arch/arm/mm/mmu.c

   1 /*
   2  *  linux/arch/arm/mm/mmu.c
   3  *
   4  *  Copyright (C) 1995-2005 Russell King
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  */
  10 #include <linux/module.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/init.h>
  14 #include <linux/bootmem.h>
  15 #include <linux/mman.h>
  16 #include <linux/nodemask.h>
  17
  18 #include <asm/cputype.h>
  19 #include <asm/mach-types.h>
  20 #include <asm/sections.h>
  21 #include <asm/cachetype.h>
  22 #include <asm/setup.h>
  23 #include <asm/sizes.h>
  24 #include <asm/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 void __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                         *p += len;
 115                         break;
 116                 }
 117         }
 118         if (i == ARRAY_SIZE(cache_policies))
 119                 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
 120         /*
 121          * This restriction is partly to do with the way we boot; it is
 122          * unpredictable to have memory mapped using two different sets of
 123          * memory attributes (shared, type, and cache attribs).  We can not
 124          * change these attributes once the initial assembly has setup the
 125          * page tables.
 126          */
 127         if (cpu_architecture() >= CPU_ARCH_ARMv6) {
 128                 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
 129                 cachepolicy = CPOLICY_WRITEBACK;
 130         }
 131         flush_cache_all();
 132         set_cr(cr_alignment);
 133 }
 134 __early_param("cachepolicy=", early_cachepolicy);
 135
 136 static void __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 }
 142 __early_param("nocache", early_nocache);
 143
 144 static void __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 }
 150 __early_param("nowb", early_nowrite);
 151
 152 static void __init early_ecc(char **p)
 153 {
 154         if (memcmp(*p, "on", 2) == 0) {
 155                 ecc_mask = PMD_PROTECTION;
 156                 *p += 2;
 157         } else if (memcmp(*p, "off", 3) == 0) {
 158                 ecc_mask = 0;
 159                 *p += 3;
 160         }
 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_MEMORY].prot_sect |= PMD_SECT_S;
 424                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
 425 #endif
 426         }
 427
 428         /*
 429          * Non-cacheable Normal - intended for memory areas that must
 430          * not cause dirty cache line writebacks when used
 431          */
 432         if (cpu_arch >= CPU_ARCH_ARMv6) {
 433                 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
 434                         /* Non-cacheable Normal is XCB = 001 */
 435                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
 436                                 PMD_SECT_BUFFERED;
 437                 } else {
 438                         /* For both ARMv6 and non-TEX-remapping ARMv7 */
 439                         mem_types[MT_MEMORY_NONCACHED].prot_sect |=
 440                                 PMD_SECT_TEX(1);
 441                 }
 442         } else {
 443                 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
 444         }
 445
 446         for (i = 0; i < 16; i++) {
 447                 unsigned long v = pgprot_val(protection_map[i]);
 448                 protection_map[i] = __pgprot(v | user_pgprot);
 449         }
 450
 451         mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
 452         mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
 453
 454         pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
 455         pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
 456                                  L_PTE_DIRTY | L_PTE_WRITE |
 457                                  L_PTE_EXEC | kern_pgprot);
 458
 459         mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
 460         mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
 461         mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
 462         mem_types[MT_ROM].prot_sect |= cp->pmd;
 463
 464         switch (cp->pmd) {
 465         case PMD_SECT_WT:
 466                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
 467                 break;
 468         case PMD_SECT_WB:
 469         case PMD_SECT_WBWA:
 470                 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
 471                 break;
 472         }
 473         printk("Memory policy: ECC %sabled, Data cache %s\n",
 474                 ecc_mask ? "en" : "dis", cp->policy);
 475
 476         for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
 477                 struct mem_type *t = &mem_types[i];
 478                 if (t->prot_l1)
 479                         t->prot_l1 |= PMD_DOMAIN(t->domain);
 480                 if (t->prot_sect)
 481                         t->prot_sect |= PMD_DOMAIN(t->domain);
 482         }
 483 }
 484
 485 #define vectors_base()  (vectors_high() ? 0xffff0000 : 0)
 486
 487 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
 488                                   unsigned long end, unsigned long pfn,
 489                                   const struct mem_type *type)
 490 {
 491         pte_t *pte;
 492
 493         if (pmd_none(*pmd)) {
 494                 pte = alloc_bootmem_low_pages(2 * PTRS_PER_PTE * sizeof(pte_t));
 495                 __pmd_populate(pmd, __pa(pte) | type->prot_l1);
 496         }
 497
 498         pte = pte_offset_kernel(pmd, addr);
 499         do {
 500                 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
 501                 pfn++;
 502         } while (pte++, addr += PAGE_SIZE, addr != end);
 503 }
 504
 505 static void __init alloc_init_section(pgd_t *pgd, unsigned long addr,
 506                                       unsigned long end, unsigned long phys,
 507                                       const struct mem_type *type)
 508 {
 509         pmd_t *pmd = pmd_offset(pgd, addr);
 510
 511         /*
 512          * Try a section mapping - end, addr and phys must all be aligned
 513          * to a section boundary.  Note that PMDs refer to the individual
 514          * L1 entries, whereas PGDs refer to a group of L1 entries making
 515          * up one logical pointer to an L2 table.
