Merge branch 'stable-3.2' into pandora-3.2
[pandora-kernel.git] / 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/mman.h>
15 #include <linux/nodemask.h>
16 #include <linux/memblock.h>
17 #include <linux/fs.h>
18 #include <linux/vmalloc.h>
19
20 #include <asm/cputype.h>
21 #include <asm/sections.h>
22 #include <asm/cachetype.h>
23 #include <asm/setup.h>
24 #include <asm/sizes.h>
25 #include <asm/smp_plat.h>
26 #include <asm/tlb.h>
27 #include <asm/highmem.h>
28 #include <asm/traps.h>
29
30 #include <asm/mach/arch.h>
31 #include <asm/mach/map.h>
32
33 #include "mm.h"
34
35 /*
36  * empty_zero_page is a special page that is used for
37  * zero-initialized data and COW.
38  */
39 struct page *empty_zero_page;
40 EXPORT_SYMBOL(empty_zero_page);
41
42 /*
43  * The pmd table for the upper-most set of pages.
44  */
45 pmd_t *top_pmd;
46
47 #define CPOLICY_UNCACHED        0
48 #define CPOLICY_BUFFERED        1
49 #define CPOLICY_WRITETHROUGH    2
50 #define CPOLICY_WRITEBACK       3
51 #define CPOLICY_WRITEALLOC      4
52
53 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
54 static unsigned int ecc_mask __initdata = 0;
55 pgprot_t pgprot_user;
56 pgprot_t pgprot_kernel;
57
58 EXPORT_SYMBOL(pgprot_user);
59 EXPORT_SYMBOL(pgprot_kernel);
60
61 struct cachepolicy {
62         const char      policy[16];
63         unsigned int    cr_mask;
64         pmdval_t        pmd;
65         pteval_t        pte;
66 };
67
68 static struct cachepolicy cache_policies[] __initdata = {
69         {
70                 .policy         = "uncached",
71                 .cr_mask        = CR_W|CR_C,
72                 .pmd            = PMD_SECT_UNCACHED,
73                 .pte            = L_PTE_MT_UNCACHED,
74         }, {
75                 .policy         = "buffered",
76                 .cr_mask        = CR_C,
77                 .pmd            = PMD_SECT_BUFFERED,
78                 .pte            = L_PTE_MT_BUFFERABLE,
79         }, {
80                 .policy         = "writethrough",
81                 .cr_mask        = 0,
82                 .pmd            = PMD_SECT_WT,
83                 .pte            = L_PTE_MT_WRITETHROUGH,
84         }, {
85                 .policy         = "writeback",
86                 .cr_mask        = 0,
87                 .pmd            = PMD_SECT_WB,
88                 .pte            = L_PTE_MT_WRITEBACK,
89         }, {
90                 .policy         = "writealloc",
91                 .cr_mask        = 0,
92                 .pmd            = PMD_SECT_WBWA,
93                 .pte            = L_PTE_MT_WRITEALLOC,
94         }
95 };
96
97 /*
98  * These are useful for identifying cache coherency
99  * problems by allowing the cache or the cache and
100  * writebuffer to be turned off.  (Note: the write
101  * buffer should not be on and the cache off).
102  */
103 static int __init early_cachepolicy(char *p)
104 {
105         int i;
106
107         for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
108                 int len = strlen(cache_policies[i].policy);
109
110                 if (memcmp(p, cache_policies[i].policy, len) == 0) {
111                         cachepolicy = i;
112                         cr_alignment &= ~cache_policies[i].cr_mask;
113                         cr_no_alignment &= ~cache_policies[i].cr_mask;
114                         break;
115                 }
116         }
117         if (i == ARRAY_SIZE(cache_policies))
118                 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
119         /*
120          * This restriction is partly to do with the way we boot; it is
121          * unpredictable to have memory mapped using two different sets of
122          * memory attributes (shared, type, and cache attribs).  We can not
123          * change these attributes once the initial assembly has setup the
124          * page tables.
