ARM: integrate CMA with DMA-mapping subsystem
[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
19 #include <asm/cputype.h>
20 #include <asm/sections.h>
21 #include <asm/cachetype.h>
22 #include <asm/setup.h>
23 #include <asm/sizes.h>
24 #include <asm/smp_plat.h>
25 #include <asm/tlb.h>
26 #include <asm/highmem.h>
27 #include <asm/traps.h>
28
29 #include <asm/mach/arch.h>
30 #include <asm/mach/map.h>
31
32 #include "mm.h"
33
34 /*
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         pmdval_t        pmd;
64         pteval_t        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 int __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                         break;
114                 }
115         }
116         if (i == ARRAY_SIZE(cache_policies))
117                 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
118         /*
119          * This restriction is partly to do with the way we boot; it is
120          * unpredictable to have memory mapped using two different sets of
121          * memory attributes (shared, type, and cache attribs).  We can not
122          * change these attributes once the initial assembly has setup the
123          * page tables.
124          */
125         if (cpu_architecture() >= CPU_ARCH_ARMv6) {
126                 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
127                 cachepolicy = CPOLICY_WRITEBACK;
128         }
129         flush_cache_all();
130         set_cr(cr_alignment);
131         return 0;
132 }
133 early_param("cachepolicy", early_cachepolicy);
134
135 static int __init early_nocache(char *__unused)
136 {
137         char *p = "buffered";
138         printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
139         early_cachepolicy(p);
140         return 0;
141 }
142 early_param("nocache", early_nocache);
143
144 static int __init early_nowrite(char *__unused)
145 {
146         char *p = "uncached";
147         printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
148         early_cachepolicy(p);
149         return 0;
150 }
151 early_param("nowb", early_nowrite);
152
153 #ifndef CONFIG_ARM_LPAE
154 static int __init early_ecc(char *p)
155 {
156         if (memcmp(p, "on", 2) == 0)
157                 ecc_mask = PMD_PROTECTION;
158         else if (memcmp(p, "off", 3) == 0)
159                 ecc_mask = 0;
160         return 0;
161 }
162 early_param("ecc", early_ecc);
163 #endif
164
165 static int __init noalign_setup(char *__unused)
166 {
167         cr_alignment &= ~CR_A;
168         cr_no_alignment &= ~CR_A;
169         set_cr(cr_alignment);
170         return 1;
171 }
172 __setup("noalign", noalign_setup);
173
174 #ifndef CONFIG_SMP
175 void adjust_cr(unsigned long mask, unsigned long set)
176 {
177         unsigned long flags;
178
179         mask &= ~CR_A;
180
181         set &= mask;
182
183         local_irq_save(flags);
184
185         cr_no_alignment = (cr_no_alignment & ~mask) | set;
186         cr_alignment = (cr_alignment & ~mask) | set;
187
188         set_cr((get_cr() & ~mask) | set);
189
190         local_irq_restore(flags);
191 }
192 #endif
193
194 #define PROT_PTE_DEVICE         L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
195 #define PROT_SECT_DEVICE        PMD_TYPE_SECT|PMD_SECT_AP_WRITE
196
197 static struct mem_type mem_types[] = {
198         [MT_DEVICE] = {           /* Strongly ordered / ARMv6 shared device */
199                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
200                                   L_PTE_SHARED,
201                 .prot_l1        = PMD_TYPE_TABLE,
202                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_S,
203                 .domain         = DOMAIN_IO,
204         },
205         [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
206                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
207                 .prot_l1        = PMD_TYPE_TABLE,
208                 .prot_sect      = PROT_SECT_DEVICE,
209                 .domain         = DOMAIN_IO,
210         },
211         [MT_DEVICE_CACHED] = {    /* ioremap_cached */
212                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
213                 .prot_l1        = PMD_TYPE_TABLE,
214                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_WB,
215                 .domain         = DOMAIN_IO,
216         },      
217         [MT_DEVICE_WC] = {      /* ioremap_wc */
218                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
219                 .prot_l1        = PMD_TYPE_TABLE,
220                 .prot_sect      = PROT_SECT_DEVICE,
221                 .domain         = DOMAIN_IO,
222         },
223         [MT_UNCACHED] = {
224                 .prot_pte       = PROT_PTE_DEVICE,
225                 .prot_l1        = PMD_TYPE_TABLE,
