Merge branch 'for-3.5/drivers' of git://git.kernel.dk/linux-block
[sfrench/cifs-2.6.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/cp15.h>
21 #include <asm/cputype.h>
22 #include <asm/sections.h>
23 #include <asm/cachetype.h>
24 #include <asm/setup.h>
25 #include <asm/sizes.h>
26 #include <asm/smp_plat.h>
27 #include <asm/tlb.h>
28 #include <asm/highmem.h>
29 #include <asm/system_info.h>
30 #include <asm/traps.h>
31
32 #include <asm/mach/arch.h>
33 #include <asm/mach/map.h>
34
35 #include "mm.h"
36
37 /*
38  * empty_zero_page is a special page that is used for
39  * zero-initialized data and COW.
40  */
41 struct page *empty_zero_page;
42 EXPORT_SYMBOL(empty_zero_page);
43
44 /*
45  * The pmd table for the upper-most set of pages.
46  */
47 pmd_t *top_pmd;
48
49 #define CPOLICY_UNCACHED        0
50 #define CPOLICY_BUFFERED        1
51 #define CPOLICY_WRITETHROUGH    2
52 #define CPOLICY_WRITEBACK       3
53 #define CPOLICY_WRITEALLOC      4
54
55 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
56 static unsigned int ecc_mask __initdata = 0;
57 pgprot_t pgprot_user;
58 pgprot_t pgprot_kernel;
59
60 EXPORT_SYMBOL(pgprot_user);
61 EXPORT_SYMBOL(pgprot_kernel);
62
63 struct cachepolicy {
64         const char      policy[16];
65         unsigned int    cr_mask;
66         pmdval_t        pmd;
67         pteval_t        pte;
68 };
69
70 static struct cachepolicy cache_policies[] __initdata = {
71         {
72                 .policy         = "uncached",
73                 .cr_mask        = CR_W|CR_C,
74                 .pmd            = PMD_SECT_UNCACHED,
75                 .pte            = L_PTE_MT_UNCACHED,
76         }, {
77                 .policy         = "buffered",
78                 .cr_mask        = CR_C,
79                 .pmd            = PMD_SECT_BUFFERED,
80                 .pte            = L_PTE_MT_BUFFERABLE,
81         }, {
82                 .policy         = "writethrough",
83                 .cr_mask        = 0,
84                 .pmd            = PMD_SECT_WT,
85                 .pte            = L_PTE_MT_WRITETHROUGH,
86         }, {
87                 .policy         = "writeback",
88                 .cr_mask        = 0,
89                 .pmd            = PMD_SECT_WB,
90                 .pte            = L_PTE_MT_WRITEBACK,
91         }, {
92                 .policy         = "writealloc",
93                 .cr_mask        = 0,
94                 .pmd            = PMD_SECT_WBWA,
95                 .pte            = L_PTE_MT_WRITEALLOC,
96         }
97 };
98
99 /*
100  * These are useful for identifying cache coherency
101  * problems by allowing the cache or the cache and
102  * writebuffer to be turned off.  (Note: the write
103  * buffer should not be on and the cache off).
104  */
105 static int __init early_cachepolicy(char *p)
106 {
107         int i;
108
109         for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
110                 int len = strlen(cache_policies[i].policy);
111
112                 if (memcmp(p, cache_policies[i].policy, len) == 0) {
113                         cachepolicy = i;
114                         cr_alignment &= ~cache_policies[i].cr_mask;
115                         cr_no_alignment &= ~cache_policies[i].cr_mask;
116                         break;
117                 }
118         }
119         if (i == ARRAY_SIZE(cache_policies))
120                 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
121         /*
122          * This restriction is partly to do with the way we boot; it is
123          * unpredictable to have memory mapped using two different sets of
124          * memory attributes (shared, type, and cache attribs).  We can not
125          * change these attributes once the initial assembly has setup the
126          * page tables.