125          */
126         if (cpu_architecture() >= CPU_ARCH_ARMv6) {
127                 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
128                 cachepolicy = CPOLICY_WRITEBACK;
129         }
130         flush_cache_all();
131         set_cr(cr_alignment);
132         return 0;
133 }
134 early_param("cachepolicy", early_cachepolicy);
135
136 static int __init early_nocache(char *__unused)
137 {
138         char *p = "buffered";
139         printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
140         early_cachepolicy(p);
141         return 0;
142 }
143 early_param("nocache", early_nocache);
144
145 static int __init early_nowrite(char *__unused)
146 {
147         char *p = "uncached";
148         printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
149         early_cachepolicy(p);
150         return 0;
151 }
152 early_param("nowb", early_nowrite);
153
154 #ifndef CONFIG_ARM_LPAE
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 #endif
165
166 static int __init noalign_setup(char *__unused)
167 {
168         cr_alignment &= ~CR_A;
169         cr_no_alignment &= ~CR_A;
170         set_cr(cr_alignment);
171         return 1;
172 }
173 __setup("noalign", noalign_setup);
174
175 #ifndef CONFIG_SMP
176 void adjust_cr(unsigned long mask, unsigned long set)
177 {
178         unsigned long flags;
179
180         mask &= ~CR_A;
181
182         set &= mask;
183
184         local_irq_save(flags);
185
186         cr_no_alignment = (cr_no_alignment & ~mask) | set;
187         cr_alignment = (cr_alignment & ~mask) | set;
188
189         set_cr((get_cr() & ~mask) | set);
190
191         local_irq_restore(flags);
192 }
193 #endif
194
195 #define PROT_PTE_DEVICE         L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
196 #define PROT_SECT_DEVICE        PMD_TYPE_SECT|PMD_SECT_AP_WRITE
197
198 static struct mem_type mem_types[] = {
199         [MT_DEVICE] = {           /* Strongly ordered / ARMv6 shared device */
200                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
201                                   L_PTE_SHARED,
202                 .prot_l1        = PMD_TYPE_TABLE,
203                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_S,
204                 .domain         = DOMAIN_IO,
205         },
206         [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
207                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
208                 .prot_l1        = PMD_TYPE_TABLE,
209                 .prot_sect      = PROT_SECT_DEVICE,
210                 .domain         = DOMAIN_IO,
211         },
212         [MT_DEVICE_CACHED] = {    /* ioremap_cached */
213                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
214                 .prot_l1        = PMD_TYPE_TABLE,
215                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_WB,
216                 .domain         = DOMAIN_IO,
217         },      
218         [MT_DEVICE_WC] = {      /* ioremap_wc */
219                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
220                 .prot_l1        = PMD_TYPE_TABLE,
221                 .prot_sect      = PROT_SECT_DEVICE,
222                 .domain         = DOMAIN_IO,
223         },
224         [MT_UNCACHED] = {
225                 .prot_pte       = PROT_PTE_DEVICE,
226                 .prot_l1        = PMD_TYPE_TABLE,
227                 .prot_sect      = PMD_TYPE_SECT | PMD_SECT_XN,
228                 .domain         = DOMAIN_IO,
229         },
230         [MT_CACHECLEAN] = {
231                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
232                 .domain    = DOMAIN_KERNEL,
233         },
234 #ifndef CONFIG_ARM_LPAE
235         [MT_MINICLEAN] = {
236                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
237                 .domain    = DOMAIN_KERNEL,
238         },
239 #endif
240         [MT_LOW_VECTORS] = {
241                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
242                                 L_PTE_RDONLY,
243                 .prot_l1   = PMD_TYPE_TABLE,
244                 .domain    = DOMAIN_USER,
245         },
246         [MT_HIGH_VECTORS] = {
247                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
248                                 L_PTE_USER | L_PTE_RDONLY,
249                 .prot_l1   = PMD_TYPE_TABLE,
250                 .domain    = DOMAIN_USER,
251         },
252         [MT_MEMORY] = {
253                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
254                 .prot_l1   = PMD_TYPE_TABLE,
255                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
256                 .domain    = DOMAIN_KERNEL,
257         },
258         [MT_ROM] = {
259                 .prot_sect = PMD_TYPE_SECT,
260                 .domain    = DOMAIN_KERNEL,
261         },
262         [MT_MEMORY_NONCACHED] = {
263                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
264                                 L_PTE_MT_BUFFERABLE,
265                 .prot_l1   = PMD_TYPE_TABLE,
266                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
267                 .domain    = DOMAIN_KERNEL,
268         },
269         [MT_MEMORY_DTCM] = {
270                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
271                                 L_PTE_XN,
272                 .prot_l1   = PMD_TYPE_TABLE,
273                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
274                 .domain    = DOMAIN_KERNEL,
275         },
276         [MT_MEMORY_ITCM] = {
277                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
278                 .prot_l1   = PMD_TYPE_TABLE,
279                 .domain    = DOMAIN_KERNEL,
280         },
281         [MT_MEMORY_SO] = {
282                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
283                                 L_PTE_MT_UNCACHED,
284                 .prot_l1   = PMD_TYPE_TABLE,
285                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
286                                 PMD_SECT_UNCACHED | PMD_SECT_XN,
287                 .domain    = DOMAIN_KERNEL,
288         },
289         [MT_MEMORY_DMA_READY] = {
290                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
291                 .prot_l1   = PMD_TYPE_TABLE,
292                 .domain    = DOMAIN_KERNEL,
293         },
294 };
295
296 const struct mem_type *get_mem_type(unsigned int type)
297 {
298         return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
299 }
300 EXPORT_SYMBOL(get_mem_type);
301
302 /*
303  * Adjust the PMD section entries according to the CPU in use.