226                 .prot_sect      = PMD_TYPE_SECT | PMD_SECT_XN,
227                 .domain         = DOMAIN_IO,
228         },
229         [MT_CACHECLEAN] = {
230                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
231                 .domain    = DOMAIN_KERNEL,
232         },
233 #ifndef CONFIG_ARM_LPAE
234         [MT_MINICLEAN] = {
235                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
236                 .domain    = DOMAIN_KERNEL,
237         },
238 #endif
239         [MT_LOW_VECTORS] = {
240                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
241                                 L_PTE_RDONLY,
242                 .prot_l1   = PMD_TYPE_TABLE,
243                 .domain    = DOMAIN_USER,
244         },
245         [MT_HIGH_VECTORS] = {
246                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
247                                 L_PTE_USER | L_PTE_RDONLY,
248                 .prot_l1   = PMD_TYPE_TABLE,
249                 .domain    = DOMAIN_USER,
250         },
251         [MT_MEMORY] = {
252                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
253                 .prot_l1   = PMD_TYPE_TABLE,
254                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
255                 .domain    = DOMAIN_KERNEL,
256         },
257         [MT_ROM] = {
258                 .prot_sect = PMD_TYPE_SECT,
259                 .domain    = DOMAIN_KERNEL,
260         },
261         [MT_MEMORY_NONCACHED] = {
262                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
263                                 L_PTE_MT_BUFFERABLE,
264                 .prot_l1   = PMD_TYPE_TABLE,
265                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
266                 .domain    = DOMAIN_KERNEL,
267         },
268         [MT_MEMORY_DTCM] = {
269                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
270                                 L_PTE_XN,
271                 .prot_l1   = PMD_TYPE_TABLE,
272                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
273                 .domain    = DOMAIN_KERNEL,
274         },
275         [MT_MEMORY_ITCM] = {
276                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
277                 .prot_l1   = PMD_TYPE_TABLE,
278                 .domain    = DOMAIN_KERNEL,
279         },
280         [MT_MEMORY_SO] = {
281                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
282                                 L_PTE_MT_UNCACHED,
283                 .prot_l1   = PMD_TYPE_TABLE,
284                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
285                                 PMD_SECT_UNCACHED | PMD_SECT_XN,
286                 .domain    = DOMAIN_KERNEL,
287         },
288         [MT_MEMORY_DMA_READY] = {
289                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
290                 .prot_l1   = PMD_TYPE_TABLE,
291                 .domain    = DOMAIN_KERNEL,
292         },
293 };
294
295 const struct mem_type *get_mem_type(unsigned int type)
296 {
297         return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
298 }
299 EXPORT_SYMBOL(get_mem_type);
300
301 /*
302  * If the system supports huge pages and we are running with short descriptors,
303  * then compute the pmd and linux pte prot values for a huge page.
304  *
305  * These values are used by both the HugeTLB and THP code.
306  */
307 #if defined(CONFIG_SYS_SUPPORTS_HUGETLBFS) && !defined(CONFIG_ARM_LPAE)
308 pmdval_t arm_hugepmdprotval;
309 EXPORT_SYMBOL(arm_hugepmdprotval);
310
311 pteval_t arm_hugepteprotval;
312 EXPORT_SYMBOL(arm_hugepteprotval);
313 #endif
314
315
316 /*
317  * Adjust the PMD section entries according to the CPU in use.
318  */
319 static void __init build_mem_type_table(void)
320 {
321         struct cachepolicy *cp;
322         unsigned int cr = get_cr();
323         pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
324         int cpu_arch = cpu_architecture();
325         int i;
326
327         if (cpu_arch < CPU_ARCH_ARMv6) {
328 #if defined(CONFIG_CPU_DCACHE_DISABLE)
329                 if (cachepolicy > CPOLICY_BUFFERED)
330                         cachepolicy = CPOLICY_BUFFERED;
331 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
332                 if (cachepolicy > CPOLICY_WRITETHROUGH)
333                         cachepolicy = CPOLICY_WRITETHROUGH;
334 #endif
335         }
336         if (cpu_arch < CPU_ARCH_ARMv5) {
337                 if (cachepolicy >= CPOLICY_WRITEALLOC)
338                         cachepolicy = CPOLICY_WRITEBACK;
339                 ecc_mask = 0;
340         }
341         if (is_smp())
342                 cachepolicy = CPOLICY_WRITEALLOC;
343
344         /*
345          * Strip out features not present on earlier architectures.
