xref: /openbmc/linux/arch/arm/mm/mmu.c (revision 54525552)
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 
19 #include <asm/cputype.h>
20 #include <asm/sections.h>
21 #include <asm/cachetype.h>
22 #include <asm/setup.h>
23 #include <asm/sizes.h>
24 #include <asm/smp_plat.h>
25 #include <asm/tlb.h>
26 #include <asm/highmem.h>
27 #include <asm/traps.h>
28 
29 #include <asm/mach/arch.h>
30 #include <asm/mach/map.h>
31 
32 #include "mm.h"
33 
34 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
35 
36 /*
37  * empty_zero_page is a special page that is used for
38  * zero-initialized data and COW.
39  */
40 struct page *empty_zero_page;
41 EXPORT_SYMBOL(empty_zero_page);
42 
43 /*
44  * The pmd table for the upper-most set of pages.
45  */
46 pmd_t *top_pmd;
47 
48 #define CPOLICY_UNCACHED	0
49 #define CPOLICY_BUFFERED	1
50 #define CPOLICY_WRITETHROUGH	2
51 #define CPOLICY_WRITEBACK	3
52 #define CPOLICY_WRITEALLOC	4
53 
54 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
55 static unsigned int ecc_mask __initdata = 0;
56 pgprot_t pgprot_user;
57 pgprot_t pgprot_kernel;
58 
59 EXPORT_SYMBOL(pgprot_user);
60 EXPORT_SYMBOL(pgprot_kernel);
61 
62 struct cachepolicy {
63 	const char	policy[16];
64 	unsigned int	cr_mask;
65 	unsigned int	pmd;
66 	pteval_t	pte;
67 };
68 
69 static struct cachepolicy cache_policies[] __initdata = {
70 	{
71 		.policy		= "uncached",
72 		.cr_mask	= CR_W|CR_C,
73 		.pmd		= PMD_SECT_UNCACHED,
74 		.pte		= L_PTE_MT_UNCACHED,
75 	}, {
76 		.policy		= "buffered",
77 		.cr_mask	= CR_C,
78 		.pmd		= PMD_SECT_BUFFERED,
79 		.pte		= L_PTE_MT_BUFFERABLE,
80 	}, {
81 		.policy		= "writethrough",
82 		.cr_mask	= 0,
83 		.pmd		= PMD_SECT_WT,
84 		.pte		= L_PTE_MT_WRITETHROUGH,
85 	}, {
86 		.policy		= "writeback",
87 		.cr_mask	= 0,
88 		.pmd		= PMD_SECT_WB,
89 		.pte		= L_PTE_MT_WRITEBACK,
90 	}, {
91 		.policy		= "writealloc",
92 		.cr_mask	= 0,
93 		.pmd		= PMD_SECT_WBWA,
94 		.pte		= L_PTE_MT_WRITEALLOC,
95 	}
96 };
97 
98 /*
99  * These are useful for identifying cache coherency
100  * problems by allowing the cache or the cache and
101  * writebuffer to be turned off.  (Note: the write
102  * buffer should not be on and the cache off).
103  */
104 static int __init early_cachepolicy(char *p)
105 {
106 	int i;
107 
108 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
109 		int len = strlen(cache_policies[i].policy);
110 
111 		if (memcmp(p, cache_policies[i].policy, len) == 0) {
112 			cachepolicy = i;
113 			cr_alignment &= ~cache_policies[i].cr_mask;
114 			cr_no_alignment &= ~cache_policies[i].cr_mask;
115 			break;
116 		}
117 	}
118 	if (i == ARRAY_SIZE(cache_policies))
119 		printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
120 	/*
121 	 * This restriction is partly to do with the way we boot; it is
122 	 * unpredictable to have memory mapped using two different sets of
123 	 * memory attributes (shared, type, and cache attribs).  We can not
124 	 * change these attributes once the initial assembly has setup the
125 	 * page tables.