304  */
305 static void __init build_mem_type_table(void)
306 {
307         struct cachepolicy *cp;
308         unsigned int cr = get_cr();
309         pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
310         int cpu_arch = cpu_architecture();
311         int i;
312
313         if (cpu_arch < CPU_ARCH_ARMv6) {
314 #if defined(CONFIG_CPU_DCACHE_DISABLE)
315                 if (cachepolicy > CPOLICY_BUFFERED)
316                         cachepolicy = CPOLICY_BUFFERED;
317 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
318                 if (cachepolicy > CPOLICY_WRITETHROUGH)
319                         cachepolicy = CPOLICY_WRITETHROUGH;
320 #endif
321         }
322         if (cpu_arch < CPU_ARCH_ARMv5) {
323                 if (cachepolicy >= CPOLICY_WRITEALLOC)
324                         cachepolicy = CPOLICY_WRITEBACK;
325                 ecc_mask = 0;
326         }
327         if (is_smp())
328                 cachepolicy = CPOLICY_WRITEALLOC;
329
330         /*
331          * Strip out features not present on earlier architectures.
332          * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
333          * without extended page tables don't have the 'Shared' bit.
334          */
335         if (cpu_arch < CPU_ARCH_ARMv5)
336                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
337                         mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
338         if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
339                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
340                         mem_types[i].prot_sect &= ~PMD_SECT_S;
341
342         /*
343          * ARMv5 and lower, bit 4 must be set for page tables (was: cache
344          * "update-able on write" bit on ARM610).  However, Xscale and
345          * Xscale3 require this bit to be cleared.
346          */
347         if (cpu_is_xscale() || cpu_is_xsc3()) {
348                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
349                         mem_types[i].prot_sect &= ~PMD_BIT4;
350                         mem_types[i].prot_l1 &= ~PMD_BIT4;
351                 }
352         } else if (cpu_arch < CPU_ARCH_ARMv6) {
353                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
354                         if (mem_types[i].prot_l1)
355                                 mem_types[i].prot_l1 |= PMD_BIT4;
356                         if (mem_types[i].prot_sect)
357                                 mem_types[i].prot_sect |= PMD_BIT4;
358                 }
359         }
360
361         /*
362          * Mark the device areas according to the CPU/architecture.
363          */
364         if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
365                 if (!cpu_is_xsc3()) {
366                         /*
367                          * Mark device regions on ARMv6+ as execute-never
368                          * to prevent speculative instruction fetches.
369                          */
370                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
371                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
372                         mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
373                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
374                 }
375                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
376                         /*
377                          * For ARMv7 with TEX remapping,
378                          * - shared device is SXCB=1100
379                          * - nonshared device is SXCB=0100
380                          * - write combine device mem is SXCB=0001
381                          * (Uncached Normal memory)
382                          */
383                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
384                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
385                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
386                 } else if (cpu_is_xsc3()) {
387                         /*
388                          * For Xscale3,
389                          * - shared device is TEXCB=00101
390                          * - nonshared device is TEXCB=01000
391                          * - write combine device mem is TEXCB=00100
392                          * (Inner/Outer Uncacheable in xsc3 parlance)
393                          */
394                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
395                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
396                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
397                 } else {
398                         /*
399                          * For ARMv6 and ARMv7 without TEX remapping,
400                          * - shared device is TEXCB=00001
401                          * - nonshared device is TEXCB=01000
402                          * - write combine device mem is TEXCB=00100
403                          * (Uncached Normal in ARMv6 parlance).
404                          */
405                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
406                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
407                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
408                 }
409         } else {
410                 /*
411                  * On others, write combining is "Uncached/Buffered"
412                  */
413                 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
414         }
415
416         /*
417          * Now deal with the memory-type mappings
418          */
419         cp = &cache_policies[cachepolicy];
420         vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
421
422         /*
423          * Only use write-through for non-SMP systems
424          */
425         if (!is_smp() && cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
426                 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
427
428         /*
429          * Enable CPU-specific coherency if supported.
