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/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 #include <asm/smp.h>
  53
  54 unsigned int __VMALLOC_RESERVE = 128 << 20;
  55
  56 unsigned long max_low_pfn_mapped;
  57 unsigned long max_pfn_mapped;
  58
  59 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  60 unsigned long highstart_pfn, highend_pfn;
  61
  62 static noinline int do_test_wp_bit(void);
  63
  64
  65 static unsigned long __initdata table_start;
  66 static unsigned long __meminitdata table_end;
  67 static unsigned long __meminitdata table_top;
  68
  69 static int __initdata after_init_bootmem;
  70
  71 static __init void *alloc_low_page(void)
  72 {
  73         unsigned long pfn = table_end++;
  74         void *adr;
  75
  76         if (pfn >= table_top)
  77                 panic("alloc_low_page: ran out of memory");
  78
  79         adr = __va(pfn * PAGE_SIZE);
  80         memset(adr, 0, PAGE_SIZE);
  81         return adr;
  82 }
  83
  84 /*
  85  * Creates a middle page table and puts a pointer to it in the
  86  * given global directory entry. This only returns the gd entry
  87  * in non-PAE compilation mode, since the middle layer is folded.
  88  */
  89 static pmd_t * __init one_md_table_init(pgd_t *pgd)
  90 {
  91         pud_t *pud;
  92         pmd_t *pmd_table;
  93
  94 #ifdef CONFIG_X86_PAE
  95         if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
  96                 if (after_init_bootmem)
  97                         pmd_table = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
  98                 else
  99                         pmd_table = (pmd_t *)alloc_low_page();
 100                 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
 101                 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
 102                 pud = pud_offset(pgd, 0);
 103                 BUG_ON(pmd_table != pmd_offset(pud, 0));
 104
 105                 return pmd_table;
 106         }
 107 #endif
 108         pud = pud_offset(pgd, 0);
 109         pmd_table = pmd_offset(pud, 0);
 110
 111         return pmd_table;
 112 }
 113
 114 /*
 115  * Create a page table and place a pointer to it in a middle page
 116  * directory entry:
 117  */
 118 static pte_t * __init one_page_table_init(pmd_t *pmd)
 119 {
 120         if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
 121                 pte_t *page_table = NULL;
 122
 123                 if (after_init_bootmem) {
 124 #ifdef CONFIG_DEBUG_PAGEALLOC
 125                         page_table = (pte_t *) alloc_bootmem_pages(PAGE_SIZE);
 126 #endif
 127                         if (!page_table)
 128                                 page_table =
 129                                 (pte_t *)alloc_bootmem_low_pages(PAGE_SIZE);
 130                 } else
 131                         page_table = (pte_t *)alloc_low_page();
 132
 133                 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
 134                 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
 135                 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
 136         }
 137
 138         return pte_offset_kernel(pmd, 0);
 139 }
 140
 141 /*
 142  * This function initializes a certain range of kernel virtual memory
 143  * with new bootmem page tables, everywhere page tables are missing in
 144  * the given range.
 145  *
 146  * NOTE: The pagetables are allocated contiguous on the physical space
 147  * so we can cache the place of the first one and move around without
 148  * checking the pgd every time.
 149  */
 150 static void __init
 151 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
 152 {
 153         int pgd_idx, pmd_idx;
 154         unsigned long vaddr;
 155         pgd_t *pgd;
 156         pmd_t *pmd;
 157
 158         vaddr = start;
 159         pgd_idx = pgd_index(vaddr);
 160         pmd_idx = pmd_index(vaddr);
 161         pgd = pgd_base + pgd_idx;
 162
 163         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
 164                 pmd = one_md_table_init(pgd);
 165                 pmd = pmd + pmd_index(vaddr);
 166                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 167                                                         pmd++, pmd_idx++) {
 168                         one_page_table_init(pmd);
 169
 170                         vaddr += PMD_SIZE;
 171                 }
 172                 pmd_idx = 0;
 173         }
 174 }
 175
 176 static inline int is_kernel_text(unsigned long addr)
 177 {
 178         if (addr >= PAGE_OFFSET && addr <= (unsigned long)__init_end)
 179                 return 1;
 180         return 0;
 181 }
 182
 183 /*
 184  * This maps the physical memory to kernel virtual address space, a total
 185  * of max_low_pfn pages, by creating page tables starting from address
 186  * PAGE_OFFSET:
 187  */
 188 static void __init kernel_physical_mapping_init(pgd_t *pgd_base,
 189                                                 unsigned long start_pfn,
 190                                                 unsigned long end_pfn,
 191                                                 int use_pse)
 192 {
 193         int pgd_idx, pmd_idx, pte_ofs;
 194         unsigned long pfn;
 195         pgd_t *pgd;
 196         pmd_t *pmd;
 197         pte_t *pte;
 198         unsigned pages_2m, pages_4k;
 199         int mapping_iter;
 200
 201         /*
 202          * First iteration will setup identity mapping using large/small pages
 203          * based on use_pse, with other attributes same as set by
 204          * the early code in head_32.S
 205          *
 206          * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
 207          * as desired for the kernel identity mapping.
