cr_mask;
111 			*p += len;
112 			break;
113 		}
114 	}
115 	if (i == ARRAY_SIZE(cache_policies))
116 		printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
117 	flush_cache_all();
118 	set_cr(cr_alignment);
119 }
120 __early_param("cachepolicy=", early_cachepolicy);
121 
122 static void __init early_nocache(char **__unused)
123 {
124 	char *p = "buffered";
125 	printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
126 	early_cachepolicy(&p);
127 }
128 __early_param("nocache", early_nocache);
129 
130 static void __init early_nowrite(char **__unused)
131 {
132 	char *p = "uncached";
133 	printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
134 	early_cachepolicy(&p);
135 }
136 __early_param("nowb", early_nowrite);
137 
138 static void __init early_ecc(char **p)
139 {
140 	if (memcmp(*p, "on", 2) == 0) {
141 		ecc_mask = PMD_PROTECTION;
142 		*p += 2;
143 	} else if (memcmp(*p, "off", 3) == 0) {
144 		ecc_mask = 0;
145 		*p += 3;
146 	}
147 }
148 __early_param("ecc=", early_ecc);
149 
150 static int __init noalign_setup(char *__unused)
151 {
152 	cr_alignment &= ~CR_A;
153 	cr_no_alignment &= ~CR_A;
154 	set_cr(cr_alignment);
155 	return 1;
156 }
157 __setup("noalign", noalign_setup);
158 
159 #ifndef CONFIG_SMP
160 void adjust_cr(unsigned long mask, unsigned long set)
161 {
162 	unsigned long flags;
163 
164 	mask &= ~CR_A;
165 
166 	set &= mask;
167 
168 	local_irq_save(flags);
169 
170 	cr_no_alignment = (cr_no_alignment & ~mask) | set;
171 	cr_alignment = (cr_alignment & ~mask) | set;
172 
173 	set_cr((get_cr() & ~mask) | set);
174 
175 	local_irq_restore(flags);
176 }
177 #endif
178 
179 struct mem_types {
180 	unsigned int	prot_pte;
181 	unsigned int	prot_l1;
182 	unsigned int	prot_sect;
183 	unsigned int	domain;
184 };
185 
186 static struct mem_types mem_types[] __initdata = {
187 	[MT_DEVICE] = {
188 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
189 				L_PTE_WRITE,
190 		.prot_l1   = PMD_TYPE_TABLE,
191 		.prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_UNCACHED |
192 				PMD_SECT_AP_WRITE,
193 		.domain    = DOMAIN_IO,
194 	},
195 	[MT_CACHECLEAN] = {
196 		.prot_sect = PMD_TYPE_SECT | PMD_BIT4,
197 		.domain    = DOMAIN_KERNEL,
198 	},
199 	[MT_MINICLEAN] = {
200 		.prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_MINICACHE,
201 		.domain    = DOMAIN_KERNEL,
202 	},
203 	[MT_LOW_VECTORS] = {
204 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
205 				L_PTE_EXEC,
206 		.prot_l1   = PMD_TYPE_TABLE,
207 		.domain    = DOMAIN_USER,
208 	},
209 	[MT_HIGH_VECTORS] = {
210 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
211 				L_PTE_USER | L_PTE_EXEC,
212 		.prot_l1   = PMD_TYPE_TABLE,
213 		.domain    = DOMAIN_USER,
214 	},
215 	[MT_MEMORY] = {
216 		.prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_AP_WRITE,
217 		.domain    = DOMAIN_KERNEL,
218 	},
219 	[MT_ROM] = {
220 		.prot_sect = PMD_TYPE_SECT | PMD_BIT4,
221 		.domain    = DOMAIN_KERNEL,
222 	},
223 	[MT_IXP2000_DEVICE] = { /* IXP2400 requires XCB=101 for on-chip I/O */
224 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
225 				L_PTE_WRITE,
226 		.prot_l1   = PMD_TYPE_TABLE,
227 		.prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_UNCACHED |
228 				PMD_SECT_AP_WRITE | PMD_SECT_BUFFERABLE |
229 				PMD_SECT_TEX(1),
230 		.domain    = DOMAIN_IO,
231 	},
232 	[MT_NONSHARED_DEVICE] = {
233 		.prot_l1   = PMD_TYPE_TABLE,
234 		.prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_NONSHARED_DEV |
235 				PMD_SECT_AP_WRITE,
236 		.domain    = DOMAIN_IO,
237 	}
238 };
239 
240 /*
241  * Adjust the PMD section entries according to the CPU in use.