126 	 */
127 	if (cpu_architecture() >= CPU_ARCH_ARMv6) {
128 		printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
129 		cachepolicy = CPOLICY_WRITEBACK;
130 	}
131 	flush_cache_all();
132 	set_cr(cr_alignment);
133 	return 0;
134 }
135 early_param("cachepolicy", early_cachepolicy);
136 
137 static int __init early_nocache(char *__unused)
138 {
139 	char *p = "buffered";
140 	printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
141 	early_cachepolicy(p);
142 	return 0;
143 }
144 early_param("nocache", early_nocache);
145 
146 static int __init early_nowrite(char *__unused)
147 {
148 	char *p = "uncached";
149 	printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
150 	early_cachepolicy(p);
151 	return 0;
152 }
153 early_param("nowb", early_nowrite);
154 
155 static int __init early_ecc(char *p)
156 {
157 	if (memcmp(p, "on", 2) == 0)
158 		ecc_mask = PMD_PROTECTION;
159 	else if (memcmp(p, "off", 3) == 0)
160 		ecc_mask = 0;
161 	return 0;
162 }
163 early_param("ecc", early_ecc);
164 
165 static int __init noalign_setup(char *__unused)
166 {
167 	cr_alignment &= ~CR_A;
168 	cr_no_alignment &= ~CR_A;
169 	set_cr(cr_alignment);
170 	return 1;
171 }
172 __setup("noalign", noalign_setup);
173 
174 #ifndef CONFIG_SMP
175 void adjust_cr(unsigned long mask, unsigned long set)
176 {
177 	unsigned long flags;
178 
179 	mask &= ~CR_A;
180 
181 	set &= mask;
182 
183 	local_irq_save(flags);
184 
185 	cr_no_alignment = (cr_no_alignment & ~mask) | set;
186 	cr_alignment = (cr_alignment & ~mask) | set;
187 
188 	set_cr((get_cr() & ~mask) | set);
189 
190 	local_irq_restore(flags);
191 }
192 #endif
193 
194 #define PROT_PTE_DEVICE		L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
195 #define PROT_SECT_DEVICE	PMD_TYPE_SECT|PMD_SECT_AP_WRITE
196 
197 static struct mem_type mem_types[] = {
198 	[MT_DEVICE] = {		  /* Strongly ordered / ARMv6 shared device */
199 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
200 				  L_PTE_SHARED,
201 		.prot_l1	= PMD_TYPE_TABLE,
202 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_S,
203 		.domain		= DOMAIN_IO,
204 	},
205 	[MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
206 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
207 		.prot_l1	= PMD_TYPE_TABLE,
208 		.prot_sect	= PROT_SECT_DEVICE,
209 		.domain		= DOMAIN_IO,
210 	},
211 	[MT_DEVICE_CACHED] = {	  /* ioremap_cached */
212 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
213 		.prot_l1	= PMD_TYPE_TABLE,
214 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_WB,
215 		.domain		= DOMAIN_IO,
216 	},
217 	[MT_DEVICE_WC] = {	/* ioremap_wc */
218 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
219 		.prot_l1	= PMD_TYPE_TABLE,
220 		.prot_sect	= PROT_SECT_DEVICE,
221 		.domain		= DOMAIN_IO,
222 	},
223 	[MT_UNCACHED] = {
224 		.prot_pte	= PROT_PTE_DEVICE,
225 		.prot_l1	= PMD_TYPE_TABLE,
226 		.prot_sect	= PMD_TYPE_SECT | PMD_SECT_XN,
227 		.domain		= DOMAIN_IO,
228 	},
229 	[MT_CACHECLEAN] = {
230 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
231 		.domain    = DOMAIN_KERNEL,
232 	},
233 	[MT_MINICLEAN] = {
234 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
235 		.domain    = DOMAIN_KERNEL,
236 	},
237 	[MT_LOW_VECTORS] = {
238 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
239 				L_PTE_RDONLY,
240 		.prot_l1   = PMD_TYPE_TABLE,
241 		.domain    = DOMAIN_USER,
242 	},
243 	[MT_HIGH_VECTORS] = {
244 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
245 				L_PTE_USER | L_PTE_RDONLY,
246 		.prot_l1   = PMD_TYPE_TABLE,
247 		.domain    = DOMAIN_USER,
248 	},
249 	[MT_MEMORY] = {
250 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
251 		.prot_l1   = PMD_TYPE_TABLE,
252 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
253 		.domain    = DOMAIN_KERNEL,
254 	},
255 	[MT_ROM] = {
256 		.prot_sect = PMD_TYPE_SECT,
257 		.domain    = DOMAIN_KERNEL,
258 	},
259 	[MT_MEMORY_NONCACHED] = {
260 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
261 				L_PTE_MT_BUFFERABLE,
262 		.prot_l1   = PMD_TYPE_TABLE,
263 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
264 		.domain    = DOMAIN_KERNEL,
265 	},
266 	[MT_MEMORY_DTCM] = {
267 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
268 				L_PTE_XN,
269 		.prot_l1   = PMD_TYPE_TABLE,
270 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
271 		.domain    = DOMAIN_KERNEL,
272 	},
273 	[MT_MEMORY_ITCM] = {
274 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
275 		.prot_l1   = PMD_TYPE_TABLE,
276 		.domain    = DOMAIN_KERNEL,
277 	},
278 };
279 
280 const struct mem_type *get_mem_type(unsigned int type)
281 {
282 	return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
283 }
284 EXPORT_SYMBOL(get_mem_type);
285 
286 /*
287  * Adjust the PMD section entries according to the CPU in use.