346          * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
347          * without extended page tables don't have the 'Shared' bit.
348          */
349         if (cpu_arch < CPU_ARCH_ARMv5)
350                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
351                         mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
352         if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
353                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
354                         mem_types[i].prot_sect &= ~PMD_SECT_S;
355
356         /*
357          * ARMv5 and lower, bit 4 must be set for page tables (was: cache
358          * "update-able on write" bit on ARM610).  However, Xscale and
359          * Xscale3 require this bit to be cleared.
360          */
361         if (cpu_is_xscale() || cpu_is_xsc3()) {
362                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
363                         mem_types[i].prot_sect &= ~PMD_BIT4;
364                         mem_types[i].prot_l1 &= ~PMD_BIT4;
365                 }
366         } else if (cpu_arch < CPU_ARCH_ARMv6) {
367                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
368                         if (mem_types[i].prot_l1)
369                                 mem_types[i].prot_l1 |= PMD_BIT4;
370                         if (mem_types[i].prot_sect)
371                                 mem_types[i].prot_sect |= PMD_BIT4;
372                 }
373         }
374
375         /*
376          * Mark the device areas according to the CPU/architecture.
377          */
378         if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
379                 if (!cpu_is_xsc3()) {
380                         /*
381                          * Mark device regions on ARMv6+ as execute-never
382                          * to prevent speculative instruction fetches.
383                          */
384                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
385                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
386                         mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
387                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
388                 }
389                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
390                         /*
391                          * For ARMv7 with TEX remapping,
392                          * - shared device is SXCB=1100
393                          * - nonshared device is SXCB=0100
394                          * - write combine device mem is SXCB=0001
395                          * (Uncached Normal memory)
396                          */
397                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
398                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
399                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
400                 } else if (cpu_is_xsc3()) {
401                         /*
402                          * For Xscale3,
403                          * - shared device is TEXCB=00101
404                          * - nonshared device is TEXCB=01000
405                          * - write combine device mem is TEXCB=00100
406                          * (Inner/Outer Uncacheable in xsc3 parlance)
407                          */
408                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
409                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
410                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
411                 } else {
412                         /*
413                          * For ARMv6 and ARMv7 without TEX remapping,
414                          * - shared device is TEXCB=00001
415                          * - nonshared device is TEXCB=01000
416                          * - write combine device mem is TEXCB=00100
417                          * (Uncached Normal in ARMv6 parlance).
418                          */
419                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
420                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
421                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
422                 }
423         } else {
424                 /*
425                  * On others, write combining is "Uncached/Buffered"
426                  */
427                 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
428         }
429
430         /*
431          * Now deal with the memory-type mappings
432          */
433         cp = &cache_policies[cachepolicy];
434         vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
435
436         /*
437          * Only use write-through for non-SMP systems
438          */
439         if (!is_smp() && cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
440                 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
441
442         /*
443          * Enable CPU-specific coherency if supported.
444          * (Only available on XSC3 at the moment.)
445          */
446         if (arch_is_coherent() && cpu_is_xsc3()) {
447                 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
448                 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
449                 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
450                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
451                 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
452         }
453         /*
454          * ARMv6 and above have extended page tables.
455          */
456         if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
457 #ifndef CONFIG_ARM_LPAE
458                 /*
459                  * Mark cache clean areas and XIP ROM read only
460                  * from SVC mode and no access from userspace.