127          */
128         if (cpu_architecture() >= CPU_ARCH_ARMv6) {
129                 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
130                 cachepolicy = CPOLICY_WRITEBACK;
131         }
132         flush_cache_all();
133         set_cr(cr_alignment);
134         return 0;
135 }
136 early_param("cachepolicy", early_cachepolicy);
137
138 static int __init early_nocache(char *__unused)
139 {
140         char *p = "buffered";
141         printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
142         early_cachepolicy(p);
143         return 0;
144 }
145 early_param("nocache", early_nocache);
146
147 static int __init early_nowrite(char *__unused)
148 {
149         char *p = "uncached";
150         printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
151         early_cachepolicy(p);
152         return 0;
153 }
154 early_param("nowb", early_nowrite);
155
156 #ifndef CONFIG_ARM_LPAE
157 static int __init early_ecc(char *p)
158 {
159         if (memcmp(p, "on", 2) == 0)
160                 ecc_mask = PMD_PROTECTION;
161         else if (memcmp(p, "off", 3) == 0)
162                 ecc_mask = 0;
163         return 0;
164 }
165 early_param("ecc", early_ecc);
166 #endif
167
168 static int __init noalign_setup(char *__unused)
169 {
170         cr_alignment &= ~CR_A;
171         cr_no_alignment &= ~CR_A;
172         set_cr(cr_alignment);
173         return 1;
174 }
175 __setup("noalign", noalign_setup);
176
177 #ifndef CONFIG_SMP
178 void adjust_cr(unsigned long mask, unsigned long set)
179 {
180         unsigned long flags;
181
182         mask &= ~CR_A;
183
184         set &= mask;
185
186         local_irq_save(flags);
187
188         cr_no_alignment = (cr_no_alignment & ~mask) | set;
189         cr_alignment = (cr_alignment & ~mask) | set;
190
191         set_cr((get_cr() & ~mask) | set);
192
193         local_irq_restore(flags);
194 }
195 #endif
196
197 #define PROT_PTE_DEVICE         L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
198 #define PROT_SECT_DEVICE        PMD_TYPE_SECT|PMD_SECT_AP_WRITE
199
200 static struct mem_type mem_types[] = {
201         [MT_DEVICE] = {           /* Strongly ordered / ARMv6 shared device */
202                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
203                                   L_PTE_SHARED,
204                 .prot_l1        = PMD_TYPE_TABLE,
205                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_S,
206                 .domain         = DOMAIN_IO,
207         },
208         [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
209                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
210                 .prot_l1        = PMD_TYPE_TABLE,
211                 .prot_sect      = PROT_SECT_DEVICE,
212                 .domain         = DOMAIN_IO,
213         },
214         [MT_DEVICE_CACHED] = {    /* ioremap_cached */
215                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
216                 .prot_l1        = PMD_TYPE_TABLE,
217                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_WB,
218                 .domain         = DOMAIN_IO,
219         },      
220         [MT_DEVICE_WC] = {      /* ioremap_wc */
221                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
222                 .prot_l1        = PMD_TYPE_TABLE,
223                 .prot_sect      = PROT_SECT_DEVICE,
224                 .domain         = DOMAIN_IO,
225         },
226         [MT_UNCACHED] = {
227                 .prot_pte       = PROT_PTE_DEVICE,
228                 .prot_l1        = PMD_TYPE_TABLE,
229                 .prot_sect      = PMD_TYPE_SECT | PMD_SECT_XN,
230                 .domain         = DOMAIN_IO,
231         },
232         [MT_CACHECLEAN] = {
233                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
234                 .domain    = DOMAIN_KERNEL,
235         },
236 #ifndef CONFIG_ARM_LPAE
237         [MT_MINICLEAN] = {
238                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
239                 .domain    = DOMAIN_KERNEL,
240         },
241 #endif
242         [MT_LOW_VECTORS] = {
243                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
244                                 L_PTE_RDONLY,
245                 .prot_l1   = PMD_TYPE_TABLE,
246                 .domain    = DOMAIN_USER,
247         },
248         [MT_HIGH_VECTORS] = {
249                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
250                                 L_PTE_USER | L_PTE_RDONLY,
251                 .prot_l1   = PMD_TYPE_TABLE,
252                 .domain    = DOMAIN_USER,
253         },
254         [MT_MEMORY] = {
255                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
256                 .prot_l1   = PMD_TYPE_TABLE,
257                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
258                 .domain    = DOMAIN_KERNEL,
259         },
260         [MT_ROM] = {
261                 .prot_sect = PMD_TYPE_SECT,
262                 .domain    = DOMAIN_KERNEL,
263         },
264         [MT_MEMORY_NONCACHED] = {
265                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
266                                 L_PTE_MT_BUFFERABLE,
267                 .prot_l1   = PMD_TYPE_TABLE,
268                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
269                 .domain    = DOMAIN_KERNEL,
270         },
271         [MT_MEMORY_DTCM] = {
272                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
273                                 L_PTE_XN,
274                 .prot_l1   = PMD_TYPE_TABLE,
275                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
276                 .domain    = DOMAIN_KERNEL,
277         },
278         [MT_MEMORY_ITCM] = {
279                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
280                 .prot_l1   = PMD_TYPE_TABLE,
281                 .domain    = DOMAIN_KERNEL,
282         },
283         [MT_MEMORY_SO] = {
284                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
285                                 L_PTE_MT_UNCACHED,
286                 .prot_l1   = PMD_TYPE_TABLE,
287                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
288                                 PMD_SECT_UNCACHED | PMD_SECT_XN,
289                 .domain    = DOMAIN_KERNEL,
290         },
291         [MT_MEMORY_DMA_READY] = {
292                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
293                 .prot_l1   = PMD_TYPE_TABLE,
294                 .domain    = DOMAIN_KERNEL,
295         },
296 };
297
298 const struct mem_type *get_mem_type(unsigned int type)
299 {
300         return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
301 }
302 EXPORT_SYMBOL(get_mem_type);
303
304 /*
305  * Adjust the PMD section entries according to the CPU in use.
