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 (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 332                 return 0;
 333         if (!page_is_ram(pagenr))
 334                 return 1;
 335         return 0;
 336 }
 337
 338 pte_t *kmap_pte;
 339 pgprot_t kmap_prot;
 340
 341 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
 342 {
 343         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
 344                         vaddr), vaddr), vaddr);
 345 }
 346
 347 static void __init kmap_init(void)
 348 {
 349         unsigned long kmap_vstart;
 350
 351         /*
 352          * Cache the first kmap pte:
 353          */
 354         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
 355         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
 356
 357         kmap_prot = PAGE_KERNEL;
 358 }
 359
 360 #ifdef CONFIG_HIGHMEM
 361 static void __init permanent_kmaps_init(pgd_t *pgd_base)
 362 {
 363         unsigned long vaddr;
 364         pgd_t *pgd;
 365         pud_t *pud;
 366         pmd_t *pmd;
 367         pte_t *pte;
 368
 369         vaddr = PKMAP_BASE;
 370         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
 371
 372         pgd = swapper_pg_dir + pgd_index(vaddr);
 373         pud = pud_offset(pgd, vaddr);
 374         pmd = pmd_offset(pud, vaddr);
 375         pte = pte_offset_kernel(pmd, vaddr);
 376         pkmap_page_table = pte;
 377 }
 378
 379 static void __init add_one_highpage_init(struct page *page, int pfn)
 380 {
 381         ClearPageReserved(page);
 382         init_page_count(page);
 383         __free_page(page);
 384         totalhigh_pages++;
 385 }
 386
 387 struct add_highpages_data {
 388         unsigned long start_pfn;
 389         unsigned long end_pfn;
 390 };
 391
 392 static int __init add_highpages_work_fn(unsigned long start_pfn,
 393                                          unsigned long end_pfn, void *datax)
 394 {
 395         int node_pfn;
 396         struct page *page;
 397         unsigned long final_start_pfn, final_end_pfn;
 398         struct add_highpages_data *data;
 399
 400         data = (struct add_highpages_data *)datax;
 401
 402         final_start_pfn = max(start_pfn, data->start_pfn);
 403         final_end_pfn = min(end_pfn, data->end_pfn);
 404         if (final_start_pfn >= final_end_pfn)
 405                 return 0;
 406
 407         for (node_pfn = final_start_pfn; node_pfn < final_end_pfn;
 408              node_pfn++) {
 409                 if (!pfn_valid(node_pfn))
 410                         continue;
 411                 page = pfn_to_page(node_pfn);
 412                 add_one_highpage_init(page, node_pfn);
 413         }
 414
 415         return 0;
 416
 417 }
 418
 419 void __init add_highpages_with_active_regions(int nid, unsigned long start_pfn,
 420                                               unsigned long end_pfn)
 421 {
 422         struct add_highpages_data data;
 423
 424         data.start_pfn = start_pfn;
 425         data.end_pfn = end_pfn;
 426
 427         work_with_active_regions(nid, add_highpages_work_fn, &data);
 428 }
 429
 430 #ifndef CONFIG_NUMA
 431 static void __init set_highmem_pages_init(void)
 432 {
 433         add_highpages_with_active_regions(0, highstart_pfn, highend_pfn);
 434
 435         totalram_pages += totalhigh_pages;
 436 }
 437 #endif /* !CONFIG_NUMA */
 438
 439 #else
 440 static inline void permanent_kmaps_init(pgd_t *pgd_base)
 441 {
 442 }
 443 static inline void set_highmem_pages_init(void)
 444 {
 445 }
 446 #endif /* CONFIG_HIGHMEM */
 447
 448 void __init native_pagetable_setup_start(pgd_t *base)
 449 {
 450         unsigned long pfn, va;
 451         pgd_t *pgd;
 452         pud_t *pud;
 453         pmd_t *pmd;
 454         pte_t *pte;
 455
 456         /*
 457          * Remove any mappings which extend past the end of physical
 458          * memory from the boot time page table:
 459          */
 460         for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
 461                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
 462                 pgd = base + pgd_index(va);
 463                 if (!pgd_present(*pgd))
 464                         break;
 465
 466                 pud = pud_offset(pgd, va);
 467                 pmd = pmd_offset(pud, va);
 468                 if (!pmd_present(*pmd))
 469                         break;
 470
 471                 pte = pte_offset_kernel(pmd, va);
 472                 if (!pte_present(*pte))
 473                         break;
 474
 475                 pte_clear(NULL, va, pte);
 476         }
 477         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
 478 }
 479
 480 void __init native_pagetable_setup_done(pgd_t *base)
 481 {
 482 }
 483
 484 /*
 485  * Build a proper pagetable for the kernel mappings.  Up until this
 486  * point, we've been running on some set of pagetables constructed by
 487  * the boot process.
