xref: /linux/arch/arm/mm/mmu.c (revision c4ee0af3fa0dc65f690fc908f02b8355f9576ea0)
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/mman.h>
15 #include <linux/nodemask.h>
16 #include <linux/memblock.h>
17 #include <linux/fs.h>
18 #include <linux/vmalloc.h>
19 #include <linux/sizes.h>
20 
21 #include <asm/cp15.h>
22 #include <asm/cputype.h>
23 #include <asm/sections.h>
24 #include <asm/cachetype.h>
25 #include <asm/setup.h>
26 #include <asm/smp_plat.h>
27 #include <asm/tlb.h>
28 #include <asm/highmem.h>
29 #include <asm/system_info.h>
30 #include <asm/traps.h>
31 #include <asm/procinfo.h>
32 #include <asm/memory.h>
33 
34 #include <asm/mach/arch.h>
35 #include <asm/mach/map.h>
36 #include <asm/mach/pci.h>
37 
38 #include "mm.h"
39 #include "tcm.h"
40 
41 /*
42  * empty_zero_page is a special page that is used for
43  * zero-initialized data and COW.
44  */
45 struct page *empty_zero_page;
46 EXPORT_SYMBOL(empty_zero_page);
47 
48 /*
49  * The pmd table for the upper-most set of pages.
50  */
51 pmd_t *top_pmd;
52 
53 #define CPOLICY_UNCACHED	0
54 #define CPOLICY_BUFFERED	1
55 #define CPOLICY_WRITETHROUGH	2
56 #define CPOLICY_WRITEBACK	3
57 #define CPOLICY_WRITEALLOC	4
58 
59 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
60 static unsigned int ecc_mask __initdata = 0;
61 pgprot_t pgprot_user;
62 pgprot_t pgprot_kernel;
63 pgprot_t pgprot_hyp_device;
64 pgprot_t pgprot_s2;
65 pgprot_t pgprot_s2_device;
66 
67 EXPORT_SYMBOL(pgprot_user);
68 EXPORT_SYMBOL(pgprot_kernel);
69 
70 struct cachepolicy {
71 	const char	policy[16];
72 	unsigned int	cr_mask;
73 	pmdval_t	pmd;
74 	pteval_t	pte;
75 	pteval_t	pte_s2;
76 };
77 
78 #ifdef CONFIG_ARM_LPAE
79 #define s2_policy(policy)	policy
80 #else
81 #define s2_policy(policy)	0
82 #endif
83 
84 static struct cachepolicy cache_policies[] __initdata = {
85 	{
86 		.policy		= "uncached",
87 		.cr_mask	= CR_W|CR_C,
88 		.pmd		= PMD_SECT_UNCACHED,
89 		.pte		= L_PTE_MT_UNCACHED,
90 		.pte_s2		= s2_policy(L_PTE_S2_MT_UNCACHED),
91 	}, {
92 		.policy		= "buffered",
93 		.cr_mask	= CR_C,
94 		.pmd		= PMD_SECT_BUFFERED,
95 		.pte		= L_PTE_MT_BUFFERABLE,
96 		.pte_s2		= s2_policy(L_PTE_S2_MT_UNCACHED),
97 	}, {
98 		.policy		= "writethrough",
99 		.cr_mask	= 0,
100 		.pmd		= PMD_SECT_WT,
101 		.pte		= L_PTE_MT_WRITETHROUGH,
102 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITETHROUGH),
103 	}, {
104 		.policy		= "writeback",
105 		.cr_mask	= 0,
106 		.pmd		= PMD_SECT_WB,
107 		.pte		= L_PTE_MT_WRITEBACK,
108 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITEBACK),
109 	}, {
110 		.policy		= "writealloc",
111 		.cr_mask	= 0,
112 		.pmd		= PMD_SECT_WBWA,
113 		.pte		= L_PTE_MT_WRITEALLOC,
114 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITEBACK),
115 	}
116 };
117 
118 #ifdef CONFIG_CPU_CP15
119 /*
120  * These are useful for identifying cache coherency
121  * problems by allowing the cache or the cache and
122  * writebuffer to be turned off.  (Note: the write
123  * buffer should not be on and the cache off).
124  */
125 static int __init early_cachepolicy(char *p)
126 {
127 	int i;
128 
129 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
130 		int len = strlen(cache_policies[i].policy);
131 
132 		if (memcmp(p, cache_policies[i].policy, len) == 0) {
133 			cachepolicy = i;
134 			cr_alignment &= ~cache_policies[i].cr_mask;
135 			cr_no_alignment &= ~cache_policies[i].cr_mask;
136 			break;
137 		}
138 	}
139 	if (i == ARRAY_SIZE(cache_policies))
140 		printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
141 	/*
142 	 * This restriction is partly to do with the way we boot; it is
143 	 * unpredictable to have memory mapped using two different sets of
144 	 * memory attributes (shared, type, and cache attribs).  We can not
145 	 * change these attributes once the initial assembly has setup the
146 	 * page tables.