306  */
307 static void __init build_mem_type_table(void)
308 {
309         struct cachepolicy *cp;
310         unsigned int cr = get_cr();
311         pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
312         int cpu_arch = cpu_architecture();
313         int i;
314
315         if (cpu_arch < CPU_ARCH_ARMv6) {
316 #if defined(CONFIG_CPU_DCACHE_DISABLE)
317                 if (cachepolicy > CPOLICY_BUFFERED)
318                         cachepolicy = CPOLICY_BUFFERED;
319 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
320                 if (cachepolicy > CPOLICY_WRITETHROUGH)
321                         cachepolicy = CPOLICY_WRITETHROUGH;
322 #endif
323         }
324         if (cpu_arch < CPU_ARCH_ARMv5) {
325                 if (cachepolicy >= CPOLICY_WRITEALLOC)
326                         cachepolicy = CPOLICY_WRITEBACK;
327                 ecc_mask = 0;
328         }
329         if (is_smp())
330                 cachepolicy = CPOLICY_WRITEALLOC;
331
332         /*
333          * Strip out features not present on earlier architectures.
334          * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
335          * without extended page tables don't have the 'Shared' bit.
336          */
337         if (cpu_arch < CPU_ARCH_ARMv5)
338                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
339                         mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
340         if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
341                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
342                         mem_types[i].prot_sect &= ~PMD_SECT_S;
343
344         /*
345          * ARMv5 and lower, bit 4 must be set for page tables (was: cache
346          * "update-able on write" bit on ARM610).  However, Xscale and
347          * Xscale3 require this bit to be cleared.
348          */
349         if (cpu_is_xscale() || cpu_is_xsc3()) {
350                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
351                         mem_types[i].prot_sect &= ~PMD_BIT4;
352                         mem_types[i].prot_l1 &= ~PMD_BIT4;
353                 }
354         } else if (cpu_arch < CPU_ARCH_ARMv6) {
355                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
356                         if (mem_types[i].prot_l1)
357                                 mem_types[i].prot_l1 |= PMD_BIT4;
358                         if (mem_types[i].prot_sect)
359                                 mem_types[i].prot_sect |= PMD_BIT4;
360                 }
361         }
362
363         /*
364          * Mark the device areas according to the CPU/architecture.
365          */
366         if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
367                 if (!cpu_is_xsc3()) {
368                         /*
369                          * Mark device regions on ARMv6+ as execute-never
370                          * to prevent speculative instruction fetches.
371                          */
372                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
373                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
374                         mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
375                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
376                 }
377                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
378                         /*
379                          * For ARMv7 with TEX remapping,
380                          * - shared device is SXCB=1100
381                          * - nonshared device is SXCB=0100
382                          * - write combine device mem is SXCB=0001
383                          * (Uncached Normal memory)
384                          */
385                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
386                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
387                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
388                 } else if (cpu_is_xsc3()) {
389                         /*
390                          * For Xscale3,
391                          * - shared device is TEXCB=00101
392                          * - nonshared device is TEXCB=01000
393                          * - write combine device mem is TEXCB=00100
394                          * (Inner/Outer Uncacheable in xsc3 parlance)
395                          */
396                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
397                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
398                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
399                 } else {
400                         /*
401                          * For ARMv6 and ARMv7 without TEX remapping,
402                          * - shared device is TEXCB=00001
403                          * - nonshared device is TEXCB=01000
404                          * - write combine device mem is TEXCB=00100
405                          * (Uncached Normal in ARMv6 parlance).