 488  *
 489  * If we're booting on native hardware, this will be a pagetable
 490  * constructed in arch/x86/kernel/head_32.S.  The root of the
 491  * pagetable will be swapper_pg_dir.
 492  *
 493  * If we're booting paravirtualized under a hypervisor, then there are
 494  * more options: we may already be running PAE, and the pagetable may
 495  * or may not be based in swapper_pg_dir.  In any case,
 496  * paravirt_pagetable_setup_start() will set up swapper_pg_dir
 497  * appropriately for the rest of the initialization to work.
 498  *
 499  * In general, pagetable_init() assumes that the pagetable may already
 500  * be partially populated, and so it avoids stomping on any existing
 501  * mappings.
 502  */
 503 static void __init early_ioremap_page_table_range_init(pgd_t *pgd_base)
 504 {
 505         unsigned long vaddr, end;
 506
 507         /*
 508          * Fixed mappings, only the page table structure has to be
 509          * created - mappings will be set by set_fixmap():
 510          */
 511         early_ioremap_clear();
 512         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 513         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 514         page_table_range_init(vaddr, end, pgd_base);
 515         early_ioremap_reset();
 516 }
 517
 518 static void __init pagetable_init(void)
 519 {
 520         pgd_t *pgd_base = swapper_pg_dir;
 521
 522         permanent_kmaps_init(pgd_base);
 523 }
 524
 525 #ifdef CONFIG_ACPI_SLEEP
 526 /*
 527  * ACPI suspend needs this for resume, because things like the intel-agp
 528  * driver might have split up a kernel 4MB mapping.
 529  */
 530 char swsusp_pg_dir[PAGE_SIZE]
 531         __attribute__ ((aligned(PAGE_SIZE)));
 532
 533 static inline void save_pg_dir(void)
 534 {
 535         memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
 536 }
 537 #else /* !CONFIG_ACPI_SLEEP */
 538 static inline void save_pg_dir(void)
 539 {
 540 }
 541 #endif /* !CONFIG_ACPI_SLEEP */
 542
 543 void zap_low_mappings(void)
 544 {
 545         int i;
 546
 547         /*
 548          * Zap initial low-memory mappings.
 549          *
 550          * Note that "pgd_clear()" doesn't do it for
 551          * us, because pgd_clear() is a no-op on i386.
 552          */
 553         for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
 554 #ifdef CONFIG_X86_PAE
 555                 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
 556 #else
 557                 set_pgd(swapper_pg_dir+i, __pgd(0));
 558 #endif
 559         }
 560         flush_tlb_all();
 561 }
 562
 563 int nx_enabled;
 564
 565 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
 566 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 567
 568 #ifdef CONFIG_X86_PAE
 569
 570 static int disable_nx __initdata;
 571
 572 /*
 573  * noexec = on|off
 574  *
 575  * Control non executable mappings.
 576  *
 577  * on      Enable
 578  * off     Disable
 579  */
 580 static int __init noexec_setup(char *str)
 581 {
 582         if (!str || !strcmp(str, "on")) {
 583                 if (cpu_has_nx) {
 584                         __supported_pte_mask |= _PAGE_NX;
 585                         disable_nx = 0;
 586                 }
 587         } else {
 588                 if (!strcmp(str, "off")) {
 589                         disable_nx = 1;
 590                         __supported_pte_mask &= ~_PAGE_NX;
 591                 } else {
 592                         return -EINVAL;
 593                 }
 594         }
 595
 596         return 0;
 597 }
 598 early_param("noexec", noexec_setup);
 599
 600 static void __init set_nx(void)
 601 {
 602         unsigned int v[4], l, h;
 603
 604         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
 605                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
 606
 607                 if ((v[3] & (1 << 20)) && !disable_nx) {
 608                         rdmsr(MSR_EFER, l, h);
 609                         l |= EFER_NX;
 610                         wrmsr(MSR_EFER, l, h);
 611                         nx_enabled = 1;
 612                         __supported_pte_mask |= _PAGE_NX;
 613                 }
 614         }
 615 }
 616 #endif
 617
 618 /* user-defined highmem size */
 619 static unsigned int highmem_pages = -1;
 620
 621 /*
 622  * highmem=size forces highmem to be exactly 'size' bytes.
