Merge branch 'devel-stable' into for-next
[linux-2.6.git] / arch / arm / mm / mmu.c
blob01c03366c45e5d944566b4013ba99b3960d95b49
1 /*
2 * linux/arch/arm/mm/mmu.c
4 * Copyright (C) 1995-2005 Russell King
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>
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>
32 #include <asm/mach/arch.h>
33 #include <asm/mach/map.h>
34 #include <asm/mach/pci.h>
36 #include "mm.h"
37 #include "tcm.h"
40 * empty_zero_page is a special page that is used for
41 * zero-initialized data and COW.
43 struct page *empty_zero_page;
44 EXPORT_SYMBOL(empty_zero_page);
47 * The pmd table for the upper-most set of pages.
49 pmd_t *top_pmd;
51 #define CPOLICY_UNCACHED 0
52 #define CPOLICY_BUFFERED 1
53 #define CPOLICY_WRITETHROUGH 2
54 #define CPOLICY_WRITEBACK 3
55 #define CPOLICY_WRITEALLOC 4
57 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
58 static unsigned int ecc_mask __initdata = 0;
59 pgprot_t pgprot_user;
60 pgprot_t pgprot_kernel;
61 pgprot_t pgprot_hyp_device;
62 pgprot_t pgprot_s2;
63 pgprot_t pgprot_s2_device;
65 EXPORT_SYMBOL(pgprot_user);
66 EXPORT_SYMBOL(pgprot_kernel);
68 struct cachepolicy {
69 const char policy[16];
70 unsigned int cr_mask;
71 pmdval_t pmd;
72 pteval_t pte;
73 pteval_t pte_s2;
76 #ifdef CONFIG_ARM_LPAE
77 #define s2_policy(policy) policy
78 #else
79 #define s2_policy(policy) 0
80 #endif
82 static struct cachepolicy cache_policies[] __initdata = {
84 .policy = "uncached",
85 .cr_mask = CR_W|CR_C,
86 .pmd = PMD_SECT_UNCACHED,
87 .pte = L_PTE_MT_UNCACHED,
88 .pte_s2 = s2_policy(L_PTE_S2_MT_UNCACHED),
89 }, {
90 .policy = "buffered",
91 .cr_mask = CR_C,
92 .pmd = PMD_SECT_BUFFERED,
93 .pte = L_PTE_MT_BUFFERABLE,
94 .pte_s2 = s2_policy(L_PTE_S2_MT_UNCACHED),
95 }, {
96 .policy = "writethrough",
97 .cr_mask = 0,
98 .pmd = PMD_SECT_WT,
99 .pte = L_PTE_MT_WRITETHROUGH,
100 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITETHROUGH),
101 }, {
102 .policy = "writeback",
103 .cr_mask = 0,
104 .pmd = PMD_SECT_WB,
105 .pte = L_PTE_MT_WRITEBACK,
106 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITEBACK),
107 }, {
108 .policy = "writealloc",
109 .cr_mask = 0,
110 .pmd = PMD_SECT_WBWA,
111 .pte = L_PTE_MT_WRITEALLOC,
112 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITEBACK),
116 #ifdef CONFIG_CPU_CP15
118 * These are useful for identifying cache coherency
119 * problems by allowing the cache or the cache and
120 * writebuffer to be turned off. (Note: the write
121 * buffer should not be on and the cache off).
123 static int __init early_cachepolicy(char *p)
125 int i;
127 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
128 int len = strlen(cache_policies[i].policy);
130 if (memcmp(p, cache_policies[i].policy, len) == 0) {
131 cachepolicy = i;
132 cr_alignment &= ~cache_policies[i].cr_mask;
133 cr_no_alignment &= ~cache_policies[i].cr_mask;
134 break;
137 if (i == ARRAY_SIZE(cache_policies))
138 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
140 * This restriction is partly to do with the way we boot; it is
141 * unpredictable to have memory mapped using two different sets of
142 * memory attributes (shared, type, and cache attribs). We can not
143 * change these attributes once the initial assembly has setup the
144 * page tables.
