ARM: early_pte_alloc()
[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/bootmem.h>
15 #include <linux/mman.h>
16 #include <linux/nodemask.h>
17 #include <linux/sort.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
28 #include <asm/mach/arch.h>
29 #include <asm/mach/map.h>
30
31 #include "mm.h"
32
33 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
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         unsigned int    pmd;
65         unsigned int    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 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
164 static int __init noalign_setup(char *__unused)
165 {
166         cr_alignment &= ~CR_A;
167         cr_no_alignment &= ~CR_A;
168         set_cr(cr_alignment);
169         return 1;
170 }
171 __setup("noalign", noalign_setup);
172
173 #ifndef CONFIG_SMP
174 void adjust_cr(unsigned long mask, unsigned long set)
175 {
176         unsigned long flags;
177
178         mask &= ~CR_A;
179
180         set &= mask;
181
182         local_irq_save(flags);
183
184         cr_no_alignment = (cr_no_alignment & ~mask) | set;
185         cr_alignment = (cr_alignment & ~mask) | set;
186
187         set_cr((get_cr() & ~mask) | set);
188
189         local_irq_restore(flags);
190 }
191 #endif
192
193 #define PROT_PTE_DEVICE         L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_WRITE
194 #define PROT_SECT_DEVICE        PMD_TYPE_SECT|PMD_SECT_AP_WRITE
195
196 static struct mem_type mem_types[] = {
197         [MT_DEVICE] = {           /* Strongly ordered / ARMv6 shared device */
198                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
199                                   L_PTE_SHARED,
200                 .prot_l1        = PMD_TYPE_TABLE,
201                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_S,
202                 .domain         = DOMAIN_IO,
203         },
204         [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
205                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
206                 .prot_l1        = PMD_TYPE_TABLE,
207                 .prot_sect      = PROT_SECT_DEVICE,
208                 .domain         = DOMAIN_IO,
209         },
210         [MT_DEVICE_CACHED] = {    /* ioremap_cached */
211                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
212                 .prot_l1        = PMD_TYPE_TABLE,
213                 .prot_sect      = PROT_SECT_DEVICE | PMD_SECT_WB,
214                 .domain         = DOMAIN_IO,
215         },      
216         [MT_DEVICE_WC] = {      /* ioremap_wc */
217                 .prot_pte       = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
218                 .prot_l1        = PMD_TYPE_TABLE,
219                 .prot_sect      = PROT_SECT_DEVICE,
220                 .domain         = DOMAIN_IO,
221         },
222         [MT_UNCACHED] = {
223                 .prot_pte       = PROT_PTE_DEVICE,
224                 .prot_l1        = PMD_TYPE_TABLE,
225                 .prot_sect      = PMD_TYPE_SECT | PMD_SECT_XN,
226                 .domain         = DOMAIN_IO,
227         },
228         [MT_CACHECLEAN] = {
229                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
230                 .domain    = DOMAIN_KERNEL,
231         },
232         [MT_MINICLEAN] = {
233                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
234                 .domain    = DOMAIN_KERNEL,
235         },
236         [MT_LOW_VECTORS] = {
237                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
238                                 L_PTE_EXEC,
239                 .prot_l1   = PMD_TYPE_TABLE,
240                 .domain    = DOMAIN_USER,
241         },
242         [MT_HIGH_VECTORS] = {
243                 .prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
244                                 L_PTE_USER | L_PTE_EXEC,
245                 .prot_l1   = PMD_TYPE_TABLE,
246                 .domain    = DOMAIN_USER,
247         },
248         [MT_MEMORY] = {
249                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
250                 .domain    = DOMAIN_KERNEL,
251         },
252         [MT_ROM] = {
253                 .prot_sect = PMD_TYPE_SECT,
254                 .domain    = DOMAIN_KERNEL,
255         },
256         [MT_MEMORY_NONCACHED] = {
257                 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
258                 .domain    = DOMAIN_KERNEL,
259         },
260 };
261
262 const struct mem_type *get_mem_type(unsigned int type)
263 {
264         return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
265 }
266 EXPORT_SYMBOL(get_mem_type);
267
268 /*
269  * Adjust the PMD section entries according to the CPU in use.