 623  * This works even on boxes that have no highmem otherwise.
 624  * This also works to reduce highmem size on bigger boxes.
 625  */
 626 static int __init parse_highmem(char *arg)
 627 {
 628         if (!arg)
 629                 return -EINVAL;
 630
 631         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
 632         return 0;
 633 }
 634 early_param("highmem", parse_highmem);
 635
 636 /*
 637  * Determine low and high memory ranges:
 638  */
 639 void __init find_low_pfn_range(void)
 640 {
 641         /* it could update max_pfn */
 642
 643         /* max_low_pfn is 0, we already have early_res support */
 644
 645         max_low_pfn = max_pfn;
 646         if (max_low_pfn > MAXMEM_PFN) {
 647                 if (highmem_pages == -1)
 648                         highmem_pages = max_pfn - MAXMEM_PFN;
 649                 if (highmem_pages + MAXMEM_PFN < max_pfn)
 650                         max_pfn = MAXMEM_PFN + highmem_pages;
 651                 if (highmem_pages + MAXMEM_PFN > max_pfn) {
 652                         printk(KERN_WARNING "only %luMB highmem pages "
 653                                 "available, ignoring highmem size of %uMB.\n",
 654                                 pages_to_mb(max_pfn - MAXMEM_PFN),
 655                                 pages_to_mb(highmem_pages));
 656                         highmem_pages = 0;
 657                 }
 658                 max_low_pfn = MAXMEM_PFN;
 659 #ifndef CONFIG_HIGHMEM
 660                 /* Maximum memory usable is what is directly addressable */
 661                 printk(KERN_WARNING "Warning only %ldMB will be used.\n",
 662                                         MAXMEM>>20);
 663                 if (max_pfn > MAX_NONPAE_PFN)
 664                         printk(KERN_WARNING
 665                                  "Use a HIGHMEM64G enabled kernel.\n");
 666                 else
 667                         printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
 668                 max_pfn = MAXMEM_PFN;
 669 #else /* !CONFIG_HIGHMEM */
 670 #ifndef CONFIG_HIGHMEM64G
 671                 if (max_pfn > MAX_NONPAE_PFN) {
 672                         max_pfn = MAX_NONPAE_PFN;
 673                         printk(KERN_WARNING "Warning only 4GB will be used."
 674                                 "Use a HIGHMEM64G enabled kernel.\n");
 675                 }
 676 #endif /* !CONFIG_HIGHMEM64G */
 677 #endif /* !CONFIG_HIGHMEM */
 678         } else {
 679                 if (highmem_pages == -1)
 680                         highmem_pages = 0;
 681 #ifdef CONFIG_HIGHMEM
 682                 if (highmem_pages >= max_pfn) {
 683                         printk(KERN_ERR "highmem size specified (%uMB) is "
 684                                 "bigger than pages available (%luMB)!.\n",
 685                                 pages_to_mb(highmem_pages),
 686                                 pages_to_mb(max_pfn));
 687                         highmem_pages = 0;
 688                 }
 689                 if (highmem_pages) {
 690                         if (max_low_pfn - highmem_pages <
 691                             64*1024*1024/PAGE_SIZE){
 692                                 printk(KERN_ERR "highmem size %uMB results in "
 693                                 "smaller than 64MB lowmem, ignoring it.\n"
 694                                         , pages_to_mb(highmem_pages));
 695                                 highmem_pages = 0;
 696                         }
 697                         max_low_pfn -= highmem_pages;
 698                 }
 699 #else
 700                 if (highmem_pages)
 701                         printk(KERN_ERR "ignoring highmem size on non-highmem"
 702                                         " kernel!\n");
 703 #endif
 704         }
 705 }
 706
 707 #ifndef CONFIG_NEED_MULTIPLE_NODES
 708 void __init initmem_init(unsigned long start_pfn,
 709                                   unsigned long end_pfn)
 710 {
 711 #ifdef CONFIG_HIGHMEM
 712         highstart_pfn = highend_pfn = max_pfn;
