Merge branch 'reiserfs-updates' from Jeff Mahoney
[linux-2.6] / arch / arm / mm / mmu.c
1 /*
2  *  linux/arch/arm/mm/mmu.c
3  *
4  *  Copyright (C) 1995-2005 Russell King
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  */
10 #include <linux/module.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/init.h>
14 #include <linux/bootmem.h>
15 #include <linux/mman.h>
16 #include <linux/nodemask.h>
17
18 #include <asm/cputype.h>
19 #include <asm/mach-types.h>
20 #include <asm/sections.h>
21 #include <asm/cachetype.h>
22 #include <asm/setup.h>
23 #include <asm/sizes.h>
24 #include <asm/tlb.h>
25 #include <asm/highmem.h>
26
27 #include <asm/mach/arch.h>
28 #include <asm/mach/map.h>
29
30 #include "mm.h"
31
32 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
33
34 /*
35  * empty_zero_page is a special page that is used for
36  * zero-initialized data and COW.
37  */
38 struct page *empty_zero_page;
39 EXPORT_SYMBOL(empty_zero_page);
40
41 /*
42  * The pmd table for the upper-most set of pages.
43  */
44 pmd_t *top_pmd;
45
46 #define CPOLICY_UNCACHED        0
47 #define CPOLICY_BUFFERED        1
48 #define CPOLICY_WRITETHROUGH    2
49 #define CPOLICY_WRITEBACK       3
50 #define CPOLICY_WRITEALLOC      4
51
52 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
53 static unsigned int ecc_mask __initdata = 0;
54 pgprot_t pgprot_user;
55 pgprot_t pgprot_kernel;
56
57 EXPORT_SYMBOL(pgprot_user);
58 EXPORT_SYMBOL(pgprot_kernel);
59
60 struct cachepolicy {
61         const char      policy[16];
62         unsigned int    cr_mask;
63         unsigned int    pmd;
64         unsigned int    pte;
65 };
66
67 static struct cachepolicy cache_policies[] __initdata = {
68         {
69                 .policy         = "uncached",
70                 .cr_mask        = CR_W|CR_C,
71                 .pmd            = PMD_SECT_UNCACHED,
72                 .pte            = L_PTE_MT_UNCACHED,
73         }, {
74                 .policy         = "buffered",
75                 .cr_mask        = CR_C,
76                 .pmd            = PMD_SECT_BUFFERED,
77                 .pte            = L_PTE_MT_BUFFERABLE,
78         }, {
79                 .policy         = "writethrough",
80                 .cr_mask        = 0,
81                 .pmd            = PMD_SECT_WT,
82                 .pte            = L_PTE_MT_WRITETHROUGH,
83         }, {
84                 .policy         = "writeback",
85                 .cr_mask        = 0,
86                 .pmd            = PMD_SECT_WB,
87                 .pte            = L_PTE_MT_WRITEBACK,
88         }, {
89                 .policy         = "writealloc",
90                 .cr_mask        = 0,
91                 .pmd            = PMD_SECT_WBWA,
92                 .pte            = L_PTE_MT_WRITEALLOC,
93         }
94 };
95
96 /*
97  * These are useful for identifying cache coherency
98  * problems by allowing the cache or the cache and
99  * writebuffer to be turned off.  (Note: the write
100  * buffer should not be on and the cache off).
