git.oblomov.eu Git - linux-2.6/blob - arch/x86/mm/init_32.c

   1 /*
   2  *
   3  *  Copyright (C) 1995  Linus Torvalds
   4  *
   5  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/string.h>
  14 #include <linux/types.h>
  15 #include <linux/ptrace.h>
  16 #include <linux/mman.h>
  17 #include <linux/mm.h>
  18 #include <linux/hugetlb.h>
  19 #include <linux/swap.h>
  20 #include <linux/smp.h>
  21 #include <linux/init.h>
  22 #include <linux/highmem.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/pfn.h>
  25 #include <linux/poison.h>
  26 #include <linux/bootmem.h>
  27 #include <linux/slab.h>
  28 #include <linux/proc_fs.h>
  29 #include <linux/memory_hotplug.h>
  30 #include <linux/initrd.h>
  31 #include <linux/cpumask.h>
  32
  33 #include <asm/asm.h>
  34 #include <asm/bios_ebda.h>
  35 #include <asm/processor.h>
  36 #include <asm/system.h>
  37 #include <asm/uaccess.h>
  38 #include <asm/pgtable.h>
  39 #include <asm/dma.h>
  40 #include <asm/fixmap.h>
  41 #include <asm/e820.h>
  42 #include <asm/apic.h>
  43 #include <asm/bugs.h>
  44 #include <asm/tlb.h>
  45 #include <asm/tlbflush.h>
  46 #include <asm/pgalloc.h>
  47 #include <asm/sections.h>
  48 #include <asm/paravirt.h>
  49 #include <asm/setup.h>
  50 #include <asm/cacheflush.h>
  51 #include <asm/smp.h>
  52
  53 unsigned int __VMALLOC_RESERVE = 128 << 20;
  54
  55 unsigned long max_low_pfn_mapped;
  56 unsigned long max_pfn_mapped;
  57
  58 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  59 unsigned long highstart_pfn, highend_pfn;
  60
  61 static noinline int do_test_wp_bit(void);
  62
  63
  64 static unsigned long __initdata table_start;
  65 static unsigned long __meminitdata table_end;
  66 static unsigned long __meminitdata table_top;
  67
  68 static int __initdata after_init_bootmem;
  69
  70 static __init void *alloc_low_page(void)
  71 {
  72         unsigned long pfn = table_end++;
  73         void *adr;
  74
  75         if (pfn >= table_top)
  76                 panic("alloc_low_page: ran out of memory");
  77
  78         adr = __va(pfn * PAGE_SIZE);
  79         memset(adr, 0, PAGE_SIZE);
  80         return adr;
  81 }
  82
  83 /*
  84  * Creates a middle page table and puts a pointer to it in the
  85  * given global directory entry. This only returns the gd entry
  86  * in non-PAE compilation mode, since the middle layer is folded.
  87  */
  88 static pmd_t * __init one_md_table_init(pgd_t *pgd)
  89 {
  90         pud_t *pud;
  91         pmd_t *pmd_table;
  92
  93 #ifdef CONFIG_X86_PAE
  94         if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
  95                 if (after_init_bootmem)
  96                         pmd_table = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
  97                 else
  98                         pmd_table = (pmd_t *)alloc_low_page();
  99                 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
 100                 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
 101                 pud = pud_offset(pgd, 0);
 102                 BUG_ON(pmd_table != pmd_offset(pud, 0));
 103
 104                 return pmd_table;
 105         }
 106 #endif
 107         pud = pud_offset(pgd, 0);
 108         pmd_table = pmd_offset(pud, 0);
 109
 110         return pmd_table;
 111 }
 112
 113 /*
 114  * Create a page table and place a pointer to it in a middle page
 115  * directory entry:
 116  */
 117 static pte_t * __init one_page_table_init(pmd_t *pmd)
 118 {
 119         if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
 120                 pte_t *page_table = NULL;
 121
 122                 if (after_init_bootmem) {
 123 #ifdef CONFIG_DEBUG_PAGEALLOC
 124                         page_table = (pte_t *) alloc_bootmem_pages(PAGE_SIZE);
 125 #endif
 126                         if (!page_table)
 127                                 page_table =
 128                                 (pte_t *)alloc_bootmem_low_pages(PAGE_SIZE);
 129                 } else
 130                         page_table = (pte_t *)alloc_low_page();
 131
 132                 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
 133                 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
 134                 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
 135         }
 136
 137         return pte_offset_kernel(pmd, 0);
 138 }
 139
 140 /*
 141  * This function initializes a certain range of kernel virtual memory
 142  * with new bootmem page tables, everywhere page tables are missing in
 143  * the given range.
 144  *
 145  * NOTE: The pagetables are allocated contiguous on the physical space
 146  * so we can cache the place of the first one and move around without
 147  * checking the pgd every time.