146 if (cpu_architecture() >= CPU_ARCH_ARMv6) {
147 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
148 cachepolicy = CPOLICY_WRITEBACK;
150 flush_cache_all();
151 set_cr(cr_alignment);
152 return 0;
154 early_param("cachepolicy", early_cachepolicy);
156 static int __init early_nocache(char *__unused)
158 char *p = "buffered";
159 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
160 early_cachepolicy(p);
161 return 0;
163 early_param("nocache", early_nocache);
165 static int __init early_nowrite(char *__unused)
167 char *p = "uncached";
168 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
169 early_cachepolicy(p);
170 return 0;
172 early_param("nowb", early_nowrite);
174 #ifndef CONFIG_ARM_LPAE
175 static int __init early_ecc(char *p)
177 if (memcmp(p, "on", 2) == 0)
178 ecc_mask = PMD_PROTECTION;
179 else if (memcmp(p, "off", 3) == 0)
180 ecc_mask = 0;
181 return 0;
183 early_param("ecc", early_ecc);
184 #endif
186 static int __init noalign_setup(char *__unused)
188 cr_alignment &= ~CR_A;
189 cr_no_alignment &= ~CR_A;
190 set_cr(cr_alignment);
191 return 1;
193 __setup("noalign", noalign_setup);
195 #ifndef CONFIG_SMP
196 void adjust_cr(unsigned long mask, unsigned long set)
198 unsigned long flags;
200 mask &= ~CR_A;
202 set &= mask;
204 local_irq_save(flags);
206 cr_no_alignment = (cr_no_alignment & ~mask) | set;
207 cr_alignment = (cr_alignment & ~mask) | set;
209 set_cr((get_cr() & ~mask) | set);
211 local_irq_restore(flags);
213 #endif
215 #else /* ifdef CONFIG_CPU_CP15 */
217 static int __init early_cachepolicy(char *p)
219 pr_warning("cachepolicy kernel parameter not supported without cp15\n");
221 early_param("cachepolicy", early_cachepolicy);
223 static int __init noalign_setup(char *__unused)
225 pr_warning("noalign kernel parameter not supported without cp15\n");
227 __setup("noalign", noalign_setup);
229 #endif /* ifdef CONFIG_CPU_CP15 / else */
231 #define PROT_PTE_DEVICE L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
232 #define PROT_SECT_DEVICE PMD_TYPE_SECT|PMD_SECT_AP_WRITE
234 static struct mem_type mem_types[] = {
235 [MT_DEVICE] = { /* Strongly ordered / ARMv6 shared device */
236 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
237 L_PTE_SHARED,
238 .prot_l1 = PMD_TYPE_TABLE,
239 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_S,
240 .domain = DOMAIN_IO,
242 [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
243 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
244 .prot_l1 = PMD_TYPE_TABLE,
245 .prot_sect = PROT_SECT_DEVICE,
246 .domain = DOMAIN_IO,
248 [MT_DEVICE_CACHED] = { /* ioremap_cached */
249 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
250 .prot_l1 = PMD_TYPE_TABLE,
251 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_WB,
252 .domain = DOMAIN_IO,
254 [MT_DEVICE_WC] = { /* ioremap_wc */
255 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
256 .prot_l1 = PMD_TYPE_TABLE,
257 .prot_sect = PROT_SECT_DEVICE,
258 .domain = DOMAIN_IO,
260 [MT_UNCACHED] = {
261 .prot_pte = PROT_PTE_DEVICE,
262 .prot_l1 = PMD_TYPE_TABLE,
263 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
264 .domain = DOMAIN_IO,
266 [MT_CACHECLEAN] = {
267 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
268 .domain = DOMAIN_KERNEL,
270 #ifndef CONFIG_ARM_LPAE
271 [MT_MINICLEAN] = {
272 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
273 .domain = DOMAIN_KERNEL,
275 #endif
276 [MT_LOW_VECTORS] = {
277 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
278 L_PTE_RDONLY,
279 .prot_l1 = PMD_TYPE_TABLE,
280 .domain = DOMAIN_USER,
282 [MT_HIGH_VECTORS] = {
283 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
284 L_PTE_USER | L_PTE_RDONLY,
285 .prot_l1 = PMD_TYPE_TABLE,
286 .domain = DOMAIN_USER,
288 [MT_MEMORY] = {
289 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
290 .prot_l1 = PMD_TYPE_TABLE,
291 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
292 .domain = DOMAIN_KERNEL,
294 [MT_ROM] = {
295 .prot_sect = PMD_TYPE_SECT,
296 .domain = DOMAIN_KERNEL,
298 [MT_MEMORY_NONCACHED] = {
299 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
300 L_PTE_MT_BUFFERABLE,
301 .prot_l1 = PMD_TYPE_TABLE,
302 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
303 .domain = DOMAIN_KERNEL,
305 [MT_MEMORY_DTCM] = {
306 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
307 L_PTE_XN,
308 .prot_l1 = PMD_TYPE_TABLE,
309 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
310 .domain = DOMAIN_KERNEL,
312 [MT_MEMORY_ITCM] = {
313 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
314 .prot_l1 = PMD_TYPE_TABLE,
315 .domain = DOMAIN_KERNEL,
317 [MT_MEMORY_SO] = {
318 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
319 L_PTE_MT_UNCACHED | L_PTE_XN,
320 .prot_l1 = PMD_TYPE_TABLE,
321 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
322 PMD_SECT_UNCACHED | PMD_SECT_XN,
323 .domain = DOMAIN_KERNEL,
325 [MT_MEMORY_DMA_READY] = {
326 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
327 .prot_l1 = PMD_TYPE_TABLE,
328 .domain = DOMAIN_KERNEL,
332 const struct mem_type *get_mem_type(unsigned int type)
334 return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
336 EXPORT_SYMBOL(get_mem_type);
339 * Adjust the PMD section entries according to the CPU in use.