 208          *
 209          * This two pass mechanism conforms to the TLB app note which says:
 210          *
 211          *     "Software should not write to a paging-structure entry in a way
 212          *      that would change, for any linear address, both the page size
 213          *      and either the page frame or attributes."
 214          */
 215         mapping_iter = 1;
 216
 217         if (!cpu_has_pse)
 218                 use_pse = 0;
 219
 220 repeat:
 221         pages_2m = pages_4k = 0;
 222         pfn = start_pfn;
 223         pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 224         pgd = pgd_base + pgd_idx;
 225         for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
 226                 pmd = one_md_table_init(pgd);
 227
 228                 if (pfn >= end_pfn)
 229                         continue;
 230 #ifdef CONFIG_X86_PAE
 231                 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 232                 pmd += pmd_idx;
 233 #else
 234                 pmd_idx = 0;
 235 #endif
 236                 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
 237                      pmd++, pmd_idx++) {
 238                         unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
 239
 240                         /*
 241                          * Map with big pages if possible, otherwise
 242                          * create normal page tables:
 243                          */
 244                         if (use_pse) {
 245                                 unsigned int addr2;
 246                                 pgprot_t prot = PAGE_KERNEL_LARGE;
 247                                 /*
 248                                  * first pass will use the same initial
 249                                  * identity mapping attribute + _PAGE_PSE.
 250                                  */
 251                                 pgprot_t init_prot =
 252                                         __pgprot(PTE_IDENT_ATTR |
 253                                                  _PAGE_PSE);
 254
 255                                 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
 256                                         PAGE_OFFSET + PAGE_SIZE-1;
 257
 258                                 if (is_kernel_text(addr) ||
 259                                     is_kernel_text(addr2))
 260                                         prot = PAGE_KERNEL_LARGE_EXEC;
 261
 262                                 pages_2m++;
 263                                 if (mapping_iter == 1)
 264                                         set_pmd(pmd, pfn_pmd(pfn, init_prot));
 265                                 else
 266                                         set_pmd(pmd, pfn_pmd(pfn, prot));
 267
 268                                 pfn += PTRS_PER_PTE;
 269                                 continue;
 270                         }
 271                         pte = one_page_table_init(pmd);
 272
 273                         pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 274                         pte += pte_ofs;
 275                         for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
 276                              pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
 277                                 pgprot_t prot = PAGE_KERNEL;
 278                                 /*
 279                                  * first pass will use the same initial
 280                                  * identity mapping attribute.
 281                                  */
 282                                 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
 283
 284                                 if (is_kernel_text(addr))
 285                                         prot = PAGE_KERNEL_EXEC;
 286
 287                                 pages_4k++;
 288                                 if (mapping_iter == 1)
 289                                         set_pte(pte, pfn_pte(pfn, init_prot));
 290                                 else
 291                                         set_pte(pte, pfn_pte(pfn, prot));
 292                         }
 293                 }
 294         }
 295         if (mapping_iter == 1) {
 296                 /*
 297                  * update direct mapping page count only in the first
 298                  * iteration.
 299                  */
 300                 update_page_count(PG_LEVEL_2M, pages_2m);
 301                 update_page_count(PG_LEVEL_4K, pages_4k);
 302
 303                 /*
 304                  * local global flush tlb, which will flush the previous
 305                  * mappings present in both small and large page TLB's.
 306                  */
 307                 __flush_tlb_all();
 308
 309                 /*
 310                  * Second iteration will set the actual desired PTE attributes.
 311                  */
 312                 mapping_iter = 2;
 313                 goto repeat;
 314         }
 315 }
 316
 317 /*
 318  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 319  * is valid. The argument is a physical page number.
 320  *
 321  *
 322  * On x86, access has to be given to the first megabyte of ram because that area
 323  * contains bios code and data regions used by X and dosemu and similar apps.
 324  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 325  * mmio resources as well as potential bios/acpi data regions.