406                          */
407                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
408                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
409                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
410                 }
411         } else {
412                 /*
413                  * On others, write combining is "Uncached/Buffered"
414                  */
415                 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
416         }
417
418         /*
419          * Now deal with the memory-type mappings
420          */
421         cp = &cache_policies[cachepolicy];
422         vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
423
424         /*
425          * Only use write-through for non-SMP systems
426          */
427         if (!is_smp() && cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
428                 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
429
430         /*
431          * Enable CPU-specific coherency if supported.
432          * (Only available on XSC3 at the moment.)
433          */
434         if (arch_is_coherent() && cpu_is_xsc3()) {
435                 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
436                 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
437                 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
438                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
439                 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
440         }
441         /*
442          * ARMv6 and above have extended page tables.
443          */
444         if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
445 #ifndef CONFIG_ARM_LPAE
446                 /*
447                  * Mark cache clean areas and XIP ROM read only
448                  * from SVC mode and no access from userspace.
449                  */
450                 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
451                 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
452                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
453 #endif
454
455                 if (is_smp()) {
456                         /*
457                          * Mark memory with the "shared" attribute
458                          * for SMP systems
459                          */
460                         user_pgprot |= L_PTE_SHARED;
461                         kern_pgprot |= L_PTE_SHARED;
462                         vecs_pgprot |= L_PTE_SHARED;
463                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
464                         mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
465                         mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
466                         mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
467                         mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
468                         mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
469                         mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
470                         mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
471                         mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
472                 }
473         }
474
475         /*
476          * Non-cacheable Normal - intended for memory areas that must
477          * not cause dirty cache line writebacks when used
478          */
479         if (cpu_arch >= CPU_ARCH_ARMv6) {
480                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
481                         /* Non-cacheable Normal is XCB = 001 */
482                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
483                                 PMD_SECT_BUFFERED;
484                 } else {
485                         /* For both ARMv6 and non-TEX-remapping ARMv7 */
486                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
487                                 PMD_SECT_TEX(1);
488                 }
489         } else {
490                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
491         }
492
493 #ifdef CONFIG_ARM_LPAE
494         /*
495          * Do not generate access flag faults for the kernel mappings.
496          */
497         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
498                 mem_types[i].prot_pte |= PTE_EXT_AF;
499                 if (mem_types[i].prot_sect)
500                         mem_types[i].prot_sect |= PMD_SECT_AF;
501         }
502         kern_pgprot |= PTE_EXT_AF;
503         vecs_pgprot |= PTE_EXT_AF;
504 #endif
505
506         for (i = 0; i < 16; i++) {
507                 unsigned long v = pgprot_val(protection_map[i]);
508                 protection_map[i] = __pgprot(v | user_pgprot);
509         }
510
511         mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
512         mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
513
514         pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
515         pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
516                                  L_PTE_DIRTY | kern_pgprot);
517
518         mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
519         mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
520         mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
521         mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
522         mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
523         mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
524         mem_types[MT_ROM].prot_sect |= cp->pmd;
525
526         switch (cp->pmd) {
527         case PMD_SECT_WT:
528                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
529                 break;
530         case PMD_SECT_WB:
531         case PMD_SECT_WBWA:
532                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
533                 break;
534         }
535         printk("Memory policy: ECC %sabled, Data cache %s\n",
536                 ecc_mask ? "en" : "dis", cp->policy);
537
538         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
539                 struct mem_type *t = &mem_types[i];
540                 if (t->prot_l1)
541                         t->prot_l1 |= PMD_DOMAIN(t->domain);
542                 if (t->prot_sect)
543                         t->prot_sect |= PMD_DOMAIN(t->domain);
544         }
545 }
546
547 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
548 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
549                               unsigned long size, pgprot_t vma_prot)
550 {
551         if (!pfn_valid(pfn))
552                 return pgprot_noncached(vma_prot);
553         else if (file->f_flags & O_SYNC)
554                 return pgprot_writecombine(vma_prot);
555         return vma_prot;
556 }
557 EXPORT_SYMBOL(phys_mem_access_prot);
558 #endif
559
560 #define vectors_base()  (vectors_high() ? 0xffff0000 : 0)
561
562 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
563 {
564         void *ptr = __va(memblock_alloc(sz, align));
565         memset(ptr, 0, sz);
566         return ptr;
567 }
568
569 static void __init *early_alloc(unsigned long sz)
570 {
571         return early_alloc_aligned(sz, sz);
572 }
573
574 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
575 {
576         if (pmd_none(*pmd)) {
577                 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
578                 __pmd_populate(pmd, __pa(pte), prot);
579         }
580         BUG_ON(pmd_bad(*pmd));
581         return pte_offset_kernel(pmd, addr);
582 }
583
584 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
585                                   unsigned long end, unsigned long pfn,
586                                   const struct mem_type *type)
587 {
588         pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
589         do {
590                 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
591                 pfn++;
592         } while (pte++, addr += PAGE_SIZE, addr != end);
593 }
594
595 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
596                                       unsigned long end, phys_addr_t phys,
597                                       const struct mem_type *type)
598 {
599         pmd_t *pmd = pmd_offset(pud, addr);
600
601         /*
602          * Try a section mapping - end, addr and phys must all be aligned
603          * to a section boundary.  Note that PMDs refer to the individual
604          * L1 entries, whereas PGDs refer to a group of L1 entries making
605          * up one logical pointer to an L2 table.
