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/pci.h>
  25 #include <linux/pfn.h>
  26 #include <linux/poison.h>
  27 #include <linux/bootmem.h>
  28 #include <linux/slab.h>
  29 #include <linux/proc_fs.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/initrd.h>
  32 #include <linux/cpumask.h>
  33
  34 #include <asm/asm.h>
  35 #include <asm/bios_ebda.h>
  36 #include <asm/processor.h>
  37 #include <asm/system.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/dma.h>
  41 #include <asm/fixmap.h>
  42 #include <asm/e820.h>
  43 #include <asm/apic.h>
  44 #include <asm/bugs.h>
  45 #include <asm/tlb.h>
  46 #include <asm/tlbflush.h>
  47 #include <asm/pgalloc.h>
  48 #include <asm/sections.h>
  49 #include <asm/paravirt.h>
  50 #include <asm/setup.h>
  51 #include <asm/cacheflush.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 (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 331                 return 0;
 332         if (!page_is_ram(pagenr))
 333                 return 1;
 334         return 0;
 335 }
 336
 337 pte_t *kmap_pte;
 338 pgprot_t kmap_prot;
 339
 340 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
 341 {
 342         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
 343                         vaddr), vaddr), vaddr);
 344 }
 345
 346 static void __init kmap_init(void)
 347 {
 348         unsigned long kmap_vstart;
 349
 350         /*
 351          * Cache the first kmap pte:
 352          */
 353         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
 354         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
 355
 356         kmap_prot = PAGE_KERNEL;
 357 }
 358
 359 #ifdef CONFIG_HIGHMEM
 360 static void __init permanent_kmaps_init(pgd_t *pgd_base)
 361 {
 362         unsigned long vaddr;
 363         pgd_t *pgd;
 364         pud_t *pud;
 365         pmd_t *pmd;
 366         pte_t *pte;
 367
 368         vaddr = PKMAP_BASE;
 369         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
 370
 371         pgd = swapper_pg_dir + pgd_index(vaddr);
 372         pud = pud_offset(pgd, vaddr);
 373         pmd = pmd_offset(pud, vaddr);
 374         pte = pte_offset_kernel(pmd, vaddr);
 375         pkmap_page_table = pte;
 376 }
 377
 378 static void __init add_one_highpage_init(struct page *page, int pfn)
 379 {
 380         ClearPageReserved(page);
 381         init_page_count(page);
 382         __free_page(page);
 383         totalhigh_pages++;
 384 }
 385
 386 struct add_highpages_data {
 387         unsigned long start_pfn;
 388         unsigned long end_pfn;
 389 };
 390
 391 static int __init add_highpages_work_fn(unsigned long start_pfn,
 392                                          unsigned long end_pfn, void *datax)
 393 {
 394         int node_pfn;
 395         struct page *page;
 396         unsigned long final_start_pfn, final_end_pfn;
 397         struct add_highpages_data *data;
 398
 399         data = (struct add_highpages_data *)datax;
 400
 401         final_start_pfn = max(start_pfn, data->start_pfn);
 402         final_end_pfn = min(end_pfn, data->end_pfn);
 403         if (final_start_pfn >= final_end_pfn)
 404                 return 0;
 405
 406         for (node_pfn = final_start_pfn; node_pfn < final_end_pfn;
 407              node_pfn++) {
 408                 if (!pfn_valid(node_pfn))
 409                         continue;
 410                 page = pfn_to_page(node_pfn);
 411                 add_one_highpage_init(page, node_pfn);
 412         }
 413
 414         return 0;
 415
 416 }
 417
 418 void __init add_highpages_with_active_regions(int nid, unsigned long start_pfn,
 419                                               unsigned long end_pfn)
 420 {
 421         struct add_highpages_data data;
 422
 423         data.start_pfn = start_pfn;
 424         data.end_pfn = end_pfn;
 425
 426         work_with_active_regions(nid, add_highpages_work_fn, &data);
 427 }
 428
 429 #ifndef CONFIG_NUMA
 430 static void __init set_highmem_pages_init(void)
 431 {
 432         add_highpages_with_active_regions(0, highstart_pfn, highend_pfn);
 433
 434         totalram_pages += totalhigh_pages;
 435 }
 436 #endif /* !CONFIG_NUMA */
 437
 438 #else
 439 static inline void permanent_kmaps_init(pgd_t *pgd_base)
 440 {
 441 }
 442 static inline void set_highmem_pages_init(void)