270  */
271 static void __init build_mem_type_table(void)
272 {
273         struct cachepolicy *cp;
274         unsigned int cr = get_cr();
275         unsigned int user_pgprot, kern_pgprot, vecs_pgprot;
276         int cpu_arch = cpu_architecture();
277         int i;
278
279         if (cpu_arch < CPU_ARCH_ARMv6) {
280 #if defined(CONFIG_CPU_DCACHE_DISABLE)
281                 if (cachepolicy > CPOLICY_BUFFERED)
282                         cachepolicy = CPOLICY_BUFFERED;
283 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
284                 if (cachepolicy > CPOLICY_WRITETHROUGH)
285                         cachepolicy = CPOLICY_WRITETHROUGH;
286 #endif
287         }
288         if (cpu_arch < CPU_ARCH_ARMv5) {
289                 if (cachepolicy >= CPOLICY_WRITEALLOC)
290                         cachepolicy = CPOLICY_WRITEBACK;
291                 ecc_mask = 0;
292         }
293 #ifdef CONFIG_SMP
294         cachepolicy = CPOLICY_WRITEALLOC;
295 #endif
296
297         /*
298          * Strip out features not present on earlier architectures.
299          * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
300          * without extended page tables don't have the 'Shared' bit.
301          */
302         if (cpu_arch < CPU_ARCH_ARMv5)
303                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
304                         mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
305         if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
306                 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
307                         mem_types[i].prot_sect &= ~PMD_SECT_S;
308
309         /*
310          * ARMv5 and lower, bit 4 must be set for page tables (was: cache
311          * "update-able on write" bit on ARM610).  However, Xscale and
312          * Xscale3 require this bit to be cleared.
313          */
314         if (cpu_is_xscale() || cpu_is_xsc3()) {
315                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
316                         mem_types[i].prot_sect &= ~PMD_BIT4;
317                         mem_types[i].prot_l1 &= ~PMD_BIT4;
318                 }
319         } else if (cpu_arch < CPU_ARCH_ARMv6) {
320                 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
321                         if (mem_types[i].prot_l1)
322                                 mem_types[i].prot_l1 |= PMD_BIT4;
323                         if (mem_types[i].prot_sect)
324                                 mem_types[i].prot_sect |= PMD_BIT4;
325                 }
326         }
327
328         /*
329          * Mark the device areas according to the CPU/architecture.
330          */
331         if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
332                 if (!cpu_is_xsc3()) {
333                         /*
334                          * Mark device regions on ARMv6+ as execute-never
335                          * to prevent speculative instruction fetches.
336                          */
337                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
338                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
339                         mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
340                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
341                 }
342                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
343                         /*
344                          * For ARMv7 with TEX remapping,
345                          * - shared device is SXCB=1100
346                          * - nonshared device is SXCB=0100
347                          * - write combine device mem is SXCB=0001
348                          * (Uncached Normal memory)
349                          */
350                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
351                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
352                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
353                 } else if (cpu_is_xsc3()) {
354                         /*
355                          * For Xscale3,
356                          * - shared device is TEXCB=00101
357                          * - nonshared device is TEXCB=01000
358                          * - write combine device mem is TEXCB=00100
359                          * (Inner/Outer Uncacheable in xsc3 parlance)
360                          */
361                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
362                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
363                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
364                 } else {
365                         /*
366                          * For ARMv6 and ARMv7 without TEX remapping,
367                          * - shared device is TEXCB=00001
368                          * - nonshared device is TEXCB=01000
369                          * - write combine device mem is TEXCB=00100
370                          * (Uncached Normal in ARMv6 parlance).
371                          */
372                         mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
373                         mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
374                         mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
375                 }
376         } else {
377                 /*
378                  * On others, write combining is "Uncached/Buffered"
379                  */
380                 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
381         }
382
383         /*
384          * Now deal with the memory-type mappings
385          */
386         cp = &cache_policies[cachepolicy];
387         vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
388
389 #ifndef CONFIG_SMP
390         /*
391          * Only use write-through for non-SMP systems
392          */
393         if (cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
394                 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
395 #endif
396
397         /*
398          * Enable CPU-specific coherency if supported.
399          * (Only available on XSC3 at the moment.)
400          */
401         if (arch_is_coherent() && cpu_is_xsc3())
402                 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
403
404         /*
405          * ARMv6 and above have extended page tables.