147 	 */
148 	if (cpu_architecture() >= CPU_ARCH_ARMv6) {
149 		printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
150 		cachepolicy = CPOLICY_WRITEBACK;
151 	}
152 	flush_cache_all();
153 	set_cr(cr_alignment);
154 	return 0;
155 }
156 early_param("cachepolicy", early_cachepolicy);
157 
158 static int __init early_nocache(char *__unused)
159 {
160 	char *p = "buffered";
161 	printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
162 	early_cachepolicy(p);
163 	return 0;
164 }
165 early_param("nocache", early_nocache);
166 
167 static int __init early_nowrite(char *__unused)
168 {
169 	char *p = "uncached";
170 	printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
171 	early_cachepolicy(p);
172 	return 0;
173 }
174 early_param("nowb", early_nowrite);
175 
176 #ifndef CONFIG_ARM_LPAE
177 static int __init early_ecc(char *p)
178 {
179 	if (memcmp(p, "on", 2) == 0)
180 		ecc_mask = PMD_PROTECTION;
181 	else if (memcmp(p, "off", 3) == 0)
182 		ecc_mask = 0;
183 	return 0;
184 }
185 early_param("ecc", early_ecc);
186 #endif
187 
188 static int __init noalign_setup(char *__unused)
189 {
190 	cr_alignment &= ~CR_A;
191 	cr_no_alignment &= ~CR_A;
192 	set_cr(cr_alignment);
193 	return 1;
194 }
195 __setup("noalign", noalign_setup);
196 
197 #ifndef CONFIG_SMP
198 void adjust_cr(unsigned long mask, unsigned long set)
199 {
200 	unsigned long flags;
201 
202 	mask &= ~CR_A;
203 
204 	set &= mask;
205 
206 	local_irq_save(flags);
207 
208 	cr_no_alignment = (cr_no_alignment & ~mask) | set;
209 	cr_alignment = (cr_alignment & ~mask) | set;
210 
211 	set_cr((get_cr() & ~mask) | set);
212 
213 	local_irq_restore(flags);
214 }
215 #endif
216 
217 #else /* ifdef CONFIG_CPU_CP15 */
218 
219 static int __init early_cachepolicy(char *p)
220 {
221 	pr_warning("cachepolicy kernel parameter not supported without cp15\n");
222 }
223 early_param("cachepolicy", early_cachepolicy);
224 
225 static int __init noalign_setup(char *__unused)
226 {
227 	pr_warning("noalign kernel parameter not supported without cp15\n");
228 }
229 __setup("noalign", noalign_setup);
230 
231 #endif /* ifdef CONFIG_CPU_CP15 / else */
232 
233 #define PROT_PTE_DEVICE		L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
234 #define PROT_SECT_DEVICE	PMD_TYPE_SECT|PMD_SECT_AP_WRITE
235 
236 static struct mem_type mem_types[] = {
237 	[MT_DEVICE] = {		  /* Strongly ordered / ARMv6 shared device */
238 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
239 				  L_PTE_SHARED,
240 		.prot_l1	= PMD_TYPE_TABLE,
241 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_S,
242 		.domain		= DOMAIN_IO,
243 	},
244 	[MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
245 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
246 		.prot_l1	= PMD_TYPE_TABLE,
247 		.prot_sect	= PROT_SECT_DEVICE,
248 		.domain		= DOMAIN_IO,
249 	},
250 	[MT_DEVICE_CACHED] = {	  /* ioremap_cached */
251 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
252 		.prot_l1	= PMD_TYPE_TABLE,
253 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_WB,
254 		.domain		= DOMAIN_IO,
255 	},
256 	[MT_DEVICE_WC] = {	/* ioremap_wc */
257 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
258 		.prot_l1	= PMD_TYPE_TABLE,
259 		.prot_sect	= PROT_SECT_DEVICE,
260 		.domain		= DOMAIN_IO,
261 	},
262 	[MT_UNCACHED] = {
263 		.prot_pte	= PROT_PTE_DEVICE,
264 		.prot_l1	= PMD_TYPE_TABLE,
265 		.prot_sect	= PMD_TYPE_SECT | PMD_SECT_XN,
266 		.domain		= DOMAIN_IO,
267 	},
268 	[MT_CACHECLEAN] = {
269 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
270 		.domain    = DOMAIN_KERNEL,
271 	},
272 #ifndef CONFIG_ARM_LPAE
273 	[MT_MINICLEAN] = {
274 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
275 		.domain    = DOMAIN_KERNEL,
276 	},
277 #endif
278 	[MT_LOW_VECTORS] = {
279 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
280 				L_PTE_RDONLY,
281 		.prot_l1   = PMD_TYPE_TABLE,
282 		.domain    = DOMAIN_USER,
283 	},
284 	[MT_HIGH_VECTORS] = {
285 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
286 				L_PTE_USER | L_PTE_RDONLY,
287 		.prot_l1   = PMD_TYPE_TABLE,
288 		.domain    = DOMAIN_USER,
289 	},
290 	[MT_MEMORY] = {
291 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
292 		.prot_l1   = PMD_TYPE_TABLE,
293 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
294 		.domain    = DOMAIN_KERNEL,
295 	},
296 	[MT_ROM] = {
297 		.prot_sect = PMD_TYPE_SECT,
298 		.domain    = DOMAIN_KERNEL,
299 	},
300 	[MT_MEMORY_NONCACHED] = {
301 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
302 				L_PTE_MT_BUFFERABLE,
303 		.prot_l1   = PMD_TYPE_TABLE,
304 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
305 		.domain    = DOMAIN_KERNEL,
306 	},
307 	[MT_MEMORY_DTCM] = {
308 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
309 				L_PTE_XN,
310 		.prot_l1   = PMD_TYPE_TABLE,
311 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
312 		.domain    = DOMAIN_KERNEL,
313 	},
314 	[MT_MEMORY_ITCM] = {
315 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
316 		.prot_l1   = PMD_TYPE_TABLE,
317 		.domain    = DOMAIN_KERNEL,
318 	},
319 	[MT_MEMORY_SO] = {
320 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
321 				L_PTE_MT_UNCACHED | L_PTE_XN,
322 		.prot_l1   = PMD_TYPE_TABLE,
323 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
324 				PMD_SECT_UNCACHED | PMD_SECT_XN,
325 		.domain    = DOMAIN_KERNEL,
326 	},
327 	[MT_MEMORY_DMA_READY] = {
328 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
329 		.prot_l1   = PMD_TYPE_TABLE,
330 		.domain    = DOMAIN_KERNEL,
331 	},
332 };
333 
334 const struct mem_type *get_mem_type(unsigned int type)
335 {
336 	return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
337 }
338 EXPORT_SYMBOL(get_mem_type);
339 
340 /*
341  * Adjust the PMD section entries according to the CPU in use.