242  */
243 static void __init build_mem_type_table(void)
244 {
245 	struct cachepolicy *cp;
246 	unsigned int cr = get_cr();
247 	unsigned int user_pgprot, kern_pgprot;
248 	int cpu_arch = cpu_architecture();
249 	int i;
250 
251 #if defined(CONFIG_CPU_DCACHE_DISABLE)
252 	if (cachepolicy > CPOLICY_BUFFERED)
253 		cachepolicy = CPOLICY_BUFFERED;
254 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
255 	if (cachepolicy > CPOLICY_WRITETHROUGH)
256 		cachepolicy = CPOLICY_WRITETHROUGH;
257 #endif
258 	if (cpu_arch < CPU_ARCH_ARMv5) {
259 		if (cachepolicy >= CPOLICY_WRITEALLOC)
260 			cachepolicy = CPOLICY_WRITEBACK;
261 		ecc_mask = 0;
262 	}
263 
264 	/*
265 	 * Xscale must not have PMD bit 4 set for section mappings.
266 	 */
267 	if (cpu_is_xscale())
268 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
269 			mem_types[i].prot_sect &= ~PMD_BIT4;
270 
271 	/*
272 	 * ARMv5 and lower, excluding Xscale, bit 4 must be set for
273 	 * page tables.
274 	 */
275 	if (cpu_arch < CPU_ARCH_ARMv6 && !cpu_is_xscale())
276 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
277 			if (mem_types[i].prot_l1)
278 				mem_types[i].prot_l1 |= PMD_BIT4;
279 
280 	cp = &cache_policies[cachepolicy];
281 	kern_pgprot = user_pgprot = cp->pte;
282 
283 	/*
284 	 * Enable CPU-specific coherency if supported.
285 	 * (Only available on XSC3 at the moment.)
286 	 */
287 	if (arch_is_coherent()) {
288 		if (cpu_is_xsc3()) {
289 			mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
290 			mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
291 		}
292 	}
293 
294 	/*
295 	 * ARMv6 and above have extended page tables.
296 	 */
297 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
298 		/*
299 		 * bit 4 becomes XN which we must clear for the
300 		 * kernel memory mapping.
301 		 */
302 		mem_types[MT_MEMORY].prot_sect &= ~PMD_SECT_XN;
303 		mem_types[MT_ROM].prot_sect &= ~PMD_SECT_XN;
304 
305 		/*
306 		 * Mark cache clean areas and XIP ROM read only
307 		 * from SVC mode and no access from userspace.
308 		 */
309 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
310 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
311 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
312 
313 		/*
314 		 * Mark the device area as "shared device"
315 		 */
316 		mem_types[MT_DEVICE].prot_pte |= L_PTE_BUFFERABLE;
317 		mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
318 
319 #ifdef CONFIG_SMP
320 		/*
321 		 * Mark memory with the "shared" attribute for SMP systems
322 		 */
323 		user_pgprot |= L_PTE_SHARED;
324 		kern_pgprot |= L_PTE_SHARED;
325 		mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
326 #endif
327 	}
328 
329 	for (i = 0; i < 16; i++) {
330 		unsigned long v = pgprot_val(protection_map[i]);
331 		v = (v & ~(L_PTE_BUFFERABLE|L_PTE_CACHEABLE)) | user_pgprot;
332 		protection_map[i] = __pgprot(v);
333 	}
334 
335 	mem_types[MT_LOW_VECTORS].prot_pte |= kern_pgprot;
336 	mem_types[MT_HIGH_VECTORS].prot_pte |= kern_pgprot;
337 
338 	if (cpu_arch >= CPU_ARCH_ARMv5) {
339 #ifndef CONFIG_SMP
340 		/*
341 		 * Only use write-through for non-SMP systems
342 		 */
343 		mem_types[MT_LOW_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
344 		mem_types[MT_HIGH_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
345 #endif
346 	} else {
347 		mem_types[MT_MINICLEAN].prot_sect &= ~PMD_SECT_TEX(1);
348 	}
349 
350 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
351 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
352 				 L_PTE_DIRTY | L_PTE_WRITE |
353 				 L_PTE_EXEC | kern_pgprot);
354 
355 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
356 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
357 	mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
358 	mem_types[MT_ROM].prot_sect |= cp->pmd;
359 
360 	switch (cp->pmd) {
361 	case PMD_SECT_WT:
362 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
363 		break;
364 	case PMD_SECT_WB:
365 	case PMD_SECT_WBWA:
366 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
367 		break;
368 	}
369 	printk("Memory policy: ECC %sabled, Data cache %s\n",
370 		ecc_mask ? "en" : "dis", cp->policy);
371 }
372 
373 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
374 
375 /*
376  * Create a SECTION PGD between VIRT and PHYS in domain
377  * DOMAIN with protection PROT.  This operates on half-
378  * pgdir entry increments.