101  */
102 static void __init early_cachepolicy(char **p)
103 {
104         int i;
105
106         for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
107                 int len = strlen(cache_policies[i].policy);
108
109                 if (memcmp(*p, cache_policies[i].policy, len) == 0) {
110                         cachepolicy = i;
111                         cr_alignment &= ~cache_policies[i].cr_mask;
112                         cr_no_alignment &= ~cache_policies[i].cr_mask;
113                         *p += len;
114                         break;
115                 }
116         }
117         if (i == ARRAY_SIZE(cache_policies))
118                 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
119         if (cpu_architecture() >= CPU_ARCH_ARMv6) {
120                 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
121                 cachepolicy = CPOLICY_WRITEBACK;
122         }
123         flush_cache_all();
124         set_cr(cr_alignment);
125 }
126 __early_param("cachepolicy=", early_cachepolicy);
127
128 static void __init early_nocache(char **__unused)
129 {
130         char *p = "buffered";
131         printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
132         early_cachepolicy(&p);
133 }
134 __early_param("nocache", early_nocache);
135
136 static void __init early_nowrite(char **__unused)
137 {
138         char *p = "uncached";
139         printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
140         early_cachepolicy(&p);
141 }
142 __early_param("nowb", early_nowrite);
143
144 static void __init early_ecc(char **p)
145 {
146         if (memcmp(*p, "on", 2) == 0) {
147                 ecc_mask = PMD_PROTECTION;
148                 *p += 2;
149         } else if (memcmp(*p, "off", 3) == 0) {
150                 ecc_mask = 0;
151                 *p += 3;
152         }
153 }
154 __early_param("ecc=", early_ecc);
155
156 static int __init noalign_setup(char *__unused)
157 {
158         cr_alignment &= ~CR_A;
159         cr_no_alignment &= ~CR_A;
160         set_cr(cr_alignment);
161         return 1;
162 }
163 __setup("noalign", noalign_setup);
164
165 #ifndef CONFIG_SMP
166 void adjust_cr(unsigned long mask, unsigned long set)
167 {
168         unsigned long flags;
169
170         mask &= ~CR_A;
171
172         set &= mask;
173
174         local_irq_save(flags);
175
176         cr_no_alignment = (cr_no_alignment & ~mask) | set;
177         cr_alignment = (cr_alignment & ~mask) | set;
178
179         set_cr((get_cr() & ~mask) | set);
180
181         local_irq_restore(flags);
182 }
183 #endif
184
185 #define PROT_PTE_DEVICE         L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_WRITE
186 #define PROT_SECT_DEVICE        PMD_TYPE_SECT|PMD_SECT_AP_WRITE
187
188 static struct mem_type mem_types[] = {
189         [MT_DEVICE] = {           /* Strongly ordered / ARMv6 shared device */
190                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
191                                   L_PTE_SHARED,
192                 .prot_l1        = PMD_TYPE_TABLE,
193                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_S,
194                 .domain         = DOMAIN_IO,
195         },
196         [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
197                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
198                 .prot_l1        = PMD_TYPE_TABLE,
199                 .prot_sect      = PROT_SECT_DEVICE,
200                 .domain         = DOMAIN_IO,
201         },
202         [MT_DEVICE_CACHED] = {    /* ioremap_cached */
203                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
204                 .prot_l1        = PMD_TYPE_TABLE,
205                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_WB,
206                 .domain         = DOMAIN_IO,
207         },      
208         [MT_DEVICE_WC] = {      /* ioremap_wc */
209                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
210                 .prot_l1        = PMD_TYPE_TABLE,
211                 .prot_sect      = PROT_SECT_DEVICE,
212                 .domain         = DOMAIN_IO,
213         },
214         [MT_UNCACHED] = {
215                 .prot_pte       = PROT_PTE_DEVICE,
216                 .prot_l1        = PMD_TYPE_TABLE,
217                 .prot_sect      = PMD_TYPE_SECT | PMD_SECT_XN,
218                 .domain         = DOMAIN_IO,
219         },
220         [MT_CACHECLEAN] = {
221                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
222                 .domain    = DOMAIN_KERNEL,
223         },
224         [MT_MINICLEAN] = {
225                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
226                 .domain    = DOMAIN_KERNEL,
227         },
228         [MT_LOW_VECTORS] = {
229                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
230                                 L_PTE_EXEC,
231                 .prot_l1   = PMD_TYPE_TABLE,
232                 .domain    = DOMAIN_USER,
233         },
234         [MT_HIGH_VECTORS] = {
235                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
236                                 L_PTE_USER | L_PTE_EXEC,
237                 .prot_l1   = PMD_TYPE_TABLE,
238                 .domain    = DOMAIN_USER,
239         },
240         [MT_MEMORY] = {
241                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
242                 .domain    = DOMAIN_KERNEL,
243         },
244         [MT_ROM] = {
245                 .prot_sect = PMD_TYPE_SECT,
246                 .domain    = DOMAIN_KERNEL,
247         },
248         [MT_MEMORY_NONCACHED] = {
249                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
250                 .domain    = DOMAIN_KERNEL,
251         },
252 };
253
254 const struct mem_type *get_mem_type(unsigned int type)
255 {
256         return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
257 }
258
259 /*
260  * Adjust the PMD section entries according to the CPU in use.