342  */
343 static void __init build_mem_type_table(void)
344 {
345 	struct cachepolicy *cp;
346 	unsigned int cr = get_cr();
347 	pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
348 	pteval_t hyp_device_pgprot, s2_pgprot, s2_device_pgprot;
349 	int cpu_arch = cpu_architecture();
350 	int i;
351 
352 	if (cpu_arch < CPU_ARCH_ARMv6) {
353 #if defined(CONFIG_CPU_DCACHE_DISABLE)
354 		if (cachepolicy > CPOLICY_BUFFERED)
355 			cachepolicy = CPOLICY_BUFFERED;
356 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
357 		if (cachepolicy > CPOLICY_WRITETHROUGH)
358 			cachepolicy = CPOLICY_WRITETHROUGH;
359 #endif
360 	}
361 	if (cpu_arch < CPU_ARCH_ARMv5) {
362 		if (cachepolicy >= CPOLICY_WRITEALLOC)
363 			cachepolicy = CPOLICY_WRITEBACK;
364 		ecc_mask = 0;
365 	}
366 	if (is_smp())
367 		cachepolicy = CPOLICY_WRITEALLOC;
368 
369 	/*
370 	 * Strip out features not present on earlier architectures.
371 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
372 	 * without extended page tables don't have the 'Shared' bit.
373 	 */
374 	if (cpu_arch < CPU_ARCH_ARMv5)
375 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
376 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
377 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
378 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
379 			mem_types[i].prot_sect &= ~PMD_SECT_S;
380 
381 	/*
382 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
383 	 * "update-able on write" bit on ARM610).  However, Xscale and
384 	 * Xscale3 require this bit to be cleared.
385 	 */
386 	if (cpu_is_xscale() || cpu_is_xsc3()) {
387 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
388 			mem_types[i].prot_sect &= ~PMD_BIT4;
389 			mem_types[i].prot_l1 &= ~PMD_BIT4;
390 		}
391 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
392 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
393 			if (mem_types[i].prot_l1)
394 				mem_types[i].prot_l1 |= PMD_BIT4;
395 			if (mem_types[i].prot_sect)
396 				mem_types[i].prot_sect |= PMD_BIT4;
397 		}
398 	}
399 
400 	/*
401 	 * Mark the device areas according to the CPU/architecture.
402 	 */
403 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
404 		if (!cpu_is_xsc3()) {
405 			/*
406 			 * Mark device regions on ARMv6+ as execute-never
407 			 * to prevent speculative instruction fetches.
408 			 */
409 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
410 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
411 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
412 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
413 		}
414 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
415 			/*
416 			 * For ARMv7 with TEX remapping,
417 			 * - shared device is SXCB=1100
418 			 * - nonshared device is SXCB=0100
419 			 * - write combine device mem is SXCB=0001
420 			 * (Uncached Normal memory)
421 			 */
422 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
423 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
424 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
425 		} else if (cpu_is_xsc3()) {
426 			/*
427 			 * For Xscale3,
428 			 * - shared device is TEXCB=00101
429 			 * - nonshared device is TEXCB=01000
430 			 * - write combine device mem is TEXCB=00100
431 			 * (Inner/Outer Uncacheable in xsc3 parlance)
432 			 */
433 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
434 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
435 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
436 		} else {
437 			/*
438 			 * For ARMv6 and ARMv7 without TEX remapping,
439 			 * - shared device is TEXCB=00001
440 			 * - nonshared device is TEXCB=01000
441 			 * - write combine device mem is TEXCB=00100
442 			 * (Uncached Normal in ARMv6 parlance).
443 			 */
444 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
445 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
446 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
447 		}
448 	} else {
449 		/*
450 		 * On others, write combining is "Uncached/Buffered"
451 		 */
452 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
453 	}
454 
455 	/*
456 	 * Now deal with the memory-type mappings
457 	 */
458 	cp = &cache_policies[cachepolicy];
459 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
460 	s2_pgprot = cp->pte_s2;
461 	hyp_device_pgprot = s2_device_pgprot = mem_types[MT_DEVICE].prot_pte;
462 
463 	/*
464 	 * ARMv6 and above have extended page tables.
465 	 */
466 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
467 #ifndef CONFIG_ARM_LPAE
468 		/*
469 		 * Mark cache clean areas and XIP ROM read only
470 		 * from SVC mode and no access from userspace.
471 		 */
472 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
473 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
474 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
475 #endif
476 
477 		if (is_smp()) {
478 			/*
479 			 * Mark memory with the "shared" attribute
480 			 * for SMP systems
481 			 */
482 			user_pgprot |= L_PTE_SHARED;
483 			kern_pgprot |= L_PTE_SHARED;
484 			vecs_pgprot |= L_PTE_SHARED;
485 			s2_pgprot |= L_PTE_SHARED;
486 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
487 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
488 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
489 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
490 			mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
491 			mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
492 			mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
493 			mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
494 			mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
495 		}
496 	}
497 
498 	/*
499 	 * Non-cacheable Normal - intended for memory areas that must
500 	 * not cause dirty cache line writebacks when used
501 	 */
502 	if (cpu_arch >= CPU_ARCH_ARMv6) {
503 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
504 			/* Non-cacheable Normal is XCB = 001 */
505 			mem_types[MT_MEMORY_NONCACHED].prot_sect |=
506 				PMD_SECT_BUFFERED;
507 		} else {
508 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
509 			mem_types[MT_MEMORY_NONCACHED].prot_sect |=
510 				PMD_SECT_TEX(1);
511 		}
512 	} else {
513 		mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
514 	}
515 
516 #ifdef CONFIG_ARM_LPAE
517 	/*
518 	 * Do not generate access flag faults for the kernel mappings.