 326  */
 327 int devmem_is_allowed(unsigned long pagenr)
 328 {
 329         if (pagenr <= 256)
 330                 return 1;
 331         if (!page_is_ram(pagenr))
 332                 return 1;
 333         return 0;
 334 }
 335
 336 pte_t *kmap_pte;
 337 pgprot_t kmap_prot;
 338
 339 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
 340 {
 341         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
 342                         vaddr), vaddr), vaddr);
 343 }
 344
 345 static void __init kmap_init(void)
 346 {
 347         unsigned long kmap_vstart;
 348
 349         /*
 350          * Cache the first kmap pte:
 351          */
 352         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
 353         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
 354
 355         kmap_prot = PAGE_KERNEL;
 356 }
 357
 358 #ifdef CONFIG_HIGHMEM
 359 static void __init permanent_kmaps_init(pgd_t *pgd_base)
 360 {
 361         unsigned long vaddr;
 362         pgd_t *pgd;
 363         pud_t *pud;
 364         pmd_t *pmd;
 365         pte_t *pte;
 366
 367         vaddr = PKMAP_BASE;
 368         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
 369
 370         pgd = swapper_pg_dir + pgd_index(vaddr);
 371         pud = pud_offset(pgd, vaddr);
 372         pmd = pmd_offset(pud, vaddr);
 373         pte = pte_offset_kernel(pmd, vaddr);
 374         pkmap_page_table = pte;
 375 }
 376
 377 static void __init add_one_highpage_init(struct page *page, int pfn)
 378 {
 379         ClearPageReserved(page);
 380         init_page_count(page);
 381         __free_page(page);
 382         totalhigh_pages++;
 383 }
 384
 385 struct add_highpages_data {
 386         unsigned long start_pfn;
 387         unsigned long end_pfn;
 388 };
 389
 390 static int __init add_highpages_work_fn(unsigned long start_pfn,
 391                                          unsigned long end_pfn, void *datax)
 392 {
 393         int node_pfn;
 394         struct page *page;
 395         unsigned long final_start_pfn, final_end_pfn;
 396         struct add_highpages_data *data;
 397
 398         data = (struct add_highpages_data *)datax;
 399
 400         final_start_pfn = max(start_pfn, data->start_pfn);
 401         final_end_pfn = min(end_pfn, data->end_pfn);
 402         if (final_start_pfn >= final_end_pfn)
 403                 return 0;
 404
 405         for (node_pfn = final_start_pfn; node_pfn < final_end_pfn;
 406              node_pfn++) {
 407                 if (!pfn_valid(node_pfn))
 408                         continue;
 409                 page = pfn_to_page(node_pfn);
 410                 add_one_highpage_init(page, node_pfn);
 411         }
 412
 413         return 0;
 414
 415 }
 416
 417 void __init add_highpages_with_active_regions(int nid, unsigned long start_pfn,
 418                                               unsigned long end_pfn)
 419 {
 420         struct add_highpages_data data;
 421
 422         data.start_pfn = start_pfn;
 423         data.end_pfn = end_pfn;
 424
 425         work_with_active_regions(nid, add_highpages_work_fn, &data);
 426 }
 427
 428 #ifndef CONFIG_NUMA
 429 static void __init set_highmem_pages_init(void)
 430 {
 431         add_highpages_with_active_regions(0, highstart_pfn, highend_pfn);
 432
 433         totalram_pages += totalhigh_pages;
 434 }
 435 #endif /* !CONFIG_NUMA */
 436
 437 #else
 438 # define permanent_kmaps_init(pgd_base)         do { } while (0)
 439 # define set_highmem_pages_init()       do { } while (0)
 440 #endif /* CONFIG_HIGHMEM */
 441
 442 void __init native_pagetable_setup_start(pgd_t *base)
 443 {
 444         unsigned long pfn, va;
 445         pgd_t *pgd;
 446         pud_t *pud;
 447         pmd_t *pmd;
 448         pte_t *pte;
 449
 450         /*
 451          * Remove any mappings which extend past the end of physical
 452          * memory from the boot time page table:
 453          */
 454         for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
 455                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
 456                 pgd = base + pgd_index(va);
 457                 if (!pgd_present(*pgd))
 458                         break;
 459
 460                 pud = pud_offset(pgd, va);
 461                 pmd = pmd_offset(pud, va);
 462                 if (!pmd_present(*pmd))
 463                         break;
 464
 465                 pte = pte_offset_kernel(pmd, va);
 466                 if (!pte_present(*pte))
 467                         break;
 468
 469                 pte_clear(NULL, va, pte);
 470         }
 471         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
 472 }
 473
 474 void __init native_pagetable_setup_done(pgd_t *base)
 475 {
 476 }
 477
 478 /*
 479  * Build a proper pagetable for the kernel mappings.  Up until this
 480  * point, we've been running on some set of pagetables constructed by
 481  * the boot process.
 482  *
 483  * If we're booting on native hardware, this will be a pagetable
 484  * constructed in arch/x86/kernel/head_32.S.  The root of the
 485  * pagetable will be swapper_pg_dir.
 486  *
 487  * If we're booting paravirtualized under a hypervisor, then there are
 488  * more options: we may already be running PAE, and the pagetable may
 489  * or may not be based in swapper_pg_dir.  In any case,
 490  * paravirt_pagetable_setup_start() will set up swapper_pg_dir
 491  * appropriately for the rest of the initialization to work.