606          */
607         if (type->prot_sect && ((addr | end | phys) & ~SECTION_MASK) == 0) {
608                 pmd_t *p = pmd;
609
610 #ifndef CONFIG_ARM_LPAE
611                 if (addr & SECTION_SIZE)
612                         pmd++;
613 #endif
614
615                 do {
616                         *pmd = __pmd(phys | type->prot_sect);
617                         phys += SECTION_SIZE;
618                 } while (pmd++, addr += SECTION_SIZE, addr != end);
619
620                 flush_pmd_entry(p);
621         } else {
622                 /*
623                  * No need to loop; pte's aren't interested in the
624                  * individual L1 entries.
625                  */
626                 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
627         }
628 }
629
630 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
631         unsigned long end, unsigned long phys, const struct mem_type *type)
632 {
633         pud_t *pud = pud_offset(pgd, addr);
634         unsigned long next;
635
636         do {
637                 next = pud_addr_end(addr, end);
638                 alloc_init_section(pud, addr, next, phys, type);
639                 phys += next - addr;
640         } while (pud++, addr = next, addr != end);
641 }
642
643 #ifndef CONFIG_ARM_LPAE
644 static void __init create_36bit_mapping(struct map_desc *md,
645                                         const struct mem_type *type)
646 {
647         unsigned long addr, length, end;
648         phys_addr_t phys;
649         pgd_t *pgd;
650
651         addr = md->virtual;
652         phys = __pfn_to_phys(md->pfn);
653         length = PAGE_ALIGN(md->length);
654
655         if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
656                 printk(KERN_ERR "MM: CPU does not support supersection "
657                        "mapping for 0x%08llx at 0x%08lx\n",
658                        (long long)__pfn_to_phys((u64)md->pfn), addr);
659                 return;
660         }
661
662         /* N.B. ARMv6 supersections are only defined to work with domain 0.
663          *      Since domain assignments can in fact be arbitrary, the
664          *      'domain == 0' check below is required to insure that ARMv6
665          *      supersections are only allocated for domain 0 regardless
666          *      of the actual domain assignments in use.
667          */
668         if (type->domain) {
669                 printk(KERN_ERR "MM: invalid domain in supersection "
670                        "mapping for 0x%08llx at 0x%08lx\n",
671                        (long long)__pfn_to_phys((u64)md->pfn), addr);
672                 return;
673         }
674
675         if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
676                 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
677                        " at 0x%08lx invalid alignment\n",
678                        (long long)__pfn_to_phys((u64)md->pfn), addr);
679                 return;
680         }
681
682         /*
683          * Shift bits [35:32] of address into bits [23:20] of PMD
684          * (See ARMv6 spec).
685          */
686         phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
687
688         pgd = pgd_offset_k(addr);
689         end = addr + length;
690         do {
691                 pud_t *pud = pud_offset(pgd, addr);
692                 pmd_t *pmd = pmd_offset(pud, addr);
693                 int i;
694
695                 for (i = 0; i < 16; i++)
696                         *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
697
698                 addr += SUPERSECTION_SIZE;
699                 phys += SUPERSECTION_SIZE;
700                 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
701         } while (addr != end);
702 }
703 #endif  /* !CONFIG_ARM_LPAE */
704
705 /*
706  * Create the page directory entries and any necessary
707  * page tables for the mapping specified by `md'.  We
708  * are able to cope here with varying sizes and address
709  * offsets, and we take full advantage of sections and
710  * supersections.