 443 {
 444 }
 445 #endif /* CONFIG_HIGHMEM */
 446
 447 void __init native_pagetable_setup_start(pgd_t *base)
 448 {
 449         unsigned long pfn, va;
 450         pgd_t *pgd;
 451         pud_t *pud;
 452         pmd_t *pmd;
 453         pte_t *pte;
 454
 455         /*
 456          * Remove any mappings which extend past the end of physical
 457          * memory from the boot time page table:
 458          */
 459         for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
 460                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
 461                 pgd = base + pgd_index(va);
 462                 if (!pgd_present(*pgd))
 463                         break;
 464
 465                 pud = pud_offset(pgd, va);
 466                 pmd = pmd_offset(pud, va);
 467                 if (!pmd_present(*pmd))
 468                         break;
 469
 470                 pte = pte_offset_kernel(pmd, va);
 471                 if (!pte_present(*pte))
 472                         break;
 473
 474                 pte_clear(NULL, va, pte);
 475         }
 476         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
 477 }
 478
 479 void __init native_pagetable_setup_done(pgd_t *base)
 480 {
 481 }
 482
 483 /*
 484  * Build a proper pagetable for the kernel mappings.  Up until this
 485  * point, we've been running on some set of pagetables constructed by
 486  * the boot process.
 487  *
 488  * If we're booting on native hardware, this will be a pagetable
 489  * constructed in arch/x86/kernel/head_32.S.  The root of the
 490  * pagetable will be swapper_pg_dir.
 491  *
 492  * If we're booting paravirtualized under a hypervisor, then there are
 493  * more options: we may already be running PAE, and the pagetable may
 494  * or may not be based in swapper_pg_dir.  In any case,
 495  * paravirt_pagetable_setup_start() will set up swapper_pg_dir
 496  * appropriately for the rest of the initialization to work.
 497  *
 498  * In general, pagetable_init() assumes that the pagetable may already
 499  * be partially populated, and so it avoids stomping on any existing
 500  * mappings.
 501  */
 502 static void __init early_ioremap_page_table_range_init(pgd_t *pgd_base)
 503 {
 504         unsigned long vaddr, end;
 505
 506         /*
 507          * Fixed mappings, only the page table structure has to be
 508          * created - mappings will be set by set_fixmap():
 509          */
 510         early_ioremap_clear();
 511         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 512         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 513         page_table_range_init(vaddr, end, pgd_base);
 514         early_ioremap_reset();
 515 }
 516
 517 static void __init pagetable_init(void)
 518 {
 519         pgd_t *pgd_base = swapper_pg_dir;
 520
 521         permanent_kmaps_init(pgd_base);
 522 }
 523
 524 #ifdef CONFIG_ACPI_SLEEP
 525 /*
 526  * ACPI suspend needs this for resume, because things like the intel-agp
 527  * driver might have split up a kernel 4MB mapping.
 528  */
 529 char swsusp_pg_dir[PAGE_SIZE]
 530         __attribute__ ((aligned(PAGE_SIZE)));
 531
 532 static inline void save_pg_dir(void)
 533 {
 534         memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
 535 }
 536 #else /* !CONFIG_ACPI_SLEEP */
 537 static inline void save_pg_dir(void)
 538 {
 539 }
 540 #endif /* !CONFIG_ACPI_SLEEP */
 541
 542 void zap_low_mappings(void)
 543 {
 544         int i;
 545
 546         /*
 547          * Zap initial low-memory mappings.
 548          *
 549          * Note that "pgd_clear()" doesn't do it for
 550          * us, because pgd_clear() is a no-op on i386.
 551          */
 552         for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
 553 #ifdef CONFIG_X86_PAE
 554                 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
 555 #else
 556                 set_pgd(swapper_pg_dir+i, __pgd(0));
 557 #endif
 558         }
 559         flush_tlb_all();
 560 }
 561
 562 int nx_enabled;
 563
 564 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
 565 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 566
 567 #ifdef CONFIG_X86_PAE
 568
 569 static int disable_nx __initdata;
 570
 571 /*
 572  * noexec = on|off
 573  *
 574  * Control non executable mappings.