288  */
289 static void __init build_mem_type_table(void)
290 {
291 	struct cachepolicy *cp;
292 	unsigned int cr = get_cr();
293 	unsigned int user_pgprot, kern_pgprot, vecs_pgprot;
294 	int cpu_arch = cpu_architecture();
295 	int i;
296 
297 	if (cpu_arch < CPU_ARCH_ARMv6) {
298 #if defined(CONFIG_CPU_DCACHE_DISABLE)
299 		if (cachepolicy > CPOLICY_BUFFERED)
300 			cachepolicy = CPOLICY_BUFFERED;
301 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
302 		if (cachepolicy > CPOLICY_WRITETHROUGH)
303 			cachepolicy = CPOLICY_WRITETHROUGH;
304 #endif
305 	}
306 	if (cpu_arch < CPU_ARCH_ARMv5) {
307 		if (cachepolicy >= CPOLICY_WRITEALLOC)
308 			cachepolicy = CPOLICY_WRITEBACK;
309 		ecc_mask = 0;
310 	}
311 	if (is_smp())
312 		cachepolicy = CPOLICY_WRITEALLOC;
313 
314 	/*
315 	 * Strip out features not present on earlier architectures.
316 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
317 	 * without extended page tables don't have the 'Shared' bit.
318 	 */
319 	if (cpu_arch < CPU_ARCH_ARMv5)
320 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
321 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
322 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
323 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
324 			mem_types[i].prot_sect &= ~PMD_SECT_S;
325 
326 	/*
327 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
328 	 * "update-able on write" bit on ARM610).  However, Xscale and
329 	 * Xscale3 require this bit to be cleared.
330 	 */
331 	if (cpu_is_xscale() || cpu_is_xsc3()) {
332 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
333 			mem_types[i].prot_sect &= ~PMD_BIT4;
334 			mem_types[i].prot_l1 &= ~PMD_BIT4;
335 		}
336 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
337 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
338 			if (mem_types[i].prot_l1)
339 				mem_types[i].prot_l1 |= PMD_BIT4;
340 			if (mem_types[i].prot_sect)
341 				mem_types[i].prot_sect |= PMD_BIT4;
342 		}
343 	}
344 
345 	/*
346 	 * Mark the device areas according to the CPU/architecture.
347 	 */
348 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
349 		if (!cpu_is_xsc3()) {
350 			/*
351 			 * Mark device regions on ARMv6+ as execute-never
352 			 * to prevent speculative instruction fetches.
353 			 */
354 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
355 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
356 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
357 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
358 		}
359 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
360 			/*
361 			 * For ARMv7 with TEX remapping,
362 			 * - shared device is SXCB=1100
363 			 * - nonshared device is SXCB=0100
364 			 * - write combine device mem is SXCB=0001
365 			 * (Uncached Normal memory)
366 			 */
367 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
368 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
369 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
370 		} else if (cpu_is_xsc3()) {
371 			/*
372 			 * For Xscale3,
373 			 * - shared device is TEXCB=00101
374 			 * - nonshared device is TEXCB=01000
375 			 * - write combine device mem is TEXCB=00100
376 			 * (Inner/Outer Uncacheable in xsc3 parlance)
377 			 */
378 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
379 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
380 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
381 		} else {
382 			/*
383 			 * For ARMv6 and ARMv7 without TEX remapping,
384 			 * - shared device is TEXCB=00001
385 			 * - nonshared device is TEXCB=01000
386 			 * - write combine device mem is TEXCB=00100
387 			 * (Uncached Normal in ARMv6 parlance).