261  */
262 static void __init build_mem_type_table(void)
263 {
264         struct cachepolicy *cp;
265         unsigned int cr = get_cr();
266         unsigned int user_pgprot, kern_pgprot, vecs_pgprot;
267         int cpu_arch = cpu_architecture();
268         int i;
269
270         if (cpu_arch < CPU_ARCH_ARMv6) {
271 #if defined(CONFIG_CPU_DCACHE_DISABLE)
272                 if (cachepolicy > CPOLICY_BUFFERED)
273                         cachepolicy = CPOLICY_BUFFERED;
274 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
275                 if (cachepolicy > CPOLICY_WRITETHROUGH)
276                         cachepolicy = CPOLICY_WRITETHROUGH;
277 #endif
278         }
279         if (cpu_arch < CPU_ARCH_ARMv5) {
280                 if (cachepolicy >= CPOLICY_WRITEALLOC)
281                         cachepolicy = CPOLICY_WRITEBACK;
282                 ecc_mask = 0;
283         }
284 #ifdef CONFIG_SMP
285         cachepolicy = CPOLICY_WRITEALLOC;
286 #endif
287
288         /*
289          * Strip out features not present on earlier architectures.
290          * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
291          * without extended page tables don't have the 'Shared' bit.
292          */
293         if (cpu_arch < CPU_ARCH_ARMv5)
294                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
295                         mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
296         if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
297                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
298                         mem_types[i].prot_sect &= ~PMD_SECT_S;
299
300         /*
301          * ARMv5 and lower, bit 4 must be set for page tables (was: cache
302          * "update-able on write" bit on ARM610).  However, Xscale and
303          * Xscale3 require this bit to be cleared.
304          */
305         if (cpu_is_xscale() || cpu_is_xsc3()) {
306                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
307                         mem_types[i].prot_sect &= ~PMD_BIT4;
308                         mem_types[i].prot_l1 &= ~PMD_BIT4;
309                 }
310         } else if (cpu_arch < CPU_ARCH_ARMv6) {
311                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
312                         if (mem_types[i].prot_l1)
313                                 mem_types[i].prot_l1 |= PMD_BIT4;
314                         if (mem_types[i].prot_sect)
315                                 mem_types[i].prot_sect |= PMD_BIT4;
316                 }
317         }
318
319         /*
320          * Mark the device areas according to the CPU/architecture.
321          */
322         if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
323                 if (!cpu_is_xsc3()) {
324                         /*
325                          * Mark device regions on ARMv6+ as execute-never
326                          * to prevent speculative instruction fetches.
327                          */
328                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
329                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
330                         mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
331                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
332                 }
333                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
334                         /*
335                          * For ARMv7 with TEX remapping,
336                          * - shared device is SXCB=1100
337                          * - nonshared device is SXCB=0100
338                          * - write combine device mem is SXCB=0001
339                          * (Uncached Normal memory)
340                          */
341                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
342                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
343                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
344                 } else if (cpu_is_xsc3()) {
345                         /*
346                          * For Xscale3,
347                          * - shared device is TEXCB=00101
348                          * - nonshared device is TEXCB=01000
349                          * - write combine device mem is TEXCB=00100
350                          * (Inner/Outer Uncacheable in xsc3 parlance)
351                          */
352                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
353                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
354                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
355                 } else {
356                         /*
357                          * For ARMv6 and ARMv7 without TEX remapping,
358                          * - shared device is TEXCB=00001
359                          * - nonshared device is TEXCB=01000
360                          * - write combine device mem is TEXCB=00100
361                          * (Uncached Normal in ARMv6 parlance).