461                  */
462                 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
463                 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
464                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
465 #endif
466
467                 if (is_smp()) {
468                         /*
469                          * Mark memory with the "shared" attribute
470                          * for SMP systems
471                          */
472                         user_pgprot |= L_PTE_SHARED;
473                         kern_pgprot |= L_PTE_SHARED;
474                         vecs_pgprot |= L_PTE_SHARED;
475                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
476                         mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
477                         mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
478                         mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
479                         mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
480                         mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
481                         mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
482                         mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
483                         mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
484                 }
485         }
486
487         /*
488          * Non-cacheable Normal - intended for memory areas that must
489          * not cause dirty cache line writebacks when used
490          */
491         if (cpu_arch >= CPU_ARCH_ARMv6) {
492                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
493                         /* Non-cacheable Normal is XCB = 001 */
494                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
495                                 PMD_SECT_BUFFERED;
496                 } else {
497                         /* For both ARMv6 and non-TEX-remapping ARMv7 */
498                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
499                                 PMD_SECT_TEX(1);
500                 }
501         } else {
502                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
503         }
504
505 #ifdef CONFIG_ARM_LPAE
506         /*
507          * Do not generate access flag faults for the kernel mappings.
508          */
509         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
510                 mem_types[i].prot_pte |= PTE_EXT_AF;
511                 mem_types[i].prot_sect |= PMD_SECT_AF;
512         }
513         kern_pgprot |= PTE_EXT_AF;
514         vecs_pgprot |= PTE_EXT_AF;
515 #endif
516
517         for (i = 0; i < 16; i++) {
518                 pteval_t v = pgprot_val(protection_map[i]);
519                 protection_map[i] = __pgprot(v | user_pgprot);
520         }
521
522         mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
523         mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
524
525         pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
526         pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
527                                  L_PTE_DIRTY | kern_pgprot);
528
529         mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
530         mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
531         mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
532         mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
533         mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
534         mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
535         mem_types[MT_ROM].prot_sect |= cp->pmd;
536
537         switch (cp->pmd) {
538         case PMD_SECT_WT:
539                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
540                 break;
541         case PMD_SECT_WB:
542         case PMD_SECT_WBWA:
543                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
544                 break;
545         }
546         printk("Memory policy: ECC %sabled, Data cache %s\n",
547                 ecc_mask ? "en" : "dis", cp->policy);
548
549         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
550                 struct mem_type *t = &mem_types[i];
551                 if (t->prot_l1)
552                         t->prot_l1 |= PMD_DOMAIN(t->domain);
553                 if (t->prot_sect)
554                         t->prot_sect |= PMD_DOMAIN(t->domain);
555         }
556
557 #if defined(CONFIG_SYS_SUPPORTS_HUGETLBFS) && !defined(CONFIG_ARM_LPAE)
558         /*
559          * we assume all huge pages are user pages and that hardware access
560          * flag updates are disabled (i.e. SCTLR.AFE == 0b).
561          */
562         arm_hugepteprotval = mem_types[MT_MEMORY].prot_pte | L_PTE_USER | L_PTE_VALID;
563
564         arm_hugepmdprotval = mem_types[MT_MEMORY].prot_sect | PMD_SECT_AP_READ
565                                 | PMD_SECT_nG;
566 #endif
567
568 }
569
570 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
571 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
572                               unsigned long size, pgprot_t vma_prot)
573 {
574         if (!pfn_valid(pfn))
575                 return pgprot_noncached(vma_prot);
576         else if (file->f_flags & O_SYNC)
577                 return pgprot_writecombine(vma_prot);
578         return vma_prot;
579 }
580 EXPORT_SYMBOL(phys_mem_access_prot);
581 #endif
582
583 #define vectors_base()  (vectors_high() ? 0xffff0000 : 0)
584
585 static void __init *early_alloc(unsigned long sz)
586 {
587         void *ptr = __va(memblock_alloc(sz, sz));
588         memset(ptr, 0, sz);
589         return ptr;
590 }
591
592 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
593 {
594         if (pmd_none(*pmd)) {
595                 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
596                 __pmd_populate(pmd, __pa(pte), prot);
597         }
598         BUG_ON(pmd_bad(*pmd));
599         return pte_offset_kernel(pmd, addr);
600 }
601
602 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
603                                   unsigned long end, unsigned long pfn,
604                                   const struct mem_type *type)
605 {
606         pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
607         do {
608                 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
609                 pfn++;
610         } while (pte++, addr += PAGE_SIZE, addr != end);
611 }
612
613 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
614                                       unsigned long end, phys_addr_t phys,
615                                       const struct mem_type *type)
616 {
617         pmd_t *pmd = pmd_offset(pud, addr);
618
619         /*
620          * Try a section mapping - end, addr and phys must all be aligned
621          * to a section boundary.  Note that PMDs refer to the individual
622          * L1 entries, whereas PGDs refer to a group of L1 entries making
623          * up one logical pointer to an L2 table.