388 			 */
389 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
390 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
391 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
392 		}
393 	} else {
394 		/*
395 		 * On others, write combining is "Uncached/Buffered"
396 		 */
397 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
398 	}
399 
400 	/*
401 	 * Now deal with the memory-type mappings
402 	 */
403 	cp = &cache_policies[cachepolicy];
404 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
405 
406 	/*
407 	 * Only use write-through for non-SMP systems
408 	 */
409 	if (!is_smp() && cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
410 		vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
411 
412 	/*
413 	 * Enable CPU-specific coherency if supported.
414 	 * (Only available on XSC3 at the moment.)
415 	 */
416 	if (arch_is_coherent() && cpu_is_xsc3()) {
417 		mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
418 		mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
419 		mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
420 		mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
421 	}
422 	/*
423 	 * ARMv6 and above have extended page tables.
424 	 */
425 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
426 		/*
427 		 * Mark cache clean areas and XIP ROM read only
428 		 * from SVC mode and no access from userspace.
429 		 */
430 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
431 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
432 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
433 
434 		if (is_smp()) {
435 			/*
436 			 * Mark memory with the "shared" attribute
437 			 * for SMP systems
438 			 */
439 			user_pgprot |= L_PTE_SHARED;
440 			kern_pgprot |= L_PTE_SHARED;
441 			vecs_pgprot |= L_PTE_SHARED;
442 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
443 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
444 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
445 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
446 			mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
447 			mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
448 			mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
449 			mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
450 		}
451 	}
452 
453 	/*
454 	 * Non-cacheable Normal - intended for memory areas that must
455 	 * not cause dirty cache line writebacks when used
456 	 */
457 	if (cpu_arch >= CPU_ARCH_ARMv6) {
458 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
459 			/* Non-cacheable Normal is XCB = 001 */
460 			mem_types[MT_MEMORY_NONCACHED].prot_sect |=
461 				PMD_SECT_BUFFERED;
462 		} else {
463 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
464 			mem_types[MT_MEMORY_NONCACHED].prot_sect |=
465 				PMD_SECT_TEX(1);
466 		}
467 	} else {
468 		mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
469 	}
470 
471 	for (i = 0; i < 16; i++) {
472 		unsigned long v = pgprot_val(protection_map[i]);
473 		protection_map[i] = __pgprot(v | user_pgprot);
474 	}
475 
476 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
477 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
478 
479 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
480 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
481 				 L_PTE_DIRTY | kern_pgprot);
482 
483 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
484 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
485 	mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
486 	mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
487 	mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
488 	mem_types[MT_ROM].prot_sect |= cp->pmd;
489 
490 	switch (cp->pmd) {
491 	case PMD_SECT_WT:
492 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
493 		break;
494 	case PMD_SECT_WB:
495 	case PMD_SECT_WBWA:
496 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
497 		break;
498 	}
499 	printk("Memory policy: ECC %sabled, Data cache %s\n",
500 		ecc_mask ? "en" : "dis", cp->policy);
501 
502 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
503 		struct mem_type *t = &mem_types[i];
504 		if (t->prot_l1)
505 			t->prot_l1 |= PMD_DOMAIN(t->domain);
506 		if (t->prot_sect)
507 			t->prot_sect |= PMD_DOMAIN(t->domain);
508 	}
509 }
510 
511 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
512 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
513 			      unsigned long size, pgprot_t vma_prot)
514 {
515 	if (!pfn_valid(pfn))
516 		return pgprot_noncached(vma_prot);
517 	else if (file->f_flags & O_SYNC)
518 		return pgprot_writecombine(vma_prot);
519 	return vma_prot;
520 }
521 EXPORT_SYMBOL(phys_mem_access_prot);
522 #endif
523 
524 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
525 
526 static void __init *early_alloc(unsigned long sz)
527 {
528 	void *ptr = __va(memblock_alloc(sz, sz));
529 	memset(ptr, 0, sz);
530 	return ptr;
531 }
532 
533 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
534 {
535 	if (pmd_none(*pmd)) {
536 		pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
537 		__pmd_populate(pmd, __pa(pte), prot);
538 	}
539 	BUG_ON(pmd_bad(*pmd));
540 	return pte_offset_kernel(pmd, addr);
541 }
542 
543 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
544 				  unsigned long end, unsigned long pfn,
545 				  const struct mem_type *type)
546 {
547 	pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
548 	do {
549 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
550 		pfn++;
551 	} while (pte++, addr += PAGE_SIZE, addr != end);
552 }
553 
554 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
555 				      unsigned long end, phys_addr_t phys,
556 				      const struct mem_type *type)
557 {
558 	pmd_t *pmd = pmd_offset(pud, addr);
559 
560 	/*
561 	 * Try a section mapping - end, addr and phys must all be aligned
562 	 * to a section boundary.  Note that PMDs refer to the individual
563 	 * L1 entries, whereas PGDs refer to a group of L1 entries making
564 	 * up one logical pointer to an L2 table.