 492  *
 493  * In general, pagetable_init() assumes that the pagetable may already
 494  * be partially populated, and so it avoids stomping on any existing
 495  * mappings.
 496  */
 497 static void __init early_ioremap_page_table_range_init(pgd_t *pgd_base)
 498 {
 499         unsigned long vaddr, end;
 500
 501         /*
 502          * Fixed mappings, only the page table structure has to be
 503          * created - mappings will be set by set_fixmap():
 504          */
 505         early_ioremap_clear();
 506         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 507         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 508         page_table_range_init(vaddr, end, pgd_base);
 509         early_ioremap_reset();
 510 }
 511
 512 static void __init pagetable_init(void)
 513 {
 514         pgd_t *pgd_base = swapper_pg_dir;
 515
 516         permanent_kmaps_init(pgd_base);
 517 }
 518
 519 #ifdef CONFIG_ACPI_SLEEP
 520 /*
 521  * ACPI suspend needs this for resume, because things like the intel-agp
 522  * driver might have split up a kernel 4MB mapping.
 523  */
 524 char swsusp_pg_dir[PAGE_SIZE]
 525         __attribute__ ((aligned(PAGE_SIZE)));
 526
 527 static inline void save_pg_dir(void)
 528 {
 529         memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
 530 }
 531 #else /* !CONFIG_ACPI_SLEEP */
 532 static inline void save_pg_dir(void)
 533 {
 534 }
 535 #endif /* !CONFIG_ACPI_SLEEP */
 536
 537 void zap_low_mappings(void)
 538 {
 539         int i;
 540
 541         /*
 542          * Zap initial low-memory mappings.
 543          *
 544          * Note that "pgd_clear()" doesn't do it for
 545          * us, because pgd_clear() is a no-op on i386.
 546          */
 547         for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
 548 #ifdef CONFIG_X86_PAE
 549                 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
 550 #else
 551                 set_pgd(swapper_pg_dir+i, __pgd(0));
 552 #endif
 553         }
 554         flush_tlb_all();
 555 }
 556
 557 int nx_enabled;
 558
 559 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
 560 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 561
 562 #ifdef CONFIG_X86_PAE
 563
 564 static int disable_nx __initdata;
 565
 566 /*
 567  * noexec = on|off
 568  *
 569  * Control non executable mappings.
 570  *
 571  * on      Enable
 572  * off     Disable
 573  */
 574 static int __init noexec_setup(char *str)
 575 {
 576         if (!str || !strcmp(str, "on")) {
 577                 if (cpu_has_nx) {
 578                         __supported_pte_mask |= _PAGE_NX;
 579                         disable_nx = 0;
 580                 }
 581         } else {
 582                 if (!strcmp(str, "off")) {
 583                         disable_nx = 1;
 584                         __supported_pte_mask &= ~_PAGE_NX;
 585                 } else {
 586                         return -EINVAL;
 587                 }
 588         }
 589
 590         return 0;
 591 }
 592 early_param("noexec", noexec_setup);
 593
 594 static void __init set_nx(void)
 595 {
 596         unsigned int v[4], l, h;
 597
 598         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
 599                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
 600
 601                 if ((v[3] & (1 << 20)) && !disable_nx) {
 602                         rdmsr(MSR_EFER, l, h);
 603                         l |= EFER_NX;
 604                         wrmsr(MSR_EFER, l, h);
 605                         nx_enabled = 1;
 606                         __supported_pte_mask |= _PAGE_NX;
 607                 }
 608         }
 609 }
 610 #endif
 611
 612 /* user-defined highmem size */
 613 static unsigned int highmem_pages = -1;
 614
 615 /*
 616  * highmem=size forces highmem to be exactly 'size' bytes.
 617  * This works even on boxes that have no highmem otherwise.
 618  * This also works to reduce highmem size on bigger boxes.