430          * (Only available on XSC3 at the moment.)
431          */
432         if (arch_is_coherent() && cpu_is_xsc3()) {
433                 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
434                 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
435                 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
436                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
437                 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
438         }
439         /*
440          * We don't use domains on ARMv6 (since this causes problems with
441          * v6/v7 kernels), so we must use a separate memory type for user
442          * r/o, kernel r/w to map the vectors page.
443          */
444         if (cpu_arch == CPU_ARCH_ARMv6)
445                 vecs_pgprot |= L_PTE_MT_VECTORS;
446
447         /*
448          * ARMv6 and above have extended page tables.
449          */
450         if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
451 #ifndef CONFIG_ARM_LPAE
452                 /*
453                  * Mark cache clean areas and XIP ROM read only
454                  * from SVC mode and no access from userspace.
455                  */
456                 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
457                 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
458                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
459 #endif
460
461                 if (is_smp()) {
462                         /*
463                          * Mark memory with the "shared" attribute
464                          * for SMP systems
465                          */
466                         user_pgprot |= L_PTE_SHARED;
467                         kern_pgprot |= L_PTE_SHARED;
468                         vecs_pgprot |= L_PTE_SHARED;
469                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
470                         mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
471                         mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
472                         mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
473                         mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
474                         mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
475                         mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
476                         mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
477                         mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
478                 }
479         }
480
481         /*
482          * Non-cacheable Normal - intended for memory areas that must
483          * not cause dirty cache line writebacks when used
484          */
485         if (cpu_arch >= CPU_ARCH_ARMv6) {
486                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
487                         /* Non-cacheable Normal is XCB = 001 */
488                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
489                                 PMD_SECT_BUFFERED;
490                 } else {
491                         /* For both ARMv6 and non-TEX-remapping ARMv7 */
492                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
493                                 PMD_SECT_TEX(1);
494                 }
495         } else {
496                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
497         }
498
499 #ifdef CONFIG_ARM_LPAE
500         /*
501          * Do not generate access flag faults for the kernel mappings.
502          */
503         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
504                 mem_types[i].prot_pte |= PTE_EXT_AF;
505                 mem_types[i].prot_sect |= PMD_SECT_AF;
506         }
507         kern_pgprot |= PTE_EXT_AF;
508         vecs_pgprot |= PTE_EXT_AF;
509 #endif
510
511         for (i = 0; i < 16; i++) {
512                 pteval_t v = pgprot_val(protection_map[i]);
513                 protection_map[i] = __pgprot(v | user_pgprot);
514         }
515
516         mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
517         mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
518
519         pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
520         pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
521                                  L_PTE_DIRTY | kern_pgprot);
522
523         mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
524         mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
525         mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
526         mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
527         mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
528         mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
529         mem_types[MT_ROM].prot_sect |= cp->pmd;
530
531         switch (cp->pmd) {
532         case PMD_SECT_WT:
533                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
534                 break;
535         case PMD_SECT_WB:
536         case PMD_SECT_WBWA:
537                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
538                 break;
539         }
540         printk("Memory policy: ECC %sabled, Data cache %s\n",
541                 ecc_mask ? "en" : "dis", cp->policy);
542
543         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
544                 struct mem_type *t = &mem_types[i];
545                 if (t->prot_l1)
546                         t->prot_l1 |= PMD_DOMAIN(t->domain);
547                 if (t->prot_sect)
548                         t->prot_sect |= PMD_DOMAIN(t->domain);
549         }
550 }
551
552 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
553 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
554                               unsigned long size, pgprot_t vma_prot)
555 {
556         if (!pfn_valid(pfn))
557                 return pgprot_noncached(vma_prot);
558         else if (file->f_flags & O_SYNC)
559                 return pgprot_writecombine(vma_prot);
560         return vma_prot;
561 }
562 EXPORT_SYMBOL(phys_mem_access_prot);
563 #endif
564
565 #define vectors_base()  (vectors_high() ? 0xffff0000 : 0)
566
567 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
568 {
569         void *ptr = __va(memblock_alloc(sz, align));
570         memset(ptr, 0, sz);
571         return ptr;
572 }
573
574 static void __init *early_alloc(unsigned long sz)
575 {
576         return early_alloc_aligned(sz, sz);
577 }
578
579 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
580 {
581         if (pmd_none(*pmd)) {
582                 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
583                 __pmd_populate(pmd, __pa(pte), prot);
584         }
585         BUG_ON(pmd_bad(*pmd));
586         return pte_offset_kernel(pmd, addr);
587 }
588
589 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
590                                   unsigned long end, unsigned long pfn,
591                                   const struct mem_type *type)
592 {
593         pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
594         do {
595                 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
596                 pfn++;
597         } while (pte++, addr += PAGE_SIZE, addr != end);
598 }
599
600 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
601                                       unsigned long end, phys_addr_t phys,
602                                       const struct mem_type *type)
603 {
604         pmd_t *pmd = pmd_offset(pud, addr);
605
606         /*
607          * Try a section mapping - end, addr and phys must all be aligned
608          * to a section boundary.  Note that PMDs refer to the individual
609          * L1 entries, whereas PGDs refer to a group of L1 entries making
610          * up one logical pointer to an L2 table.