 516          */
 517         if (((addr | end | phys) & ~SECTION_MASK) == 0) {
 518                 pmd_t *p = pmd;
 519
 520                 if (addr & SECTION_SIZE)
 521                         pmd++;
 522
 523                 do {
 524                         *pmd = __pmd(phys | type->prot_sect);
 525                         phys += SECTION_SIZE;
 526                 } while (pmd++, addr += SECTION_SIZE, addr != end);
 527
 528                 flush_pmd_entry(p);
 529         } else {
 530                 /*
 531                  * No need to loop; pte's aren't interested in the
 532                  * individual L1 entries.
 533                  */
 534                 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
 535         }
 536 }
 537
 538 static void __init create_36bit_mapping(struct map_desc *md,
 539                                         const struct mem_type *type)
 540 {
 541         unsigned long phys, addr, length, end;
 542         pgd_t *pgd;
 543
 544         addr = md->virtual;
 545         phys = (unsigned long)__pfn_to_phys(md->pfn);
 546         length = PAGE_ALIGN(md->length);
 547
 548         if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
 549                 printk(KERN_ERR "MM: CPU does not support supersection "
 550                        "mapping for 0x%08llx at 0x%08lx\n",
 551                        __pfn_to_phys((u64)md->pfn), addr);
 552                 return;
 553         }
 554
 555         /* N.B. ARMv6 supersections are only defined to work with domain 0.
 556          *      Since domain assignments can in fact be arbitrary, the
 557          *      'domain == 0' check below is required to insure that ARMv6
 558          *      supersections are only allocated for domain 0 regardless
 559          *      of the actual domain assignments in use.
 560          */
 561         if (type->domain) {
 562                 printk(KERN_ERR "MM: invalid domain in supersection "
 563                        "mapping for 0x%08llx at 0x%08lx\n",
 564                        __pfn_to_phys((u64)md->pfn), addr);
 565                 return;
 566         }
 567
 568         if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
 569                 printk(KERN_ERR "MM: cannot create mapping for "
 570                        "0x%08llx at 0x%08lx invalid alignment\n",
 571                        __pfn_to_phys((u64)md->pfn), addr);
 572                 return;
 573         }
 574
 575         /*
 576          * Shift bits [35:32] of address into bits [23:20] of PMD
 577          * (See ARMv6 spec).
 578          */
 579         phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
 580
 581         pgd = pgd_offset_k(addr);
 582         end = addr + length;
 583         do {
 584                 pmd_t *pmd = pmd_offset(pgd, addr);
 585                 int i;
 586
 587                 for (i = 0; i < 16; i++)
 588                         *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
 589
 590                 addr += SUPERSECTION_SIZE;
 591                 phys += SUPERSECTION_SIZE;
 592                 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
 593         } while (addr != end);
 594 }
 595
 596 /*
 597  * Create the page directory entries and any necessary
 598  * page tables for the mapping specified by `md'.  We
 599  * are able to cope here with varying sizes and address
 600  * offsets, and we take full advantage of sections and
 601  * supersections.