 148  */
 149 static void __init
 150 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
 151 {
 152         int pgd_idx, pmd_idx;
 153         unsigned long vaddr;
 154         pgd_t *pgd;
 155         pmd_t *pmd;
 156
 157         vaddr = start;
 158         pgd_idx = pgd_index(vaddr);
 159         pmd_idx = pmd_index(vaddr);
 160         pgd = pgd_base + pgd_idx;
 161
 162         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
 163                 pmd = one_md_table_init(pgd);
 164                 pmd = pmd + pmd_index(vaddr);
 165                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 166                                                         pmd++, pmd_idx++) {
 167                         one_page_table_init(pmd);
 168
 169                         vaddr += PMD_SIZE;
 170                 }
 171                 pmd_idx = 0;
 172         }
 173 }
 174
 175 static inline int is_kernel_text(unsigned long addr)
 176 {
 177         if (addr >= PAGE_OFFSET && addr <= (unsigned long)__init_end)
 178                 return 1;
 179         return 0;
 180 }
 181
 182 /*
 183  * This maps the physical memory to kernel virtual address space, a total
 184  * of max_low_pfn pages, by creating page tables starting from address
 185  * PAGE_OFFSET:
 186  */
 187 static void __init kernel_physical_mapping_init(pgd_t *pgd_base,
 188                                                 unsigned long start_pfn,
 189                                                 unsigned long end_pfn,
 190                                                 int use_pse)
 191 {
 192         int pgd_idx, pmd_idx, pte_ofs;
 193         unsigned long pfn;
 194         pgd_t *pgd;
 195         pmd_t *pmd;
 196         pte_t *pte;
 197         unsigned pages_2m, pages_4k;
 198         int mapping_iter;
 199
 200         /*
 201          * First iteration will setup identity mapping using large/small pages
 202          * based on use_pse, with other attributes same as set by
 203          * the early code in head_32.S
 204          *
 205          * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
 206          * as desired for the kernel identity mapping.
 207          *
 208          * This two pass mechanism conforms to the TLB app note which says:
 209          *
 210          *     "Software should not write to a paging-structure entry in a way
 211          *      that would change, for any linear address, both the page size
 212          *      and either the page frame or attributes."
 213          */
 214         mapping_iter = 1;
 215
 216         if (!cpu_has_pse)
 217                 use_pse = 0;
 218
 219 repeat:
 220         pages_2m = pages_4k = 0;
 221         pfn = start_pfn;
 222         pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 223         pgd = pgd_base + pgd_idx;
 224         for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
 225                 pmd = one_md_table_init(pgd);
 226
 227                 if (pfn >= end_pfn)
 228                         continue;
 229 #ifdef CONFIG_X86_PAE
 230                 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 231                 pmd += pmd_idx;
 232 #else
 233                 pmd_idx = 0;
 234 #endif
 235                 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
 236                      pmd++, pmd_idx++) {
 237                         unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
 238
 239                         /*
 240                          * Map with big pages if possible, otherwise
 241                          * create normal page tables:
 242                          */
 243                         if (use_pse) {
 244                                 unsigned int addr2;
 245                                 pgprot_t prot = PAGE_KERNEL_LARGE;
 246                                 /*
 247                                  * first pass will use the same initial
 248                                  * identity mapping attribute + _PAGE_PSE.
 249                                  */
 250                                 pgprot_t init_prot =
 251                                         __pgprot(PTE_IDENT_ATTR |
 252                                                  _PAGE_PSE);
 253
 254                                 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
 255                                         PAGE_OFFSET + PAGE_SIZE-1;
 256
 257                                 if (is_kernel_text(addr) ||
 258                                     is_kernel_text(addr2))
 259                                         prot = PAGE_KERNEL_LARGE_EXEC;
 260
 261                                 pages_2m++;
 262                                 if (mapping_iter == 1)
 263                                         set_pmd(pmd, pfn_pmd(pfn, init_prot));
 264                                 else
 265                                         set_pmd(pmd, pfn_pmd(pfn, prot));
 266
 267                                 pfn += PTRS_PER_PTE;
 268                                 continue;
 269                         }
 270                         pte = one_page_table_init(pmd);
 271
 272                         pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 273                         pte += pte_ofs;
 274                         for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
 275                              pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
 276                                 pgprot_t prot = PAGE_KERNEL;
 277                                 /*
 278                                  * first pass will use the same initial
 279                                  * identity mapping attribute.
 280                                  */
 281                                 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
 282
 283                                 if (is_kernel_text(addr))
 284                                         prot = PAGE_KERNEL_EXEC;
 285
 286                                 pages_4k++;
 287                                 if (mapping_iter == 1)
 288                                         set_pte(pte, pfn_pte(pfn, init_prot));
 289                                 else
 290                                         set_pte(pte, pfn_pte(pfn, prot));
 291                         }
 292                 }
 293         }
 294         if (mapping_iter == 1) {
 295                 /*
 296                  * update direct mapping page count only in the first
 297                  * iteration.
 298                  */
 299                 update_page_count(PG_LEVEL_2M, pages_2m);
 300                 update_page_count(PG_LEVEL_4K, pages_4k);
 301
 302                 /*
 303                  * local global flush tlb, which will flush the previous
 304                  * mappings present in both small and large page TLB's.