519 	 */
520 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
521 		mem_types[i].prot_pte |= PTE_EXT_AF;
522 		if (mem_types[i].prot_sect)
523 			mem_types[i].prot_sect |= PMD_SECT_AF;
524 	}
525 	kern_pgprot |= PTE_EXT_AF;
526 	vecs_pgprot |= PTE_EXT_AF;
527 #endif
528 
529 	for (i = 0; i < 16; i++) {
530 		pteval_t v = pgprot_val(protection_map[i]);
531 		protection_map[i] = __pgprot(v | user_pgprot);
532 	}
533 
534 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
535 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
536 
537 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
538 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
539 				 L_PTE_DIRTY | kern_pgprot);
540 	pgprot_s2  = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | s2_pgprot);
541 	pgprot_s2_device  = __pgprot(s2_device_pgprot);
542 	pgprot_hyp_device  = __pgprot(hyp_device_pgprot);
543 
544 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
545 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
546 	mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
547 	mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
548 	mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
549 	mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
550 	mem_types[MT_ROM].prot_sect |= cp->pmd;
551 
552 	switch (cp->pmd) {
553 	case PMD_SECT_WT:
554 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
555 		break;
556 	case PMD_SECT_WB:
557 	case PMD_SECT_WBWA:
558 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
559 		break;
560 	}
561 	pr_info("Memory policy: %sData cache %s\n",
562 		ecc_mask ? "ECC enabled, " : "", cp->policy);
563 
564 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
565 		struct mem_type *t = &mem_types[i];
566 		if (t->prot_l1)
567 			t->prot_l1 |= PMD_DOMAIN(t->domain);
568 		if (t->prot_sect)
569 			t->prot_sect |= PMD_DOMAIN(t->domain);
570 	}
571 }
572 
573 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
574 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
575 			      unsigned long size, pgprot_t vma_prot)
576 {
577 	if (!pfn_valid(pfn))
578 		return pgprot_noncached(vma_prot);
579 	else if (file->f_flags & O_SYNC)
580 		return pgprot_writecombine(vma_prot);
581 	return vma_prot;
582 }
583 EXPORT_SYMBOL(phys_mem_access_prot);
584 #endif
585 
586 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
587 
588 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
589 {
590 	void *ptr = __va(memblock_alloc(sz, align));
591 	memset(ptr, 0, sz);
592 	return ptr;
593 }
594 
595 static void __init *early_alloc(unsigned long sz)
596 {
597 	return early_alloc_aligned(sz, sz);
598 }
599 
600 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
601 {
602 	if (pmd_none(*pmd)) {
603 		pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
604 		__pmd_populate(pmd, __pa(pte), prot);
605 	}
606 	BUG_ON(pmd_bad(*pmd));
607 	return pte_offset_kernel(pmd, addr);
608 }
609 
610 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
611 				  unsigned long end, unsigned long pfn,
612 				  const struct mem_type *type)
613 {
614 	pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
615 	do {
616 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
617 		pfn++;
618 	} while (pte++, addr += PAGE_SIZE, addr != end);
619 }
620 
621 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
622 			unsigned long end, phys_addr_t phys,
623 			const struct mem_type *type)
624 {
625 	pmd_t *p = pmd;
626 
627 #ifndef CONFIG_ARM_LPAE
628 	/*
629 	 * In classic MMU format, puds and pmds are folded in to
630 	 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
631 	 * group of L1 entries making up one logical pointer to
632 	 * an L2 table (2MB), where as PMDs refer to the individual
633 	 * L1 entries (1MB). Hence increment to get the correct
634 	 * offset for odd 1MB sections.
635 	 * (See arch/arm/include/asm/pgtable-2level.h)
636 	 */
637 	if (addr & SECTION_SIZE)
638 		pmd++;
639 #endif
640 	do {
641 		*pmd = __pmd(phys | type->prot_sect);
642 		phys += SECTION_SIZE;
643 	} while (pmd++, addr += SECTION_SIZE, addr != end);
644 
645 	flush_pmd_entry(p);
646 }
647 
648 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
649 				      unsigned long end, phys_addr_t phys,
650 				      const struct mem_type *type)
651 {
652 	pmd_t *pmd = pmd_offset(pud, addr);
653 	unsigned long next;
654 
655 	do {
656 		/*
657 		 * With LPAE, we must loop over to map
658 		 * all the pmds for the given range.
659 		 */
660 		next = pmd_addr_end(addr, end);
661 
662 		/*
663 		 * Try a section mapping - addr, next and phys must all be
664 		 * aligned to a section boundary.
665 		 */
666 		if (type->prot_sect &&
667 				((addr | next | phys) & ~SECTION_MASK) == 0) {
668 			__map_init_section(pmd, addr, next, phys, type);
669 		} else {
670 			alloc_init_pte(pmd, addr, next,
671 						__phys_to_pfn(phys), type);
672 		}
673 
674 		phys += next - addr;
675 
676 	} while (pmd++, addr = next, addr != end);
677 }
678 
679 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
680 				  unsigned long end, phys_addr_t phys,
681 				  const struct mem_type *type)
682 {
683 	pud_t *pud = pud_offset(pgd, addr);
684 	unsigned long next;
685 
686 	do {
687 		next = pud_addr_end(addr, end);
688 		alloc_init_pmd(pud, addr, next, phys, type);
689 		phys += next - addr;
690 	} while (pud++, addr = next, addr != end);
691 }
692 
693 #ifndef CONFIG_ARM_LPAE
694 static void __init create_36bit_mapping(struct map_desc *md,
695 					const struct mem_type *type)
696 {
697 	unsigned long addr, length, end;
698 	phys_addr_t phys;
699 	pgd_t *pgd;
700 
701 	addr = md->virtual;
702 	phys = __pfn_to_phys(md->pfn);
703 	length = PAGE_ALIGN(md->length);
704 
705 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
706 		printk(KERN_ERR "MM: CPU does not support supersection "
707 		       "mapping for 0x%08llx at 0x%08lx\n",
708 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
709 		return;
710 	}
711 
712 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
713 	 *	Since domain assignments can in fact be arbitrary, the
714 	 *	'domain == 0' check below is required to insure that ARMv6
715 	 *	supersections are only allocated for domain 0 regardless
716 	 *	of the actual domain assignments in use.