406          */
407         if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
408                 /*
409                  * Mark cache clean areas and XIP ROM read only
410                  * from SVC mode and no access from userspace.
411                  */
412                 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
413                 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
414                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
415
416 #ifdef CONFIG_SMP
417                 /*
418                  * Mark memory with the "shared" attribute for SMP systems
419                  */
420                 user_pgprot |= L_PTE_SHARED;
421                 kern_pgprot |= L_PTE_SHARED;
422                 vecs_pgprot |= L_PTE_SHARED;
423                 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
424                 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
425                 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
426                 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
427                 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
428                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
429 #endif
430         }
431
432         /*
433          * Non-cacheable Normal - intended for memory areas that must
434          * not cause dirty cache line writebacks when used
435          */
436         if (cpu_arch >= CPU_ARCH_ARMv6) {
437                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
438                         /* Non-cacheable Normal is XCB = 001 */
439                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
440                                 PMD_SECT_BUFFERED;
441                 } else {
442                         /* For both ARMv6 and non-TEX-remapping ARMv7 */
443                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
444                                 PMD_SECT_TEX(1);
445                 }
446         } else {
447                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
448         }
449
450         for (i = 0; i < 16; i++) {
451                 unsigned long v = pgprot_val(protection_map[i]);
452                 protection_map[i] = __pgprot(v | user_pgprot);
453         }
454
455         mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
456         mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
457
458         pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
459         pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
460                                  L_PTE_DIRTY | L_PTE_WRITE | kern_pgprot);
461
462         mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
463         mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
464         mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
465         mem_types[MT_ROM].prot_sect |= cp->pmd;
466
467         switch (cp->pmd) {
468         case PMD_SECT_WT:
469                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
470                 break;
471         case PMD_SECT_WB:
472         case PMD_SECT_WBWA:
473                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
474                 break;
475         }
476         printk("Memory policy: ECC %sabled, Data cache %s\n",
477                 ecc_mask ? "en" : "dis", cp->policy);
478
479         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
480                 struct mem_type *t = &mem_types[i];
481                 if (t->prot_l1)
482                         t->prot_l1 |= PMD_DOMAIN(t->domain);
483                 if (t->prot_sect)
484                         t->prot_sect |= PMD_DOMAIN(t->domain);
485         }
486 }
487
488 #define vectors_base()  (vectors_high() ? 0xffff0000 : 0)
489
490 static void __init *early_alloc(unsigned long sz)
491 {
492         return alloc_bootmem_low_pages(sz);
493 }
494
495 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
496 {
497         if (pmd_none(*pmd)) {
498                 pte_t *pte = early_alloc(2 * PTRS_PER_PTE * sizeof(pte_t));
499                 __pmd_populate(pmd, __pa(pte) | prot);
500         }
501         BUG_ON(pmd_bad(*pmd));
502         return pte_offset_kernel(pmd, addr);
503 }
504
505 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
506                                   unsigned long end, unsigned long pfn,
507                                   const struct mem_type *type)
508 {
509         pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
510         do {
511                 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
512                 pfn++;
513         } while (pte++, addr += PAGE_SIZE, addr != end);
514 }
515
516 static void __init alloc_init_section(pgd_t *pgd, unsigned long addr,
517                                       unsigned long end, unsigned long phys,
518                                       const struct mem_type *type)
519 {
520         pmd_t *pmd = pmd_offset(pgd, addr);
521
522         /*
523          * Try a section mapping - end, addr and phys must all be aligned
524          * to a section boundary.  Note that PMDs refer to the individual
525          * L1 entries, whereas PGDs refer to a group of L1 entries making
526          * up one logical pointer to an L2 table.
527          */
528         if (((addr | end | phys) & ~SECTION_MASK) == 0) {
529                 pmd_t *p = pmd;
530
531                 if (addr & SECTION_SIZE)
532                         pmd++;
533
534                 do {
535                         *pmd = __pmd(phys | type->prot_sect);
536                         phys += SECTION_SIZE;
537                 } while (pmd++, addr += SECTION_SIZE, addr != end);
538
539                 flush_pmd_entry(p);
540         } else {
541                 /*
542                  * No need to loop; pte's aren't interested in the
543                  * individual L1 entries.