 602  */
 603 void __init create_mapping(struct map_desc *md)
 604 {
 605         unsigned long phys, addr, length, end;
 606         const struct mem_type *type;
 607         pgd_t *pgd;
 608
 609         if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
 610                 printk(KERN_WARNING "BUG: not creating mapping for "
 611                        "0x%08llx at 0x%08lx in user region\n",
 612                        __pfn_to_phys((u64)md->pfn), md->virtual);
 613                 return;
 614         }
 615
 616         if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
 617             md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
 618                 printk(KERN_WARNING "BUG: mapping for 0x%08llx at 0x%08lx "
 619                        "overlaps vmalloc space\n",
 620                        __pfn_to_phys((u64)md->pfn), md->virtual);
 621         }
 622
 623         type = &mem_types[md->type];
 624
 625         /*
 626          * Catch 36-bit addresses
 627          */
 628         if (md->pfn >= 0x100000) {
 629                 create_36bit_mapping(md, type);
 630                 return;
 631         }
 632
 633         addr = md->virtual & PAGE_MASK;
 634         phys = (unsigned long)__pfn_to_phys(md->pfn);
 635         length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
 636
 637         if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
 638                 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
 639                        "be mapped using pages, ignoring.\n",
 640                        __pfn_to_phys(md->pfn), addr);
 641                 return;
 642         }
 643
 644         pgd = pgd_offset_k(addr);
 645         end = addr + length;
 646         do {
 647                 unsigned long next = pgd_addr_end(addr, end);
 648
 649                 alloc_init_section(pgd, addr, next, phys, type);
 650
 651                 phys += next - addr;
 652                 addr = next;
 653         } while (pgd++, addr != end);
 654 }
 655
 656 /*
 657  * Create the architecture specific mappings
 658  */
 659 void __init iotable_init(struct map_desc *io_desc, int nr)
 660 {
 661         int i;
 662
 663         for (i = 0; i < nr; i++)
 664                 create_mapping(io_desc + i);
 665 }
 666
 667 static unsigned long __initdata vmalloc_reserve = SZ_128M;
 668
 669 /*
 670  * vmalloc=size forces the vmalloc area to be exactly 'size'
 671  * bytes. This can be used to increase (or decrease) the vmalloc
 672  * area - the default is 128m.
 673  */
 674 static void __init early_vmalloc(char **arg)
 675 {
 676         vmalloc_reserve = memparse(*arg, arg);
 677
 678         if (vmalloc_reserve < SZ_16M) {
 679                 vmalloc_reserve = SZ_16M;
 680                 printk(KERN_WARNING
 681                         "vmalloc area too small, limiting to %luMB\n",
 682                         vmalloc_reserve >> 20);
 683         }
 684
 685         if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
 686                 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
 687                 printk(KERN_WARNING
 688                         "vmalloc area is too big, limiting to %luMB\n",
 689                         vmalloc_reserve >> 20);
 690         }
 691 }
 692 __early_param("vmalloc=", early_vmalloc);
 693
 694 #define VMALLOC_MIN     (void *)(VMALLOC_END - vmalloc_reserve)
 695
 696 static void __init sanity_check_meminfo(void)
 697 {
 698         int i, j, highmem = 0;
 699
 700         for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
 701                 struct membank *bank = &meminfo.bank[j];
 702                 *bank = meminfo.bank[i];
 703
 704 #ifdef CONFIG_HIGHMEM
 705                 if (__va(bank->start) > VMALLOC_MIN ||
 706                     __va(bank->start) < (void *)PAGE_OFFSET)
 707                         highmem = 1;
 708
 709                 bank->highmem = highmem;
 710
 711                 /*
 712                  * Split those memory banks which are partially overlapping
 713                  * the vmalloc area greatly simplifying things later.
 714                  */
 715                 if (__va(bank->start) < VMALLOC_MIN &&
 716                     bank->size > VMALLOC_MIN - __va(bank->start)) {
 717                         if (meminfo.nr_banks >= NR_BANKS) {
 718                                 printk(KERN_CRIT "NR_BANKS too low, "
 719                                                  "ignoring high memory\n");
 720                         } else {
 721                                 memmove(bank + 1, bank,
 722                                         (meminfo.nr_banks - i) * sizeof(*bank));
 723                                 meminfo.nr_banks++;
 724                                 i++;
 725                                 bank[1].size -= VMALLOC_MIN - __va(bank->start);
 726                                 bank[1].start = __pa(VMALLOC_MIN - 1) + 1;
 727                                 bank[1].highmem = highmem = 1;
 728                                 j++;
 729                         }
 730                         bank->size = VMALLOC_MIN - __va(bank->start);
 731                 }
 732 #else
 733                 bank->highmem = highmem;
 734
 735                 /*
 736                  * Check whether this memory bank would entirely overlap
 737                  * the vmalloc area.