717 	 */
718 	if (type->domain) {
719 		printk(KERN_ERR "MM: invalid domain in supersection "
720 		       "mapping for 0x%08llx at 0x%08lx\n",
721 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
722 		return;
723 	}
724 
725 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
726 		printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
727 		       " at 0x%08lx invalid alignment\n",
728 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
729 		return;
730 	}
731 
732 	/*
733 	 * Shift bits [35:32] of address into bits [23:20] of PMD
734 	 * (See ARMv6 spec).
735 	 */
736 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
737 
738 	pgd = pgd_offset_k(addr);
739 	end = addr + length;
740 	do {
741 		pud_t *pud = pud_offset(pgd, addr);
742 		pmd_t *pmd = pmd_offset(pud, addr);
743 		int i;
744 
745 		for (i = 0; i < 16; i++)
746 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
747 
748 		addr += SUPERSECTION_SIZE;
749 		phys += SUPERSECTION_SIZE;
750 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
751 	} while (addr != end);
752 }
753 #endif	/* !CONFIG_ARM_LPAE */
754 
755 /*
756  * Create the page directory entries and any necessary
757  * page tables for the mapping specified by `md'.  We
758  * are able to cope here with varying sizes and address
759  * offsets, and we take full advantage of sections and
760  * supersections.
761  */
762 static void __init create_mapping(struct map_desc *md)
763 {
764 	unsigned long addr, length, end;
765 	phys_addr_t phys;
766 	const struct mem_type *type;
767 	pgd_t *pgd;
768 
769 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
770 		printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
771 		       " at 0x%08lx in user region\n",
772 		       (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
773 		return;
774 	}
775 
776 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
777 	    md->virtual >= PAGE_OFFSET &&
778 	    (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
779 		printk(KERN_WARNING "BUG: mapping for 0x%08llx"
780 		       " at 0x%08lx out of vmalloc space\n",
781 		       (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
782 	}
783 
784 	type = &mem_types[md->type];
785 
786 #ifndef CONFIG_ARM_LPAE
787 	/*
788 	 * Catch 36-bit addresses
789 	 */
790 	if (md->pfn >= 0x100000) {
791 		create_36bit_mapping(md, type);
792 		return;
793 	}
794 #endif
795 
796 	addr = md->virtual & PAGE_MASK;
797 	phys = __pfn_to_phys(md->pfn);
798 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
799 
800 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
801 		printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
802 		       "be mapped using pages, ignoring.\n",
803 		       (long long)__pfn_to_phys(md->pfn), addr);
804 		return;
805 	}
806 
807 	pgd = pgd_offset_k(addr);
808 	end = addr + length;
809 	do {
810 		unsigned long next = pgd_addr_end(addr, end);
811 
812 		alloc_init_pud(pgd, addr, next, phys, type);
813 
814 		phys += next - addr;
815 		addr = next;
816 	} while (pgd++, addr != end);
817 }
818 
819 /*
820  * Create the architecture specific mappings
821  */
822 void __init iotable_init(struct map_desc *io_desc, int nr)
823 {
824 	struct map_desc *md;
825 	struct vm_struct *vm;
826 	struct static_vm *svm;
827 
828 	if (!nr)
829 		return;
830 
831 	svm = early_alloc_aligned(sizeof(*svm) * nr, __alignof__(*svm));
832 
833 	for (md = io_desc; nr; md++, nr--) {
834 		create_mapping(md);
835 
836 		vm = &svm->vm;
837 		vm->addr = (void *)(md->virtual & PAGE_MASK);
838 		vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
839 		vm->phys_addr = __pfn_to_phys(md->pfn);
840 		vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
841 		vm->flags |= VM_ARM_MTYPE(md->type);
842 		vm->caller = iotable_init;
843 		add_static_vm_early(svm++);
844 	}
845 }
846 
847 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
848 				  void *caller)
849 {
850 	struct vm_struct *vm;
851 	struct static_vm *svm;
852 
853 	svm = early_alloc_aligned(sizeof(*svm), __alignof__(*svm));
854 
855 	vm = &svm->vm;
856 	vm->addr = (void *)addr;
857 	vm->size = size;
858 	vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
859 	vm->caller = caller;
860 	add_static_vm_early(svm);
861 }
862 
863 #ifndef CONFIG_ARM_LPAE
864 
865 /*
866  * The Linux PMD is made of two consecutive section entries covering 2MB
867  * (see definition in include/asm/pgtable-2level.h).  However a call to
868  * create_mapping() may optimize static mappings by using individual
869  * 1MB section mappings.  This leaves the actual PMD potentially half
870  * initialized if the top or bottom section entry isn't used, leaving it
871  * open to problems if a subsequent ioremap() or vmalloc() tries to use
872  * the virtual space left free by that unused section entry.
873  *
874  * Let's avoid the issue by inserting dummy vm entries covering the unused
875  * PMD halves once the static mappings are in place.
876  */
877 
878 static void __init pmd_empty_section_gap(unsigned long addr)
879 {
880 	vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
881 }
882 
883 static void __init fill_pmd_gaps(void)
884 {
885 	struct static_vm *svm;
886 	struct vm_struct *vm;
887 	unsigned long addr, next = 0;
888 	pmd_t *pmd;
889 
890 	list_for_each_entry(svm, &static_vmlist, list) {
891 		vm = &svm->vm;
892 		addr = (unsigned long)vm->addr;
893 		if (addr < next)
894 			continue;
895 
896 		/*
897 		 * Check if this vm starts on an odd section boundary.
898 		 * If so and the first section entry for this PMD is free
899 		 * then we block the corresponding virtual address.
900 		 */
901 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
902 			pmd = pmd_off_k(addr);
903 			if (pmd_none(*pmd))
904 				pmd_empty_section_gap(addr & PMD_MASK);
905 		}
906 
907 		/*
908 		 * Then check if this vm ends on an odd section boundary.
909 		 * If so and the second section entry for this PMD is empty
910 		 * then we block the corresponding virtual address.