341 static void __init build_mem_type_table(void)
343 struct cachepolicy *cp;
344 unsigned int cr = get_cr();
345 pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
346 pteval_t hyp_device_pgprot, s2_pgprot, s2_device_pgprot;
347 int cpu_arch = cpu_architecture();
348 int i;
350 if (cpu_arch < CPU_ARCH_ARMv6) {
351 #if defined(CONFIG_CPU_DCACHE_DISABLE)
352 if (cachepolicy > CPOLICY_BUFFERED)
353 cachepolicy = CPOLICY_BUFFERED;
354 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
355 if (cachepolicy > CPOLICY_WRITETHROUGH)
356 cachepolicy = CPOLICY_WRITETHROUGH;
357 #endif
359 if (cpu_arch < CPU_ARCH_ARMv5) {
360 if (cachepolicy >= CPOLICY_WRITEALLOC)
361 cachepolicy = CPOLICY_WRITEBACK;
362 ecc_mask = 0;
364 if (is_smp())
365 cachepolicy = CPOLICY_WRITEALLOC;
368 * Strip out features not present on earlier architectures.
369 * Pre-ARMv5 CPUs don't have TEX bits. Pre-ARMv6 CPUs or those
370 * without extended page tables don't have the 'Shared' bit.
372 if (cpu_arch < CPU_ARCH_ARMv5)
373 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
374 mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
375 if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
376 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
377 mem_types[i].prot_sect &= ~PMD_SECT_S;
380 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
381 * "update-able on write" bit on ARM610). However, Xscale and
382 * Xscale3 require this bit to be cleared.
384 if (cpu_is_xscale() || cpu_is_xsc3()) {
385 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
386 mem_types[i].prot_sect &= ~PMD_BIT4;
387 mem_types[i].prot_l1 &= ~PMD_BIT4;
389 } else if (cpu_arch < CPU_ARCH_ARMv6) {
390 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
391 if (mem_types[i].prot_l1)
392 mem_types[i].prot_l1 |= PMD_BIT4;
393 if (mem_types[i].prot_sect)
394 mem_types[i].prot_sect |= PMD_BIT4;
399 * Mark the device areas according to the CPU/architecture.
401 if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
402 if (!cpu_is_xsc3()) {
404 * Mark device regions on ARMv6+ as execute-never
405 * to prevent speculative instruction fetches.
407 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
408 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
409 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
410 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
412 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
414 * For ARMv7 with TEX remapping,
415 * - shared device is SXCB=1100
416 * - nonshared device is SXCB=0100
417 * - write combine device mem is SXCB=0001
418 * (Uncached Normal memory)
420 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
421 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
422 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
423 } else if (cpu_is_xsc3()) {
425 * For Xscale3,
426 * - shared device is TEXCB=00101
427 * - nonshared device is TEXCB=01000
428 * - write combine device mem is TEXCB=00100
429 * (Inner/Outer Uncacheable in xsc3 parlance)
431 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
432 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
433 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
434 } else {
436 * For ARMv6 and ARMv7 without TEX remapping,
437 * - shared device is TEXCB=00001
438 * - nonshared device is TEXCB=01000
439 * - write combine device mem is TEXCB=00100
440 * (Uncached Normal in ARMv6 parlance).
442 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
443 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
444 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
446 } else {
448 * On others, write combining is "Uncached/Buffered"
450 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
454 * Now deal with the memory-type mappings
456 cp = &cache_policies[cachepolicy];
457 vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
458 s2_pgprot = cp->pte_s2;
459 hyp_device_pgprot = s2_device_pgprot = mem_types[MT_DEVICE].prot_pte;
462 * ARMv6 and above have extended page tables.