624          */
625         if (type->prot_sect && ((addr | end | phys) & ~SECTION_MASK) == 0) {
626                 pmd_t *p = pmd;
627
628 #ifndef CONFIG_ARM_LPAE
629                 if (addr & SECTION_SIZE)
630                         pmd++;
631 #endif
632
633                 do {
634                         *pmd = __pmd(phys | type->prot_sect);
635                         phys += SECTION_SIZE;
636                 } while (pmd++, addr += SECTION_SIZE, addr != end);
637
638                 flush_pmd_entry(p);
639         } else {
640                 /*
641                  * No need to loop; pte's aren't interested in the
642                  * individual L1 entries.
643                  */
644                 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
645         }
646 }
647
648 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
649         unsigned long phys, const struct mem_type *type)
650 {
651         pud_t *pud = pud_offset(pgd, addr);
652         unsigned long next;
653
654         do {
655                 next = pud_addr_end(addr, end);
656                 alloc_init_section(pud, addr, next, phys, type);
657                 phys += next - addr;
658         } while (pud++, addr = next, addr != end);
659 }
660
661 #ifndef CONFIG_ARM_LPAE
662 static void __init create_36bit_mapping(struct map_desc *md,
663                                         const struct mem_type *type)
664 {
665         unsigned long addr, length, end;
666         phys_addr_t phys;
667         pgd_t *pgd;
668
669         addr = md->virtual;
670         phys = __pfn_to_phys(md->pfn);
671         length = PAGE_ALIGN(md->length);
672
673         if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
674                 printk(KERN_ERR "MM: CPU does not support 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         /* N.B. ARMv6 supersections are only defined to work with domain 0.
681          *      Since domain assignments can in fact be arbitrary, the
682          *      'domain == 0' check below is required to insure that ARMv6
683          *      supersections are only allocated for domain 0 regardless
684          *      of the actual domain assignments in use.
685          */
686         if (type->domain) {
687                 printk(KERN_ERR "MM: invalid domain in supersection "
688                        "mapping for 0x%08llx at 0x%08lx\n",
689                        (long long)__pfn_to_phys((u64)md->pfn), addr);
690                 return;
691         }
692
693         if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
694                 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
695                        " at 0x%08lx invalid alignment\n",
696                        (long long)__pfn_to_phys((u64)md->pfn), addr);
697                 return;
698         }
699
700         /*
701          * Shift bits [35:32] of address into bits [23:20] of PMD
702          * (See ARMv6 spec).
703          */
704         phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
705
706         pgd = pgd_offset_k(addr);
707         end = addr + length;
708         do {
709                 pud_t *pud = pud_offset(pgd, addr);
710                 pmd_t *pmd = pmd_offset(pud, addr);
711                 int i;
712
713                 for (i = 0; i < 16; i++)
714                         *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
715
716                 addr += SUPERSECTION_SIZE;
717                 phys += SUPERSECTION_SIZE;
718                 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
719         } while (addr != end);
720 }
721 #endif  /* !CONFIG_ARM_LPAE */
722
723 /*
724  * Create the page directory entries and any necessary
725  * page tables for the mapping specified by `md'.  We
726  * are able to cope here with varying sizes and address
727  * offsets, and we take full advantage of sections and
728  * supersections.
