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