 713         if (max_pfn > max_low_pfn)
 714                 highstart_pfn = max_low_pfn;
 715         memory_present(0, 0, highend_pfn);
 716         e820_register_active_regions(0, 0, highend_pfn);
 717         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 718                 pages_to_mb(highend_pfn - highstart_pfn));
 719         num_physpages = highend_pfn;
 720         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 721 #else
 722         memory_present(0, 0, max_low_pfn);
 723         e820_register_active_regions(0, 0, max_low_pfn);
 724         num_physpages = max_low_pfn;
 725         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 726 #endif
 727 #ifdef CONFIG_FLATMEM
 728         max_mapnr = num_physpages;
 729 #endif
 730         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 731                         pages_to_mb(max_low_pfn));
 732
 733         setup_bootmem_allocator();
 734 }
 735 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
 736
 737 static void __init zone_sizes_init(void)
 738 {
 739         unsigned long max_zone_pfns[MAX_NR_ZONES];
 740         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 741         max_zone_pfns[ZONE_DMA] =
 742                 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
 743         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
 744 #ifdef CONFIG_HIGHMEM
 745         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
 746 #endif
 747
 748         free_area_init_nodes(max_zone_pfns);
 749 }
 750
 751 void __init setup_bootmem_allocator(void)
 752 {
 753         int i;
 754         unsigned long bootmap_size, bootmap;
 755         /*
 756          * Initialize the boot-time allocator (with low memory only):
 757          */
 758         bootmap_size = bootmem_bootmap_pages(max_low_pfn)<<PAGE_SHIFT;
 759         bootmap = find_e820_area(min_low_pfn<<PAGE_SHIFT,
 760                                  max_pfn_mapped<<PAGE_SHIFT, bootmap_size,
 761                                  PAGE_SIZE);
 762         if (bootmap == -1L)
 763                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 764         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
 765
 766         /* don't touch min_low_pfn */
 767         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 768                                          min_low_pfn, max_low_pfn);
 769         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
 770                  max_pfn_mapped<<PAGE_SHIFT);
 771         printk(KERN_INFO "  low ram: %08lx - %08lx\n",
 772                  min_low_pfn<<PAGE_SHIFT, max_low_pfn<<PAGE_SHIFT);
 773         printk(KERN_INFO "  bootmap %08lx - %08lx\n",
 774                  bootmap, bootmap + bootmap_size);
 775         for_each_online_node(i)
 776                 free_bootmem_with_active_regions(i, max_low_pfn);
 777         early_res_to_bootmem(0, max_low_pfn<<PAGE_SHIFT);
 778
 779         after_init_bootmem = 1;
 780 }
 781
 782 static void __init find_early_table_space(unsigned long end, int use_pse)
 783 {
 784         unsigned long puds, pmds, ptes, tables, start;
 785
 786         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 787         tables = PAGE_ALIGN(puds * sizeof(pud_t));
 788
 789         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 790         tables += PAGE_ALIGN(pmds * sizeof(pmd_t));
 791
 792         if (use_pse) {
 793                 unsigned long extra;
 794
 795                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 796                 extra += PMD_SIZE;
 797                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 798         } else
 799                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 800
 801         tables += PAGE_ALIGN(ptes * sizeof(pte_t));
 802
 803         /* for fixmap */
 804         tables += PAGE_SIZE * 2;
 805
 806         /*
 807          * RED-PEN putting page tables only on node 0 could
 808          * cause a hotspot and fill up ZONE_DMA. The page tables
 809          * need roughly 0.5KB per GB.