 619  */
 620 static int __init parse_highmem(char *arg)
 621 {
 622         if (!arg)
 623                 return -EINVAL;
 624
 625         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
 626         return 0;
 627 }
 628 early_param("highmem", parse_highmem);
 629
 630 /*
 631  * Determine low and high memory ranges:
 632  */
 633 void __init find_low_pfn_range(void)
 634 {
 635         /* it could update max_pfn */
 636
 637         /* max_low_pfn is 0, we already have early_res support */
 638
 639         max_low_pfn = max_pfn;
 640         if (max_low_pfn > MAXMEM_PFN) {
 641                 if (highmem_pages == -1)
 642                         highmem_pages = max_pfn - MAXMEM_PFN;
 643                 if (highmem_pages + MAXMEM_PFN < max_pfn)
 644                         max_pfn = MAXMEM_PFN + highmem_pages;
 645                 if (highmem_pages + MAXMEM_PFN > max_pfn) {
 646                         printk(KERN_WARNING "only %luMB highmem pages "
 647                                 "available, ignoring highmem size of %uMB.\n",
 648                                 pages_to_mb(max_pfn - MAXMEM_PFN),
 649                                 pages_to_mb(highmem_pages));
 650                         highmem_pages = 0;
 651                 }
 652                 max_low_pfn = MAXMEM_PFN;
 653 #ifndef CONFIG_HIGHMEM
 654                 /* Maximum memory usable is what is directly addressable */
 655                 printk(KERN_WARNING "Warning only %ldMB will be used.\n",
 656                                         MAXMEM>>20);
 657                 if (max_pfn > MAX_NONPAE_PFN)
 658                         printk(KERN_WARNING
 659                                  "Use a HIGHMEM64G enabled kernel.\n");
 660                 else
 661                         printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
 662                 max_pfn = MAXMEM_PFN;
 663 #else /* !CONFIG_HIGHMEM */
 664 #ifndef CONFIG_HIGHMEM64G
 665                 if (max_pfn > MAX_NONPAE_PFN) {
 666                         max_pfn = MAX_NONPAE_PFN;
 667                         printk(KERN_WARNING "Warning only 4GB will be used."
 668                                 "Use a HIGHMEM64G enabled kernel.\n");
 669                 }
 670 #endif /* !CONFIG_HIGHMEM64G */
 671 #endif /* !CONFIG_HIGHMEM */
 672         } else {
 673                 if (highmem_pages == -1)
 674                         highmem_pages = 0;
 675 #ifdef CONFIG_HIGHMEM
 676                 if (highmem_pages >= max_pfn) {
 677                         printk(KERN_ERR "highmem size specified (%uMB) is "
 678                                 "bigger than pages available (%luMB)!.\n",
 679                                 pages_to_mb(highmem_pages),
 680                                 pages_to_mb(max_pfn));
 681                         highmem_pages = 0;
 682                 }
 683                 if (highmem_pages) {
 684                         if (max_low_pfn - highmem_pages <
 685                             64*1024*1024/PAGE_SIZE){
 686                                 printk(KERN_ERR "highmem size %uMB results in "
 687                                 "smaller than 64MB lowmem, ignoring it.\n"
 688                                         , pages_to_mb(highmem_pages));
 689                                 highmem_pages = 0;
 690                         }
 691                         max_low_pfn -= highmem_pages;
 692                 }
 693 #else
 694                 if (highmem_pages)
 695                         printk(KERN_ERR "ignoring highmem size on non-highmem"
 696                                         " kernel!\n");
 697 #endif
 698         }
 699 }
 700
 701 #ifndef CONFIG_NEED_MULTIPLE_NODES
 702 void __init initmem_init(unsigned long start_pfn,
 703                                   unsigned long end_pfn)
 704 {
 705 #ifdef CONFIG_HIGHMEM
 706         highstart_pfn = highend_pfn = max_pfn;
 707         if (max_pfn > max_low_pfn)
 708                 highstart_pfn = max_low_pfn;
 709         memory_present(0, 0, highend_pfn);
 710         e820_register_active_regions(0, 0, highend_pfn);
 711         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 712                 pages_to_mb(highend_pfn - highstart_pfn));
 713         num_physpages = highend_pfn;
 714         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 715 #else
 716         memory_present(0, 0, max_low_pfn);
 717         e820_register_active_regions(0, 0, max_low_pfn);
 718         num_physpages = max_low_pfn;
 719         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 720 #endif
 721 #ifdef CONFIG_FLATMEM
 722         max_mapnr = num_physpages;
 723 #endif
 724         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 725                         pages_to_mb(max_low_pfn));
 726
 727         setup_bootmem_allocator();
 728 }
 729 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
 730
 731 static void __init zone_sizes_init(void)
 732 {
 733         unsigned long max_zone_pfns[MAX_NR_ZONES];
 734         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 735         max_zone_pfns[ZONE_DMA] =