711  */
712 static void __init create_mapping(struct map_desc *md)
713 {
714         unsigned long addr, length, end;
715         phys_addr_t phys;
716         const struct mem_type *type;
717         pgd_t *pgd;
718
719         if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
720                 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
721                        " at 0x%08lx in user region\n",
722                        (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
723                 return;
724         }
725
726         if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
727             md->virtual >= PAGE_OFFSET &&
728             (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
729                 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
730                        " at 0x%08lx out of vmalloc space\n",
731                        (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
732         }
733
734         type = &mem_types[md->type];
735
736 #ifndef CONFIG_ARM_LPAE
737         /*
738          * Catch 36-bit addresses
739          */
740         if (md->pfn >= 0x100000) {
741                 create_36bit_mapping(md, type);
742                 return;
743         }
744 #endif
745
746         addr = md->virtual & PAGE_MASK;
747         phys = __pfn_to_phys(md->pfn);
748         length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
749
750         if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
751                 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
752                        "be mapped using pages, ignoring.\n",
753                        (long long)__pfn_to_phys(md->pfn), addr);
754                 return;
755         }
756
757         pgd = pgd_offset_k(addr);
758         end = addr + length;
759         do {
760                 unsigned long next = pgd_addr_end(addr, end);
761
762                 alloc_init_pud(pgd, addr, next, phys, type);
763
764                 phys += next - addr;
765                 addr = next;
766         } while (pgd++, addr != end);
767 }
768
769 /*
770  * Create the architecture specific mappings
771  */
772 void __init iotable_init(struct map_desc *io_desc, int nr)
773 {
774         struct map_desc *md;
775         struct vm_struct *vm;
776
777         if (!nr)
778                 return;
779
780         vm = early_alloc_aligned(sizeof(*vm) * nr, __alignof__(*vm));
781
782         for (md = io_desc; nr; md++, nr--) {
783                 create_mapping(md);
784                 vm->addr = (void *)(md->virtual & PAGE_MASK);
785                 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
786                 vm->phys_addr = __pfn_to_phys(md->pfn); 
787                 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING; 
788                 vm->flags |= VM_ARM_MTYPE(md->type);
789                 vm->caller = iotable_init;
790                 vm_area_add_early(vm++);
791         }
792 }
793
794 static void * __initdata vmalloc_min =
795         (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
796
797 /*
798  * vmalloc=size forces the vmalloc area to be exactly 'size'
799  * bytes. This can be used to increase (or decrease) the vmalloc
800  * area - the default is 240m.
801  */
802 static int __init early_vmalloc(char *arg)
803 {
804         unsigned long vmalloc_reserve = memparse(arg, NULL);
805
806         if (vmalloc_reserve < SZ_16M) {
807                 vmalloc_reserve = SZ_16M;
808                 printk(KERN_WARNING
809                         "vmalloc area too small, limiting to %luMB\n",
810                         vmalloc_reserve >> 20);
811         }
812
813         if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
814                 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
815                 printk(KERN_WARNING
816                         "vmalloc area is too big, limiting to %luMB\n",
817                         vmalloc_reserve >> 20);
818         }
819
820         vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
821         return 0;
822 }
823 early_param("vmalloc", early_vmalloc);
824
825 phys_addr_t arm_lowmem_limit __initdata = 0;
826
827 void __init sanity_check_meminfo(void)
828 {
829         int i, j, highmem = 0;
830
831         for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
832                 struct membank *bank = &meminfo.bank[j];
833                 *bank = meminfo.bank[i];
834
835                 if (bank->start > ULONG_MAX)
836                         highmem = 1;
837
838 #ifdef CONFIG_HIGHMEM
839                 if (__va(bank->start) >= vmalloc_min ||
840                     __va(bank->start) < (void *)PAGE_OFFSET)
841                         highmem = 1;
842
843                 bank->highmem = highmem;
844
845                 /*
846                  * Split those memory banks which are partially overlapping
847                  * the vmalloc area greatly simplifying things later.
848                  */
849                 if (!highmem && __va(bank->start) < vmalloc_min &&
850                     bank->size > vmalloc_min - __va(bank->start)) {
851                         if (meminfo.nr_banks >= NR_BANKS) {
852                                 printk(KERN_CRIT "NR_BANKS too low, "
853                                                  "ignoring high memory\n");
854                         } else {
855                                 memmove(bank + 1, bank,
856                                         (meminfo.nr_banks - i) * sizeof(*bank));
857                                 meminfo.nr_banks++;
858                                 i++;
859                                 bank[1].size -= vmalloc_min - __va(bank->start);
860                                 bank[1].start = __pa(vmalloc_min - 1) + 1;
861                                 bank[1].highmem = highmem = 1;
862                                 j++;
863                         }
864                         bank->size = vmalloc_min - __va(bank->start);
865                 }
866 #else
867                 bank->highmem = highmem;
868
869                 /*
870                  * Highmem banks not allowed with !CONFIG_HIGHMEM.