565 	 */
566 	if (((addr | end | phys) & ~SECTION_MASK) == 0) {
567 		pmd_t *p = pmd;
568 
569 		if (addr & SECTION_SIZE)
570 			pmd++;
571 
572 		do {
573 			*pmd = __pmd(phys | type->prot_sect);
574 			phys += SECTION_SIZE;
575 		} while (pmd++, addr += SECTION_SIZE, addr != end);
576 
577 		flush_pmd_entry(p);
578 	} else {
579 		/*
580 		 * No need to loop; pte's aren't interested in the
581 		 * individual L1 entries.
582 		 */
583 		alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
584 	}
585 }
586 
587 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
588 	unsigned long phys, const struct mem_type *type)
589 {
590 	pud_t *pud = pud_offset(pgd, addr);
591 	unsigned long next;
592 
593 	do {
594 		next = pud_addr_end(addr, end);
595 		alloc_init_section(pud, addr, next, phys, type);
596 		phys += next - addr;
597 	} while (pud++, addr = next, addr != end);
598 }
599 
600 static void __init create_36bit_mapping(struct map_desc *md,
601 					const struct mem_type *type)
602 {
603 	unsigned long addr, length, end;
604 	phys_addr_t phys;
605 	pgd_t *pgd;
606 
607 	addr = md->virtual;
608 	phys = __pfn_to_phys(md->pfn);
609 	length = PAGE_ALIGN(md->length);
610 
611 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
612 		printk(KERN_ERR "MM: CPU does not support supersection "
613 		       "mapping for 0x%08llx at 0x%08lx\n",
614 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
615 		return;
616 	}
617 
618 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
619 	 *	Since domain assignments can in fact be arbitrary, the
620 	 *	'domain == 0' check below is required to insure that ARMv6
621 	 *	supersections are only allocated for domain 0 regardless
622 	 *	of the actual domain assignments in use.
623 	 */
624 	if (type->domain) {
625 		printk(KERN_ERR "MM: invalid domain in supersection "
626 		       "mapping for 0x%08llx at 0x%08lx\n",
627 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
628 		return;
629 	}
630 
631 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
632 		printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
633 		       " at 0x%08lx invalid alignment\n",
634 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
635 		return;
636 	}
637 
638 	/*
639 	 * Shift bits [35:32] of address into bits [23:20] of PMD
640 	 * (See ARMv6 spec).
641 	 */
642 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
643 
644 	pgd = pgd_offset_k(addr);
645 	end = addr + length;
646 	do {
647 		pud_t *pud = pud_offset(pgd, addr);
648 		pmd_t *pmd = pmd_offset(pud, addr);
649 		int i;
650 
651 		for (i = 0; i < 16; i++)
652 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
653 
654 		addr += SUPERSECTION_SIZE;
655 		phys += SUPERSECTION_SIZE;
656 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
657 	} while (addr != end);
658 }
659 
660 /*
661  * Create the page directory entries and any necessary
662  * page tables for the mapping specified by `md'.  We
663  * are able to cope here with varying sizes and address
664  * offsets, and we take full advantage of sections and
665  * supersections.