 305                  */
 306                 __flush_tlb_all();
 307
 308                 /*
 309                  * Second iteration will set the actual desired PTE attributes.
 310                  */
 311                 mapping_iter = 2;
 312                 goto repeat;
 313         }
 314 }
 315
 316 /*
 317  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 318  * is valid. The argument is a physical page number.
 319  *
 320  *
 321  * On x86, access has to be given to the first megabyte of ram because that area
 322  * contains bios code and data regions used by X and dosemu and similar apps.
 323  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 324  * mmio resources as well as potential bios/acpi data regions.
 325  */
 326 int devmem_is_allowed(unsigned long pagenr)
 327 {
 328         if (pagenr <= 256)
 329                 return 1;
 330         if (!page_is_ram(pagenr))
 331                 return 1;
 332         return 0;
 333 }
 334
 335 pte_t *kmap_pte;
 336 pgprot_t kmap_prot;
 337
 338 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
 339 {
 340         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
 341                         vaddr), vaddr), vaddr);
 342 }
 343
 344 static void __init kmap_init(void)
 345 {
 346         unsigned long kmap_vstart;
 347
 348         /*
 349          * Cache the first kmap pte:
 350          */
 351         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
 352         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
 353
 354         kmap_prot = PAGE_KERNEL;
 355 }
 356
 357 #ifdef CONFIG_HIGHMEM
 358 static void __init permanent_kmaps_init(pgd_t *pgd_base)
 359 {
 360         unsigned long vaddr;
 361         pgd_t *pgd;
 362         pud_t *pud;
 363         pmd_t *pmd;
 364         pte_t *pte;
 365
 366         vaddr = PKMAP_BASE;
 367         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
 368
 369         pgd = swapper_pg_dir + pgd_index(vaddr);
 370         pud = pud_offset(pgd, vaddr);
 371         pmd = pmd_offset(pud, vaddr);
 372         pte = pte_offset_kernel(pmd, vaddr);
 373         pkmap_page_table = pte;
 374 }
 375
 376 static void __init add_one_highpage_init(struct page *page, int pfn)
 377 {
 378         ClearPageReserved(page);
 379         init_page_count(page);
 380         __free_page(page);
 381         totalhigh_pages++;
 382 }
 383
 384 struct add_highpages_data {
 385         unsigned long start_pfn;
 386         unsigned long end_pfn;
 387 };
 388
 389 static int __init add_highpages_work_fn(unsigned long start_pfn,
 390                                          unsigned long end_pfn, void *datax)
 391 {
 392         int node_pfn;
 393         struct page *page;
 394         unsigned long final_start_pfn, final_end_pfn;
 395         struct add_highpages_data *data;
 396
 397         data = (struct add_highpages_data *)datax;
 398
 399         final_start_pfn = max(start_pfn, data->start_pfn);
 400         final_end_pfn = min(end_pfn, data->end_pfn);
 401         if (final_start_pfn >= final_end_pfn)
 402                 return 0;
 403
 404         for (node_pfn = final_start_pfn; node_pfn < final_end_pfn;
 405              node_pfn++) {
 406                 if (!pfn_valid(node_pfn))
 407                         continue;
 408                 page = pfn_to_page(node_pfn);
 409                 add_one_highpage_init(page, node_pfn);
 410         }
 411
 412         return 0;
 413
 414 }
 415
 416 void __init add_highpages_with_active_regions(int nid, unsigned long start_pfn,
 417                                               unsigned long end_pfn)
 418 {
 419         struct add_highpages_data data;
 420
 421         data.start_pfn = start_pfn;
 422         data.end_pfn = end_pfn;
 423
 424         work_with_active_regions(nid, add_highpages_work_fn, &data);
 425 }
 426
 427 #ifndef CONFIG_NUMA
 428 static void __init set_highmem_pages_init(void)
 429 {
 430         add_highpages_with_active_regions(0, highstart_pfn, highend_pfn);
 431
 432         totalram_pages += totalhigh_pages;
 433 }
 434 #endif /* !CONFIG_NUMA */
 435
 436 #else
 437 # define permanent_kmaps_init(pgd_base)         do { } while (0)
 438 # define set_highmem_pages_init()       do { } while (0)
 439 #endif /* CONFIG_HIGHMEM */
 440
 441 void __init native_pagetable_setup_start(pgd_t *base)
 442 {
 443         unsigned long pfn, va;
 444         pgd_t *pgd;
 445         pud_t *pud;
 446         pmd_t *pmd;
 447         pte_t *pte;
 448
 449         /*
 450          * Remove any mappings which extend past the end of physical
 451          * memory from the boot time page table:
 452          */
 453         for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
 454                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
 455                 pgd = base + pgd_index(va);
 456                 if (!pgd_present(*pgd))
 457                         break;
 458
 459                 pud = pud_offset(pgd, va);
 460                 pmd = pmd_offset(pud, va);
 461                 if (!pmd_present(*pmd))
 462                         break;
 463
 464                 pte = pte_offset_kernel(pmd, va);
 465                 if (!pte_present(*pte))
 466                         break;
 467
 468                 pte_clear(NULL, va, pte);
 469         }
 470         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
 471 }
 472
 473 void __init native_pagetable_setup_done(pgd_t *base)
 474 {
 475 }
 476
 477 /*
 478  * Build a proper pagetable for the kernel mappings.  Up until this
 479  * point, we've been running on some set of pagetables constructed by
 480  * the boot process.