379  */
380 static inline void
381 alloc_init_section(unsigned long virt, unsigned long phys, int prot)
382 {
383 	pmd_t *pmdp = pmd_off_k(virt);
384 
385 	if (virt & (1 << 20))
386 		pmdp++;
387 
388 	*pmdp = __pmd(phys | prot);
389 	flush_pmd_entry(pmdp);
390 }
391 
392 /*
393  * Create a SUPER SECTION PGD between VIRT and PHYS with protection PROT
394  */
395 static inline void
396 alloc_init_supersection(unsigned long virt, unsigned long phys, int prot)
397 {
398 	int i;
399 
400 	for (i = 0; i < 16; i += 1) {
401 		alloc_init_section(virt, phys, prot | PMD_SECT_SUPER);
402 
403 		virt += (PGDIR_SIZE / 2);
404 	}
405 }
406 
407 /*
408  * Add a PAGE mapping between VIRT and PHYS in domain
409  * DOMAIN with protection PROT.  Note that due to the
410  * way we map the PTEs, we must allocate two PTE_SIZE'd
411  * blocks - one for the Linux pte table, and one for
412  * the hardware pte table.
413  */
414 static inline void
415 alloc_init_page(unsigned long virt, unsigned long phys, unsigned int prot_l1, pgprot_t prot)
416 {
417 	pmd_t *pmdp = pmd_off_k(virt);
418 	pte_t *ptep;
419 
420 	if (pmd_none(*pmdp)) {
421 		ptep = alloc_bootmem_low_pages(2 * PTRS_PER_PTE *
422 					       sizeof(pte_t));
423 
424 		__pmd_populate(pmdp, __pa(ptep) | prot_l1);
425 	}
426 	ptep = pte_offset_kernel(pmdp, virt);
427 
428 	set_pte_ext(ptep, pfn_pte(phys >> PAGE_SHIFT, prot), 0);
429 }
430 
431 /*
432  * Create the page directory entries and any necessary
433  * page tables for the mapping specified by `md'.  We
434  * are able to cope here with varying sizes and address
435  * offsets, and we take full advantage of sections and
436  * supersections.