729  */
730 static void __init create_mapping(struct map_desc *md)
731 {
732         unsigned long addr, length, end;
733         phys_addr_t phys;
734         const struct mem_type *type;
735         pgd_t *pgd;
736
737         if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
738                 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
739                        " at 0x%08lx in user region\n",
740                        (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
741                 return;
742         }
743
744         if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
745             md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
746                 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
747                        " at 0x%08lx overlaps vmalloc space\n",
748                        (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
749         }
750
751         type = &mem_types[md->type];
752
753 #ifndef CONFIG_ARM_LPAE
754         /*
755          * Catch 36-bit addresses
756          */
757         if (md->pfn >= 0x100000) {
758                 create_36bit_mapping(md, type);
759                 return;
760         }
761 #endif
762
763         addr = md->virtual & PAGE_MASK;
764         phys = __pfn_to_phys(md->pfn);
765         length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
766
767         if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
768                 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
769                        "be mapped using pages, ignoring.\n",
770                        (long long)__pfn_to_phys(md->pfn), addr);
771                 return;
772         }
773
774         pgd = pgd_offset_k(addr);
775         end = addr + length;
776         do {
777                 unsigned long next = pgd_addr_end(addr, end);
778
779                 alloc_init_pud(pgd, addr, next, phys, type);
780
781                 phys += next - addr;
782                 addr = next;
783         } while (pgd++, addr != end);
784 }
785
786 /*
787  * Create the architecture specific mappings
788  */
789 void __init iotable_init(struct map_desc *io_desc, int nr)
790 {
791         int i;
792
793         for (i = 0; i < nr; i++)
794                 create_mapping(io_desc + i);
795 }
796
797 static void * __initdata vmalloc_min = (void *)(VMALLOC_END - SZ_128M);
798
799 /*
800  * vmalloc=size forces the vmalloc area to be exactly 'size'
801  * bytes. This can be used to increase (or decrease) the vmalloc
802  * area - the default is 128m.
803  */
804 static int __init early_vmalloc(char *arg)
805 {
806         unsigned long vmalloc_reserve = memparse(arg, NULL);
807
808         if (vmalloc_reserve < SZ_16M) {
809                 vmalloc_reserve = SZ_16M;
810                 printk(KERN_WARNING
811                         "vmalloc area too small, limiting to %luMB\n",
812                         vmalloc_reserve >> 20);
813         }
814
815         if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
816                 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
817                 printk(KERN_WARNING
818                         "vmalloc area is too big, limiting to %luMB\n",
819                         vmalloc_reserve >> 20);
820         }
821
822         vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
823         return 0;
824 }
825 early_param("vmalloc", early_vmalloc);
826
827 phys_addr_t arm_lowmem_limit __initdata = 0;
828
829 void __init sanity_check_meminfo(void)
830 {
831         int i, j, highmem = 0;
832
833         for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
834                 struct membank *bank = &meminfo.bank[j];
835                 *bank = meminfo.bank[i];
836
837                 if (bank->start > ULONG_MAX)
838                         highmem = 1;
839
840 #ifdef CONFIG_HIGHMEM
841                 if (__va(bank->start) >= vmalloc_min ||
842                     __va(bank->start) < (void *)PAGE_OFFSET)
843                         highmem = 1;
844
845                 bank->highmem = highmem;
846
847                 /*
848                  * Split those memory banks which are partially overlapping
849                  * the vmalloc area greatly simplifying things later.
850                  */
851                 if (!highmem && __va(bank->start) < vmalloc_min &&
852                     bank->size > vmalloc_min - __va(bank->start)) {
853                         if (meminfo.nr_banks >= NR_BANKS) {
854                                 printk(KERN_CRIT "NR_BANKS too low, "
855                                                  "ignoring high memory\n");
856                         } else {
857                                 memmove(bank + 1, bank,
858                                         (meminfo.nr_banks - i) * sizeof(*bank));
859                                 meminfo.nr_banks++;
860                                 i++;
861                                 bank[1].size -= vmalloc_min - __va(bank->start);
862                                 bank[1].start = __pa(vmalloc_min - 1) + 1;
863                                 bank[1].highmem = highmem = 1;
864                                 j++;
865                         }
866                         bank->size = vmalloc_min - __va(bank->start);
867                 }
868 #else
869                 bank->highmem = highmem;
870
871                 /*
872                  * Highmem banks not allowed with !CONFIG_HIGHMEM.
873                  */
874                 if (highmem) {
875                         printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
876                                "(!CONFIG_HIGHMEM).\n",
877                                (unsigned long long)bank->start,
878                                (unsigned long long)bank->start + bank->size - 1);
879                         continue;
880                 }
881
882                 /*
883                  * Check whether this memory bank would entirely overlap
884                  * the vmalloc area.