464 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
465 #ifndef CONFIG_ARM_LPAE
467 * Mark cache clean areas and XIP ROM read only
468 * from SVC mode and no access from userspace.
470 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
471 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
472 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
473 #endif
475 if (is_smp()) {
477 * Mark memory with the "shared" attribute
478 * for SMP systems
480 user_pgprot |= L_PTE_SHARED;
481 kern_pgprot |= L_PTE_SHARED;
482 vecs_pgprot |= L_PTE_SHARED;
483 s2_pgprot |= L_PTE_SHARED;
484 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
485 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
486 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
487 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
488 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
489 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
490 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
491 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
492 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
497 * Non-cacheable Normal - intended for memory areas that must
498 * not cause dirty cache line writebacks when used
500 if (cpu_arch >= CPU_ARCH_ARMv6) {
501 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
502 /* Non-cacheable Normal is XCB = 001 */
503 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
504 PMD_SECT_BUFFERED;
505 } else {
506 /* For both ARMv6 and non-TEX-remapping ARMv7 */
507 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
508 PMD_SECT_TEX(1);
510 } else {
511 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
514 #ifdef CONFIG_ARM_LPAE
516 * Do not generate access flag faults for the kernel mappings.
518 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
519 mem_types[i].prot_pte |= PTE_EXT_AF;
520 if (mem_types[i].prot_sect)
521 mem_types[i].prot_sect |= PMD_SECT_AF;
523 kern_pgprot |= PTE_EXT_AF;
524 vecs_pgprot |= PTE_EXT_AF;
525 #endif
527 for (i = 0; i < 16; i++) {
528 pteval_t v = pgprot_val(protection_map[i]);
529 protection_map[i] = __pgprot(v | user_pgprot);
532 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
533 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
535 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
536 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
537 L_PTE_DIRTY | kern_pgprot);
538 pgprot_s2 = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | s2_pgprot);
539 pgprot_s2_device = __pgprot(s2_device_pgprot);
540 pgprot_hyp_device = __pgprot(hyp_device_pgprot);
542 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
543 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
544 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
545 mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
546 mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
547 mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
548 mem_types[MT_ROM].prot_sect |= cp->pmd;
550 switch (cp->pmd) {
551 case PMD_SECT_WT:
552 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
553 break;
554 case PMD_SECT_WB:
555 case PMD_SECT_WBWA:
556 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
557 break;
559 printk("Memory policy: ECC %sabled, Data cache %s\n",
560 ecc_mask ? "en" : "dis", cp->policy);
562 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
563 struct mem_type *t = &mem_types[i];
564 if (t->prot_l1)
565 t->prot_l1 |= PMD_DOMAIN(t->domain);
566 if (t->prot_sect)
567 t->prot_sect |= PMD_DOMAIN(t->domain);
571 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
572 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
573 unsigned long size, pgprot_t vma_prot)
575 if (!pfn_valid(pfn))
576 return pgprot_noncached(vma_prot);
577 else if (file->f_flags & O_SYNC)
578 return pgprot_writecombine(vma_prot);
579 return vma_prot;
581 EXPORT_SYMBOL(phys_mem_access_prot);
582 #endif
584 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
586 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
588 void *ptr = __va(memblock_alloc(sz, align));
589 memset(ptr, 0, sz);
590 return ptr;
593 static void __init *early_alloc(unsigned long sz)
595 return early_alloc_aligned(sz, sz);
598 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
600 if (pmd_none(*pmd)) {
601 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
602 __pmd_populate(pmd, __pa(pte), prot);
604 BUG_ON(pmd_bad(*pmd));
605 return pte_offset_kernel(pmd, addr);
608 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
609 unsigned long end, unsigned long pfn,
610 const struct mem_type *type)
612 pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
613 do {
614 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
615 pfn++;
616 } while (pte++, addr += PAGE_SIZE, addr != end);
619 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
620 unsigned long end, phys_addr_t phys,
621 const struct mem_type *type)
623 pmd_t *p = pmd;
625 #ifndef CONFIG_ARM_LPAE
627 * In classic MMU format, puds and pmds are folded in to
628 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
629 * group of L1 entries making up one logical pointer to
630 * an L2 table (2MB), where as PMDs refer to the individual
631 * L1 entries (1MB). Hence increment to get the correct
632 * offset for odd 1MB sections.