911 		 */
912 		addr += vm->size;
913 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
914 			pmd = pmd_off_k(addr) + 1;
915 			if (pmd_none(*pmd))
916 				pmd_empty_section_gap(addr);
917 		}
918 
919 		/* no need to look at any vm entry until we hit the next PMD */
920 		next = (addr + PMD_SIZE - 1) & PMD_MASK;
921 	}
922 }
923 
924 #else
925 #define fill_pmd_gaps() do { } while (0)
926 #endif
927 
928 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
929 static void __init pci_reserve_io(void)
930 {
931 	struct static_vm *svm;
932 
933 	svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
934 	if (svm)
935 		return;
936 
937 	vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
938 }
939 #else
940 #define pci_reserve_io() do { } while (0)
941 #endif
942 
943 #ifdef CONFIG_DEBUG_LL
944 void __init debug_ll_io_init(void)
945 {
946 	struct map_desc map;
947 
948 	debug_ll_addr(&map.pfn, &map.virtual);
949 	if (!map.pfn || !map.virtual)
950 		return;
951 	map.pfn = __phys_to_pfn(map.pfn);
952 	map.virtual &= PAGE_MASK;
953 	map.length = PAGE_SIZE;
954 	map.type = MT_DEVICE;
955 	iotable_init(&map, 1);
956 }
957 #endif
958 
959 static void * __initdata vmalloc_min =
960 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
961 
962 /*
963  * vmalloc=size forces the vmalloc area to be exactly 'size'
964  * bytes. This can be used to increase (or decrease) the vmalloc
965  * area - the default is 240m.
966  */
967 static int __init early_vmalloc(char *arg)
968 {
969 	unsigned long vmalloc_reserve = memparse(arg, NULL);
970 
971 	if (vmalloc_reserve < SZ_16M) {
972 		vmalloc_reserve = SZ_16M;
973 		printk(KERN_WARNING
974 			"vmalloc area too small, limiting to %luMB\n",
975 			vmalloc_reserve >> 20);
976 	}
977 
978 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
979 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
980 		printk(KERN_WARNING
981 			"vmalloc area is too big, limiting to %luMB\n",
982 			vmalloc_reserve >> 20);
983 	}
984 
985 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
986 	return 0;
987 }
988 early_param("vmalloc", early_vmalloc);
989 
990 phys_addr_t arm_lowmem_limit __initdata = 0;
991 
992 void __init sanity_check_meminfo(void)
993 {
994 	phys_addr_t memblock_limit = 0;
995 	int i, j, highmem = 0;
996 	phys_addr_t vmalloc_limit = __pa(vmalloc_min - 1) + 1;
997 
998 	for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
999 		struct membank *bank = &meminfo.bank[j];
1000 		phys_addr_t size_limit;
1001 
1002 		*bank = meminfo.bank[i];
1003 		size_limit = bank->size;
1004 
1005 		if (bank->start >= vmalloc_limit)
1006 			highmem = 1;
1007 		else
1008 			size_limit = vmalloc_limit - bank->start;
1009 
1010 		bank->highmem = highmem;
1011 
1012 #ifdef CONFIG_HIGHMEM
1013 		/*
1014 		 * Split those memory banks which are partially overlapping
1015 		 * the vmalloc area greatly simplifying things later.
1016 		 */
1017 		if (!highmem && bank->size > size_limit) {
1018 			if (meminfo.nr_banks >= NR_BANKS) {
1019 				printk(KERN_CRIT "NR_BANKS too low, "
1020 						 "ignoring high memory\n");
1021 			} else {
1022 				memmove(bank + 1, bank,
1023 					(meminfo.nr_banks - i) * sizeof(*bank));
1024 				meminfo.nr_banks++;
1025 				i++;
1026 				bank[1].size -= size_limit;
1027 				bank[1].start = vmalloc_limit;
1028 				bank[1].highmem = highmem = 1;
1029 				j++;
1030 			}
1031 			bank->size = size_limit;
1032 		}
1033 #else
1034 		/*
1035 		 * Highmem banks not allowed with !CONFIG_HIGHMEM.
1036 		 */
1037 		if (highmem) {
1038 			printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
1039 			       "(!CONFIG_HIGHMEM).\n",
1040 			       (unsigned long long)bank->start,
1041 			       (unsigned long long)bank->start + bank->size - 1);
1042 			continue;
1043 		}
1044 
1045 		/*
1046 		 * Check whether this memory bank would partially overlap
1047 		 * the vmalloc area.
1048 		 */
1049 		if (bank->size > size_limit) {
1050 			printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
1051 			       "to -%.8llx (vmalloc region overlap).\n",
1052 			       (unsigned long long)bank->start,
1053 			       (unsigned long long)bank->start + bank->size - 1,
1054 			       (unsigned long long)bank->start + size_limit - 1);
1055 			bank->size = size_limit;
1056 		}
1057 #endif
1058 		if (!bank->highmem) {
1059 			phys_addr_t bank_end = bank->start + bank->size;
1060 
1061 			if (bank_end > arm_lowmem_limit)
1062 				arm_lowmem_limit = bank_end;
1063 
1064 			/*
1065 			 * Find the first non-section-aligned page, and point
1066 			 * memblock_limit at it. This relies on rounding the
1067 			 * limit down to be section-aligned, which happens at
1068 			 * the end of this function.
1069 			 *
1070 			 * With this algorithm, the start or end of almost any
1071 			 * bank can be non-section-aligned. The only exception
1072 			 * is that the start of the bank 0 must be section-
1073 			 * aligned, since otherwise memory would need to be
1074 			 * allocated when mapping the start of bank 0, which
1075 			 * occurs before any free memory is mapped.
1076 			 */
1077 			if (!memblock_limit) {
1078 				if (!IS_ALIGNED(bank->start, SECTION_SIZE))
1079 					memblock_limit = bank->start;
1080 				else if (!IS_ALIGNED(bank_end, SECTION_SIZE))
1081 					memblock_limit = bank_end;
1082 			}
1083 		}
1084 		j++;
1085 	}
1086 #ifdef CONFIG_HIGHMEM
1087 	if (highmem) {
1088 		const char *reason = NULL;
1089 
1090 		if (cache_is_vipt_aliasing()) {
1091 			/*
1092 			 * Interactions between kmap and other mappings
1093 			 * make highmem support with aliasing VIPT caches
1094 			 * rather difficult.