611          */
612         if (type->prot_sect && ((addr | end | phys) & ~SECTION_MASK) == 0) {
613                 pmd_t *p = pmd;
614
615 #ifndef CONFIG_ARM_LPAE
616                 if (addr & SECTION_SIZE)
617                         pmd++;
618 #endif
619
620                 do {
621                         *pmd = __pmd(phys | type->prot_sect);
622                         phys += SECTION_SIZE;
623                 } while (pmd++, addr += SECTION_SIZE, addr != end);
624
625                 flush_pmd_entry(p);
626         } else {
627                 /*
628                  * No need to loop; pte's aren't interested in the
629                  * individual L1 entries.
630                  */
631                 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
632         }
633 }
634
635 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
636         unsigned long phys, const struct mem_type *type)
637 {
638         pud_t *pud = pud_offset(pgd, addr);
639         unsigned long next;
640
641         do {
642                 next = pud_addr_end(addr, end);
643                 alloc_init_section(pud, addr, next, phys, type);
644                 phys += next - addr;
645         } while (pud++, addr = next, addr != end);
646 }
647
648 #ifndef CONFIG_ARM_LPAE
649 static void __init create_36bit_mapping(struct map_desc *md,
650                                         const struct mem_type *type)
651 {
652         unsigned long addr, length, end;
653         phys_addr_t phys;
654         pgd_t *pgd;
655
656         addr = md->virtual;
657         phys = __pfn_to_phys(md->pfn);
658         length = PAGE_ALIGN(md->length);
659
660         if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
661                 printk(KERN_ERR "MM: CPU does not support supersection "
662                        "mapping for 0x%08llx at 0x%08lx\n",
663                        (long long)__pfn_to_phys((u64)md->pfn), addr);
664                 return;
665         }
666
667         /* N.B. ARMv6 supersections are only defined to work with domain 0.
668          *      Since domain assignments can in fact be arbitrary, the
669          *      'domain == 0' check below is required to insure that ARMv6
670          *      supersections are only allocated for domain 0 regardless
671          *      of the actual domain assignments in use.
672          */
673         if (type->domain) {
674                 printk(KERN_ERR "MM: invalid domain in supersection "
675                        "mapping for 0x%08llx at 0x%08lx\n",
676                        (long long)__pfn_to_phys((u64)md->pfn), addr);
677                 return;
678         }
679
680         if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
681                 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
682                        " at 0x%08lx invalid alignment\n",
683                        (long long)__pfn_to_phys((u64)md->pfn), addr);
684                 return;
685         }
686
687         /*
688          * Shift bits [35:32] of address into bits [23:20] of PMD
689          * (See ARMv6 spec).
690          */
691         phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
692
693         pgd = pgd_offset_k(addr);
694         end = addr + length;
695         do {
696                 pud_t *pud = pud_offset(pgd, addr);
697                 pmd_t *pmd = pmd_offset(pud, addr);
698                 int i;
699
700                 for (i = 0; i < 16; i++)
701                         *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
702
703                 addr += SUPERSECTION_SIZE;
704                 phys += SUPERSECTION_SIZE;
705                 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
706         } while (addr != end);
707 }
708 #endif  /* !CONFIG_ARM_LPAE */
709
710 /*
711  * Create the page directory entries and any necessary
712  * page tables for the mapping specified by `md'.  We
713  * are able to cope here with varying sizes and address
714  * offsets, and we take full advantage of sections and
715  * supersections.