437  */
438 void __init create_mapping(struct map_desc *md)
439 {
440 	unsigned long virt, length;
441 	int prot_sect, prot_l1, domain;
442 	pgprot_t prot_pte;
443 	unsigned long off = (u32)__pfn_to_phys(md->pfn);
444 
445 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
446 		printk(KERN_WARNING "BUG: not creating mapping for "
447 		       "0x%08llx at 0x%08lx in user region\n",
448 		       __pfn_to_phys((u64)md->pfn), md->virtual);
449 		return;
450 	}
451 
452 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
453 	    md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
454 		printk(KERN_WARNING "BUG: mapping for 0x%08llx at 0x%08lx "
455 		       "overlaps vmalloc space\n",
456 		       __pfn_to_phys((u64)md->pfn), md->virtual);
457 	}
458 
459 	domain	  = mem_types[md->type].domain;
460 	prot_pte  = __pgprot(mem_types[md->type].prot_pte);
461 	prot_l1   = mem_types[md->type].prot_l1 | PMD_DOMAIN(domain);
462 	prot_sect = mem_types[md->type].prot_sect | PMD_DOMAIN(domain);
463 
464 	/*
465 	 * Catch 36-bit addresses
466 	 */
467 	if(md->pfn >= 0x100000) {
468 		if(domain) {
469 			printk(KERN_ERR "MM: invalid domain in supersection "
470 				"mapping for 0x%08llx at 0x%08lx\n",
471 				__pfn_to_phys((u64)md->pfn), md->virtual);
472 			return;
473 		}
474 		if((md->virtual | md->length | __pfn_to_phys(md->pfn))
475 			& ~SUPERSECTION_MASK) {
476 			printk(KERN_ERR "MM: cannot create mapping for "
477 				"0x%08llx at 0x%08lx invalid alignment\n",
478 				__pfn_to_phys((u64)md->pfn), md->virtual);
479 			return;
480 		}
481 
482 		/*
483 		 * Shift bits [35:32] of address into bits [23:20] of PMD
484 		 * (See ARMv6 spec).
485 		 */
486 		off |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
487 	}
488 
489 	virt   = md->virtual;
490 	off   -= virt;
491 	length = md->length;
492 
493 	if (mem_types[md->type].prot_l1 == 0 &&
494 	    (virt & 0xfffff || (virt + off) & 0xfffff || (virt + length) & 0xfffff)) {
495 		printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
496 		       "be mapped using pages, ignoring.\n",
497 		       __pfn_to_phys(md->pfn), md->virtual);
498 		return;
499 	}
500 
501 	while ((virt & 0xfffff || (virt + off) & 0xfffff) && length >= PAGE_SIZE) {
502 		alloc_init_page(virt, virt + off, prot_l1, prot_pte);
503 
504 		virt   += PAGE_SIZE;
505 		length -= PAGE_SIZE;
506 	}
507 
508 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
509 	 *	Since domain assignments can in fact be arbitrary, the
510 	 *	'domain == 0' check below is required to insure that ARMv6
511 	 *	supersections are only allocated for domain 0 regardless
512 	 *	of the actual domain assignments in use.
513 	 */
514 	if ((cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())
515 		&& domain == 0) {
516 		/*
517 		 * Align to supersection boundary if !high pages.
518 		 * High pages have already been checked for proper
519 		 * alignment above and they will fail the SUPSERSECTION_MASK
520 		 * check because of the way the address is encoded into
521 		 * offset.
522 		 */
523 		if (md->pfn <= 0x100000) {
524 			while ((virt & ~SUPERSECTION_MASK ||
525 			        (virt + off) & ~SUPERSECTION_MASK) &&
526 				length >= (PGDIR_SIZE / 2)) {
527 				alloc_init_section(virt, virt + off, prot_sect);
528 
529 				virt   += (PGDIR_SIZE / 2);
530 				length -= (PGDIR_SIZE / 2);
531 			}
532 		}
533 
534 		while (length >= SUPERSECTION_SIZE) {
535 			alloc_init_supersection(virt, virt + off, prot_sect);
536 
537 			virt   += SUPERSECTION_SIZE;
538 			length -= SUPERSECTION_SIZE;
539 		}
540 	}
541 
542 	/*
543 	 * A section mapping covers half a "pgdir" entry.
544 	 */
545 	while (length >= (PGDIR_SIZE / 2)) {
546 		alloc_init_section(virt, virt + off, prot_sect);
547 
548 		virt   += (PGDIR_SIZE / 2);
549 		length -= (PGDIR_SIZE / 2);
550 	}
551 
552 	while (length >= PAGE_SIZE) {
553 		alloc_init_page(virt, virt + off, prot_l1, prot_pte);
554 
555 		virt   += PAGE_SIZE;
556 		length -= PAGE_SIZE;
557 	}
558 }
559 
560 /*
561  * Create the architecture specific mappings
562  */
563 void __init iotable_init(struct map_desc *io_desc, int nr)
564 {
565 	int i;
566 
567 	for (i = 0; i < nr; i++)
568 		create_mapping(io_desc + i);
569 }
570 
571 static inline void prepare_page_table(struct meminfo *mi)
572 {
573 	unsigned long addr;
574 
575 	/*
576 	 * Clear out all the mappings below the kernel image.