1095 			 */
1096 			reason = "with VIPT aliasing cache";
1097 		}
1098 		if (reason) {
1099 			printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
1100 				reason);
1101 			while (j > 0 && meminfo.bank[j - 1].highmem)
1102 				j--;
1103 		}
1104 	}
1105 #endif
1106 	meminfo.nr_banks = j;
1107 	high_memory = __va(arm_lowmem_limit - 1) + 1;
1108 
1109 	/*
1110 	 * Round the memblock limit down to a section size.  This
1111 	 * helps to ensure that we will allocate memory from the
1112 	 * last full section, which should be mapped.
1113 	 */
1114 	if (memblock_limit)
1115 		memblock_limit = round_down(memblock_limit, SECTION_SIZE);
1116 	if (!memblock_limit)
1117 		memblock_limit = arm_lowmem_limit;
1118 
1119 	memblock_set_current_limit(memblock_limit);
1120 }
1121 
1122 static inline void prepare_page_table(void)
1123 {
1124 	unsigned long addr;
1125 	phys_addr_t end;
1126 
1127 	/*
1128 	 * Clear out all the mappings below the kernel image.
1129 	 */
1130 	for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1131 		pmd_clear(pmd_off_k(addr));
1132 
1133 #ifdef CONFIG_XIP_KERNEL
1134 	/* The XIP kernel is mapped in the module area -- skip over it */
1135 	addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
1136 #endif
1137 	for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1138 		pmd_clear(pmd_off_k(addr));
1139 
1140 	/*
1141 	 * Find the end of the first block of lowmem.
1142 	 */
1143 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1144 	if (end >= arm_lowmem_limit)
1145 		end = arm_lowmem_limit;
1146 
1147 	/*
1148 	 * Clear out all the kernel space mappings, except for the first
1149 	 * memory bank, up to the vmalloc region.
1150 	 */
1151 	for (addr = __phys_to_virt(end);
1152 	     addr < VMALLOC_START; addr += PMD_SIZE)
1153 		pmd_clear(pmd_off_k(addr));
1154 }
1155 
1156 #ifdef CONFIG_ARM_LPAE
1157 /* the first page is reserved for pgd */
1158 #define SWAPPER_PG_DIR_SIZE	(PAGE_SIZE + \
1159 				 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1160 #else
1161 #define SWAPPER_PG_DIR_SIZE	(PTRS_PER_PGD * sizeof(pgd_t))
1162 #endif
1163 
1164 /*
1165  * Reserve the special regions of memory
1166  */
1167 void __init arm_mm_memblock_reserve(void)
1168 {
1169 	/*
1170 	 * Reserve the page tables.  These are already in use,
1171 	 * and can only be in node 0.
1172 	 */
1173 	memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1174 
1175 #ifdef CONFIG_SA1111
1176 	/*
1177 	 * Because of the SA1111 DMA bug, we want to preserve our
1178 	 * precious DMA-able memory...
1179 	 */
1180 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1181 #endif
1182 }
1183 
1184 /*
1185  * Set up the device mappings.  Since we clear out the page tables for all
1186  * mappings above VMALLOC_START, we will remove any debug device mappings.
1187  * This means you have to be careful how you debug this function, or any
1188  * called function.  This means you can't use any function or debugging
1189  * method which may touch any device, otherwise the kernel _will_ crash.
1190  */
1191 static void __init devicemaps_init(const struct machine_desc *mdesc)
1192 {
1193 	struct map_desc map;
1194 	unsigned long addr;
1195 	void *vectors;
1196 
1197 	/*
1198 	 * Allocate the vector page early.
1199 	 */
1200 	vectors = early_alloc(PAGE_SIZE * 2);
1201 
1202 	early_trap_init(vectors);
1203 
1204 	for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1205 		pmd_clear(pmd_off_k(addr));
1206 
1207 	/*
1208 	 * Map the kernel if it is XIP.
1209 	 * It is always first in the modulearea.
1210 	 */
1211 #ifdef CONFIG_XIP_KERNEL
1212 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1213 	map.virtual = MODULES_VADDR;
1214 	map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1215 	map.type = MT_ROM;
1216 	create_mapping(&map);
1217 #endif
1218 
1219 	/*
1220 	 * Map the cache flushing regions.
1221 	 */
1222 #ifdef FLUSH_BASE
1223 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1224 	map.virtual = FLUSH_BASE;
1225 	map.length = SZ_1M;
1226 	map.type = MT_CACHECLEAN;
1227 	create_mapping(&map);
1228 #endif
1229 #ifdef FLUSH_BASE_MINICACHE
1230 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1231 	map.virtual = FLUSH_BASE_MINICACHE;
1232 	map.length = SZ_1M;
1233 	map.type = MT_MINICLEAN;
1234 	create_mapping(&map);
1235 #endif
1236 
1237 	/*
1238 	 * Create a mapping for the machine vectors at the high-vectors
1239 	 * location (0xffff0000).  If we aren't using high-vectors, also
1240 	 * create a mapping at the low-vectors virtual address.
1241 	 */
1242 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1243 	map.virtual = 0xffff0000;
1244 	map.length = PAGE_SIZE;
1245 #ifdef CONFIG_KUSER_HELPERS
1246 	map.type = MT_HIGH_VECTORS;
1247 #else
1248 	map.type = MT_LOW_VECTORS;
1249 #endif
1250 	create_mapping(&map);
1251 
1252 	if (!vectors_high()) {
1253 		map.virtual = 0;
1254 		map.length = PAGE_SIZE * 2;
1255 		map.type = MT_LOW_VECTORS;
1256 		create_mapping(&map);
1257 	}
1258 
1259 	/* Now create a kernel read-only mapping */
1260 	map.pfn += 1;
1261 	map.virtual = 0xffff0000 + PAGE_SIZE;
1262 	map.length = PAGE_SIZE;
1263 	map.type = MT_LOW_VECTORS;
1264 	create_mapping(&map);
1265 
1266 	/*
1267 	 * Ask the machine support to map in the statically mapped devices.