716  */
717 static void __init create_mapping(struct map_desc *md)
718 {
719         unsigned long addr, length, end;
720         phys_addr_t phys;
721         const struct mem_type *type;
722         pgd_t *pgd;
723
724         if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
725                 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
726                        " at 0x%08lx in user region\n",
727                        (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
728                 return;
729         }
730
731         if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
732             md->virtual >= PAGE_OFFSET &&
733             (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
734                 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
735                        " at 0x%08lx out of vmalloc space\n",
736                        (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
737         }
738
739         type = &mem_types[md->type];
740
741 #ifndef CONFIG_ARM_LPAE
742         /*
743          * Catch 36-bit addresses
744          */
745         if (md->pfn >= 0x100000) {
746                 create_36bit_mapping(md, type);
747                 return;
748         }
749 #endif
750
751         addr = md->virtual & PAGE_MASK;
752         phys = __pfn_to_phys(md->pfn);
753         length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
754
755         if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
756                 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
757                        "be mapped using pages, ignoring.\n",
758                        (long long)__pfn_to_phys(md->pfn), addr);
759                 return;
760         }
761
762         pgd = pgd_offset_k(addr);
763         end = addr + length;
764         do {
765                 unsigned long next = pgd_addr_end(addr, end);
766
767                 alloc_init_pud(pgd, addr, next, phys, type);
768
769                 phys += next - addr;
770                 addr = next;
771         } while (pgd++, addr != end);
772 }
773
774 /*
775  * Create the architecture specific mappings
776  */
777 void __init iotable_init(struct map_desc *io_desc, int nr)
778 {
779         struct map_desc *md;
780         struct vm_struct *vm;
781
782         if (!nr)
783                 return;
784
785         vm = early_alloc_aligned(sizeof(*vm) * nr, __alignof__(*vm));
786
787         for (md = io_desc; nr; md++, nr--) {
788                 create_mapping(md);
789                 vm->addr = (void *)(md->virtual & PAGE_MASK);
790                 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
791                 vm->phys_addr = __pfn_to_phys(md->pfn); 
792                 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING; 
793                 vm->flags |= VM_ARM_MTYPE(md->type);
794                 vm->caller = iotable_init;
795                 vm_area_add_early(vm++);
796         }
797 }
798
799 static void * __initdata vmalloc_min =
800         (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
801
802 /*
803  * vmalloc=size forces the vmalloc area to be exactly 'size'
804  * bytes. This can be used to increase (or decrease) the vmalloc
805  * area - the default is 240m.
806  */
807 static int __init early_vmalloc(char *arg)
808 {
809         unsigned long vmalloc_reserve = memparse(arg, NULL);
810
811         if (vmalloc_reserve < SZ_16M) {
812                 vmalloc_reserve = SZ_16M;
813                 printk(KERN_WARNING
814                         "vmalloc area too small, limiting to %luMB\n",
815                         vmalloc_reserve >> 20);
816         }
817
818         if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
819                 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
820                 printk(KERN_WARNING
821                         "vmalloc area is too big, limiting to %luMB\n",
822                         vmalloc_reserve >> 20);
823         }
824
825         vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
826         return 0;
827 }
828 early_param("vmalloc", early_vmalloc);
829
830 phys_addr_t arm_lowmem_limit __initdata = 0;
831
832 void __init sanity_check_meminfo(void)
833 {
834         int i, j, highmem = 0;
835
836         for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
837                 struct membank *bank = &meminfo.bank[j];
838                 *bank = meminfo.bank[i];
839
840                 if (bank->start > ULONG_MAX)
841                         highmem = 1;
842
843 #ifdef CONFIG_HIGHMEM
844                 if (__va(bank->start) >= vmalloc_min ||
845                     __va(bank->start) < (void *)PAGE_OFFSET)
846                         highmem = 1;
847
848                 bank->highmem = highmem;
849
850                 /*
851                  * Split those memory banks which are partially overlapping
852                  * the vmalloc area greatly simplifying things later.
853                  */
854                 if (!highmem && __va(bank->start) < vmalloc_min &&
855                     bank->size > vmalloc_min - __va(bank->start)) {
856                         if (meminfo.nr_banks >= NR_BANKS) {
857                                 printk(KERN_CRIT "NR_BANKS too low, "
858                                                  "ignoring high memory\n");
859                         } else {
860                                 memmove(bank + 1, bank,
861                                         (meminfo.nr_banks - i) * sizeof(*bank));
862                                 meminfo.nr_banks++;
863                                 i++;
864                                 bank[1].size -= vmalloc_min - __va(bank->start);
865                                 bank[1].start = __pa(vmalloc_min - 1) + 1;
866                                 bank[1].highmem = highmem = 1;
867                                 j++;
868                         }
869                         bank->size = vmalloc_min - __va(bank->start);
870                 }
871 #else
872                 bank->highmem = highmem;
873
874                 /*
875                  * Highmem banks not allowed with !CONFIG_HIGHMEM.