362                          */
363                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
364                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
365                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
366                 }
367         } else {
368                 /*
369                  * On others, write combining is "Uncached/Buffered"
370                  */
371                 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
372         }
373
374         /*
375          * Now deal with the memory-type mappings
376          */
377         cp = &cache_policies[cachepolicy];
378         vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
379
380 #ifndef CONFIG_SMP
381         /*
382          * Only use write-through for non-SMP systems
383          */
384         if (cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
385                 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
386 #endif
387
388         /*
389          * Enable CPU-specific coherency if supported.
390          * (Only available on XSC3 at the moment.)
391          */
392         if (arch_is_coherent() && cpu_is_xsc3())
393                 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
394
395         /*
396          * ARMv6 and above have extended page tables.
397          */
398         if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
399                 /*
400                  * Mark cache clean areas and XIP ROM read only
401                  * from SVC mode and no access from userspace.
402                  */
403                 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
404                 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
405                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
406
407 #ifdef CONFIG_SMP
408                 /*
409                  * Mark memory with the "shared" attribute for SMP systems
410                  */
411                 user_pgprot |= L_PTE_SHARED;
412                 kern_pgprot |= L_PTE_SHARED;
413                 vecs_pgprot |= L_PTE_SHARED;
414                 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
415                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
416 #endif
417         }
418
419         /*
420          * Non-cacheable Normal - intended for memory areas that must
421          * not cause dirty cache line writebacks when used
422          */
423         if (cpu_arch >= CPU_ARCH_ARMv6) {
424                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
425                         /* Non-cacheable Normal is XCB = 001 */
426                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
427                                 PMD_SECT_BUFFERED;
428                 } else {
429                         /* For both ARMv6 and non-TEX-remapping ARMv7 */
430                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
431                                 PMD_SECT_TEX(1);
432                 }
433         } else {
434                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
435         }
436
437         for (i = 0; i < 16; i++) {
438                 unsigned long v = pgprot_val(protection_map[i]);
439                 protection_map[i] = __pgprot(v | user_pgprot);
440         }
441
442         mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
443         mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
444
445         pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
446         pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
447                                  L_PTE_DIRTY | L_PTE_WRITE |
448                                  L_PTE_EXEC | kern_pgprot);
449
450         mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
451         mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
452         mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
453         mem_types[MT_ROM].prot_sect |= cp->pmd;
454
455         switch (cp->pmd) {
456         case PMD_SECT_WT:
457                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
458                 break;
459         case PMD_SECT_WB:
460         case PMD_SECT_WBWA:
461                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
462                 break;
463         }
464         printk("Memory policy: ECC %sabled, Data cache %s\n",
465                 ecc_mask ? "en" : "dis", cp->policy);
466
467         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
468                 struct mem_type *t = &mem_types[i];
469                 if (t->prot_l1)
470                         t->prot_l1 |= PMD_DOMAIN(t->domain);
471                 if (t->prot_sect)
472                         t->prot_sect |= PMD_DOMAIN(t->domain);
473         }
474 }
475
476 #define vectors_base()  (vectors_high() ? 0xffff0000 : 0)
477
478 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
479                                   unsigned long end, unsigned long pfn,
480                                   const struct mem_type *type)
481 {
482         pte_t *pte;
483
484         if (pmd_none(*pmd)) {
485                 pte = alloc_bootmem_low_pages(2 * PTRS_PER_PTE * sizeof(pte_t));
486                 __pmd_populate(pmd, __pa(pte) | type->prot_l1);
487         }
488
489         pte = pte_offset_kernel(pmd, addr);
490         do {
491                 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
492                 pfn++;
493         } while (pte++, addr += PAGE_SIZE, addr != end);
494 }
495
496 static void __init alloc_init_section(pgd_t *pgd, unsigned long addr,
497                                       unsigned long end, unsigned long phys,
498                                       const struct mem_type *type)
499 {
500         pmd_t *pmd = pmd_offset(pgd, addr);
501
502         /*
503          * Try a section mapping - end, addr and phys must all be aligned
504          * to a section boundary.  Note that PMDs refer to the individual
505          * L1 entries, whereas PGDs refer to a group of L1 entries making
506          * up one logical pointer to an L2 table.