544                  */
545                 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
546         }
547 }
548
549 static void __init create_36bit_mapping(struct map_desc *md,
550                                         const struct mem_type *type)
551 {
552         unsigned long phys, addr, length, end;
553         pgd_t *pgd;
554
555         addr = md->virtual;
556         phys = (unsigned long)__pfn_to_phys(md->pfn);
557         length = PAGE_ALIGN(md->length);
558
559         if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
560                 printk(KERN_ERR "MM: CPU does not support supersection "
561                        "mapping for 0x%08llx at 0x%08lx\n",
562                        __pfn_to_phys((u64)md->pfn), addr);
563                 return;
564         }
565
566         /* N.B. ARMv6 supersections are only defined to work with domain 0.
567          *      Since domain assignments can in fact be arbitrary, the
568          *      'domain == 0' check below is required to insure that ARMv6
569          *      supersections are only allocated for domain 0 regardless
570          *      of the actual domain assignments in use.
571          */
572         if (type->domain) {
573                 printk(KERN_ERR "MM: invalid domain in supersection "
574                        "mapping for 0x%08llx at 0x%08lx\n",
575                        __pfn_to_phys((u64)md->pfn), addr);
576                 return;
577         }
578
579         if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
580                 printk(KERN_ERR "MM: cannot create mapping for "
581                        "0x%08llx at 0x%08lx invalid alignment\n",
582                        __pfn_to_phys((u64)md->pfn), addr);
583                 return;
584         }
585
586         /*
587          * Shift bits [35:32] of address into bits [23:20] of PMD
588          * (See ARMv6 spec).
589          */
590         phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
591
592         pgd = pgd_offset_k(addr);
593         end = addr + length;
594         do {
595                 pmd_t *pmd = pmd_offset(pgd, addr);
596                 int i;
597
598                 for (i = 0; i < 16; i++)
599                         *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
600
601                 addr += SUPERSECTION_SIZE;
602                 phys += SUPERSECTION_SIZE;
603                 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
604         } while (addr != end);
605 }
606
607 /*
608  * Create the page directory entries and any necessary
609  * page tables for the mapping specified by `md'.  We
610  * are able to cope here with varying sizes and address
611  * offsets, and we take full advantage of sections and
612  * supersections.
613  */
614 static void __init create_mapping(struct map_desc *md)
615 {
616         unsigned long phys, addr, length, end;
617         const struct mem_type *type;
618         pgd_t *pgd;
619
620         if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
621                 printk(KERN_WARNING "BUG: not creating mapping for "
622                        "0x%08llx at 0x%08lx in user region\n",
623                        __pfn_to_phys((u64)md->pfn), md->virtual);
624                 return;
625         }
626
627         if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
628             md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
629                 printk(KERN_WARNING "BUG: mapping for 0x%08llx at 0x%08lx "
630                        "overlaps vmalloc space\n",
631                        __pfn_to_phys((u64)md->pfn), md->virtual);
632         }
633
634         type = &mem_types[md->type];
635
636         /*
637          * Catch 36-bit addresses
638          */
639         if (md->pfn >= 0x100000) {
640                 create_36bit_mapping(md, type);
641                 return;
642         }
643
644         addr = md->virtual & PAGE_MASK;
645         phys = (unsigned long)__pfn_to_phys(md->pfn);
646         length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
647
648         if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
649                 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
650                        "be mapped using pages, ignoring.\n",
651                        __pfn_to_phys(md->pfn), addr);
652                 return;
653         }
654
655         pgd = pgd_offset_k(addr);
656         end = addr + length;
657         do {
658                 unsigned long next = pgd_addr_end(addr, end);
659
660                 alloc_init_section(pgd, addr, next, phys, type);
661
662                 phys += next - addr;
663                 addr = next;
664         } while (pgd++, addr != end);
665 }
666
667 /*
668  * Create the architecture specific mappings
669  */
670 void __init iotable_init(struct map_desc *io_desc, int nr)
671 {
672         int i;
673
674         for (i = 0; i < nr; i++)
675                 create_mapping(io_desc + i);
676 }
677
678 static void * __initdata vmalloc_min = (void *)(VMALLOC_END - SZ_128M);
679
680 /*
681  * vmalloc=size forces the vmalloc area to be exactly 'size'
682  * bytes. This can be used to increase (or decrease) the vmalloc
683  * area - the default is 128m.