 738                  */
 739                 if (__va(bank->start) >= VMALLOC_MIN ||
 740                     __va(bank->start) < (void *)PAGE_OFFSET) {
 741                         printk(KERN_NOTICE "Ignoring RAM at %.8lx-%.8lx "
 742                                "(vmalloc region overlap).\n",
 743                                bank->start, bank->start + bank->size - 1);
 744                         continue;
 745                 }
 746
 747                 /*
 748                  * Check whether this memory bank would partially overlap
 749                  * the vmalloc area.
 750                  */
 751                 if (__va(bank->start + bank->size) > VMALLOC_MIN ||
 752                     __va(bank->start + bank->size) < __va(bank->start)) {
 753                         unsigned long newsize = VMALLOC_MIN - __va(bank->start);
 754                         printk(KERN_NOTICE "Truncating RAM at %.8lx-%.8lx "
 755                                "to -%.8lx (vmalloc region overlap).\n",
 756                                bank->start, bank->start + bank->size - 1,
 757                                bank->start + newsize - 1);
 758                         bank->size = newsize;
 759                 }
 760 #endif
 761                 j++;
 762         }
 763 #ifdef CONFIG_HIGHMEM
 764         if (highmem) {
 765                 const char *reason = NULL;
 766
 767                 if (cache_is_vipt_aliasing()) {
 768                         /*
 769                          * Interactions between kmap and other mappings
 770                          * make highmem support with aliasing VIPT caches
 771                          * rather difficult.
 772                          */
 773                         reason = "with VIPT aliasing cache";
 774 #ifdef CONFIG_SMP
 775                 } else if (tlb_ops_need_broadcast()) {
 776                         /*
 777                          * kmap_high needs to occasionally flush TLB entries,
 778                          * however, if the TLB entries need to be broadcast
 779                          * we may deadlock:
 780                          *  kmap_high(irqs off)->flush_all_zero_pkmaps->
 781                          *  flush_tlb_kernel_range->smp_call_function_many
 782                          *   (must not be called with irqs off)
 783                          */
 784                         reason = "without hardware TLB ops broadcasting";
 785 #endif
 786                 }
 787                 if (reason) {
 788                         printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
 789                                 reason);
 790                         while (j > 0 && meminfo.bank[j - 1].highmem)
 791                                 j--;
 792                 }
 793         }
 794 #endif
 795         meminfo.nr_banks = j;
 796 }
 797
 798 static inline void prepare_page_table(void)
 799 {
 800         unsigned long addr;
 801
 802         /*
 803          * Clear out all the mappings below the kernel image.
 804          */
 805         for (addr = 0; addr < MODULES_VADDR; addr += PGDIR_SIZE)
 806                 pmd_clear(pmd_off_k(addr));
 807
 808 #ifdef CONFIG_XIP_KERNEL
 809         /* The XIP kernel is mapped in the module area -- skip over it */
 810         addr = ((unsigned long)_etext + PGDIR_SIZE - 1) & PGDIR_MASK;
 811 #endif
 812         for ( ; addr < PAGE_OFFSET; addr += PGDIR_SIZE)
 813                 pmd_clear(pmd_off_k(addr));
 814
 815         /*
 816          * Clear out all the kernel space mappings, except for the first
 817          * memory bank, up to the end of the vmalloc region.
 818          */
 819         for (addr = __phys_to_virt(bank_phys_end(&meminfo.bank[0]));
 820              addr < VMALLOC_END; addr += PGDIR_SIZE)
 821                 pmd_clear(pmd_off_k(addr));
 822 }
 823
 824 /*
 825  * Reserve the various regions of node 0
 826  */
 827 void __init reserve_node_zero(pg_data_t *pgdat)
 828 {
 829         unsigned long res_size = 0;
 830
 831         /*
 832          * Register the kernel text and data with bootmem.
 833          * Note that this can only be in node 0.
 834          */
 835 #ifdef CONFIG_XIP_KERNEL
 836         reserve_bootmem_node(pgdat, __pa(_data), _end - _data,
 837                         BOOTMEM_DEFAULT);
 838 #else
 839         reserve_bootmem_node(pgdat, __pa(_stext), _end - _stext,
 840                         BOOTMEM_DEFAULT);
 841 #endif
 842
 843         /*
 844          * Reserve the page tables.  These are already in use,
 845          * and can only be in node 0.