 481  *
 482  * If we're booting on native hardware, this will be a pagetable
 483  * constructed in arch/x86/kernel/head_32.S.  The root of the
 484  * pagetable will be swapper_pg_dir.
 485  *
 486  * If we're booting paravirtualized under a hypervisor, then there are
 487  * more options: we may already be running PAE, and the pagetable may
 488  * or may not be based in swapper_pg_dir.  In any case,
 489  * paravirt_pagetable_setup_start() will set up swapper_pg_dir
 490  * appropriately for the rest of the initialization to work.
 491  *
 492  * In general, pagetable_init() assumes that the pagetable may already
 493  * be partially populated, and so it avoids stomping on any existing
 494  * mappings.
 495  */
 496 static void __init early_ioremap_page_table_range_init(pgd_t *pgd_base)
 497 {
 498         unsigned long vaddr, end;
 499
 500         /*
 501          * Fixed mappings, only the page table structure has to be
 502          * created - mappings will be set by set_fixmap():
 503          */
 504         early_ioremap_clear();
 505         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 506         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 507         page_table_range_init(vaddr, end, pgd_base);
 508         early_ioremap_reset();
 509 }
 510
 511 static void __init pagetable_init(void)
 512 {
 513         pgd_t *pgd_base = swapper_pg_dir;
 514
 515         permanent_kmaps_init(pgd_base);
 516 }
 517
 518 #ifdef CONFIG_ACPI_SLEEP
 519 /*
 520  * ACPI suspend needs this for resume, because things like the intel-agp
 521  * driver might have split up a kernel 4MB mapping.
 522  */
 523 char swsusp_pg_dir[PAGE_SIZE]
 524         __attribute__ ((aligned(PAGE_SIZE)));
 525
 526 static inline void save_pg_dir(void)
 527 {
 528         memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
 529 }
 530 #else /* !CONFIG_ACPI_SLEEP */
 531 static inline void save_pg_dir(void)
 532 {
 533 }
 534 #endif /* !CONFIG_ACPI_SLEEP */
 535
 536 void zap_low_mappings(void)
 537 {
 538         int i;
 539
 540         /*
 541          * Zap initial low-memory mappings.
 542          *
 543          * Note that "pgd_clear()" doesn't do it for
 544          * us, because pgd_clear() is a no-op on i386.
 545          */
 546         for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
 547 #ifdef CONFIG_X86_PAE
 548                 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
 549 #else
 550                 set_pgd(swapper_pg_dir+i, __pgd(0));
 551 #endif
 552         }
 553         flush_tlb_all();
 554 }
 555
 556 int nx_enabled;
 557
 558 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
 559 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 560
 561 #ifdef CONFIG_X86_PAE
 562
 563 static int disable_nx __initdata;
 564
 565 /*
 566  * noexec = on|off
 567  *
 568  * Control non executable mappings.
 569  *
 570  * on      Enable
 571  * off     Disable
 572  */
 573 static int __init noexec_setup(char *str)
 574 {
 575         if (!str || !strcmp(str, "on")) {
 576                 if (cpu_has_nx) {
 577                         __supported_pte_mask |= _PAGE_NX;
 578                         disable_nx = 0;
 579                 }
 580         } else {
 581                 if (!strcmp(str, "off")) {
 582                         disable_nx = 1;
 583                         __supported_pte_mask &= ~_PAGE_NX;
 584                 } else {
 585                         return -EINVAL;
 586                 }
 587         }
 588
 589         return 0;
 590 }
 591 early_param("noexec", noexec_setup);
 592
 593 static void __init set_nx(void)
 594 {
 595         unsigned int v[4], l, h;
 596
 597         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
 598                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
 599
 600                 if ((v[3] & (1 << 20)) && !disable_nx) {
 601                         rdmsr(MSR_EFER, l, h);
 602                         l |= EFER_NX;
 603                         wrmsr(MSR_EFER, l, h);
 604                         nx_enabled = 1;
 605                         __supported_pte_mask |= _PAGE_NX;
 606                 }
 607         }
 608 }
 609 #endif
 610
 611 /* user-defined highmem size */
 612 static unsigned int highmem_pages = -1;
 613
 614 /*
 615  * highmem=size forces highmem to be exactly 'size' bytes.
 616  * This works even on boxes that have no highmem otherwise.
 617  * This also works to reduce highmem size on bigger boxes.