666  */
667 static void __init create_mapping(struct map_desc *md)
668 {
669 	unsigned long addr, length, end;
670 	phys_addr_t phys;
671 	const struct mem_type *type;
672 	pgd_t *pgd;
673 
674 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
675 		printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
676 		       " at 0x%08lx in user region\n",
677 		       (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
678 		return;
679 	}
680 
681 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
682 	    md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
683 		printk(KERN_WARNING "BUG: mapping for 0x%08llx"
684 		       " at 0x%08lx overlaps vmalloc space\n",
685 		       (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
686 	}
687 
688 	type = &mem_types[md->type];
689 
690 	/*
691 	 * Catch 36-bit addresses
692 	 */
693 	if (md->pfn >= 0x100000) {
694 		create_36bit_mapping(md, type);
695 		return;
696 	}
697 
698 	addr = md->virtual & PAGE_MASK;
699 	phys = __pfn_to_phys(md->pfn);
700 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
701 
702 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
703 		printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
704 		       "be mapped using pages, ignoring.\n",
705 		       (long long)__pfn_to_phys(md->pfn), addr);
706 		return;
707 	}
708 
709 	pgd = pgd_offset_k(addr);
710 	end = addr + length;
711 	do {
712 		unsigned long next = pgd_addr_end(addr, end);
713 
714 		alloc_init_pud(pgd, addr, next, phys, type);
715 
716 		phys += next - addr;
717 		addr = next;
718 	} while (pgd++, addr != end);
719 }
720 
721 /*
722  * Create the architecture specific mappings
723  */
724 void __init iotable_init(struct map_desc *io_desc, int nr)
725 {
726 	int i;
727 
728 	for (i = 0; i < nr; i++)
729 		create_mapping(io_desc + i);
730 }
731 
732 static void * __initdata vmalloc_min = (void *)(VMALLOC_END - SZ_128M);
733 
734 /*
735  * vmalloc=size forces the vmalloc area to be exactly 'size'
736  * bytes. This can be used to increase (or decrease) the vmalloc
737  * area - the default is 128m.
738  */
739 static int __init early_vmalloc(char *arg)
740 {
741 	unsigned long vmalloc_reserve = memparse(arg, NULL);
742 
743 	if (vmalloc_reserve < SZ_16M) {
744 		vmalloc_reserve = SZ_16M;
745 		printk(KERN_WARNING
746 			"vmalloc area too small, limiting to %luMB\n",
747 			vmalloc_reserve >> 20);
748 	}
749 
750 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
751 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
752 		printk(KERN_WARNING
753 			"vmalloc area is too big, limiting to %luMB\n",
754 			vmalloc_reserve >> 20);
755 	}
756 
757 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
758 	return 0;
759 }
760 early_param("vmalloc", early_vmalloc);
761 
762 static phys_addr_t lowmem_limit __initdata = 0;
763 
764 static void __init sanity_check_meminfo(void)
765 {
766 	int i, j, highmem = 0;
767 
768 	lowmem_limit = __pa(vmalloc_min - 1) + 1;
769 	memblock_set_current_limit(lowmem_limit);
770 
771 	for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
772 		struct membank *bank = &meminfo.bank[j];
773 		*bank = meminfo.bank[i];
774 
775 #ifdef CONFIG_HIGHMEM
776 		if (__va(bank->start) > vmalloc_min ||
777 		    __va(bank->start) < (void *)PAGE_OFFSET)
778 			highmem = 1;
779 
780 		bank->highmem = highmem;
781 
782 		/*
783 		 * Split those memory banks which are partially overlapping
784 		 * the vmalloc area greatly simplifying things later.
785 		 */
786 		if (__va(bank->start) < vmalloc_min &&
787 		    bank->size > vmalloc_min - __va(bank->start)) {
788 			if (meminfo.nr_banks >= NR_BANKS) {
789 				printk(KERN_CRIT "NR_BANKS too low, "
790 						 "ignoring high memory\n");
791 			} else {
792 				memmove(bank + 1, bank,
793 					(meminfo.nr_banks - i) * sizeof(*bank));
794 				meminfo.nr_banks++;
795 				i++;
796 				bank[1].size -= vmalloc_min - __va(bank->start);
797 				bank[1].start = __pa(vmalloc_min - 1) + 1;
798 				bank[1].highmem = highmem = 1;
799 				j++;
800 			}
801 			bank->size = vmalloc_min - __va(bank->start);
802 		}
803 #else
804 		bank->highmem = highmem;
805 
806 		/*
807 		 * Check whether this memory bank would entirely overlap
808 		 * the vmalloc area.