 575  *
 576  * on      Enable
 577  * off     Disable
 578  */
 579 static int __init noexec_setup(char *str)
 580 {
 581         if (!str || !strcmp(str, "on")) {
 582                 if (cpu_has_nx) {
 583                         __supported_pte_mask |= _PAGE_NX;
 584                         disable_nx = 0;
 585                 }
 586         } else {
 587                 if (!strcmp(str, "off")) {
 588                         disable_nx = 1;
 589                         __supported_pte_mask &= ~_PAGE_NX;
 590                 } else {
 591                         return -EINVAL;
 592                 }
 593         }
 594
 595         return 0;
 596 }
 597 early_param("noexec", noexec_setup);
 598
 599 static void __init set_nx(void)
 600 {
 601         unsigned int v[4], l, h;
 602
 603         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
 604                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
 605
 606                 if ((v[3] & (1 << 20)) && !disable_nx) {
 607                         rdmsr(MSR_EFER, l, h);
 608                         l |= EFER_NX;
 609                         wrmsr(MSR_EFER, l, h);
 610                         nx_enabled = 1;
 611                         __supported_pte_mask |= _PAGE_NX;
 612                 }
 613         }
 614 }
 615 #endif
 616
 617 /* user-defined highmem size */
 618 static unsigned int highmem_pages = -1;
 619
 620 /*
 621  * highmem=size forces highmem to be exactly 'size' bytes.
 622  * This works even on boxes that have no highmem otherwise.
 623  * This also works to reduce highmem size on bigger boxes.
 624  */
 625 static int __init parse_highmem(char *arg)
 626 {
 627         if (!arg)
 628                 return -EINVAL;
 629
 630         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
 631         return 0;
 632 }
 633 early_param("highmem", parse_highmem);
 634
 635 /*
 636  * Determine low and high memory ranges:
 637  */
 638 void __init find_low_pfn_range(void)
 639 {
 640         /* it could update max_pfn */
 641
 642         /* max_low_pfn is 0, we already have early_res support */
 643
 644         max_low_pfn = max_pfn;
 645         if (max_low_pfn > MAXMEM_PFN) {
 646                 if (highmem_pages == -1)
 647                         highmem_pages = max_pfn - MAXMEM_PFN;
 648                 if (highmem_pages + MAXMEM_PFN < max_pfn)
 649                         max_pfn = MAXMEM_PFN + highmem_pages;
 650                 if (highmem_pages + MAXMEM_PFN > max_pfn) {
 651                         printk(KERN_WARNING "only %luMB highmem pages "
 652                                 "available, ignoring highmem size of %uMB.\n",
 653                                 pages_to_mb(max_pfn - MAXMEM_PFN),
 654                                 pages_to_mb(highmem_pages));
 655                         highmem_pages = 0;
 656                 }
 657                 max_low_pfn = MAXMEM_PFN;
 658 #ifndef CONFIG_HIGHMEM
 659                 /* Maximum memory usable is what is directly addressable */
 660                 printk(KERN_WARNING "Warning only %ldMB will be used.\n",
 661                                         MAXMEM>>20);
 662                 if (max_pfn > MAX_NONPAE_PFN)
 663                         printk(KERN_WARNING
 664                                  "Use a HIGHMEM64G enabled kernel.\n");
 665                 else
 666                         printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
 667                 max_pfn = MAXMEM_PFN;
 668 #else /* !CONFIG_HIGHMEM */
 669 #ifndef CONFIG_HIGHMEM64G
 670                 if (max_pfn > MAX_NONPAE_PFN) {
 671                         max_pfn = MAX_NONPAE_PFN;
 672                         printk(KERN_WARNING "Warning only 4GB will be used."