871                  */
872                 if (highmem) {
873                         printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
874                                "(!CONFIG_HIGHMEM).\n",
875                                (unsigned long long)bank->start,
876                                (unsigned long long)bank->start + bank->size - 1);
877                         continue;
878                 }
879
880                 /*
881                  * Check whether this memory bank would entirely overlap
882                  * the vmalloc area.
883                  */
884                 if (__va(bank->start) >= vmalloc_min ||
885                     __va(bank->start) < (void *)PAGE_OFFSET) {
886                         printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
887                                "(vmalloc region overlap).\n",
888                                (unsigned long long)bank->start,
889                                (unsigned long long)bank->start + bank->size - 1);
890                         continue;
891                 }
892
893                 /*
894                  * Check whether this memory bank would partially overlap
895                  * the vmalloc area.
896                  */
897                 if (__va(bank->start + bank->size) > vmalloc_min ||
898                     __va(bank->start + bank->size) < __va(bank->start)) {
899                         unsigned long newsize = vmalloc_min - __va(bank->start);
900                         printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
901                                "to -%.8llx (vmalloc region overlap).\n",
902                                (unsigned long long)bank->start,
903                                (unsigned long long)bank->start + bank->size - 1,
904                                (unsigned long long)bank->start + newsize - 1);
905                         bank->size = newsize;
906                 }
907 #endif
908                 if (!bank->highmem && bank->start + bank->size > arm_lowmem_limit)
909                         arm_lowmem_limit = bank->start + bank->size;
910
911                 j++;
912         }
913 #ifdef CONFIG_HIGHMEM
914         if (highmem) {
915                 const char *reason = NULL;
916
917                 if (cache_is_vipt_aliasing()) {
918                         /*
919                          * Interactions between kmap and other mappings
920                          * make highmem support with aliasing VIPT caches
921                          * rather difficult.
922                          */
923                         reason = "with VIPT aliasing cache";
924                 }
925                 if (reason) {
926                         printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
927                                 reason);
928                         while (j > 0 && meminfo.bank[j - 1].highmem)
929                                 j--;
930                 }
931         }
932 #endif
933         meminfo.nr_banks = j;
934         high_memory = __va(arm_lowmem_limit - 1) + 1;
935         memblock_set_current_limit(arm_lowmem_limit);
936 }
937
938 static inline void prepare_page_table(void)
939 {
940         unsigned long addr;
941         phys_addr_t end;
942
943         /*
944          * Clear out all the mappings below the kernel image.
945          */
946         for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
947                 pmd_clear(pmd_off_k(addr));
948
949 #ifdef CONFIG_XIP_KERNEL
950         /* The XIP kernel is mapped in the module area -- skip over it */
951         addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
952 #endif
953         for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
954                 pmd_clear(pmd_off_k(addr));
955
956         /*
957          * Find the end of the first block of lowmem.
958          */
959         end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
960         if (end >= arm_lowmem_limit)
961                 end = arm_lowmem_limit;
962
963         /*
964          * Clear out all the kernel space mappings, except for the first
965          * memory bank, up to the vmalloc region.
966          */
967         for (addr = __phys_to_virt(end);
968              addr < VMALLOC_START; addr += PMD_SIZE)
969                 pmd_clear(pmd_off_k(addr));
970 }
971
972 #ifdef CONFIG_ARM_LPAE
973 /* the first page is reserved for pgd */
974 #define SWAPPER_PG_DIR_SIZE     (PAGE_SIZE + \
975                                  PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
976 #else
977 #define SWAPPER_PG_DIR_SIZE     (PTRS_PER_PGD * sizeof(pgd_t))
978 #endif
979
980 /*
981  * Reserve the special regions of memory
982  */
983 void __init arm_mm_memblock_reserve(void)
984 {
985         /*
986          * Reserve the page tables.  These are already in use,
987          * and can only be in node 0.
988          */
989         memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
990
991 #ifdef CONFIG_SA1111
992         /*
993          * Because of the SA1111 DMA bug, we want to preserve our
994          * precious DMA-able memory...
995          */
996         memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
997 #endif
998 }
999
1000 /*
1001  * Set up the device mappings.  Since we clear out the page tables for all
1002  * mappings above VMALLOC_START, we will remove any debug device mappings.