876                  */
877                 if (highmem) {
878                         printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
879                                "(!CONFIG_HIGHMEM).\n",
880                                (unsigned long long)bank->start,
881                                (unsigned long long)bank->start + bank->size - 1);
882                         continue;
883                 }
884
885                 /*
886                  * Check whether this memory bank would entirely overlap
887                  * the vmalloc area.
888                  */
889                 if (__va(bank->start) >= vmalloc_min ||
890                     __va(bank->start) < (void *)PAGE_OFFSET) {
891                         printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
892                                "(vmalloc region overlap).\n",
893                                (unsigned long long)bank->start,
894                                (unsigned long long)bank->start + bank->size - 1);
895                         continue;
896                 }
897
898                 /*
899                  * Check whether this memory bank would partially overlap
900                  * the vmalloc area.
901                  */
902                 if (__va(bank->start + bank->size) > vmalloc_min ||
903                     __va(bank->start + bank->size) < __va(bank->start)) {
904                         unsigned long newsize = vmalloc_min - __va(bank->start);
905                         printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
906                                "to -%.8llx (vmalloc region overlap).\n",
907                                (unsigned long long)bank->start,
908                                (unsigned long long)bank->start + bank->size - 1,
909                                (unsigned long long)bank->start + newsize - 1);
910                         bank->size = newsize;
911                 }
912 #endif
913                 if (!bank->highmem && bank->start + bank->size > arm_lowmem_limit)
914                         arm_lowmem_limit = bank->start + bank->size;
915
916                 j++;
917         }
918 #ifdef CONFIG_HIGHMEM
919         if (highmem) {
920                 const char *reason = NULL;
921
922                 if (cache_is_vipt_aliasing()) {
923                         /*
924                          * Interactions between kmap and other mappings
925                          * make highmem support with aliasing VIPT caches
926                          * rather difficult.
927                          */
928                         reason = "with VIPT aliasing cache";
929                 }
930                 if (reason) {
931                         printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
932                                 reason);
933                         while (j > 0 && meminfo.bank[j - 1].highmem)
934                                 j--;
935                 }
936         }
937 #endif
938         meminfo.nr_banks = j;
939         high_memory = __va(arm_lowmem_limit - 1) + 1;
940         memblock_set_current_limit(arm_lowmem_limit);
941 }
942
943 static inline void prepare_page_table(void)
944 {
945         unsigned long addr;
946         phys_addr_t end;
947
948         /*
949          * Clear out all the mappings below the kernel image.
950          */
951         for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
952                 pmd_clear(pmd_off_k(addr));
953
954 #ifdef CONFIG_XIP_KERNEL
955         /* The XIP kernel is mapped in the module area -- skip over it */
956         addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
957 #endif
958         for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
959                 pmd_clear(pmd_off_k(addr));
960
961         /*
962          * Find the end of the first block of lowmem.
963          */
964         end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
965         if (end >= arm_lowmem_limit)
966                 end = arm_lowmem_limit;
967
968         /*
969          * Clear out all the kernel space mappings, except for the first
970          * memory bank, up to the vmalloc region.
971          */
972         for (addr = __phys_to_virt(end);
973              addr < VMALLOC_START; addr += PMD_SIZE)
974                 pmd_clear(pmd_off_k(addr));
975 }
976
977 #ifdef CONFIG_ARM_LPAE
978 /* the first page is reserved for pgd */
979 #define SWAPPER_PG_DIR_SIZE     (PAGE_SIZE + \
980                                  PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
981 #else
982 #define SWAPPER_PG_DIR_SIZE     (PTRS_PER_PGD * sizeof(pgd_t))
983 #endif
984
985 /*
986  * Reserve the special regions of memory
987  */
988 void __init arm_mm_memblock_reserve(void)
989 {
990         /*
991          * Reserve the page tables.  These are already in use,
992          * and can only be in node 0.
993          */
994         memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
995
996 #ifdef CONFIG_SA1111
997         /*
998          * Because of the SA1111 DMA bug, we want to preserve our
999          * precious DMA-able memory...
1000          */
1001         memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1002 #endif
1003 }
1004
1005 /*
1006  * Set up the device mappings.  Since we clear out the page tables for all
1007  * mappings above VMALLOC_START, we will remove any debug device mappings.