1268 	 */
1269 	if (mdesc->map_io)
1270 		mdesc->map_io();
1271 	else
1272 		debug_ll_io_init();
1273 	fill_pmd_gaps();
1274 
1275 	/* Reserve fixed i/o space in VMALLOC region */
1276 	pci_reserve_io();
1277 
1278 	/*
1279 	 * Finally flush the caches and tlb to ensure that we're in a
1280 	 * consistent state wrt the writebuffer.  This also ensures that
1281 	 * any write-allocated cache lines in the vector page are written
1282 	 * back.  After this point, we can start to touch devices again.
1283 	 */
1284 	local_flush_tlb_all();
1285 	flush_cache_all();
1286 }
1287 
1288 static void __init kmap_init(void)
1289 {
1290 #ifdef CONFIG_HIGHMEM
1291 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1292 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1293 #endif
1294 }
1295 
1296 static void __init map_lowmem(void)
1297 {
1298 	struct memblock_region *reg;
1299 
1300 	/* Map all the lowmem memory banks. */
1301 	for_each_memblock(memory, reg) {
1302 		phys_addr_t start = reg->base;
1303 		phys_addr_t end = start + reg->size;
1304 		struct map_desc map;
1305 
1306 		if (end > arm_lowmem_limit)
1307 			end = arm_lowmem_limit;
1308 		if (start >= end)
1309 			break;
1310 
1311 		map.pfn = __phys_to_pfn(start);
1312 		map.virtual = __phys_to_virt(start);
1313 		map.length = end - start;
1314 		map.type = MT_MEMORY;
1315 
1316 		create_mapping(&map);
1317 	}
1318 }
1319 
1320 #ifdef CONFIG_ARM_LPAE
1321 /*
1322  * early_paging_init() recreates boot time page table setup, allowing machines
1323  * to switch over to a high (>4G) address space on LPAE systems
1324  */
1325 void __init early_paging_init(const struct machine_desc *mdesc,
1326 			      struct proc_info_list *procinfo)
1327 {
1328 	pmdval_t pmdprot = procinfo->__cpu_mm_mmu_flags;
1329 	unsigned long map_start, map_end;
1330 	pgd_t *pgd0, *pgdk;
1331 	pud_t *pud0, *pudk, *pud_start;
1332 	pmd_t *pmd0, *pmdk;
1333 	phys_addr_t phys;
1334 	int i;
1335 
1336 	if (!(mdesc->init_meminfo))
1337 		return;
1338 
1339 	/* remap kernel code and data */
1340 	map_start = init_mm.start_code;
1341 	map_end   = init_mm.brk;
1342 
1343 	/* get a handle on things... */
1344 	pgd0 = pgd_offset_k(0);
1345 	pud_start = pud0 = pud_offset(pgd0, 0);
1346 	pmd0 = pmd_offset(pud0, 0);
1347 
1348 	pgdk = pgd_offset_k(map_start);
1349 	pudk = pud_offset(pgdk, map_start);
1350 	pmdk = pmd_offset(pudk, map_start);
1351 
1352 	mdesc->init_meminfo();
1353 
1354 	/* Run the patch stub to update the constants */
1355 	fixup_pv_table(&__pv_table_begin,
1356 		(&__pv_table_end - &__pv_table_begin) << 2);
1357 
1358 	/*
1359 	 * Cache cleaning operations for self-modifying code
1360 	 * We should clean the entries by MVA but running a
1361 	 * for loop over every pv_table entry pointer would
1362 	 * just complicate the code.
1363 	 */
1364 	flush_cache_louis();
1365 	dsb();
1366 	isb();
1367 
1368 	/* remap level 1 table */
1369 	for (i = 0; i < PTRS_PER_PGD; pud0++, i++) {
1370 		set_pud(pud0,
1371 			__pud(__pa(pmd0) | PMD_TYPE_TABLE | L_PGD_SWAPPER));
1372 		pmd0 += PTRS_PER_PMD;
1373 	}
1374 
1375 	/* remap pmds for kernel mapping */
1376 	phys = __pa(map_start) & PMD_MASK;
1377 	do {
1378 		*pmdk++ = __pmd(phys | pmdprot);
1379 		phys += PMD_SIZE;
1380 	} while (phys < map_end);
1381 
1382 	flush_cache_all();
1383 	cpu_switch_mm(pgd0, &init_mm);
1384 	cpu_set_ttbr(1, __pa(pgd0) + TTBR1_OFFSET);
1385 	local_flush_bp_all();
1386 	local_flush_tlb_all();
1387 }
1388 
1389 #else
1390 
1391 void __init early_paging_init(const struct machine_desc *mdesc,
1392 			      struct proc_info_list *procinfo)
1393 {
1394 	if (mdesc->init_meminfo)
1395 		mdesc->init_meminfo();
1396 }
1397 
1398 #endif
1399 
1400 /*
1401  * paging_init() sets up the page tables, initialises the zone memory
1402  * maps, and sets up the zero page, bad page and bad page tables.
1403  */
1404 void __init paging_init(const struct machine_desc *mdesc)
1405 {
1406 	void *zero_page;
1407 
1408 	build_mem_type_table();
1409 	prepare_page_table();
1410 	map_lowmem();
1411 	dma_contiguous_remap();
1412 	devicemaps_init(mdesc);
1413 	kmap_init();
1414 	tcm_init();
1415 
1416 	top_pmd = pmd_off_k(0xffff0000);
1417 
1418 	/* allocate the zero page. */
1419 	zero_page = early_alloc(PAGE_SIZE);
1420 
1421 	bootmem_init();
1422 
1423 	empty_zero_page = virt_to_page(zero_page);
1424 	__flush_dcache_page(NULL, empty_zero_page);
1425 }
1426