633 * (See arch/arm/include/asm/pgtable-2level.h)
635 if (addr & SECTION_SIZE)
636 pmd++;
637 #endif
638 do {
639 *pmd = __pmd(phys | type->prot_sect);
640 phys += SECTION_SIZE;
641 } while (pmd++, addr += SECTION_SIZE, addr != end);
643 flush_pmd_entry(p);
646 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
647 unsigned long end, phys_addr_t phys,
648 const struct mem_type *type)
650 pmd_t *pmd = pmd_offset(pud, addr);
651 unsigned long next;
653 do {
655 * With LPAE, we must loop over to map
656 * all the pmds for the given range.
658 next = pmd_addr_end(addr, end);
661 * Try a section mapping - addr, next and phys must all be
662 * aligned to a section boundary.
664 if (type->prot_sect &&
665 ((addr | next | phys) & ~SECTION_MASK) == 0) {
666 __map_init_section(pmd, addr, next, phys, type);
667 } else {
668 alloc_init_pte(pmd, addr, next,
669 __phys_to_pfn(phys), type);
672 phys += next - addr;
674 } while (pmd++, addr = next, addr != end);
677 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
678 unsigned long end, phys_addr_t phys,
679 const struct mem_type *type)
681 pud_t *pud = pud_offset(pgd, addr);
682 unsigned long next;
684 do {
685 next = pud_addr_end(addr, end);
686 alloc_init_pmd(pud, addr, next, phys, type);
687 phys += next - addr;
688 } while (pud++, addr = next, addr != end);
691 #ifndef CONFIG_ARM_LPAE
692 static void __init create_36bit_mapping(struct map_desc *md,
693 const struct mem_type *type)
695 unsigned long addr, length, end;
696 phys_addr_t phys;
697 pgd_t *pgd;
699 addr = md->virtual;
700 phys = __pfn_to_phys(md->pfn);
701 length = PAGE_ALIGN(md->length);
703 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
704 printk(KERN_ERR "MM: CPU does not support supersection "
705 "mapping for 0x%08llx at 0x%08lx\n",
706 (long long)__pfn_to_phys((u64)md->pfn), addr);
707 return;
710 /* N.B. ARMv6 supersections are only defined to work with domain 0.
711 * Since domain assignments can in fact be arbitrary, the
712 * 'domain == 0' check below is required to insure that ARMv6
713 * supersections are only allocated for domain 0 regardless
714 * of the actual domain assignments in use.
716 if (type->domain) {
717 printk(KERN_ERR "MM: invalid domain in supersection "
718 "mapping for 0x%08llx at 0x%08lx\n",
719 (long long)__pfn_to_phys((u64)md->pfn), addr);
720 return;
723 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
724 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
725 " at 0x%08lx invalid alignment\n",
726 (long long)__pfn_to_phys((u64)md->pfn), addr);
727 return;
731 * Shift bits [35:32] of address into bits [23:20] of PMD
732 * (See ARMv6 spec).
734 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
736 pgd = pgd_offset_k(addr);
737 end = addr + length;
738 do {
739 pud_t *pud = pud_offset(pgd, addr);
740 pmd_t *pmd = pmd_offset(pud, addr);
741 int i;
743 for (i = 0; i < 16; i++)
744 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
746 addr += SUPERSECTION_SIZE;
747 phys += SUPERSECTION_SIZE;
748 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
749 } while (addr != end);
751 #endif /* !CONFIG_ARM_LPAE */
754 * Create the page directory entries and any necessary
755 * page tables for the mapping specified by `md'. We
756 * are able to cope here with varying sizes and address
757 * offsets, and we take full advantage of sections and
758 * supersections.