 736                 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
 737         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
 738 #ifdef CONFIG_HIGHMEM
 739         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
 740 #endif
 741
 742         free_area_init_nodes(max_zone_pfns);
 743 }
 744
 745 void __init setup_bootmem_allocator(void)
 746 {
 747         int i;
 748         unsigned long bootmap_size, bootmap;
 749         /*
 750          * Initialize the boot-time allocator (with low memory only):
 751          */
 752         bootmap_size = bootmem_bootmap_pages(max_low_pfn)<<PAGE_SHIFT;
 753         bootmap = find_e820_area(min_low_pfn<<PAGE_SHIFT,
 754                                  max_pfn_mapped<<PAGE_SHIFT, bootmap_size,
 755                                  PAGE_SIZE);
 756         if (bootmap == -1L)
 757                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 758         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
 759
 760         /* don't touch min_low_pfn */
 761         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 762                                          min_low_pfn, max_low_pfn);
 763         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
 764                  max_pfn_mapped<<PAGE_SHIFT);
 765         printk(KERN_INFO "  low ram: %08lx - %08lx\n",
 766                  min_low_pfn<<PAGE_SHIFT, max_low_pfn<<PAGE_SHIFT);
 767         printk(KERN_INFO "  bootmap %08lx - %08lx\n",
 768                  bootmap, bootmap + bootmap_size);
 769         for_each_online_node(i)
 770                 free_bootmem_with_active_regions(i, max_low_pfn);
 771         early_res_to_bootmem(0, max_low_pfn<<PAGE_SHIFT);
 772
 773         after_init_bootmem = 1;
 774 }
 775
 776 static void __init find_early_table_space(unsigned long end, int use_pse)
 777 {
 778         unsigned long puds, pmds, ptes, tables, start;
 779
 780         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 781         tables = PAGE_ALIGN(puds * sizeof(pud_t));
 782
 783         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 784         tables += PAGE_ALIGN(pmds * sizeof(pmd_t));
 785
 786         if (use_pse) {
 787                 unsigned long extra;
 788
 789                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 790                 extra += PMD_SIZE;
 791                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 792         } else
 793                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 794
 795         tables += PAGE_ALIGN(ptes * sizeof(pte_t));
 796
 797         /* for fixmap */
 798         tables += PAGE_SIZE * 2;
 799
 800         /*
 801          * RED-PEN putting page tables only on node 0 could
 802          * cause a hotspot and fill up ZONE_DMA. The page tables
 803          * need roughly 0.5KB per GB.
 804          */
 805         start = 0x7000;
 806         table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
 807                                         tables, PAGE_SIZE);
 808         if (table_start == -1UL)
 809                 panic("Cannot find space for the kernel page tables");
 810
 811         table_start >>= PAGE_SHIFT;
 812         table_end = table_start;
 813         table_top = table_start + (tables>>PAGE_SHIFT);
 814
 815         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 816                 end, table_start << PAGE_SHIFT,
 817                 (table_start << PAGE_SHIFT) + tables);
 818 }
 819
 820 unsigned long __init_refok init_memory_mapping(unsigned long start,
 821                                                 unsigned long end)
 822 {
 823         pgd_t *pgd_base = swapper_pg_dir;
 824         unsigned long start_pfn, end_pfn;
 825         unsigned long big_page_start;
 826 #ifdef CONFIG_DEBUG_PAGEALLOC
 827         /*
 828          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 829          * This will simplify cpa(), which otherwise needs to support splitting
 830          * large pages into small in interrupt context, etc.
 831          */
 832         int use_pse = 0;
 833 #else
 834         int use_pse = cpu_has_pse;
 835 #endif
 836
 837         /*
 838          * Find space for the kernel direct mapping tables.
 839          */
 840         if (!after_init_bootmem)
 841                 find_early_table_space(end, use_pse);
 842
 843 #ifdef CONFIG_X86_PAE
 844         set_nx();
 845         if (nx_enabled)
 846                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
 847 #endif
 848
 849         /* Enable PSE if available */
 850         if (cpu_has_pse)
 851                 set_in_cr4(X86_CR4_PSE);
 852
 853         /* Enable PGE if available */
 854         if (cpu_has_pge) {
 855                 set_in_cr4(X86_CR4_PGE);
 856                 __supported_pte_mask |= _PAGE_GLOBAL;
 857         }
 858
 859         /*
 860          * Don't use a large page for the first 2/4MB of memory
 861          * because there are often fixed size MTRRs in there
 862          * and overlapping MTRRs into large pages can cause
 863          * slowdowns.