577 	 */
578 	for (addr = 0; addr < MODULE_START; addr += PGDIR_SIZE)
579 		pmd_clear(pmd_off_k(addr));
580 
581 #ifdef CONFIG_XIP_KERNEL
582 	/* The XIP kernel is mapped in the module area -- skip over it */
583 	addr = ((unsigned long)&_etext + PGDIR_SIZE - 1) & PGDIR_MASK;
584 #endif
585 	for ( ; addr < PAGE_OFFSET; addr += PGDIR_SIZE)
586 		pmd_clear(pmd_off_k(addr));
587 
588 	/*
589 	 * Clear out all the kernel space mappings, except for the first
590 	 * memory bank, up to the end of the vmalloc region.
591 	 */
592 	for (addr = __phys_to_virt(mi->bank[0].start + mi->bank[0].size);
593 	     addr < VMALLOC_END; addr += PGDIR_SIZE)
594 		pmd_clear(pmd_off_k(addr));
595 }
596 
597 /*
598  * Reserve the various regions of node 0
599  */
600 void __init reserve_node_zero(pg_data_t *pgdat)
601 {
602 	unsigned long res_size = 0;
603 
604 	/*
605 	 * Register the kernel text and data with bootmem.
606 	 * Note that this can only be in node 0.
607 	 */
608 #ifdef CONFIG_XIP_KERNEL
609 	reserve_bootmem_node(pgdat, __pa(&__data_start), &_end - &__data_start);
610 #else
611 	reserve_bootmem_node(pgdat, __pa(&_stext), &_end - &_stext);
612 #endif
613 
614 	/*
615 	 * Reserve the page tables.  These are already in use,
616 	 * and can only be in node 0.
617 	 */
618 	reserve_bootmem_node(pgdat, __pa(swapper_pg_dir),
619 			     PTRS_PER_PGD * sizeof(pgd_t));
620 
621 	/*
622 	 * Hmm... This should go elsewhere, but we really really need to
623 	 * stop things allocating the low memory; ideally we need a better
624 	 * implementation of GFP_DMA which does not assume that DMA-able
625 	 * memory starts at zero.
626 	 */
627 	if (machine_is_integrator() || machine_is_cintegrator())
628 		res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
629 
630 	/*
631 	 * These should likewise go elsewhere.  They pre-reserve the
632 	 * screen memory region at the start of main system memory.
633 	 */
634 	if (machine_is_edb7211())
635 		res_size = 0x00020000;
636 	if (machine_is_p720t())
637 		res_size = 0x00014000;
638 
639 	/* H1940 and RX3715 need to reserve this for suspend */
640 
641 	if (machine_is_h1940() || machine_is_rx3715()) {
642 		reserve_bootmem_node(pgdat, 0x30003000, 0x1000);
643 		reserve_bootmem_node(pgdat, 0x30081000, 0x1000);
644 	}
645 
646 #ifdef CONFIG_SA1111
647 	/*
648 	 * Because of the SA1111 DMA bug, we want to preserve our
649 	 * precious DMA-able memory...
650 	 */
651 	res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
652 #endif
653 	if (res_size)
654 		reserve_bootmem_node(pgdat, PHYS_OFFSET, res_size);
655 }
656 
657 /*
658  * Set up device the mappings.  Since we clear out the page tables for all
659  * mappings above VMALLOC_END, we will remove any debug device mappings.
660  * This means you have to be careful how you debug this function, or any
661  * called function.  This means you can't use any function or debugging
662  * method which may touch any device, otherwise the kernel _will_ crash.
663  */
664 static void __init devicemaps_init(struct machine_desc *mdesc)
665 {
666 	struct map_desc map;
667 	unsigned long addr;
668 	void *vectors;
669 
670 	/*
671 	 * Allocate the vector page early.