760 static void __init create_mapping(struct map_desc *md)
762 unsigned long addr, length, end;
763 phys_addr_t phys;
764 const struct mem_type *type;
765 pgd_t *pgd;
767 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
768 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
769 " at 0x%08lx in user region\n",
770 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
771 return;
774 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
775 md->virtual >= PAGE_OFFSET &&
776 (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
777 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
778 " at 0x%08lx out of vmalloc space\n",
779 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
782 type = &mem_types[md->type];
784 #ifndef CONFIG_ARM_LPAE
786 * Catch 36-bit addresses
788 if (md->pfn >= 0x100000) {
789 create_36bit_mapping(md, type);
790 return;
792 #endif
794 addr = md->virtual & PAGE_MASK;
795 phys = __pfn_to_phys(md->pfn);
796 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
798 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
799 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
800 "be mapped using pages, ignoring.\n",
801 (long long)__pfn_to_phys(md->pfn), addr);
802 return;
805 pgd = pgd_offset_k(addr);
806 end = addr + length;
807 do {
808 unsigned long next = pgd_addr_end(addr, end);
810 alloc_init_pud(pgd, addr, next, phys, type);
812 phys += next - addr;
813 addr = next;
814 } while (pgd++, addr != end);
818 * Create the architecture specific mappings
820 void __init iotable_init(struct map_desc *io_desc, int nr)
822 struct map_desc *md;
823 struct vm_struct *vm;
824 struct static_vm *svm;
826 if (!nr)
827 return;
829 svm = early_alloc_aligned(sizeof(*svm) * nr, __alignof__(*svm));
831 for (md = io_desc; nr; md++, nr--) {
832 create_mapping(md);
834 vm = &svm->vm;
835 vm->addr = (void *)(md->virtual & PAGE_MASK);
836 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
837 vm->phys_addr = __pfn_to_phys(md->pfn);
838 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
839 vm->flags |= VM_ARM_MTYPE(md->type);
840 vm->caller = iotable_init;
841 add_static_vm_early(svm++);
845 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
846 void *caller)
848 struct vm_struct *vm;
849 struct static_vm *svm;
851 svm = early_alloc_aligned(sizeof(*svm), __alignof__(*svm));
853 vm = &svm->vm;
854 vm->addr = (void *)addr;
855 vm->size = size;
856 vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
857 vm->caller = caller;
858 add_static_vm_early(svm);
861 #ifndef CONFIG_ARM_LPAE
864 * The Linux PMD is made of two consecutive section entries covering 2MB
865 * (see definition in include/asm/pgtable-2level.h). However a call to
866 * create_mapping() may optimize static mappings by using individual
867 * 1MB section mappings. This leaves the actual PMD potentially half
868 * initialized if the top or bottom section entry isn't used, leaving it
869 * open to problems if a subsequent ioremap() or vmalloc() tries to use
870 * the virtual space left free by that unused section entry.
872 * Let's avoid the issue by inserting dummy vm entries covering the unused
873 * PMD halves once the static mappings are in place.
876 static void __init pmd_empty_section_gap(unsigned long addr)
878 vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
881 static void __init fill_pmd_gaps(void)
883 struct static_vm *svm;
884 struct vm_struct *vm;
885 unsigned long addr, next = 0;
886 pmd_t *pmd;
888 list_for_each_entry(svm, &static_vmlist, list) {
889 vm = &svm->vm;
890 addr = (unsigned long)vm->addr;
891 if (addr < next)
892 continue;
895 * Check if this vm starts on an odd section boundary.
896 * If so and the first section entry for this PMD is free
897 * then we block the corresponding virtual address.
899 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
900 pmd = pmd_off_k(addr);
901 if (pmd_none(*pmd))
902 pmd_empty_section_gap(addr & PMD_MASK);
906 * Then check if this vm ends on an odd section boundary.
907 * If so and the second section entry for this PMD is empty
908 * then we block the corresponding virtual address.
910 addr += vm->size;
911 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
912 pmd = pmd_off_k(addr) + 1;
913 if (pmd_none(*pmd))
914 pmd_empty_section_gap(addr);
917 /* no need to look at any vm entry until we hit the next PMD */
918 next = (addr + PMD_SIZE - 1) & PMD_MASK;
922 #else
923 #define fill_pmd_gaps() do { } while (0)
924 #endif
926 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
927 static void __init pci_reserve_io(void)
929 struct static_vm *svm;
931 svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
932 if (svm)
933 return;
935 vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
937 #else
938 #define pci_reserve_io() do { } while (0)
939 #endif
941 #ifdef CONFIG_DEBUG_LL
942 void __init debug_ll_io_init(void)
944 struct map_desc map;
946 debug_ll_addr(&map.pfn, &map.virtual);
947 if (!map.pfn || !map.virtual)
948 return;
949 map.pfn = __phys_to_pfn(map.pfn);
950 map.virtual &= PAGE_MASK;
951 map.length = PAGE_SIZE;
952 map.type = MT_DEVICE;
953 create_mapping(&map);
955 #endif
957 static void * __initdata vmalloc_min =
958 (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
961 * vmalloc=size forces the vmalloc area to be exactly 'size'
962 * bytes. This can be used to increase (or decrease) the vmalloc
963 * area - the default is 240m.