507          */
508         if (((addr | end | phys) & ~SECTION_MASK) == 0) {
509                 pmd_t *p = pmd;
510
511                 if (addr & SECTION_SIZE)
512                         pmd++;
513
514                 do {
515                         *pmd = __pmd(phys | type->prot_sect);
516                         phys += SECTION_SIZE;
517                 } while (pmd++, addr += SECTION_SIZE, addr != end);
518
519                 flush_pmd_entry(p);
520         } else {
521                 /*
522                  * No need to loop; pte's aren't interested in the
523                  * individual L1 entries.
524                  */
525                 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
526         }
527 }
528
529 static void __init create_36bit_mapping(struct map_desc *md,
530                                         const struct mem_type *type)
531 {
532         unsigned long phys, addr, length, end;
533         pgd_t *pgd;
534
535         addr = md->virtual;
536         phys = (unsigned long)__pfn_to_phys(md->pfn);
537         length = PAGE_ALIGN(md->length);
538
539         if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
540                 printk(KERN_ERR "MM: CPU does not support supersection "
541                        "mapping for 0x%08llx at 0x%08lx\n",
542                        __pfn_to_phys((u64)md->pfn), addr);
543                 return;
544         }
545
546         /* N.B. ARMv6 supersections are only defined to work with domain 0.
547          *      Since domain assignments can in fact be arbitrary, the
548          *      'domain == 0' check below is required to insure that ARMv6
549          *      supersections are only allocated for domain 0 regardless
550          *      of the actual domain assignments in use.
551          */
552         if (type->domain) {
553                 printk(KERN_ERR "MM: invalid domain in supersection "
554                        "mapping for 0x%08llx at 0x%08lx\n",
555                        __pfn_to_phys((u64)md->pfn), addr);
556                 return;
557         }
558
559         if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
560                 printk(KERN_ERR "MM: cannot create mapping for "
561                        "0x%08llx at 0x%08lx invalid alignment\n",
562                        __pfn_to_phys((u64)md->pfn), addr);
563                 return;
564         }
565
566         /*
567          * Shift bits [35:32] of address into bits [23:20] of PMD
568          * (See ARMv6 spec).
569          */
570         phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
571
572         pgd = pgd_offset_k(addr);
573         end = addr + length;
574         do {
575                 pmd_t *pmd = pmd_offset(pgd, addr);
576                 int i;
577
578                 for (i = 0; i < 16; i++)
579                         *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
580
581                 addr += SUPERSECTION_SIZE;
582                 phys += SUPERSECTION_SIZE;
583                 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
584         } while (addr != end);
585 }
586
587 /*
588  * Create the page directory entries and any necessary
589  * page tables for the mapping specified by `md'.  We
590  * are able to cope here with varying sizes and address
591  * offsets, and we take full advantage of sections and
592  * supersections.
593  */
594 void __init create_mapping(struct map_desc *md)
595 {
596         unsigned long phys, addr, length, end;
597         const struct mem_type *type;
598         pgd_t *pgd;
599
600         if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
601                 printk(KERN_WARNING "BUG: not creating mapping for "
602                        "0x%08llx at 0x%08lx in user region\n",
603                        __pfn_to_phys((u64)md->pfn), md->virtual);
604                 return;
605         }
606
607         if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
608             md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
609                 printk(KERN_WARNING "BUG: mapping for 0x%08llx at 0x%08lx "
610                        "overlaps vmalloc space\n",
611                        __pfn_to_phys((u64)md->pfn), md->virtual);
612         }
613
614         type = &mem_types[md->type];
615
616         /*
617          * Catch 36-bit addresses
618          */
619         if (md->pfn >= 0x100000) {
620                 create_36bit_mapping(md, type);
621                 return;
622         }
623
624         addr = md->virtual & PAGE_MASK;
625         phys = (unsigned long)__pfn_to_phys(md->pfn);
626         length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
627
628         if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
629                 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
630                        "be mapped using pages, ignoring.\n",
631                        __pfn_to_phys(md->pfn), addr);
632                 return;
633         }
634
635         pgd = pgd_offset_k(addr);
636         end = addr + length;
637         do {
638                 unsigned long next = pgd_addr_end(addr, end);
639
640                 alloc_init_section(pgd, addr, next, phys, type);
641
642                 phys += next - addr;
643                 addr = next;
644         } while (pgd++, addr != end);
645 }
646
647 /*
648  * Create the architecture specific mappings
649  */
650 void __init iotable_init(struct map_desc *io_desc, int nr)
651 {
652         int i;
653
654         for (i = 0; i < nr; i++)
655                 create_mapping(io_desc + i);
656 }
657
658 static unsigned long __initdata vmalloc_reserve = SZ_128M;
659
660 /*
661  * vmalloc=size forces the vmalloc area to be exactly 'size'
662  * bytes. This can be used to increase (or decrease) the vmalloc
663  * area - the default is 128m.