 864          */
 865         big_page_start = PMD_SIZE;
 866
 867         if (start < big_page_start) {
 868                 start_pfn = start >> PAGE_SHIFT;
 869                 end_pfn = min(big_page_start>>PAGE_SHIFT, end>>PAGE_SHIFT);
 870         } else {
 871                 /* head is not big page alignment ? */
 872                 start_pfn = start >> PAGE_SHIFT;
 873                 end_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 874                                  << (PMD_SHIFT - PAGE_SHIFT);
 875         }
 876         if (start_pfn < end_pfn)
 877                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn, 0);
 878
 879         /* big page range */
 880         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 881                          << (PMD_SHIFT - PAGE_SHIFT);
 882         if (start_pfn < (big_page_start >> PAGE_SHIFT))
 883                 start_pfn =  big_page_start >> PAGE_SHIFT;
 884         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 885         if (start_pfn < end_pfn)
 886                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn,
 887                                              use_pse);
 888
 889         /* tail is not big page alignment ? */
 890         start_pfn = end_pfn;
 891         if (start_pfn > (big_page_start>>PAGE_SHIFT)) {
 892                 end_pfn = end >> PAGE_SHIFT;
 893                 if (start_pfn < end_pfn)
 894                         kernel_physical_mapping_init(pgd_base, start_pfn,
 895                                                          end_pfn, 0);
 896         }
 897
 898         early_ioremap_page_table_range_init(pgd_base);
 899
 900         load_cr3(swapper_pg_dir);
 901
 902         __flush_tlb_all();
 903
 904         if (!after_init_bootmem)
 905                 reserve_early(table_start << PAGE_SHIFT,
 906                                  table_end << PAGE_SHIFT, "PGTABLE");
 907
 908         if (!after_init_bootmem)
 909                 early_memtest(start, end);
 910
 911         return end >> PAGE_SHIFT;
 912 }
 913
 914
 915 /*
 916  * paging_init() sets up the page tables - note that the first 8MB are
 917  * already mapped by head.S.
 918  *
 919  * This routines also unmaps the page at virtual kernel address 0, so
 920  * that we can trap those pesky NULL-reference errors in the kernel.
 921  */
 922 void __init paging_init(void)
 923 {
 924         pagetable_init();
 925
 926         __flush_tlb_all();
 927
 928         kmap_init();
 929
 930         /*
 931          * NOTE: at this point the bootmem allocator is fully available.
 932          */
 933         sparse_init();
 934         zone_sizes_init();
 935 }
 936
 937 /*
 938  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
 939  * and also on some strange 486's. All 586+'s are OK. This used to involve
 940  * black magic jumps to work around some nasty CPU bugs, but fortunately the
 941  * switch to using exceptions got rid of all that.
 942  */
 943 static void __init test_wp_bit(void)
 944 {
 945         printk(KERN_INFO
 946   "Checking if this processor honours the WP bit even in supervisor mode...");
 947
 948         /* Any page-aligned address will do, the test is non-destructive */
 949         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
 950         boot_cpu_data.wp_works_ok = do_test_wp_bit();
 951         clear_fixmap(FIX_WP_TEST);
 952
 953         if (!boot_cpu_data.wp_works_ok) {
 954                 printk(KERN_CONT "No.\n");
 955 #ifdef CONFIG_X86_WP_WORKS_OK
 956                 panic(
 957   "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
 958 #endif
 959         } else {
 960                 printk(KERN_CONT "Ok.\n");
 961         }
 962 }
 963
 964 static struct kcore_list kcore_mem, kcore_vmalloc;
 965
 966 void __init mem_init(void)
 967 {
 968         int codesize, reservedpages, datasize, initsize;
 969         int tmp;
 970
 971         pci_iommu_alloc();
 972
 973 #ifdef CONFIG_FLATMEM
 974         BUG_ON(!mem_map);
 975 #endif
 976         /* this will put all low memory onto the freelists */
 977         totalram_pages += free_all_bootmem();
 978
 979         reservedpages = 0;
 980         for (tmp = 0; tmp < max_low_pfn; tmp++)
 981                 /*
 982                  * Only count reserved RAM pages:
 983                  */
 984                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
 985                         reservedpages++;
 986
 987         set_highmem_pages_init();
 988
 989         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 990         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 991         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 992
 993         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 994         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 995                    VMALLOC_END-VMALLOC_START);
 996
 997         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
 998                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
 999                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
1000                 num_physpages << (PAGE_SHIFT-10),
1001                 codesize >> 10,
1002                 reservedpages << (PAGE_SHIFT-10),
1003                 datasize >> 10,
1004                 initsize >> 10,
1005                 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
1006                );
1007
1008         printk(KERN_INFO "virtual kernel memory layout:\n"
1009                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1010 #ifdef CONFIG_HIGHMEM
1011                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1012 #endif
1013                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1014                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1015                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1016                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1017                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
1018                 FIXADDR_START, FIXADDR_TOP,
1019                 (FIXADDR_TOP - FIXADDR_START) >> 10,
1020
1021 #ifdef CONFIG_HIGHMEM
1022                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
1023                 (LAST_PKMAP*PAGE_SIZE) >> 10,
1024 #endif
1025
1026                 VMALLOC_START, VMALLOC_END,
1027                 (VMALLOC_END - VMALLOC_START) >> 20,
1028
1029                 (unsigned long)__va(0), (unsigned long)high_memory,
1030                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
1031
1032                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
1033                 ((unsigned long)&__init_end -
1034                  (unsigned long)&__init_begin) >> 10,
1035
1036                 (unsigned long)&_etext, (unsigned long)&_edata,
1037                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
1038
1039                 (unsigned long)&_text, (unsigned long)&_etext,
1040                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
1041
1042         /*
1043          * Check boundaries twice: Some fundamental inconsistencies can
1044          * be detected at build time already.