885                  */
886                 if (__va(bank->start) >= vmalloc_min ||
887                     __va(bank->start) < (void *)PAGE_OFFSET) {
888                         printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
889                                "(vmalloc region overlap).\n",
890                                (unsigned long long)bank->start,
891                                (unsigned long long)bank->start + bank->size - 1);
892                         continue;
893                 }
894
895                 /*
896                  * Check whether this memory bank would partially overlap
897                  * the vmalloc area.
898                  */
899                 if (__va(bank->start + bank->size) > vmalloc_min ||
900                     __va(bank->start + bank->size) < __va(bank->start)) {
901                         unsigned long newsize = vmalloc_min - __va(bank->start);
902                         printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
903                                "to -%.8llx (vmalloc region overlap).\n",
904                                (unsigned long long)bank->start,
905                                (unsigned long long)bank->start + bank->size - 1,
906                                (unsigned long long)bank->start + newsize - 1);
907                         bank->size = newsize;
908                 }
909 #endif
910                 if (!bank->highmem && bank->start + bank->size > arm_lowmem_limit)
911                         arm_lowmem_limit = bank->start + bank->size;
912
913                 j++;
914         }
915 #ifdef CONFIG_HIGHMEM
916         if (highmem) {
917                 const char *reason = NULL;
918
919                 if (cache_is_vipt_aliasing()) {
920                         /*
921                          * Interactions between kmap and other mappings
922                          * make highmem support with aliasing VIPT caches
923                          * rather difficult.
924                          */
925                         reason = "with VIPT aliasing cache";
926                 }
927                 if (reason) {
928                         printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
929                                 reason);
930                         while (j > 0 && meminfo.bank[j - 1].highmem)
931                                 j--;
932                 }
933         }
934 #endif
935         meminfo.nr_banks = j;
936         high_memory = __va(arm_lowmem_limit - 1) + 1;
937         memblock_set_current_limit(arm_lowmem_limit);
938 }
939
940 static inline void prepare_page_table(void)
941 {
942         unsigned long addr;
943         phys_addr_t end;
944
945         /*
946          * Clear out all the mappings below the kernel image.
947          */
948         for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
949                 pmd_clear(pmd_off_k(addr));
950
951 #ifdef CONFIG_XIP_KERNEL
952         /* The XIP kernel is mapped in the module area -- skip over it */
953         addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
954 #endif
955         for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
956                 pmd_clear(pmd_off_k(addr));
957
958         /*
959          * Find the end of the first block of lowmem.
960          */
961         end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
962         if (end >= arm_lowmem_limit)
963                 end = arm_lowmem_limit;
964
965         /*
966          * Clear out all the kernel space mappings, except for the first
967          * memory bank, up to the end of the vmalloc region.
968          */
969         for (addr = __phys_to_virt(end);
970              addr < VMALLOC_END; addr += PMD_SIZE)
971                 pmd_clear(pmd_off_k(addr));
972 }
973
974 #ifdef CONFIG_ARM_LPAE
975 /* the first page is reserved for pgd */
976 #define SWAPPER_PG_DIR_SIZE     (PAGE_SIZE + \
977                                  PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
978 #else
979 #define SWAPPER_PG_DIR_SIZE     (PTRS_PER_PGD * sizeof(pgd_t))
980 #endif
981
982 /*
983  * Reserve the special regions of memory
984  */
985 void __init arm_mm_memblock_reserve(void)
986 {
987         /*
988          * Reserve the page tables.  These are already in use,
989          * and can only be in node 0.
990          */
991         memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
992
993 #ifdef CONFIG_SA1111
994         /*
995          * Because of the SA1111 DMA bug, we want to preserve our
996          * precious DMA-able memory...
997          */
998         memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
999 #endif
1000 }
1001
1002 /*
1003  * Set up device the mappings.  Since we clear out the page tables for all
1004  * mappings above VMALLOC_END, we will remove any debug device mappings.
1005  * This means you have to be careful how you debug this function, or any
1006  * called function.  This means you can't use any function or debugging
1007  * method which may touch any device, otherwise the kernel _will_ crash.
1008  */
1009 static void __init devicemaps_init(struct machine_desc *mdesc)
1010 {
1011         struct map_desc map;
1012         unsigned long addr;
1013
1014         /*
1015          * Allocate the vector page early.