809 		 */
810 		if (__va(bank->start) >= vmalloc_min ||
811 		    __va(bank->start) < (void *)PAGE_OFFSET) {
812 			printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
813 			       "(vmalloc region overlap).\n",
814 			       (unsigned long long)bank->start,
815 			       (unsigned long long)bank->start + bank->size - 1);
816 			continue;
817 		}
818 
819 		/*
820 		 * Check whether this memory bank would partially overlap
821 		 * the vmalloc area.
822 		 */
823 		if (__va(bank->start + bank->size) > vmalloc_min ||
824 		    __va(bank->start + bank->size) < __va(bank->start)) {
825 			unsigned long newsize = vmalloc_min - __va(bank->start);
826 			printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
827 			       "to -%.8llx (vmalloc region overlap).\n",
828 			       (unsigned long long)bank->start,
829 			       (unsigned long long)bank->start + bank->size - 1,
830 			       (unsigned long long)bank->start + newsize - 1);
831 			bank->size = newsize;
832 		}
833 #endif
834 		j++;
835 	}
836 #ifdef CONFIG_HIGHMEM
837 	if (highmem) {
838 		const char *reason = NULL;
839 
840 		if (cache_is_vipt_aliasing()) {
841 			/*
842 			 * Interactions between kmap and other mappings
843 			 * make highmem support with aliasing VIPT caches
844 			 * rather difficult.
845 			 */
846 			reason = "with VIPT aliasing cache";
847 		}
848 		if (reason) {
849 			printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
850 				reason);
851 			while (j > 0 && meminfo.bank[j - 1].highmem)
852 				j--;
853 		}
854 	}
855 #endif
856 	meminfo.nr_banks = j;
857 }
858 
859 static inline void prepare_page_table(void)
860 {
861 	unsigned long addr;
862 	phys_addr_t end;
863 
864 	/*
865 	 * Clear out all the mappings below the kernel image.
866 	 */
867 	for (addr = 0; addr < MODULES_VADDR; addr += PGDIR_SIZE)
868 		pmd_clear(pmd_off_k(addr));
869 
870 #ifdef CONFIG_XIP_KERNEL
871 	/* The XIP kernel is mapped in the module area -- skip over it */
872 	addr = ((unsigned long)_etext + PGDIR_SIZE - 1) & PGDIR_MASK;
873 #endif
874 	for ( ; addr < PAGE_OFFSET; addr += PGDIR_SIZE)
875 		pmd_clear(pmd_off_k(addr));
876 
877 	/*
878 	 * Find the end of the first block of lowmem.
879 	 */
880 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
881 	if (end >= lowmem_limit)
882 		end = lowmem_limit;
883 
884 	/*
885 	 * Clear out all the kernel space mappings, except for the first
886 	 * memory bank, up to the end of the vmalloc region.
887 	 */
888 	for (addr = __phys_to_virt(end);
889 	     addr < VMALLOC_END; addr += PGDIR_SIZE)
890 		pmd_clear(pmd_off_k(addr));
891 }
892 
893 /*
894  * Reserve the special regions of memory
895  */
896 void __init arm_mm_memblock_reserve(void)
897 {
898 	/*
899 	 * Reserve the page tables.  These are already in use,
900 	 * and can only be in node 0.
901 	 */
902 	memblock_reserve(__pa(swapper_pg_dir), PTRS_PER_PGD * sizeof(pgd_t));
903 
904 #ifdef CONFIG_SA1111
905 	/*
906 	 * Because of the SA1111 DMA bug, we want to preserve our
907 	 * precious DMA-able memory...
908 	 */
909 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
910 #endif
911 }
912 
913 /*
914  * Set up device the mappings.  Since we clear out the page tables for all
915  * mappings above VMALLOC_END, we will remove any debug device mappings.
916  * This means you have to be careful how you debug this function, or any
917  * called function.  This means you can't use any function or debugging
918  * method which may touch any device, otherwise the kernel _will_ crash.
919  */
920 static void __init devicemaps_init(struct machine_desc *mdesc)
921 {
922 	struct map_desc map;
923 	unsigned long addr;
924 
925 	/*
926 	 * Allocate the vector page early.