 618  */
 619 static int __init parse_highmem(char *arg)
 620 {
 621         if (!arg)
 622                 return -EINVAL;
 623
 624         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
 625         return 0;
 626 }
 627 early_param("highmem", parse_highmem);
 628
 629 /*
 630  * Determine low and high memory ranges:
 631  */
 632 void __init find_low_pfn_range(void)
 633 {
 634         /* it could update max_pfn */
 635
 636         /* max_low_pfn is 0, we already have early_res support */
 637
 638         max_low_pfn = max_pfn;
 639         if (max_low_pfn > MAXMEM_PFN) {
 640                 if (highmem_pages == -1)
 641                         highmem_pages = max_pfn - MAXMEM_PFN;
 642                 if (highmem_pages + MAXMEM_PFN < max_pfn)
 643                         max_pfn = MAXMEM_PFN + highmem_pages;
 644                 if (highmem_pages + MAXMEM_PFN > max_pfn) {
 645                         printk(KERN_WARNING "only %luMB highmem pages "
 646                                 "available, ignoring highmem size of %uMB.\n",
 647                                 pages_to_mb(max_pfn - MAXMEM_PFN),
 648                                 pages_to_mb(highmem_pages));
 649                         highmem_pages = 0;
 650                 }
 651                 max_low_pfn = MAXMEM_PFN;
 652 #ifndef CONFIG_HIGHMEM
 653                 /* Maximum memory usable is what is directly addressable */
 654                 printk(KERN_WARNING "Warning only %ldMB will be used.\n",
 655                                         MAXMEM>>20);
 656                 if (max_pfn > MAX_NONPAE_PFN)
 657                         printk(KERN_WARNING
 658                                  "Use a HIGHMEM64G enabled kernel.\n");
 659                 else
 660                         printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
 661                 max_pfn = MAXMEM_PFN;
 662 #else /* !CONFIG_HIGHMEM */
 663 #ifndef CONFIG_HIGHMEM64G
 664                 if (max_pfn > MAX_NONPAE_PFN) {
 665                         max_pfn = MAX_NONPAE_PFN;
 666                         printk(KERN_WARNING "Warning only 4GB will be used."
 667                                 "Use a HIGHMEM64G enabled kernel.\n");
 668                 }
 669 #endif /* !CONFIG_HIGHMEM64G */
 670 #endif /* !CONFIG_HIGHMEM */
 671         } else {
 672                 if (highmem_pages == -1)
 673                         highmem_pages = 0;
 674 #ifdef CONFIG_HIGHMEM
 675                 if (highmem_pages >= max_pfn) {
 676                         printk(KERN_ERR "highmem size specified (%uMB) is "
 677                                 "bigger than pages available (%luMB)!.\n",
 678                                 pages_to_mb(highmem_pages),
 679                                 pages_to_mb(max_pfn));
 680                         highmem_pages = 0;
 681                 }
 682                 if (highmem_pages) {
 683                         if (max_low_pfn - highmem_pages <
 684                             64*1024*1024/PAGE_SIZE){
 685                                 printk(KERN_ERR "highmem size %uMB results in "
 686                                 "smaller than 64MB lowmem, ignoring it.\n"
 687                                         , pages_to_mb(highmem_pages));
 688                                 highmem_pages = 0;
 689                         }
 690                         max_low_pfn -= highmem_pages;
 691                 }
 692 #else
 693                 if (highmem_pages)
 694                         printk(KERN_ERR "ignoring highmem size on non-highmem"
 695                                         " kernel!\n");
 696 #endif
 697         }
 698 }
 699
 700 #ifndef CONFIG_NEED_MULTIPLE_NODES
 701 void __init initmem_init(unsigned long start_pfn,
 702                                   unsigned long end_pfn)
 703 {
 704 #ifdef CONFIG_HIGHMEM
 705         highstart_pfn = highend_pfn = max_pfn;
 706         if (max_pfn > max_low_pfn)
 707                 highstart_pfn = max_low_pfn;
 708         memory_present(0, 0, highend_pfn);
 709         e820_register_active_regions(0, 0, highend_pfn);
 710         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 711                 pages_to_mb(highend_pfn - highstart_pfn));
 712         num_physpages = highend_pfn;
 713         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 714 #else
 715         memory_present(0, 0, max_low_pfn);
 716         e820_register_active_regions(0, 0, max_low_pfn);
 717         num_physpages = max_low_pfn;
 718         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 719 #endif
 720 #ifdef CONFIG_FLATMEM
 721         max_mapnr = num_physpages;
 722 #endif
 723         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 724                         pages_to_mb(max_low_pfn));
 725
 726         setup_bootmem_allocator();
 727 }
 728 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
 729
 730 static void __init zone_sizes_init(void)
 731 {
 732         unsigned long max_zone_pfns[MAX_NR_ZONES];
 733         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 734         max_zone_pfns[ZONE_DMA] =