965 static int __init early_vmalloc(char *arg)
967 unsigned long vmalloc_reserve = memparse(arg, NULL);
969 if (vmalloc_reserve < SZ_16M) {
970 vmalloc_reserve = SZ_16M;
971 printk(KERN_WARNING
972 "vmalloc area too small, limiting to %luMB\n",
973 vmalloc_reserve >> 20);
976 if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
977 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
978 printk(KERN_WARNING
979 "vmalloc area is too big, limiting to %luMB\n",
980 vmalloc_reserve >> 20);
983 vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
984 return 0;
986 early_param("vmalloc", early_vmalloc);
988 phys_addr_t arm_lowmem_limit __initdata = 0;
990 void __init sanity_check_meminfo(void)
992 int i, j, highmem = 0;
993 phys_addr_t vmalloc_limit = __pa(vmalloc_min - 1) + 1;
995 for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
996 struct membank *bank = &meminfo.bank[j];
997 phys_addr_t size_limit;
999 *bank = meminfo.bank[i];
1000 size_limit = bank->size;
1002 if (bank->start >= vmalloc_limit)
1003 highmem = 1;
1004 else
1005 size_limit = vmalloc_limit - bank->start;
1007 bank->highmem = highmem;
1009 #ifdef CONFIG_HIGHMEM
1011 * Split those memory banks which are partially overlapping
1012 * the vmalloc area greatly simplifying things later.
1014 if (!highmem && bank->size > size_limit) {
1015 if (meminfo.nr_banks >= NR_BANKS) {
1016 printk(KERN_CRIT "NR_BANKS too low, "
1017 "ignoring high memory\n");
1018 } else {
1019 memmove(bank + 1, bank,
1020 (meminfo.nr_banks - i) * sizeof(*bank));
1021 meminfo.nr_banks++;
1022 i++;
1023 bank[1].size -= size_limit;
1024 bank[1].start = vmalloc_limit;
1025 bank[1].highmem = highmem = 1;
1026 j++;
1028 bank->size = size_limit;
1030 #else
1032 * Highmem banks not allowed with !CONFIG_HIGHMEM.
1034 if (highmem) {
1035 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
1036 "(!CONFIG_HIGHMEM).\n",
1037 (unsigned long long)bank->start,
1038 (unsigned long long)bank->start + bank->size - 1);
1039 continue;
1043 * Check whether this memory bank would partially overlap
1044 * the vmalloc area.
1046 if (bank->size > size_limit) {
1047 printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
1048 "to -%.8llx (vmalloc region overlap).\n",
1049 (unsigned long long)bank->start,
1050 (unsigned long long)bank->start + bank->size - 1,
1051 (unsigned long long)bank->start + size_limit - 1);
1052 bank->size = size_limit;
1054 #endif
1055 if (!bank->highmem && bank->start + bank->size > arm_lowmem_limit)
1056 arm_lowmem_limit = bank->start + bank->size;
1058 j++;
1060 #ifdef CONFIG_HIGHMEM
1061 if (highmem) {
1062 const char *reason = NULL;
1064 if (cache_is_vipt_aliasing()) {
1066 * Interactions between kmap and other mappings
1067 * make highmem support with aliasing VIPT caches
1068 * rather difficult.
1070 reason = "with VIPT aliasing cache";
1072 if (reason) {
1073 printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
1074 reason);
1075 while (j > 0 && meminfo.bank[j - 1].highmem)
1076 j--;
1079 #endif
1080 meminfo.nr_banks = j;
1081 high_memory = __va(arm_lowmem_limit - 1) + 1;
1082 memblock_set_current_limit(arm_lowmem_limit);
1085 static inline void prepare_page_table(void)
1087 unsigned long addr;
1088 phys_addr_t end;
1091 * Clear out all the mappings below the kernel image.
1093 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1094 pmd_clear(pmd_off_k(addr));
1096 #ifdef CONFIG_XIP_KERNEL
1097 /* The XIP kernel is mapped in the module area -- skip over it */
1098 addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
1099 #endif
1100 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1101 pmd_clear(pmd_off_k(addr));
1104 * Find the end of the first block of lowmem.
1106 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1107 if (end >= arm_lowmem_limit)
1108 end = arm_lowmem_limit;
1111 * Clear out all the kernel space mappings, except for the first
1112 * memory bank, up to the vmalloc region.
1114 for (addr = __phys_to_virt(end);
1115 addr < VMALLOC_START; addr += PMD_SIZE)
1116 pmd_clear(pmd_off_k(addr));
1119 #ifdef CONFIG_ARM_LPAE
1120 /* the first page is reserved for pgd */
1121 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \
1122 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1123 #else
1124 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
1125 #endif
1128 * Reserve the special regions of memory
1130 void __init arm_mm_memblock_reserve(void)
1133 * Reserve the page tables. These are already in use,
1134 * and can only be in node 0.