 673                                 "Use a HIGHMEM64G enabled kernel.\n");
 674                 }
 675 #endif /* !CONFIG_HIGHMEM64G */
 676 #endif /* !CONFIG_HIGHMEM */
 677         } else {
 678                 if (highmem_pages == -1)
 679                         highmem_pages = 0;
 680 #ifdef CONFIG_HIGHMEM
 681                 if (highmem_pages >= max_pfn) {
 682                         printk(KERN_ERR "highmem size specified (%uMB) is "
 683                                 "bigger than pages available (%luMB)!.\n",
 684                                 pages_to_mb(highmem_pages),
 685                                 pages_to_mb(max_pfn));
 686                         highmem_pages = 0;
 687                 }
 688                 if (highmem_pages) {
 689                         if (max_low_pfn - highmem_pages <
 690                             64*1024*1024/PAGE_SIZE){
 691                                 printk(KERN_ERR "highmem size %uMB results in "
 692                                 "smaller than 64MB lowmem, ignoring it.\n"
 693                                         , pages_to_mb(highmem_pages));
 694                                 highmem_pages = 0;
 695                         }
 696                         max_low_pfn -= highmem_pages;
 697                 }
 698 #else
 699                 if (highmem_pages)
 700                         printk(KERN_ERR "ignoring highmem size on non-highmem"
 701                                         " kernel!\n");
 702 #endif
 703         }
 704 }
 705
 706 #ifndef CONFIG_NEED_MULTIPLE_NODES
 707 void __init initmem_init(unsigned long start_pfn,
 708                                   unsigned long end_pfn)
 709 {
 710 #ifdef CONFIG_HIGHMEM
 711         highstart_pfn = highend_pfn = max_pfn;
 712         if (max_pfn > max_low_pfn)
 713                 highstart_pfn = max_low_pfn;
 714         memory_present(0, 0, highend_pfn);
 715         e820_register_active_regions(0, 0, highend_pfn);
 716         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 717                 pages_to_mb(highend_pfn - highstart_pfn));
 718         num_physpages = highend_pfn;
 719         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 720 #else
 721         memory_present(0, 0, max_low_pfn);
 722         e820_register_active_regions(0, 0, max_low_pfn);
 723         num_physpages = max_low_pfn;
 724         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 725 #endif
 726 #ifdef CONFIG_FLATMEM
 727         max_mapnr = num_physpages;
 728 #endif
 729         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 730                         pages_to_mb(max_low_pfn));
 731
 732         setup_bootmem_allocator();
 733 }
 734 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
 735
 736 static void __init zone_sizes_init(void)
 737 {
 738         unsigned long max_zone_pfns[MAX_NR_ZONES];
 739         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 740         max_zone_pfns[ZONE_DMA] =
 741                 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
 742         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
 743 #ifdef CONFIG_HIGHMEM
 744         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
 745 #endif
 746
 747         free_area_init_nodes(max_zone_pfns);
 748 }
 749
 750 void __init setup_bootmem_allocator(void)
 751 {
 752         int i;
 753         unsigned long bootmap_size, bootmap;
 754         /*
 755          * Initialize the boot-time allocator (with low memory only):
 756          */
 757         bootmap_size = bootmem_bootmap_pages(max_low_pfn)<<PAGE_SHIFT;
 758         bootmap = find_e820_area(min_low_pfn<<PAGE_SHIFT,
 759                                  max_pfn_mapped<<PAGE_SHIFT, bootmap_size,
 760                                  PAGE_SIZE);
 761         if (bootmap == -1L)
 762                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 763         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
 764
 765         /* don't touch min_low_pfn */
 766         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 767                                          min_low_pfn, max_low_pfn);
 768         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
 769                  max_pfn_mapped<<PAGE_SHIFT);
 770         printk(KERN_INFO "  low ram: %08lx - %08lx\n",
 771                  min_low_pfn<<PAGE_SHIFT, max_low_pfn<<PAGE_SHIFT);
 772         printk(KERN_INFO "  bootmap %08lx - %08lx\n",
 773                  bootmap, bootmap + bootmap_size);
 774         for_each_online_node(i)
 775                 free_bootmem_with_active_regions(i, max_low_pfn);
 776         early_res_to_bootmem(0, max_low_pfn<<PAGE_SHIFT);
 777
 778         after_init_bootmem = 1;
 779 }
 780
 781 static void __init find_early_table_space(unsigned long end, int use_pse)
 782 {
 783         unsigned long puds, pmds, ptes, tables, start;
 784
 785         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 786         tables = PAGE_ALIGN(puds * sizeof(pud_t));
 787
 788         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 789         tables += PAGE_ALIGN(pmds * sizeof(pmd_t));
 790
 791         if (use_pse) {
 792                 unsigned long extra;
 793
 794                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 795                 extra += PMD_SIZE;
 796                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 797         } else
 798                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 799
 800         tables += PAGE_ALIGN(ptes * sizeof(pte_t));
 801
 802         /* for fixmap */
 803         tables += PAGE_SIZE * 2;
 804
 805         /*
 806          * RED-PEN putting page tables only on node 0 could
 807          * cause a hotspot and fill up ZONE_DMA. The page tables
 808          * need roughly 0.5KB per GB.