1003  * This means you have to be careful how you debug this function, or any
1004  * called function.  This means you can't use any function or debugging
1005  * method which may touch any device, otherwise the kernel _will_ crash.
1006  */
1007 static void __init devicemaps_init(struct machine_desc *mdesc)
1008 {
1009         struct map_desc map;
1010         unsigned long addr;
1011         void *vectors;
1012
1013         /*
1014          * Allocate the vector page early.
1015          */
1016         vectors = early_alloc(PAGE_SIZE);
1017
1018         early_trap_init(vectors);
1019
1020         for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1021                 pmd_clear(pmd_off_k(addr));
1022
1023         /*
1024          * Map the kernel if it is XIP.
1025          * It is always first in the modulearea.
1026          */
1027 #ifdef CONFIG_XIP_KERNEL
1028         map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1029         map.virtual = MODULES_VADDR;
1030         map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1031         map.type = MT_ROM;
1032         create_mapping(&map);
1033 #endif
1034
1035         /*
1036          * Map the cache flushing regions.
1037          */
1038 #ifdef FLUSH_BASE
1039         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1040         map.virtual = FLUSH_BASE;
1041         map.length = SZ_1M;
1042         map.type = MT_CACHECLEAN;
1043         create_mapping(&map);
1044 #endif
1045 #ifdef FLUSH_BASE_MINICACHE
1046         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1047         map.virtual = FLUSH_BASE_MINICACHE;
1048         map.length = SZ_1M;
1049         map.type = MT_MINICLEAN;
1050         create_mapping(&map);
1051 #endif
1052
1053         /*
1054          * Create a mapping for the machine vectors at the high-vectors
1055          * location (0xffff0000).  If we aren't using high-vectors, also
1056          * create a mapping at the low-vectors virtual address.
1057          */
1058         map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1059         map.virtual = 0xffff0000;
1060         map.length = PAGE_SIZE;
1061         map.type = MT_HIGH_VECTORS;
1062         create_mapping(&map);
1063
1064         if (!vectors_high()) {
1065                 map.virtual = 0;
1066                 map.type = MT_LOW_VECTORS;
1067                 create_mapping(&map);
1068         }
1069
1070         /*
1071          * Ask the machine support to map in the statically mapped devices.
1072          */
1073         if (mdesc->map_io)
1074                 mdesc->map_io();
1075
1076         /*
1077          * Finally flush the caches and tlb to ensure that we're in a
1078          * consistent state wrt the writebuffer.  This also ensures that
1079          * any write-allocated cache lines in the vector page are written
1080          * back.  After this point, we can start to touch devices again.
1081          */
1082         local_flush_tlb_all();
1083         flush_cache_all();
1084 }
1085
1086 static void __init kmap_init(void)
1087 {
1088 #ifdef CONFIG_HIGHMEM
1089         pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1090                 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1091 #endif
1092 }
1093
1094 static void __init map_lowmem(void)
1095 {
1096         struct memblock_region *reg;
1097
1098         /* Map all the lowmem memory banks. */
1099         for_each_memblock(memory, reg) {
1100                 phys_addr_t start = reg->base;
1101                 phys_addr_t end = start + reg->size;
1102                 struct map_desc map;
1103
1104                 if (end > arm_lowmem_limit)
1105                         end = arm_lowmem_limit;
1106                 if (start >= end)
1107                         break;
1108
1109                 map.pfn = __phys_to_pfn(start);
1110                 map.virtual = __phys_to_virt(start);
1111                 map.length = end - start;
1112                 map.type = MT_MEMORY;
1113
1114                 create_mapping(&map);
1115         }
1116 }
1117
1118 /*
1119  * paging_init() sets up the page tables, initialises the zone memory
1120  * maps, and sets up the zero page, bad page and bad page tables.
1121  */
1122 void __init paging_init(struct machine_desc *mdesc)
1123 {
1124         void *zero_page;
1125
1126         memblock_set_current_limit(arm_lowmem_limit);
1127
1128         build_mem_type_table();
1129         prepare_page_table();
1130         map_lowmem();
1131         dma_contiguous_remap();
1132         devicemaps_init(mdesc);
1133         kmap_init();
1134
1135         top_pmd = pmd_off_k(0xffff0000);
1136
1137         /* allocate the zero page. */
1138         zero_page = early_alloc(PAGE_SIZE);
1139
1140         bootmem_init();
1141
1142         empty_zero_page = virt_to_page(zero_page);
1143         __flush_dcache_page(NULL, empty_zero_page);
1144 }