 846          */
 847         reserve_bootmem_node(pgdat, __pa(swapper_pg_dir),
 848                              PTRS_PER_PGD * sizeof(pgd_t), BOOTMEM_DEFAULT);
 849
 850         /*
 851          * Hmm... This should go elsewhere, but we really really need to
 852          * stop things allocating the low memory; ideally we need a better
 853          * implementation of GFP_DMA which does not assume that DMA-able
 854          * memory starts at zero.
 855          */
 856         if (machine_is_integrator() || machine_is_cintegrator())
 857                 res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
 858
 859         /*
 860          * These should likewise go elsewhere.  They pre-reserve the
 861          * screen memory region at the start of main system memory.
 862          */
 863         if (machine_is_edb7211())
 864                 res_size = 0x00020000;
 865         if (machine_is_p720t())
 866                 res_size = 0x00014000;
 867
 868         /* H1940 and RX3715 need to reserve this for suspend */
 869
 870         if (machine_is_h1940() || machine_is_rx3715()) {
 871                 reserve_bootmem_node(pgdat, 0x30003000, 0x1000,
 872                                 BOOTMEM_DEFAULT);
 873                 reserve_bootmem_node(pgdat, 0x30081000, 0x1000,
 874                                 BOOTMEM_DEFAULT);
 875         }
 876
 877         if (machine_is_palmld() || machine_is_palmtx()) {
 878                 reserve_bootmem_node(pgdat, 0xa0000000, 0x1000,
 879                                 BOOTMEM_EXCLUSIVE);
 880                 reserve_bootmem_node(pgdat, 0xa0200000, 0x1000,
 881                                 BOOTMEM_EXCLUSIVE);
 882         }
 883
 884         if (machine_is_treo680() || machine_is_centro()) {
 885                 reserve_bootmem_node(pgdat, 0xa0000000, 0x1000,
 886                                 BOOTMEM_EXCLUSIVE);
 887                 reserve_bootmem_node(pgdat, 0xa2000000, 0x1000,
 888                                 BOOTMEM_EXCLUSIVE);
 889         }
 890
 891         if (machine_is_palmt5())
 892                 reserve_bootmem_node(pgdat, 0xa0200000, 0x1000,
 893                                 BOOTMEM_EXCLUSIVE);
 894
 895         /*
 896          * U300 - This platform family can share physical memory
 897          * between two ARM cpus, one running Linux and the other
 898          * running another OS.
 899          */
 900         if (machine_is_u300()) {
 901 #ifdef CONFIG_MACH_U300_SINGLE_RAM
 902 #if ((CONFIG_MACH_U300_ACCESS_MEM_SIZE & 1) == 1) &&    \
 903         CONFIG_MACH_U300_2MB_ALIGNMENT_FIX
 904                 res_size = 0x00100000;
 905 #endif
 906 #endif
 907         }
 908
 909 #ifdef CONFIG_SA1111
 910         /*
 911          * Because of the SA1111 DMA bug, we want to preserve our
 912          * precious DMA-able memory...
 913          */
 914         res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
 915 #endif
 916         if (res_size)
 917                 reserve_bootmem_node(pgdat, PHYS_OFFSET, res_size,
 918                                 BOOTMEM_DEFAULT);
 919 }
 920
 921 /*
 922  * Set up device the mappings.  Since we clear out the page tables for all
 923  * mappings above VMALLOC_END, we will remove any debug device mappings.
 924  * This means you have to be careful how you debug this function, or any
 925  * called function.  This means you can't use any function or debugging
 926  * method which may touch any device, otherwise the kernel _will_ crash.
 927  */
 928 static void __init devicemaps_init(struct machine_desc *mdesc)
 929 {
 930         struct map_desc map;
 931         unsigned long addr;
 932         void *vectors;
 933
 934         /*
 935          * Allocate the vector page early.
 936          */
 937         vectors = alloc_bootmem_low_pages(PAGE_SIZE);
 938
 939         for (addr = VMALLOC_END; addr; addr += PGDIR_SIZE)
 940                 pmd_clear(pmd_off_k(addr));
 941
 942         /*
 943          * Map the kernel if it is XIP.
 944          * It is always first in the modulearea.