 809          */
 810         start = 0x7000;
 811         table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
 812                                         tables, PAGE_SIZE);
 813         if (table_start == -1UL)
 814                 panic("Cannot find space for the kernel page tables");
 815
 816         table_start >>= PAGE_SHIFT;
 817         table_end = table_start;
 818         table_top = table_start + (tables>>PAGE_SHIFT);
 819
 820         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 821                 end, table_start << PAGE_SHIFT,
 822                 (table_start << PAGE_SHIFT) + tables);
 823 }
 824
 825 unsigned long __init_refok init_memory_mapping(unsigned long start,
 826                                                 unsigned long end)
 827 {
 828         pgd_t *pgd_base = swapper_pg_dir;
 829         unsigned long start_pfn, end_pfn;
 830         unsigned long big_page_start;
 831 #ifdef CONFIG_DEBUG_PAGEALLOC
 832         /*
 833          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 834          * This will simplify cpa(), which otherwise needs to support splitting
 835          * large pages into small in interrupt context, etc.
 836          */
 837         int use_pse = 0;
 838 #else
 839         int use_pse = cpu_has_pse;
 840 #endif
 841
 842         /*
 843          * Find space for the kernel direct mapping tables.
 844          */
 845         if (!after_init_bootmem)
 846                 find_early_table_space(end, use_pse);
 847
 848 #ifdef CONFIG_X86_PAE
 849         set_nx();
 850         if (nx_enabled)
 851                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
 852 #endif
 853
 854         /* Enable PSE if available */
 855         if (cpu_has_pse)
 856                 set_in_cr4(X86_CR4_PSE);
 857
 858         /* Enable PGE if available */
 859         if (cpu_has_pge) {
 860                 set_in_cr4(X86_CR4_PGE);
 861                 __supported_pte_mask |= _PAGE_GLOBAL;
 862         }
 863
 864         /*
 865          * Don't use a large page for the first 2/4MB of memory
 866          * because there are often fixed size MTRRs in there
 867          * and overlapping MTRRs into large pages can cause
 868          * slowdowns.
 869          */
 870         big_page_start = PMD_SIZE;
 871
 872         if (start < big_page_start) {
 873                 start_pfn = start >> PAGE_SHIFT;
 874                 end_pfn = min(big_page_start>>PAGE_SHIFT, end>>PAGE_SHIFT);
 875         } else {
 876                 /* head is not big page alignment ? */
 877                 start_pfn = start >> PAGE_SHIFT;
 878                 end_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 879                                  << (PMD_SHIFT - PAGE_SHIFT);
 880         }
 881         if (start_pfn < end_pfn)
 882                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn, 0);
 883
 884         /* big page range */
 885         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 886                          << (PMD_SHIFT - PAGE_SHIFT);
 887         if (start_pfn < (big_page_start >> PAGE_SHIFT))
 888                 start_pfn =  big_page_start >> PAGE_SHIFT;
 889         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 890         if (start_pfn < end_pfn)
 891                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn,
 892                                              use_pse);
 893
 894         /* tail is not big page alignment ? */
 895         start_pfn = end_pfn;
 896         if (start_pfn > (big_page_start>>PAGE_SHIFT)) {
 897                 end_pfn = end >> PAGE_SHIFT;
 898                 if (start_pfn < end_pfn)
 899                         kernel_physical_mapping_init(pgd_base, start_pfn,
 900                                                          end_pfn, 0);
 901         }
 902
 903         early_ioremap_page_table_range_init(pgd_base);
 904
 905         load_cr3(swapper_pg_dir);
 906
 907         __flush_tlb_all();
 908
 909         if (!after_init_bootmem)
 910                 reserve_early(table_start << PAGE_SHIFT,
 911                                  table_end << PAGE_SHIFT, "PGTABLE");
 912
 913         if (!after_init_bootmem)
 914                 early_memtest(start, end);
 915
 916         return end >> PAGE_SHIFT;
 917 }
 918
 919
 920 /*
 921  * paging_init() sets up the page tables - note that the first 8MB are
 922  * already mapped by head.S.