1136 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1138 #ifdef CONFIG_SA1111
1140 * Because of the SA1111 DMA bug, we want to preserve our
1141 * precious DMA-able memory...
1143 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1144 #endif
1148 * Set up the device mappings. Since we clear out the page tables for all
1149 * mappings above VMALLOC_START, we will remove any debug device mappings.
1150 * This means you have to be careful how you debug this function, or any
1151 * called function. This means you can't use any function or debugging
1152 * method which may touch any device, otherwise the kernel _will_ crash.
1154 static void __init devicemaps_init(struct machine_desc *mdesc)
1156 struct map_desc map;
1157 unsigned long addr;
1158 void *vectors;
1161 * Allocate the vector page early.
1163 vectors = early_alloc(PAGE_SIZE);
1165 early_trap_init(vectors);
1167 for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1168 pmd_clear(pmd_off_k(addr));
1171 * Map the kernel if it is XIP.
1172 * It is always first in the modulearea.
1174 #ifdef CONFIG_XIP_KERNEL
1175 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1176 map.virtual = MODULES_VADDR;
1177 map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1178 map.type = MT_ROM;
1179 create_mapping(&map);
1180 #endif
1183 * Map the cache flushing regions.
1185 #ifdef FLUSH_BASE
1186 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1187 map.virtual = FLUSH_BASE;
1188 map.length = SZ_1M;
1189 map.type = MT_CACHECLEAN;
1190 create_mapping(&map);
1191 #endif
1192 #ifdef FLUSH_BASE_MINICACHE
1193 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1194 map.virtual = FLUSH_BASE_MINICACHE;
1195 map.length = SZ_1M;
1196 map.type = MT_MINICLEAN;
1197 create_mapping(&map);
1198 #endif
1201 * Create a mapping for the machine vectors at the high-vectors
1202 * location (0xffff0000). If we aren't using high-vectors, also
1203 * create a mapping at the low-vectors virtual address.
1205 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1206 map.virtual = 0xffff0000;
1207 map.length = PAGE_SIZE;
1208 map.type = MT_HIGH_VECTORS;
1209 create_mapping(&map);
1211 if (!vectors_high()) {
1212 map.virtual = 0;
1213 map.type = MT_LOW_VECTORS;
1214 create_mapping(&map);
1218 * Ask the machine support to map in the statically mapped devices.
1220 if (mdesc->map_io)
1221 mdesc->map_io();
1222 fill_pmd_gaps();
1224 /* Reserve fixed i/o space in VMALLOC region */
1225 pci_reserve_io();
1228 * Finally flush the caches and tlb to ensure that we're in a
1229 * consistent state wrt the writebuffer. This also ensures that
1230 * any write-allocated cache lines in the vector page are written
1231 * back. After this point, we can start to touch devices again.
1233 local_flush_tlb_all();
1234 flush_cache_all();
1237 static void __init kmap_init(void)
1239 #ifdef CONFIG_HIGHMEM
1240 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1241 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1242 #endif
1245 static void __init map_lowmem(void)
1247 struct memblock_region *reg;
1249 /* Map all the lowmem memory banks. */
1250 for_each_memblock(memory, reg) {
1251 phys_addr_t start = reg->base;
1252 phys_addr_t end = start + reg->size;
1253 struct map_desc map;
1255 if (end > arm_lowmem_limit)
1256 end = arm_lowmem_limit;
1257 if (start >= end)
1258 break;
1260 map.pfn = __phys_to_pfn(start);
1261 map.virtual = __phys_to_virt(start);
1262 map.length = end - start;
1263 map.type = MT_MEMORY;
1265 create_mapping(&map);
1270 * paging_init() sets up the page tables, initialises the zone memory
1271 * maps, and sets up the zero page, bad page and bad page tables.
1273 void __init paging_init(struct machine_desc *mdesc)
1275 void *zero_page;
1277 memblock_set_current_limit(arm_lowmem_limit);
1279 build_mem_type_table();
1280 prepare_page_table();
1281 map_lowmem();
1282 dma_contiguous_remap();
1283 devicemaps_init(mdesc);
1284 kmap_init();
1285 tcm_init();
1287 top_pmd = pmd_off_k(0xffff0000);
1289 /* allocate the zero page. */
1290 zero_page = early_alloc(PAGE_SIZE);
1292 bootmem_init();
1294 empty_zero_page = virt_to_page(zero_page);
1295 __flush_dcache_page(NULL, empty_zero_page);