664  */
665 static void __init early_vmalloc(char **arg)
666 {
667         vmalloc_reserve = memparse(*arg, arg);
668
669         if (vmalloc_reserve < SZ_16M) {
670                 vmalloc_reserve = SZ_16M;
671                 printk(KERN_WARNING
672                         "vmalloc area too small, limiting to %luMB\n",
673                         vmalloc_reserve >> 20);
674         }
675
676         if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
677                 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
678                 printk(KERN_WARNING
679                         "vmalloc area is too big, limiting to %luMB\n",
680                         vmalloc_reserve >> 20);
681         }
682 }
683 __early_param("vmalloc=", early_vmalloc);
684
685 #define VMALLOC_MIN     (void *)(VMALLOC_END - vmalloc_reserve)
686
687 static void __init sanity_check_meminfo(void)
688 {
689         int i, j;
690
691         for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
692                 struct membank *bank = &meminfo.bank[j];
693                 *bank = meminfo.bank[i];
694
695 #ifdef CONFIG_HIGHMEM
696                 /*
697                  * Split those memory banks which are partially overlapping
698                  * the vmalloc area greatly simplifying things later.
699                  */
700                 if (__va(bank->start) < VMALLOC_MIN &&
701                     bank->size > VMALLOC_MIN - __va(bank->start)) {
702                         if (meminfo.nr_banks >= NR_BANKS) {
703                                 printk(KERN_CRIT "NR_BANKS too low, "
704                                                  "ignoring high memory\n");
705                         } else if (cache_is_vipt_aliasing()) {
706                                 printk(KERN_CRIT "HIGHMEM is not yet supported "
707                                                  "with VIPT aliasing cache, "
708                                                  "ignoring high memory\n");
709                         } else {
710                                 memmove(bank + 1, bank,
711                                         (meminfo.nr_banks - i) * sizeof(*bank));
712                                 meminfo.nr_banks++;
713                                 i++;
714                                 bank[1].size -= VMALLOC_MIN - __va(bank->start);
715                                 bank[1].start = __pa(VMALLOC_MIN - 1) + 1;
716                                 j++;
717                         }
718                         bank->size = VMALLOC_MIN - __va(bank->start);
719                 }
720 #else
721                 /*
722                  * Check whether this memory bank would entirely overlap
723                  * the vmalloc area.
724                  */
725                 if (__va(bank->start) >= VMALLOC_MIN ||
726                     __va(bank->start) < (void *)PAGE_OFFSET) {
727                         printk(KERN_NOTICE "Ignoring RAM at %.8lx-%.8lx "
728                                "(vmalloc region overlap).\n",
729                                bank->start, bank->start + bank->size - 1);
730                         continue;
731                 }
732
733                 /*
734                  * Check whether this memory bank would partially overlap
735                  * the vmalloc area.
736                  */
737                 if (__va(bank->start + bank->size) > VMALLOC_MIN ||
738                     __va(bank->start + bank->size) < __va(bank->start)) {
739                         unsigned long newsize = VMALLOC_MIN - __va(bank->start);
740                         printk(KERN_NOTICE "Truncating RAM at %.8lx-%.8lx "
741                                "to -%.8lx (vmalloc region overlap).\n",
742                                bank->start, bank->start + bank->size - 1,
743                                bank->start + newsize - 1);
744                         bank->size = newsize;
745                 }
746 #endif
747                 j++;
748         }
749         meminfo.nr_banks = j;
750 }
751
752 static inline void prepare_page_table(void)
753 {
754         unsigned long addr;
755
756         /*
757          * Clear out all the mappings below the kernel image.