684  */
685 static int __init early_vmalloc(char *arg)
686 {
687         unsigned long vmalloc_reserve = memparse(arg, NULL);
688
689         if (vmalloc_reserve < SZ_16M) {
690                 vmalloc_reserve = SZ_16M;
691                 printk(KERN_WARNING
692                         "vmalloc area too small, limiting to %luMB\n",
693                         vmalloc_reserve >> 20);
694         }
695
696         if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
697                 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
698                 printk(KERN_WARNING
699                         "vmalloc area is too big, limiting to %luMB\n",
700                         vmalloc_reserve >> 20);
701         }
702
703         vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
704         return 0;
705 }
706 early_param("vmalloc", early_vmalloc);
707
708 static void __init sanity_check_meminfo(void)
709 {
710         int i, j, highmem = 0;
711
712         for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
713                 struct membank *bank = &meminfo.bank[j];
714                 *bank = meminfo.bank[i];
715
716 #ifdef CONFIG_HIGHMEM
717                 if (__va(bank->start) > vmalloc_min ||
718                     __va(bank->start) < (void *)PAGE_OFFSET)
719                         highmem = 1;
720
721                 bank->highmem = highmem;
722
723                 /*
724                  * Split those memory banks which are partially overlapping
725                  * the vmalloc area greatly simplifying things later.
726                  */
727                 if (__va(bank->start) < vmalloc_min &&
728                     bank->size > vmalloc_min - __va(bank->start)) {
729                         if (meminfo.nr_banks >= NR_BANKS) {
730                                 printk(KERN_CRIT "NR_BANKS too low, "
731                                                  "ignoring high memory\n");
732                         } else {
733                                 memmove(bank + 1, bank,
734                                         (meminfo.nr_banks - i) * sizeof(*bank));
735                                 meminfo.nr_banks++;
736                                 i++;
737                                 bank[1].size -= vmalloc_min - __va(bank->start);
738                                 bank[1].start = __pa(vmalloc_min - 1) + 1;
739                                 bank[1].highmem = highmem = 1;
740                                 j++;
741                         }
742                         bank->size = vmalloc_min - __va(bank->start);
743                 }
744 #else
745                 bank->highmem = highmem;
746
747                 /*
748                  * Check whether this memory bank would entirely overlap
749                  * the vmalloc area.
750                  */
751                 if (__va(bank->start) >= vmalloc_min ||
752                     __va(bank->start) < (void *)PAGE_OFFSET) {
753                         printk(KERN_NOTICE "Ignoring RAM at %.8lx-%.8lx "
754                                "(vmalloc region overlap).\n",
755                                bank->start, bank->start + bank->size - 1);
756                         continue;
757                 }
758
759                 /*
760                  * Check whether this memory bank would partially overlap
761                  * the vmalloc area.
762                  */
763                 if (__va(bank->start + bank->size) > vmalloc_min ||
764                     __va(bank->start + bank->size) < __va(bank->start)) {
765                         unsigned long newsize = vmalloc_min - __va(bank->start);
766                         printk(KERN_NOTICE "Truncating RAM at %.8lx-%.8lx "
767                                "to -%.8lx (vmalloc region overlap).\n",
768                                bank->start, bank->start + bank->size - 1,
769                                bank->start + newsize - 1);
770                         bank->size = newsize;
771                 }
772 #endif
773                 j++;
774         }
775 #ifdef CONFIG_HIGHMEM
776         if (highmem) {
777                 const char *reason = NULL;
778
779                 if (cache_is_vipt_aliasing()) {
780                         /*
781                          * Interactions between kmap and other mappings
782                          * make highmem support with aliasing VIPT caches
783                          * rather difficult.
784                          */
785                         reason = "with VIPT aliasing cache";
786 #ifdef CONFIG_SMP
787                 } else if (tlb_ops_need_broadcast()) {
788                         /*
789                          * kmap_high needs to occasionally flush TLB entries,
790                          * however, if the TLB entries need to be broadcast
791                          * we may deadlock:
792                          *  kmap_high(irqs off)->flush_all_zero_pkmaps->
793                          *  flush_tlb_kernel_range->smp_call_function_many
794                          *   (must not be called with irqs off)
795                          */
796                         reason = "without hardware TLB ops broadcasting";
797 #endif
798                 }
799                 if (reason) {
800                         printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
801                                 reason);
802                         while (j > 0 && meminfo.bank[j - 1].highmem)
803                                 j--;
804                 }
805         }
806 #endif
807         meminfo.nr_banks = j;
808 }
809
810 static inline void prepare_page_table(void)
811 {
812         unsigned long addr;
813
814         /*
815          * Clear out all the mappings below the kernel image.