927 	 */
928 	vectors_page = early_alloc(PAGE_SIZE);
929 
930 	for (addr = VMALLOC_END; addr; addr += PGDIR_SIZE)
931 		pmd_clear(pmd_off_k(addr));
932 
933 	/*
934 	 * Map the kernel if it is XIP.
935 	 * It is always first in the modulearea.
936 	 */
937 #ifdef CONFIG_XIP_KERNEL
938 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
939 	map.virtual = MODULES_VADDR;
940 	map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
941 	map.type = MT_ROM;
942 	create_mapping(&map);
943 #endif
944 
945 	/*
946 	 * Map the cache flushing regions.
947 	 */
948 #ifdef FLUSH_BASE
949 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
950 	map.virtual = FLUSH_BASE;
951 	map.length = SZ_1M;
952 	map.type = MT_CACHECLEAN;
953 	create_mapping(&map);
954 #endif
955 #ifdef FLUSH_BASE_MINICACHE
956 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
957 	map.virtual = FLUSH_BASE_MINICACHE;
958 	map.length = SZ_1M;
959 	map.type = MT_MINICLEAN;
960 	create_mapping(&map);
961 #endif
962 
963 	/*
964 	 * Create a mapping for the machine vectors at the high-vectors
965 	 * location (0xffff0000).  If we aren't using high-vectors, also
966 	 * create a mapping at the low-vectors virtual address.
967 	 */
968 	map.pfn = __phys_to_pfn(virt_to_phys(vectors_page));
969 	map.virtual = 0xffff0000;
970 	map.length = PAGE_SIZE;
971 	map.type = MT_HIGH_VECTORS;
972 	create_mapping(&map);
973 
974 	if (!vectors_high()) {
975 		map.virtual = 0;
976 		map.type = MT_LOW_VECTORS;
977 		create_mapping(&map);
978 	}
979 
980 	/*
981 	 * Ask the machine support to map in the statically mapped devices.
982 	 */
983 	if (mdesc->map_io)
984 		mdesc->map_io();
985 
986 	/*
987 	 * Finally flush the caches and tlb to ensure that we're in a
988 	 * consistent state wrt the writebuffer.  This also ensures that
989 	 * any write-allocated cache lines in the vector page are written
990 	 * back.  After this point, we can start to touch devices again.
991 	 */
992 	local_flush_tlb_all();
993 	flush_cache_all();
994 }
995 
996 static void __init kmap_init(void)
997 {
998 #ifdef CONFIG_HIGHMEM
999 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1000 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1001 #endif
1002 }
1003 
1004 static void __init map_lowmem(void)
1005 {
1006 	struct memblock_region *reg;
1007 
1008 	/* Map all the lowmem memory banks. */
1009 	for_each_memblock(memory, reg) {
1010 		phys_addr_t start = reg->base;
1011 		phys_addr_t end = start + reg->size;
1012 		struct map_desc map;
1013 
1014 		if (end > lowmem_limit)
1015 			end = lowmem_limit;
1016 		if (start >= end)
1017 			break;
1018 
1019 		map.pfn = __phys_to_pfn(start);
1020 		map.virtual = __phys_to_virt(start);
1021 		map.length = end - start;
1022 		map.type = MT_MEMORY;
1023 
1024 		create_mapping(&map);
1025 	}
1026 }
1027 
1028 /*
1029  * paging_init() sets up the page tables, initialises the zone memory
1030  * maps, and sets up the zero page, bad page and bad page tables.
1031  */
1032 void __init paging_init(struct machine_desc *mdesc)
1033 {
1034 	void *zero_page;
1035 
1036 	build_mem_type_table();
1037 	sanity_check_meminfo();
1038 	prepare_page_table();
1039 	map_lowmem();
1040 	devicemaps_init(mdesc);
1041 	kmap_init();
1042 
1043 	top_pmd = pmd_off_k(0xffff0000);
1044 
1045 	/* allocate the zero page. */
1046 	zero_page = early_alloc(PAGE_SIZE);
1047 
1048 	bootmem_init();
1049 
1050 	empty_zero_page = virt_to_page(zero_page);
1051 	__flush_dcache_page(NULL, empty_zero_page);
1052 }
1053