758          */
759         for (addr = 0; addr < MODULES_VADDR; addr += PGDIR_SIZE)
760                 pmd_clear(pmd_off_k(addr));
761
762 #ifdef CONFIG_XIP_KERNEL
763         /* The XIP kernel is mapped in the module area -- skip over it */
764         addr = ((unsigned long)_etext + PGDIR_SIZE - 1) & PGDIR_MASK;
765 #endif
766         for ( ; addr < PAGE_OFFSET; addr += PGDIR_SIZE)
767                 pmd_clear(pmd_off_k(addr));
768
769         /*
770          * Clear out all the kernel space mappings, except for the first
771          * memory bank, up to the end of the vmalloc region.
772          */
773         for (addr = __phys_to_virt(bank_phys_end(&meminfo.bank[0]));
774              addr < VMALLOC_END; addr += PGDIR_SIZE)
775                 pmd_clear(pmd_off_k(addr));
776 }
777
778 /*
779  * Reserve the various regions of node 0
780  */
781 void __init reserve_node_zero(pg_data_t *pgdat)
782 {
783         unsigned long res_size = 0;
784
785         /*
786          * Register the kernel text and data with bootmem.
787          * Note that this can only be in node 0.
788          */
789 #ifdef CONFIG_XIP_KERNEL
790         reserve_bootmem_node(pgdat, __pa(_data), _end - _data,
791                         BOOTMEM_DEFAULT);
792 #else
793         reserve_bootmem_node(pgdat, __pa(_stext), _end - _stext,
794                         BOOTMEM_DEFAULT);
795 #endif
796
797         /*
798          * Reserve the page tables.  These are already in use,
799          * and can only be in node 0.
800          */
801         reserve_bootmem_node(pgdat, __pa(swapper_pg_dir),
802                              PTRS_PER_PGD * sizeof(pgd_t), BOOTMEM_DEFAULT);
803
804         /*
805          * Hmm... This should go elsewhere, but we really really need to
806          * stop things allocating the low memory; ideally we need a better
807          * implementation of GFP_DMA which does not assume that DMA-able
808          * memory starts at zero.
809          */
810         if (machine_is_integrator() || machine_is_cintegrator())
811                 res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
812
813         /*
814          * These should likewise go elsewhere.  They pre-reserve the
815          * screen memory region at the start of main system memory.
816          */
817         if (machine_is_edb7211())
818                 res_size = 0x00020000;
819         if (machine_is_p720t())
820                 res_size = 0x00014000;
821
822         /* H1940 and RX3715 need to reserve this for suspend */
823
824         if (machine_is_h1940() || machine_is_rx3715()) {
825                 reserve_bootmem_node(pgdat, 0x30003000, 0x1000,
826                                 BOOTMEM_DEFAULT);
827                 reserve_bootmem_node(pgdat, 0x30081000, 0x1000,
828                                 BOOTMEM_DEFAULT);
829         }
830
831 #ifdef CONFIG_SA1111
832         /*
833          * Because of the SA1111 DMA bug, we want to preserve our
834          * precious DMA-able memory...
835          */
836         res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
837 #endif
838         if (res_size)
839                 reserve_bootmem_node(pgdat, PHYS_OFFSET, res_size,
840                                 BOOTMEM_DEFAULT);
841 }
842
843 /*
844  * Set up device the mappings.  Since we clear out the page tables for all
845  * mappings above VMALLOC_END, we will remove any debug device mappings.
846  * This means you have to be careful how you debug this function, or any
847  * called function.  This means you can't use any function or debugging
848  * method which may touch any device, otherwise the kernel _will_ crash.
849  */
850 static void __init devicemaps_init(struct machine_desc *mdesc)
851 {
852         struct map_desc map;
853         unsigned long addr;
854         void *vectors;
855
856         /*
857          * Allocate the vector page early.
858          */
859         vectors = alloc_bootmem_low_pages(PAGE_SIZE);
860
861         for (addr = VMALLOC_END; addr; addr += PGDIR_SIZE)
862                 pmd_clear(pmd_off_k(addr));
863
864         /*
865          * Map the kernel if it is XIP.