1008  * This means you have to be careful how you debug this function, or any
1009  * called function.  This means you can't use any function or debugging
1010  * method which may touch any device, otherwise the kernel _will_ crash.
1011  */
1012 static void __init devicemaps_init(struct machine_desc *mdesc)
1013 {
1014         struct map_desc map;
1015         unsigned long addr;
1016
1017         /*
1018          * Allocate the vector page early.
1019          */
1020         vectors_page = early_alloc(PAGE_SIZE);
1021
1022         for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1023                 pmd_clear(pmd_off_k(addr));
1024
1025         /*
1026          * Map the kernel if it is XIP.
1027          * It is always first in the modulearea.
1028          */
1029 #ifdef CONFIG_XIP_KERNEL
1030         map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1031         map.virtual = MODULES_VADDR;
1032         map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1033         map.type = MT_ROM;
1034         create_mapping(&map);
1035 #endif
1036
1037         /*
1038          * Map the cache flushing regions.
1039          */
1040 #ifdef FLUSH_BASE
1041         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1042         map.virtual = FLUSH_BASE;
1043         map.length = SZ_1M;
1044         map.type = MT_CACHECLEAN;
1045         create_mapping(&map);
1046 #endif
1047 #ifdef FLUSH_BASE_MINICACHE
1048         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1049         map.virtual = FLUSH_BASE_MINICACHE;
1050         map.length = SZ_1M;
1051         map.type = MT_MINICLEAN;
1052         create_mapping(&map);
1053 #endif
1054
1055         /*
1056          * Create a mapping for the machine vectors at the high-vectors
1057          * location (0xffff0000).  If we aren't using high-vectors, also
1058          * create a mapping at the low-vectors virtual address.
1059          */
1060         map.pfn = __phys_to_pfn(virt_to_phys(vectors_page));
1061         map.virtual = 0xffff0000;
1062         map.length = PAGE_SIZE;
1063         map.type = MT_HIGH_VECTORS;
1064         create_mapping(&map);
1065
1066         if (!vectors_high()) {
1067                 map.virtual = 0;
1068                 map.type = MT_LOW_VECTORS;
1069                 create_mapping(&map);
1070         }
1071
1072         /*
1073          * Ask the machine support to map in the statically mapped devices.
1074          */
1075         if (mdesc->map_io)
1076                 mdesc->map_io();
1077
1078         /*
1079          * Finally flush the caches and tlb to ensure that we're in a
1080          * consistent state wrt the writebuffer.  This also ensures that
1081          * any write-allocated cache lines in the vector page are written
1082          * back.  After this point, we can start to touch devices again.
1083          */
1084         local_flush_tlb_all();
1085         flush_cache_all();
1086 }
1087
1088 static void __init kmap_init(void)
1089 {
1090 #ifdef CONFIG_HIGHMEM
1091         pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1092                 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1093 #endif
1094 }
1095
1096 static void __init map_lowmem(void)
1097 {
1098         struct memblock_region *reg;
1099
1100         /* Map all the lowmem memory banks. */
1101         for_each_memblock(memory, reg) {
1102                 phys_addr_t start = reg->base;
1103                 phys_addr_t end = start + reg->size;
1104                 struct map_desc map;
1105
1106                 if (end > arm_lowmem_limit)
1107                         end = arm_lowmem_limit;
1108                 if (start >= end)
1109                         break;
1110
1111                 map.pfn = __phys_to_pfn(start);
1112                 map.virtual = __phys_to_virt(start);
1113                 map.length = end - start;
1114                 map.type = MT_MEMORY;
1115
1116                 create_mapping(&map);
1117         }
1118 }
1119
1120 /*
1121  * paging_init() sets up the page tables, initialises the zone memory
1122  * maps, and sets up the zero page, bad page and bad page tables.
1123  */
1124 void __init paging_init(struct machine_desc *mdesc)
1125 {
1126         void *zero_page;
1127
1128         memblock_set_current_limit(arm_lowmem_limit);
1129
1130         build_mem_type_table();
1131         prepare_page_table();
1132         map_lowmem();
1133         dma_contiguous_remap();
1134         devicemaps_init(mdesc);
1135         kmap_init();
1136
1137         top_pmd = pmd_off_k(0xffff0000);
1138
1139         /* allocate the zero page. */
1140         zero_page = early_alloc(PAGE_SIZE);
1141
1142         bootmem_init();
1143
1144         empty_zero_page = virt_to_page(zero_page);
1145         __flush_dcache_page(NULL, empty_zero_page);
1146 }