 923  *
 924  * This routines also unmaps the page at virtual kernel address 0, so
 925  * that we can trap those pesky NULL-reference errors in the kernel.
 926  */
 927 void __init paging_init(void)
 928 {
 929         pagetable_init();
 930
 931         __flush_tlb_all();
 932
 933         kmap_init();
 934
 935         /*
 936          * NOTE: at this point the bootmem allocator is fully available.
 937          */
 938         sparse_init();
 939         zone_sizes_init();
 940 }
 941
 942 /*
 943  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
 944  * and also on some strange 486's. All 586+'s are OK. This used to involve
 945  * black magic jumps to work around some nasty CPU bugs, but fortunately the
 946  * switch to using exceptions got rid of all that.
 947  */
 948 static void __init test_wp_bit(void)
 949 {
 950         printk(KERN_INFO
 951   "Checking if this processor honours the WP bit even in supervisor mode...");
 952
 953         /* Any page-aligned address will do, the test is non-destructive */
 954         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
 955         boot_cpu_data.wp_works_ok = do_test_wp_bit();
 956         clear_fixmap(FIX_WP_TEST);
 957
 958         if (!boot_cpu_data.wp_works_ok) {
 959                 printk(KERN_CONT "No.\n");
 960 #ifdef CONFIG_X86_WP_WORKS_OK
 961                 panic(
 962   "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
 963 #endif
 964         } else {
 965                 printk(KERN_CONT "Ok.\n");
 966         }
 967 }
 968
 969 static struct kcore_list kcore_mem, kcore_vmalloc;
 970
 971 void __init mem_init(void)
 972 {
 973         int codesize, reservedpages, datasize, initsize;
 974         int tmp;
 975
 976         pci_iommu_alloc();
 977
 978 #ifdef CONFIG_FLATMEM
 979         BUG_ON(!mem_map);
 980 #endif
 981         /* this will put all low memory onto the freelists */
 982         totalram_pages += free_all_bootmem();
 983
 984         reservedpages = 0;
 985         for (tmp = 0; tmp < max_low_pfn; tmp++)
 986                 /*
 987                  * Only count reserved RAM pages:
 988                  */
 989                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
 990                         reservedpages++;
 991
 992         set_highmem_pages_init();
 993
 994         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 995         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 996         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 997
 998         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 999         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
1000                    VMALLOC_END-VMALLOC_START);
1001
1002         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
1003                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
1004                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
1005                 num_physpages << (PAGE_SHIFT-10),
1006                 codesize >> 10,
1007                 reservedpages << (PAGE_SHIFT-10),
1008                 datasize >> 10,
1009                 initsize >> 10,
1010                 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
1011                );
1012
1013         printk(KERN_INFO "virtual kernel memory layout:\n"
1014                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1015 #ifdef CONFIG_HIGHMEM
1016                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1017 #endif
1018                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1019                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1020                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1021                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1022                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
1023                 FIXADDR_START, FIXADDR_TOP,
1024                 (FIXADDR_TOP - FIXADDR_START) >> 10,
1025
1026 #ifdef CONFIG_HIGHMEM
1027                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
1028                 (LAST_PKMAP*PAGE_SIZE) >> 10,
1029 #endif
1030
1031                 VMALLOC_START, VMALLOC_END,
1032                 (VMALLOC_END - VMALLOC_START) >> 20,
1033
1034                 (unsigned long)__va(0), (unsigned long)high_memory,
1035                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
1036
1037                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
1038                 ((unsigned long)&__init_end -
1039                  (unsigned long)&__init_begin) >> 10,
1040
1041                 (unsigned long)&_etext, (unsigned long)&_edata,
1042                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
1043
1044                 (unsigned long)&_text, (unsigned long)&_etext,
1045                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
1046
1047         /*
1048          * Check boundaries twice: Some fundamental inconsistencies can
1049          * be detected at build time already.