 735                 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
 736         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
 737 #ifdef CONFIG_HIGHMEM
 738         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
 739 #endif
 740
 741         free_area_init_nodes(max_zone_pfns);
 742 }
 743
 744 void __init setup_bootmem_allocator(void)
 745 {
 746         int i;
 747         unsigned long bootmap_size, bootmap;
 748         /*
 749          * Initialize the boot-time allocator (with low memory only):
 750          */
 751         bootmap_size = bootmem_bootmap_pages(max_low_pfn)<<PAGE_SHIFT;
 752         bootmap = find_e820_area(min_low_pfn<<PAGE_SHIFT,
 753                                  max_pfn_mapped<<PAGE_SHIFT, bootmap_size,
 754                                  PAGE_SIZE);
 755         if (bootmap == -1L)
 756                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 757         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
 758
 759         /* don't touch min_low_pfn */
 760         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 761                                          min_low_pfn, max_low_pfn);
 762         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
 763                  max_pfn_mapped<<PAGE_SHIFT);
 764         printk(KERN_INFO "  low ram: %08lx - %08lx\n",
 765                  min_low_pfn<<PAGE_SHIFT, max_low_pfn<<PAGE_SHIFT);
 766         printk(KERN_INFO "  bootmap %08lx - %08lx\n",
 767                  bootmap, bootmap + bootmap_size);
 768         for_each_online_node(i)
 769                 free_bootmem_with_active_regions(i, max_low_pfn);
 770         early_res_to_bootmem(0, max_low_pfn<<PAGE_SHIFT);
 771
 772         after_init_bootmem = 1;
 773 }
 774
 775 static void __init find_early_table_space(unsigned long end, int use_pse)
 776 {
 777         unsigned long puds, pmds, ptes, tables, start;
 778
 779         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 780         tables = PAGE_ALIGN(puds * sizeof(pud_t));
 781
 782         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 783         tables += PAGE_ALIGN(pmds * sizeof(pmd_t));
 784
 785         if (use_pse) {
 786                 unsigned long extra;
 787
 788                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 789                 extra += PMD_SIZE;
 790                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 791         } else
 792                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 793
 794         tables += PAGE_ALIGN(ptes * sizeof(pte_t));
 795
 796         /* for fixmap */
 797         tables += PAGE_SIZE * 2;
 798
 799         /*
 800          * RED-PEN putting page tables only on node 0 could
 801          * cause a hotspot and fill up ZONE_DMA. The page tables
 802          * need roughly 0.5KB per GB.
 803          */
 804         start = 0x7000;
 805         table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
 806                                         tables, PAGE_SIZE);
 807         if (table_start == -1UL)
 808                 panic("Cannot find space for the kernel page tables");
 809
 810         table_start >>= PAGE_SHIFT;
 811         table_end = table_start;
 812         table_top = table_start + (tables>>PAGE_SHIFT);
 813
 814         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 815                 end, table_start << PAGE_SHIFT,
 816                 (table_start << PAGE_SHIFT) + tables);
 817 }
 818
 819 unsigned long __init_refok init_memory_mapping(unsigned long start,
 820                                                 unsigned long end)
 821 {
 822         pgd_t *pgd_base = swapper_pg_dir;
 823         unsigned long start_pfn, end_pfn;
 824         unsigned long big_page_start;
 825 #ifdef CONFIG_DEBUG_PAGEALLOC
 826         /*
 827          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 828          * This will simplify cpa(), which otherwise needs to support splitting
 829          * large pages into small in interrupt context, etc.
 830          */
 831         int use_pse = 0;
 832 #else
 833         int use_pse = cpu_has_pse;
 834 #endif
 835
 836         /*
 837          * Find space for the kernel direct mapping tables.
 838          */
 839         if (!after_init_bootmem)
 840                 find_early_table_space(end, use_pse);
 841
 842 #ifdef CONFIG_X86_PAE
 843         set_nx();
 844         if (nx_enabled)
 845                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
 846 #endif
 847
 848         /* Enable PSE if available */
 849         if (cpu_has_pse)
 850                 set_in_cr4(X86_CR4_PSE);
 851
 852         /* Enable PGE if available */
 853         if (cpu_has_pge) {
 854                 set_in_cr4(X86_CR4_PGE);
 855                 __supported_pte_mask |= _PAGE_GLOBAL;
 856         }
 857
 858         /*
 859          * Don't use a large page for the first 2/4MB of memory
 860          * because there are often fixed size MTRRs in there
 861          * and overlapping MTRRs into large pages can cause
 862          * slowdowns.