 945          */
 946 #ifdef CONFIG_XIP_KERNEL
 947         map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
 948         map.virtual = MODULES_VADDR;
 949         map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
 950         map.type = MT_ROM;
 951         create_mapping(&map);
 952 #endif
 953
 954         /*
 955          * Map the cache flushing regions.
 956          */
 957 #ifdef FLUSH_BASE
 958         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
 959         map.virtual = FLUSH_BASE;
 960         map.length = SZ_1M;
 961         map.type = MT_CACHECLEAN;
 962         create_mapping(&map);
 963 #endif
 964 #ifdef FLUSH_BASE_MINICACHE
 965         map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
 966         map.virtual = FLUSH_BASE_MINICACHE;
 967         map.length = SZ_1M;
 968         map.type = MT_MINICLEAN;
 969         create_mapping(&map);
 970 #endif
 971
 972         /*
 973          * Create a mapping for the machine vectors at the high-vectors
 974          * location (0xffff0000).  If we aren't using high-vectors, also
 975          * create a mapping at the low-vectors virtual address.
 976          */
 977         map.pfn = __phys_to_pfn(virt_to_phys(vectors));
 978         map.virtual = 0xffff0000;
 979         map.length = PAGE_SIZE;
 980         map.type = MT_HIGH_VECTORS;
 981         create_mapping(&map);
 982
 983         if (!vectors_high()) {
 984                 map.virtual = 0;
 985                 map.type = MT_LOW_VECTORS;
 986                 create_mapping(&map);
 987         }
 988
 989         /*
 990          * Ask the machine support to map in the statically mapped devices.
 991          */
 992         if (mdesc->map_io)
 993                 mdesc->map_io();
 994
 995         /*
 996          * Finally flush the caches and tlb to ensure that we're in a
 997          * consistent state wrt the writebuffer.  This also ensures that
 998          * any write-allocated cache lines in the vector page are written
 999          * back.  After this point, we can start to touch devices again.
1000          */
1001         local_flush_tlb_all();
1002         flush_cache_all();
1003 }
1004
1005 static void __init kmap_init(void)
1006 {
1007 #ifdef CONFIG_HIGHMEM
1008         pmd_t *pmd = pmd_off_k(PKMAP_BASE);
1009         pte_t *pte = alloc_bootmem_low_pages(2 * PTRS_PER_PTE * sizeof(pte_t));
1010         BUG_ON(!pmd_none(*pmd) || !pte);
1011         __pmd_populate(pmd, __pa(pte) | _PAGE_KERNEL_TABLE);
1012         pkmap_page_table = pte + PTRS_PER_PTE;
1013 #endif
1014 }
1015
1016 /*
1017  * paging_init() sets up the page tables, initialises the zone memory
1018  * maps, and sets up the zero page, bad page and bad page tables.
1019  */
1020 void __init paging_init(struct machine_desc *mdesc)
1021 {
1022         void *zero_page;
1023
1024         build_mem_type_table();
1025         sanity_check_meminfo();
1026         prepare_page_table();
1027         bootmem_init();
1028         devicemaps_init(mdesc);
1029         kmap_init();
1030
1031         top_pmd = pmd_off_k(0xffff0000);
1032
1033         /*
1034          * allocate the zero page.  Note that this always succeeds and
1035          * returns a zeroed result.
1036          */
1037         zero_page = alloc_bootmem_low_pages(PAGE_SIZE);
1038         empty_zero_page = virt_to_page(zero_page);
1039         __flush_dcache_page(NULL, empty_zero_page);
1040 }
1041
1042 /*
1043  * In order to soft-boot, we need to insert a 1:1 mapping in place of
1044  * the user-mode pages.  This will then ensure that we have predictable
1045  * results when turning the mmu off
1046  */
1047 void setup_mm_for_reboot(char mode)
1048 {
1049         unsigned long base_pmdval;
1050         pgd_t *pgd;
1051         int i;
1052
1053         if (current->mm && current->mm->pgd)
1054                 pgd = current->mm->pgd;
1055         else
1056                 pgd = init_mm.pgd;
1057
1058         base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
1059         if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ && !cpu_is_xscale())
1060                 base_pmdval |= PMD_BIT4;
1061
1062         for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
1063                 unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
1064                 pmd_t *pmd;
1065
1066                 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
1067                 pmd[0] = __pmd(pmdval);
1068                 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
1069                 flush_pmd_entry(pmd);
1070         }
1071 }