1050          */
1051 #define __FIXADDR_TOP (-PAGE_SIZE)
1052 #ifdef CONFIG_HIGHMEM
1053         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
1054         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
1055 #endif
1056 #define high_memory (-128UL << 20)
1057         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
1058 #undef high_memory
1059 #undef __FIXADDR_TOP
1060
1061 #ifdef CONFIG_HIGHMEM
1062         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
1063         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
1064 #endif
1065         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
1066         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
1067
1068         if (boot_cpu_data.wp_works_ok < 0)
1069                 test_wp_bit();
1070
1071         save_pg_dir();
1072         zap_low_mappings();
1073 }
1074
1075 #ifdef CONFIG_MEMORY_HOTPLUG
1076 int arch_add_memory(int nid, u64 start, u64 size)
1077 {
1078         struct pglist_data *pgdata = NODE_DATA(nid);
1079         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
1080         unsigned long start_pfn = start >> PAGE_SHIFT;
1081         unsigned long nr_pages = size >> PAGE_SHIFT;
1082
1083         return __add_pages(nid, zone, start_pfn, nr_pages);
1084 }
1085 #endif
1086
1087 /*
1088  * This function cannot be __init, since exceptions don't work in that
1089  * section.  Put this after the callers, so that it cannot be inlined.
1090  */
1091 static noinline int do_test_wp_bit(void)
1092 {
1093         char tmp_reg;
1094         int flag;
1095
1096         __asm__ __volatile__(
1097                 "       movb %0, %1     \n"
1098                 "1:     movb %1, %0     \n"
1099                 "       xorl %2, %2     \n"
1100                 "2:                     \n"
1101                 _ASM_EXTABLE(1b,2b)
1102                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
1103                  "=q" (tmp_reg),
1104                  "=r" (flag)
1105                 :"2" (1)
1106                 :"memory");
1107
1108         return flag;
1109 }
1110
1111 #ifdef CONFIG_DEBUG_RODATA
1112 const int rodata_test_data = 0xC3;
1113 EXPORT_SYMBOL_GPL(rodata_test_data);
1114
1115 void mark_rodata_ro(void)
1116 {
1117         unsigned long start = PFN_ALIGN(_text);
1118         unsigned long size = PFN_ALIGN(_etext) - start;
1119
1120 #ifndef CONFIG_DYNAMIC_FTRACE
1121         /* Dynamic tracing modifies the kernel text section */
1122         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1123         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
1124                 size >> 10);
1125
1126 #ifdef CONFIG_CPA_DEBUG
1127         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
1128                 start, start+size);
1129         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
1130
1131         printk(KERN_INFO "Testing CPA: write protecting again\n");
1132         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
1133 #endif
1134 #endif /* CONFIG_DYNAMIC_FTRACE */
1135
1136         start += size;
1137         size = (unsigned long)__end_rodata - start;
1138         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1139         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1140                 size >> 10);
1141         rodata_test();
1142
1143 #ifdef CONFIG_CPA_DEBUG
1144         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
1145         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
1146
1147         printk(KERN_INFO "Testing CPA: write protecting again\n");
1148         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1149 #endif
1150 }
1151 #endif
1152
1153 void free_init_pages(char *what, unsigned long begin, unsigned long end)
1154 {
1155 #ifdef CONFIG_DEBUG_PAGEALLOC
1156         /*
1157          * If debugging page accesses then do not free this memory but
1158          * mark them not present - any buggy init-section access will
1159          * create a kernel page fault:
1160          */
1161         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
1162                 begin, PAGE_ALIGN(end));
1163         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
1164 #else
1165         unsigned long addr;
1166
1167         /*
1168          * We just marked the kernel text read only above, now that
1169          * we are going to free part of that, we need to make that
1170          * writeable first.
1171          */
1172         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
1173
1174         for (addr = begin; addr < end; addr += PAGE_SIZE) {
1175                 ClearPageReserved(virt_to_page(addr));
1176                 init_page_count(virt_to_page(addr));
1177                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
1178                 free_page(addr);
1179                 totalram_pages++;
1180         }
1181         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
1182 #endif
1183 }
1184
1185 void free_initmem(void)
1186 {
1187         free_init_pages("unused kernel memory",
1188                         (unsigned long)(&__init_begin),
1189                         (unsigned long)(&__init_end));
1190 }
1191
1192 #ifdef CONFIG_BLK_DEV_INITRD
1193 void free_initrd_mem(unsigned long start, unsigned long end)
1194 {
1195         free_init_pages("initrd memory", start, end);
1196 }
1197 #endif
1198
1199 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1200                                    int flags)
1201 {
1202         return reserve_bootmem(phys, len, flags);
1203 }