816          */
817         for (addr = 0; addr < MODULES_VADDR; addr += PGDIR_SIZE)
818                 pmd_clear(pmd_off_k(addr));
819
820 #ifdef CONFIG_XIP_KERNEL
821         /* The XIP kernel is mapped in the module area -- skip over it */
822         addr = ((unsigned long)_etext + PGDIR_SIZE - 1) & PGDIR_MASK;
823 #endif
824         for ( ; addr < PAGE_OFFSET; addr += PGDIR_SIZE)
825                 pmd_clear(pmd_off_k(addr));
826
827         /*
828          * Clear out all the kernel space mappings, except for the first
829          * memory bank, up to the end of the vmalloc region.
830          */
831         for (addr = __phys_to_virt(bank_phys_end(&meminfo.bank[0]));
832              addr < VMALLOC_END; addr += PGDIR_SIZE)
833                 pmd_clear(pmd_off_k(addr));
834 }
835
836 /*
837  * Reserve the various regions
838  */
839 void __init reserve_special_regions(void)
840 {
841         /*
842          * Register the kernel text and data with bootmem.
843          * Note that this can only be in node 0.
844          */
845 #ifdef CONFIG_XIP_KERNEL
846         reserve_bootmem(__pa(_data), _end - _data, BOOTMEM_DEFAULT);
847 #else
848         reserve_bootmem(__pa(_stext), _end - _stext, BOOTMEM_DEFAULT);
849 #endif
850
851         /*
852          * Reserve the page tables.  These are already in use,
853          * and can only be in node 0.
854          */
855         reserve_bootmem(__pa(swapper_pg_dir),
856                         PTRS_PER_PGD * sizeof(pgd_t), BOOTMEM_DEFAULT);
857
858 #ifdef CONFIG_SA1111
859         /*
860          * Because of the SA1111 DMA bug, we want to preserve our
861          * precious DMA-able memory...
862          */
863         reserve_bootmem(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET,
864                         BOOTMEM_DEFAULT);
865 #endif
866 }
867
868 /*
869  * Set up device the mappings.  Since we clear out the page tables for all
870  * mappings above VMALLOC_END, we will remove any debug device mappings.
871  * This means you have to be careful how you debug this function, or any
872  * called function.  This means you can't use any function or debugging
873  * method which may touch any device, otherwise the kernel _will_ crash.
874  */
875 static void __init devicemaps_init(struct machine_desc *mdesc)
876 {
877         struct map_desc map;
878         unsigned long addr;
879         void *vectors;
880
881         /*
882          * Allocate the vector page early.
883          */
884         vectors = early_alloc(PAGE_SIZE);
885
886         for (addr = VMALLOC_END; addr; addr += PGDIR_SIZE)
887                 pmd_clear(pmd_off_k(addr));
888
889         /*
890          * Map the kernel if it is XIP.
891          * It is always first in the modulearea.
892          */
893 #ifdef CONFIG_XIP_KERNEL
894         map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
895         map.virtual = MODULES_VADDR;
896         map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
897         map.type = MT_ROM;
898         create_mapping(&map);
899 #endif
900
901         /*
902          * Map the cache flushing regions.
903          */
904 #ifdef FLUSH_BASE
905         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
906         map.virtual = FLUSH_BASE;
907         map.length = SZ_1M;
908         map.type = MT_CACHECLEAN;
909         create_mapping(&map);
910 #endif
911 #ifdef FLUSH_BASE_MINICACHE
912         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
913         map.virtual = FLUSH_BASE_MINICACHE;
914         map.length = SZ_1M;
915         map.type = MT_MINICLEAN;
916         create_mapping(&map);
917 #endif
918
919         /*
920          * Create a mapping for the machine vectors at the high-vectors
921          * location (0xffff0000).  If we aren't using high-vectors, also
922          * create a mapping at the low-vectors virtual address.