 863          */
 864         big_page_start = PMD_SIZE;
 865
 866         if (start < big_page_start) {
 867                 start_pfn = start >> PAGE_SHIFT;
 868                 end_pfn = min(big_page_start>>PAGE_SHIFT, end>>PAGE_SHIFT);
 869         } else {
 870                 /* head is not big page alignment ? */
 871                 start_pfn = start >> PAGE_SHIFT;
 872                 end_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 873                                  << (PMD_SHIFT - PAGE_SHIFT);
 874         }
 875         if (start_pfn < end_pfn)
 876                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn, 0);
 877
 878         /* big page range */
 879         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 880                          << (PMD_SHIFT - PAGE_SHIFT);
 881         if (start_pfn < (big_page_start >> PAGE_SHIFT))
 882                 start_pfn =  big_page_start >> PAGE_SHIFT;
 883         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 884         if (start_pfn < end_pfn)
 885                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn,
 886                                              use_pse);
 887
 888         /* tail is not big page alignment ? */
 889         start_pfn = end_pfn;
 890         if (start_pfn > (big_page_start>>PAGE_SHIFT)) {
 891                 end_pfn = end >> PAGE_SHIFT;
 892                 if (start_pfn < end_pfn)
 893                         kernel_physical_mapping_init(pgd_base, start_pfn,
 894                                                          end_pfn, 0);
 895         }
 896
 897         early_ioremap_page_table_range_init(pgd_base);
 898
 899         load_cr3(swapper_pg_dir);
 900
 901         __flush_tlb_all();
 902
 903         if (!after_init_bootmem)
 904                 reserve_early(table_start << PAGE_SHIFT,
 905                                  table_end << PAGE_SHIFT, "PGTABLE");
 906
 907         if (!after_init_bootmem)
 908                 early_memtest(start, end);
 909
 910         return end >> PAGE_SHIFT;
 911 }
 912
 913
 914 /*
 915  * paging_init() sets up the page tables - note that the first 8MB are
 916  * already mapped by head.S.
 917  *
 918  * This routines also unmaps the page at virtual kernel address 0, so
 919  * that we can trap those pesky NULL-reference errors in the kernel.
 920  */
 921 void __init paging_init(void)
 922 {
 923         pagetable_init();
 924
 925         __flush_tlb_all();
 926
 927         kmap_init();
 928
 929         /*
 930          * NOTE: at this point the bootmem allocator is fully available.
 931          */
 932         sparse_init();
 933         zone_sizes_init();
 934 }
 935
 936 /*
 937  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
 938  * and also on some strange 486's. All 586+'s are OK. This used to involve
 939  * black magic jumps to work around some nasty CPU bugs, but fortunately the
 940  * switch to using exceptions got rid of all that.
 941  */
 942 static void __init test_wp_bit(void)
 943 {
 944         printk(KERN_INFO
 945   "Checking if this processor honours the WP bit even in supervisor mode...");
 946
 947         /* Any page-aligned address will do, the test is non-destructive */
 948         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
 949         boot_cpu_data.wp_works_ok = do_test_wp_bit();
 950         clear_fixmap(FIX_WP_TEST);
 951
 952         if (!boot_cpu_data.wp_works_ok) {
 953                 printk(KERN_CONT "No.\n");
 954 #ifdef CONFIG_X86_WP_WORKS_OK
 955                 panic(
 956   "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
 957 #endif
 958         } else {
 959                 printk(KERN_CONT "Ok.\n");
 960         }
 961 }
 962
 963 static struct kcore_list kcore_mem, kcore_vmalloc;
 964
 965 void __init mem_init(void)
 966 {
 967         int codesize, reservedpages, datasize, initsize;
 968         int tmp;
 969
 970 #ifdef CONFIG_FLATMEM
 971         BUG_ON(!mem_map);
 972 #endif
 973         /* this will put all low memory onto the freelists */
 974         totalram_pages += free_all_bootmem();
 975
 976         reservedpages = 0;
 977         for (tmp = 0; tmp < max_low_pfn; tmp++)
 978                 /*
 979                  * Only count reserved RAM pages:
 980                  */
 981                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
 982                         reservedpages++;
 983
 984         set_highmem_pages_init();
 985
 986         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 987         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 988         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 989
 990         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 991         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 992                    VMALLOC_END-VMALLOC_START);
 993
 994         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
 995                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
 996                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 997                 num_physpages << (PAGE_SHIFT-10),
 998                 codesize >> 10,
 999                 reservedpages << (PAGE_SHIFT-10),
1000                 datasize >> 10,
1001                 initsize >> 10,
1002                 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
1003                );
1004
1005         printk(KERN_INFO "virtual kernel memory layout:\n"
1006                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1007 #ifdef CONFIG_HIGHMEM
1008                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1009 #endif
1010                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1011                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1012                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1013                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1014                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
1015                 FIXADDR_START, FIXADDR_TOP,
1016                 (FIXADDR_TOP - FIXADDR_START) >> 10,
1017
1018 #ifdef CONFIG_HIGHMEM
1019                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
1020                 (LAST_PKMAP*PAGE_SIZE) >> 10,
1021 #endif
1022
1023                 VMALLOC_START, VMALLOC_END,
1024                 (VMALLOC_END - VMALLOC_START) >> 20,
1025
1026                 (unsigned long)__va(0), (unsigned long)high_memory,
1027                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
1028
1029                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
1030                 ((unsigned long)&__init_end -
1031                  (unsigned long)&__init_begin) >> 10,
1032
1033                 (unsigned long)&_etext, (unsigned long)&_edata,
1034                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
1035
1036                 (unsigned long)&_text, (unsigned long)&_etext,
1037                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
1038
1039         /*
1040          * Check boundaries twice: Some fundamental inconsistencies can
1041          * be detected at build time already.