672 	 */
673 	vectors = alloc_bootmem_low_pages(PAGE_SIZE);
674 	BUG_ON(!vectors);
675 
676 	for (addr = VMALLOC_END; addr; addr += PGDIR_SIZE)
677 		pmd_clear(pmd_off_k(addr));
678 
679 	/*
680 	 * Map the kernel if it is XIP.
681 	 * It is always first in the modulearea.
682 	 */
683 #ifdef CONFIG_XIP_KERNEL
684 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
685 	map.virtual = MODULE_START;
686 	map.length = ((unsigned long)&_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
687 	map.type = MT_ROM;
688 	create_mapping(&map);
689 #endif
690 
691 	/*
692 	 * Map the cache flushing regions.
693 	 */
694 #ifdef FLUSH_BASE
695 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
696 	map.virtual = FLUSH_BASE;
697 	map.length = SZ_1M;
698 	map.type = MT_CACHECLEAN;
699 	create_mapping(&map);
700 #endif
701 #ifdef FLUSH_BASE_MINICACHE
702 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
703 	map.virtual = FLUSH_BASE_MINICACHE;
704 	map.length = SZ_1M;
705 	map.type = MT_MINICLEAN;
706 	create_mapping(&map);
707 #endif
708 
709 	/*
710 	 * Create a mapping for the machine vectors at the high-vectors
711 	 * location (0xffff0000).  If we aren't using high-vectors, also
712 	 * create a mapping at the low-vectors virtual address.
713 	 */
714 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
715 	map.virtual = 0xffff0000;
716 	map.length = PAGE_SIZE;
717 	map.type = MT_HIGH_VECTORS;
718 	create_mapping(&map);
719 
720 	if (!vectors_high()) {
721 		map.virtual = 0;
722 		map.type = MT_LOW_VECTORS;
723 		create_mapping(&map);
724 	}
725 
726 	/*
727 	 * Ask the machine support to map in the statically mapped devices.
728 	 */
729 	if (mdesc->map_io)
730 		mdesc->map_io();
731 
732 	/*
733 	 * Finally flush the caches and tlb to ensure that we're in a
734 	 * consistent state wrt the writebuffer.  This also ensures that
735 	 * any write-allocated cache lines in the vector page are written
736 	 * back.  After this point, we can start to touch devices again.
737 	 */
738 	local_flush_tlb_all();
739 	flush_cache_all();
740 }
741 
742 /*
743  * paging_init() sets up the page tables, initialises the zone memory
744  * maps, and sets up the zero page, bad page and bad page tables.
745  */
746 void __init paging_init(struct meminfo *mi, struct machine_desc *mdesc)
747 {
748 	void *zero_page;
749 
750 	build_mem_type_table();
751 	prepare_page_table(mi);
752 	bootmem_init(mi);
753 	devicemaps_init(mdesc);
754 
755 	top_pmd = pmd_off_k(0xffff0000);
756 
757 	/*
758 	 * allocate the zero page.  Note that we count on this going ok.
759 	 */
760 	zero_page = alloc_bootmem_low_pages(PAGE_SIZE);
761 	memzero(zero_page, PAGE_SIZE);
762 	empty_zero_page = virt_to_page(zero_page);
763 	flush_dcache_page(empty_zero_page);
764 }
765 
766 /*
767  * In order to soft-boot, we need to insert a 1:1 mapping in place of
768  * the user-mode pages.  This will then ensure that we have predictable
769  * results when turning the mmu off
770  */
771 void setup_mm_for_reboot(char mode)
772 {
773 	unsigned long base_pmdval;
774 	pgd_t *pgd;
775 	int i;
776 
777 	if (current->mm && current->mm->pgd)
778 		pgd = current->mm->pgd;
779 	else
780 		pgd = init_mm.pgd;
781 
782 	base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
783 	if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ && !cpu_is_xscale())
784 		base_pmdval |= PMD_BIT4;
785 
786 	for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
787 		unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
788 		pmd_t *pmd;
789 
790 		pmd = pmd_off(pgd, i << PGDIR_SHIFT);
791 		pmd[0] = __pmd(pmdval);
792 		pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
793 		flush_pmd_entry(pmd);
794 	}
795 }
796