923          */
924         map.pfn = __phys_to_pfn(virt_to_phys(vectors));
925         map.virtual = 0xffff0000;
926         map.length = PAGE_SIZE;
927         map.type = MT_HIGH_VECTORS;
928         create_mapping(&map);
929
930         if (!vectors_high()) {
931                 map.virtual = 0;
932                 map.type = MT_LOW_VECTORS;
933                 create_mapping(&map);
934         }
935
936         /*
937          * Ask the machine support to map in the statically mapped devices.
938          */
939         if (mdesc->map_io)
940                 mdesc->map_io();
941
942         /*
943          * Finally flush the caches and tlb to ensure that we're in a
944          * consistent state wrt the writebuffer.  This also ensures that
945          * any write-allocated cache lines in the vector page are written
946          * back.  After this point, we can start to touch devices again.
947          */
948         local_flush_tlb_all();
949         flush_cache_all();
950 }
951
952 static void __init kmap_init(void)
953 {
954 #ifdef CONFIG_HIGHMEM
955         pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
956                 PKMAP_BASE, _PAGE_KERNEL_TABLE);
957 #endif
958 }
959
960 static inline void map_memory_bank(struct membank *bank)
961 {
962         struct map_desc map;
963
964         map.pfn = bank_pfn_start(bank);
965         map.virtual = __phys_to_virt(bank_phys_start(bank));
966         map.length = bank_phys_size(bank);
967         map.type = MT_MEMORY;
968
969         create_mapping(&map);
970 }
971
972 static void __init map_lowmem(void)
973 {
974         struct meminfo *mi = &meminfo;
975         int i;
976
977         /* Map all the lowmem memory banks. */
978         for (i = 0; i < mi->nr_banks; i++) {
979                 struct membank *bank = &mi->bank[i];
980
981                 if (!bank->highmem)
982                         map_memory_bank(bank);
983         }
984 }
985
986 static int __init meminfo_cmp(const void *_a, const void *_b)
987 {
988         const struct membank *a = _a, *b = _b;
989         long cmp = bank_pfn_start(a) - bank_pfn_start(b);
990         return cmp < 0 ? -1 : cmp > 0 ? 1 : 0;
991 }
992
993 /*
994  * paging_init() sets up the page tables, initialises the zone memory
995  * maps, and sets up the zero page, bad page and bad page tables.
996  */
997 void __init paging_init(struct machine_desc *mdesc)
998 {
999         void *zero_page;
1000
1001         sort(&meminfo.bank, meminfo.nr_banks, sizeof(meminfo.bank[0]), meminfo_cmp, NULL);
1002
1003         build_mem_type_table();
1004         sanity_check_meminfo();
1005         prepare_page_table();
1006         map_lowmem();
1007         bootmem_init(mdesc);
1008         devicemaps_init(mdesc);
1009         kmap_init();
1010
1011         top_pmd = pmd_off_k(0xffff0000);
1012
1013         /* allocate the zero page. */
1014         zero_page = early_alloc(PAGE_SIZE);
1015         empty_zero_page = virt_to_page(zero_page);
1016         __flush_dcache_page(NULL, empty_zero_page);
1017 }
1018
1019 /*
1020  * In order to soft-boot, we need to insert a 1:1 mapping in place of
1021  * the user-mode pages.  This will then ensure that we have predictable
1022  * results when turning the mmu off
1023  */
1024 void setup_mm_for_reboot(char mode)
1025 {
1026         unsigned long base_pmdval;
1027         pgd_t *pgd;
1028         int i;
1029
1030         /*
1031          * We need to access to user-mode page tables here. For kernel threads
1032          * we don't have any user-mode mappings so we use the context that we
1033          * "borrowed".
1034          */
1035         pgd = current->active_mm->pgd;
1036
1037         base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
1038         if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ && !cpu_is_xscale())
1039                 base_pmdval |= PMD_BIT4;
1040
1041         for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
1042                 unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
1043                 pmd_t *pmd;
1044
1045                 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
1046                 pmd[0] = __pmd(pmdval);
1047                 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
1048                 flush_pmd_entry(pmd);
1049         }
1050
1051         local_flush_tlb_all();
1052 }