1045          */
1046 #define __FIXADDR_TOP (-PAGE_SIZE)
1047 #ifdef CONFIG_HIGHMEM
1048         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
1049         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
1050 #endif
1051 #define high_memory (-128UL << 20)
1052         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
1053 #undef high_memory
1054 #undef __FIXADDR_TOP
1055
1056 #ifdef CONFIG_HIGHMEM
1057         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
1058         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
1059 #endif
1060         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
1061         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
1062
1063         if (boot_cpu_data.wp_works_ok < 0)
1064                 test_wp_bit();
1065
1066         save_pg_dir();
1067         zap_low_mappings();
1068 }
1069
1070 #ifdef CONFIG_MEMORY_HOTPLUG
1071 int arch_add_memory(int nid, u64 start, u64 size)
1072 {
1073         struct pglist_data *pgdata = NODE_DATA(nid);
1074         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
1075         unsigned long start_pfn = start >> PAGE_SHIFT;
1076         unsigned long nr_pages = size >> PAGE_SHIFT;
1077
1078         return __add_pages(zone, start_pfn, nr_pages);
1079 }
1080 #endif
1081
1082 /*
1083  * This function cannot be __init, since exceptions don't work in that
1084  * section.  Put this after the callers, so that it cannot be inlined.
1085  */
1086 static noinline int do_test_wp_bit(void)
1087 {
1088         char tmp_reg;
1089         int flag;
1090
1091         __asm__ __volatile__(
1092                 "       movb %0, %1     \n"
1093                 "1:     movb %1, %0     \n"
1094                 "       xorl %2, %2     \n"
1095                 "2:                     \n"
1096                 _ASM_EXTABLE(1b,2b)
1097                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
1098                  "=q" (tmp_reg),
1099                  "=r" (flag)
1100                 :"2" (1)
1101                 :"memory");
1102
1103         return flag;
1104 }
1105
1106 #ifdef CONFIG_DEBUG_RODATA
1107 const int rodata_test_data = 0xC3;
1108 EXPORT_SYMBOL_GPL(rodata_test_data);
1109
1110 void mark_rodata_ro(void)
1111 {
1112         unsigned long start = PFN_ALIGN(_text);
1113         unsigned long size = PFN_ALIGN(_etext) - start;
1114
1115 #ifndef CONFIG_DYNAMIC_FTRACE
1116         /* Dynamic tracing modifies the kernel text section */
1117         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1118         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
1119                 size >> 10);
1120
1121 #ifdef CONFIG_CPA_DEBUG
1122         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
1123                 start, start+size);
1124         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
1125
1126         printk(KERN_INFO "Testing CPA: write protecting again\n");
1127         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
1128 #endif
1129 #endif /* CONFIG_DYNAMIC_FTRACE */
1130
1131         start += size;
1132         size = (unsigned long)__end_rodata - start;
1133         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1134         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1135                 size >> 10);
1136         rodata_test();
1137
1138 #ifdef CONFIG_CPA_DEBUG
1139         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
1140         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
1141
1142         printk(KERN_INFO "Testing CPA: write protecting again\n");
1143         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1144 #endif
1145 }
1146 #endif
1147
1148 void free_init_pages(char *what, unsigned long begin, unsigned long end)
1149 {
1150 #ifdef CONFIG_DEBUG_PAGEALLOC
1151         /*
1152          * If debugging page accesses then do not free this memory but
1153          * mark them not present - any buggy init-section access will
1154          * create a kernel page fault:
1155          */
1156         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
1157                 begin, PAGE_ALIGN(end));
1158         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
1159 #else
1160         unsigned long addr;
1161
1162         /*
1163          * We just marked the kernel text read only above, now that
1164          * we are going to free part of that, we need to make that
1165          * writeable first.
1166          */
1167         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
1168
1169         for (addr = begin; addr < end; addr += PAGE_SIZE) {
1170                 ClearPageReserved(virt_to_page(addr));
1171                 init_page_count(virt_to_page(addr));
1172                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
1173                 free_page(addr);
1174                 totalram_pages++;
1175         }
1176         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
1177 #endif
1178 }
1179
1180 void free_initmem(void)
1181 {
1182         free_init_pages("unused kernel memory",
1183                         (unsigned long)(&__init_begin),
1184                         (unsigned long)(&__init_end));
1185 }
1186
1187 #ifdef CONFIG_BLK_DEV_INITRD
1188 void free_initrd_mem(unsigned long start, unsigned long end)
1189 {
1190         free_init_pages("initrd memory", start, end);
1191 }
1192 #endif
1193
1194 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1195                                    int flags)
1196 {
1197         return reserve_bootmem(phys, len, flags);
1198 }