1016          */
1017         vectors_page = early_alloc(PAGE_SIZE);
1018
1019         for (addr = VMALLOC_END; addr; addr += PMD_SIZE)
1020                 pmd_clear(pmd_off_k(addr));
1021
1022         /*
1023          * Map the kernel if it is XIP.
1024          * It is always first in the modulearea.
1025          */
1026 #ifdef CONFIG_XIP_KERNEL
1027         map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1028         map.virtual = MODULES_VADDR;
1029         map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1030         map.type = MT_ROM;
1031         create_mapping(&map);
1032 #endif
1033
1034         /*
1035          * Map the cache flushing regions.
1036          */
1037 #ifdef FLUSH_BASE
1038         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1039         map.virtual = FLUSH_BASE;
1040         map.length = SZ_1M;
1041         map.type = MT_CACHECLEAN;
1042         create_mapping(&map);
1043 #endif
1044 #ifdef FLUSH_BASE_MINICACHE
1045         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1046         map.virtual = FLUSH_BASE_MINICACHE;
1047         map.length = SZ_1M;
1048         map.type = MT_MINICLEAN;
1049         create_mapping(&map);
1050 #endif
1051
1052         /*
1053          * Create a mapping for the machine vectors at the high-vectors
1054          * location (0xffff0000).  If we aren't using high-vectors, also
1055          * create a mapping at the low-vectors virtual address.
1056          */
1057         map.pfn = __phys_to_pfn(virt_to_phys(vectors_page));
1058         map.virtual = 0xffff0000;
1059         map.length = PAGE_SIZE;
1060         map.type = MT_HIGH_VECTORS;
1061         create_mapping(&map);
1062
1063         if (!vectors_high()) {
1064                 map.virtual = 0;
1065                 map.type = MT_LOW_VECTORS;
1066                 create_mapping(&map);
1067         }
1068
1069         /*
1070          * Ask the machine support to map in the statically mapped devices.
1071          */
1072         if (mdesc->map_io)
1073                 mdesc->map_io();
1074
1075         /*
1076          * Finally flush the caches and tlb to ensure that we're in a
1077          * consistent state wrt the writebuffer.  This also ensures that
1078          * any write-allocated cache lines in the vector page are written
1079          * back.  After this point, we can start to touch devices again.
1080          */
1081         local_flush_tlb_all();
1082         flush_cache_all();
1083 }
1084
1085 static void __init kmap_init(void)
1086 {
1087 #ifdef CONFIG_HIGHMEM
1088         pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1089                 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1090 #endif
1091 }
1092
1093 static void __init map_lowmem(void)
1094 {
1095         struct memblock_region *reg;
1096
1097         /* Map all the lowmem memory banks. */
1098         for_each_memblock(memory, reg) {
1099                 phys_addr_t start = reg->base;
1100                 phys_addr_t end = start + reg->size;
1101                 struct map_desc map;
1102
1103                 if (end > arm_lowmem_limit)
1104                         end = arm_lowmem_limit;
1105                 if (start >= end)
1106                         break;
1107
1108                 map.pfn = __phys_to_pfn(start);
1109                 map.virtual = __phys_to_virt(start);
1110                 map.length = end - start;
1111                 map.type = MT_MEMORY;
1112
1113                 create_mapping(&map);
1114         }
1115 }
1116
1117 /*
1118  * paging_init() sets up the page tables, initialises the zone memory
1119  * maps, and sets up the zero page, bad page and bad page tables.
1120  */
1121 void __init paging_init(struct machine_desc *mdesc)
1122 {
1123         void *zero_page;
1124
1125         memblock_set_current_limit(arm_lowmem_limit);
1126
1127         build_mem_type_table();
1128         prepare_page_table();
1129         map_lowmem();
1130         dma_contiguous_remap();
1131         devicemaps_init(mdesc);
1132         kmap_init();
1133
1134         top_pmd = pmd_off_k(0xffff0000);
1135
1136         /* allocate the zero page. */
1137         zero_page = early_alloc(PAGE_SIZE);
1138
1139         bootmem_init();
1140
1141         empty_zero_page = virt_to_page(zero_page);
1142         __flush_dcache_page(NULL, empty_zero_page);
1143 }