866          * It is always first in the modulearea.
867          */
868 #ifdef CONFIG_XIP_KERNEL
869         map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
870         map.virtual = MODULES_VADDR;
871         map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
872         map.type = MT_ROM;
873         create_mapping(&map);
874 #endif
875
876         /*
877          * Map the cache flushing regions.
878          */
879 #ifdef FLUSH_BASE
880         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
881         map.virtual = FLUSH_BASE;
882         map.length = SZ_1M;
883         map.type = MT_CACHECLEAN;
884         create_mapping(&map);
885 #endif
886 #ifdef FLUSH_BASE_MINICACHE
887         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
888         map.virtual = FLUSH_BASE_MINICACHE;
889         map.length = SZ_1M;
890         map.type = MT_MINICLEAN;
891         create_mapping(&map);
892 #endif
893
894         /*
895          * Create a mapping for the machine vectors at the high-vectors
896          * location (0xffff0000).  If we aren't using high-vectors, also
897          * create a mapping at the low-vectors virtual address.
898          */
899         map.pfn = __phys_to_pfn(virt_to_phys(vectors));
900         map.virtual = 0xffff0000;
901         map.length = PAGE_SIZE;
902         map.type = MT_HIGH_VECTORS;
903         create_mapping(&map);
904
905         if (!vectors_high()) {
906                 map.virtual = 0;
907                 map.type = MT_LOW_VECTORS;
908                 create_mapping(&map);
909         }
910
911         /*
912          * Ask the machine support to map in the statically mapped devices.
913          */
914         if (mdesc->map_io)
915                 mdesc->map_io();
916
917         /*
918          * Finally flush the caches and tlb to ensure that we're in a
919          * consistent state wrt the writebuffer.  This also ensures that
920          * any write-allocated cache lines in the vector page are written
921          * back.  After this point, we can start to touch devices again.
922          */
923         local_flush_tlb_all();
924         flush_cache_all();
925 }
926
927 static void __init kmap_init(void)
928 {
929 #ifdef CONFIG_HIGHMEM
930         pmd_t *pmd = pmd_off_k(PKMAP_BASE);
931         pte_t *pte = alloc_bootmem_low_pages(2 * PTRS_PER_PTE * sizeof(pte_t));
932         BUG_ON(!pmd_none(*pmd) || !pte);
933         __pmd_populate(pmd, __pa(pte) | _PAGE_KERNEL_TABLE);
934         pkmap_page_table = pte + PTRS_PER_PTE;
935 #endif
936 }
937
938 /*
939  * paging_init() sets up the page tables, initialises the zone memory
940  * maps, and sets up the zero page, bad page and bad page tables.
941  */
942 void __init paging_init(struct machine_desc *mdesc)
943 {
944         void *zero_page;
945
946         build_mem_type_table();
947         sanity_check_meminfo();
948         prepare_page_table();
949         bootmem_init();
950         devicemaps_init(mdesc);
951         kmap_init();
952
953         top_pmd = pmd_off_k(0xffff0000);
954
955         /*
956          * allocate the zero page.  Note that this always succeeds and
957          * returns a zeroed result.
958          */
959         zero_page = alloc_bootmem_low_pages(PAGE_SIZE);
960         empty_zero_page = virt_to_page(zero_page);
961         flush_dcache_page(empty_zero_page);
962 }
963
964 /*
965  * In order to soft-boot, we need to insert a 1:1 mapping in place of
966  * the user-mode pages.  This will then ensure that we have predictable
967  * results when turning the mmu off
968  */
969 void setup_mm_for_reboot(char mode)
970 {
971         unsigned long base_pmdval;
972         pgd_t *pgd;
973         int i;
974
975         if (current->mm && current->mm->pgd)
976                 pgd = current->mm->pgd;
977         else
978                 pgd = init_mm.pgd;
979
980         base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
981         if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ && !cpu_is_xscale())
982                 base_pmdval |= PMD_BIT4;
983
984         for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
985                 unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
986                 pmd_t *pmd;
987
988                 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
989                 pmd[0] = __pmd(pmdval);
990                 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
991                 flush_pmd_entry(pmd);
992         }
993 }