1042          */
1043 #define __FIXADDR_TOP (-PAGE_SIZE)
1044 #ifdef CONFIG_HIGHMEM
1045         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
1046         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
1047 #endif
1048 #define high_memory (-128UL << 20)
1049         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
1050 #undef high_memory
1051 #undef __FIXADDR_TOP
1052
1053 #ifdef CONFIG_HIGHMEM
1054         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
1055         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
1056 #endif
1057         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
1058         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
1059
1060         if (boot_cpu_data.wp_works_ok < 0)
1061                 test_wp_bit();
1062
1063         save_pg_dir();
1064         zap_low_mappings();
1065 }
1066
1067 #ifdef CONFIG_MEMORY_HOTPLUG
1068 int arch_add_memory(int nid, u64 start, u64 size)
1069 {
1070         struct pglist_data *pgdata = NODE_DATA(nid);
1071         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
1072         unsigned long start_pfn = start >> PAGE_SHIFT;
1073         unsigned long nr_pages = size >> PAGE_SHIFT;
1074
1075         return __add_pages(zone, start_pfn, nr_pages);
1076 }
1077 #endif
1078
1079 /*
1080  * This function cannot be __init, since exceptions don't work in that
1081  * section.  Put this after the callers, so that it cannot be inlined.
1082  */
1083 static noinline int do_test_wp_bit(void)
1084 {
1085         char tmp_reg;
1086         int flag;
1087
1088         __asm__ __volatile__(
1089                 "       movb %0, %1     \n"
1090                 "1:     movb %1, %0     \n"
1091                 "       xorl %2, %2     \n"
1092                 "2:                     \n"
1093                 _ASM_EXTABLE(1b,2b)
1094                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
1095                  "=q" (tmp_reg),
1096                  "=r" (flag)
1097                 :"2" (1)
1098                 :"memory");
1099
1100         return flag;
1101 }
1102
1103 #ifdef CONFIG_DEBUG_RODATA
1104 const int rodata_test_data = 0xC3;
1105 EXPORT_SYMBOL_GPL(rodata_test_data);
1106
1107 void mark_rodata_ro(void)
1108 {
1109         unsigned long start = PFN_ALIGN(_text);
1110         unsigned long size = PFN_ALIGN(_etext) - start;
1111
1112 #ifndef CONFIG_DYNAMIC_FTRACE
1113         /* Dynamic tracing modifies the kernel text section */
1114         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1115         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
1116                 size >> 10);
1117
1118 #ifdef CONFIG_CPA_DEBUG
1119         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
1120                 start, start+size);
1121         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
1122
1123         printk(KERN_INFO "Testing CPA: write protecting again\n");
1124         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
1125 #endif
1126 #endif /* CONFIG_DYNAMIC_FTRACE */
1127
1128         start += size;
1129         size = (unsigned long)__end_rodata - start;
1130         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1131         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1132                 size >> 10);
1133         rodata_test();
1134
1135 #ifdef CONFIG_CPA_DEBUG
1136         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
1137         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
1138
1139         printk(KERN_INFO "Testing CPA: write protecting again\n");
1140         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1141 #endif
1142 }
1143 #endif
1144
1145 void free_init_pages(char *what, unsigned long begin, unsigned long end)
1146 {
1147 #ifdef CONFIG_DEBUG_PAGEALLOC
1148         /*
1149          * If debugging page accesses then do not free this memory but
1150          * mark them not present - any buggy init-section access will
1151          * create a kernel page fault:
1152          */
1153         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
1154                 begin, PAGE_ALIGN(end));
1155         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
1156 #else
1157         unsigned long addr;
1158
1159         /*
1160          * We just marked the kernel text read only above, now that
1161          * we are going to free part of that, we need to make that
1162          * writeable first.
1163          */
1164         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
1165
1166         for (addr = begin; addr < end; addr += PAGE_SIZE) {
1167                 ClearPageReserved(virt_to_page(addr));
1168                 init_page_count(virt_to_page(addr));
1169                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
1170                 free_page(addr);
1171                 totalram_pages++;
1172         }
1173         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
1174 #endif
1175 }
1176
1177 void free_initmem(void)
1178 {
1179         free_init_pages("unused kernel memory",
1180                         (unsigned long)(&__init_begin),
1181                         (unsigned long)(&__init_end));
1182 }
1183
1184 #ifdef CONFIG_BLK_DEV_INITRD
1185 void free_initrd_mem(unsigned long start, unsigned long end)
1186 {
1187         free_init_pages("initrd memory", start, end);
1188 }
1189 #endif
1190
1191 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1192                                    int flags)
1193 {
1194         return reserve_bootmem(phys, len, flags);
1195 }