 810          */
 811         start = 0x7000;
 812         table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
 813                                         tables, PAGE_SIZE);
 814         if (table_start == -1UL)
 815                 panic("Cannot find space for the kernel page tables");
 816
 817         table_start >>= PAGE_SHIFT;
 818         table_end = table_start;
 819         table_top = table_start + (tables>>PAGE_SHIFT);
 820
 821         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 822                 end, table_start << PAGE_SHIFT,
 823                 (table_start << PAGE_SHIFT) + tables);
 824 }
 825
 826 unsigned long __init_refok init_memory_mapping(unsigned long start,
 827                                                 unsigned long end)
 828 {
 829         pgd_t *pgd_base = swapper_pg_dir;
 830         unsigned long start_pfn, end_pfn;
 831         unsigned long big_page_start;
 832 #ifdef CONFIG_DEBUG_PAGEALLOC
 833         /*
 834          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 835          * This will simplify cpa(), which otherwise needs to support splitting
 836          * large pages into small in interrupt context, etc.
 837          */
 838         int use_pse = 0;
 839 #else
 840         int use_pse = cpu_has_pse;
 841 #endif
 842
 843         /*
 844          * Find space for the kernel direct mapping tables.
 845          */
 846         if (!after_init_bootmem)
 847                 find_early_table_space(end, use_pse);
 848
 849 #ifdef CONFIG_X86_PAE
 850         set_nx();
 851         if (nx_enabled)
 852                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
 853 #endif
 854
 855         /* Enable PSE if available */
 856         if (cpu_has_pse)
 857                 set_in_cr4(X86_CR4_PSE);
 858
 859         /* Enable PGE if available */
 860         if (cpu_has_pge) {
 861                 set_in_cr4(X86_CR4_PGE);
 862                 __supported_pte_mask |= _PAGE_GLOBAL;
 863         }
 864
 865         /*
 866          * Don't use a large page for the first 2/4MB of memory
 867          * because there are often fixed size MTRRs in there
 868          * and overlapping MTRRs into large pages can cause
 869          * slowdowns.
 870          */
 871         big_page_start = PMD_SIZE;
 872
 873         if (start < big_page_start) {
 874                 start_pfn = start >> PAGE_SHIFT;
 875                 end_pfn = min(big_page_start>>PAGE_SHIFT, end>>PAGE_SHIFT);
 876         } else {
 877                 /* head is not big page alignment ? */
 878                 start_pfn = start >> PAGE_SHIFT;
 879                 end_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 880                                  << (PMD_SHIFT - PAGE_SHIFT);
 881         }
 882         if (start_pfn < end_pfn)
 883                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn, 0);
 884
 885         /* big page range */
 886         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 887                          << (PMD_SHIFT - PAGE_SHIFT);
 888         if (start_pfn < (big_page_start >> PAGE_SHIFT))
 889                 start_pfn =  big_page_start >> PAGE_SHIFT;
 890         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 891         if (start_pfn < end_pfn)
 892                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn,
 893                                              use_pse);
 894
 895         /* tail is not big page alignment ? */
 896         start_pfn = end_pfn;
 897         if (start_pfn > (big_page_start>>PAGE_SHIFT)) {
 898                 end_pfn = end >> PAGE_SHIFT;
 899                 if (start_pfn < end_pfn)
 900                         kernel_physical_mapping_init(pgd_base, start_pfn,
 901                                                          end_pfn, 0);
 902         }
 903
 904         early_ioremap_page_table_range_init(pgd_base);
 905
 906         load_cr3(swapper_pg_dir);
 907
 908         __flush_tlb_all();
 909
 910         if (!after_init_bootmem)
 911                 reserve_early(table_start << PAGE_SHIFT,
 912                                  table_end << PAGE_SHIFT, "PGTABLE");
 913
 914         if (!after_init_bootmem)
 915                 early_memtest(start, end);
 916
 917         return end >> PAGE_SHIFT;
 918 }
 919
 920
 921 /*
 922  * paging_init() sets up the page tables - note that the first 8MB are
 923  * already mapped by head.S.
 924  *
 925  * This routines also unmaps the page at virtual kernel address 0, so
 926  * that we can trap those pesky NULL-reference errors in the kernel.
 927  */
 928 void __init paging_init(void)
 929 {
 930         pagetable_init();
 931
 932         __flush_tlb_all();
 933
 934         kmap_init();
 935
 936         /*
 937          * NOTE: at this point the bootmem allocator is fully available.
 938          */
 939         sparse_init();
 940         zone_sizes_init();
 941 }
 942
 943 /*
 944  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
 945  * and also on some strange 486's. All 586+'s are OK. This used to involve
 946  * black magic jumps to work around some nasty CPU bugs, but fortunately the
 947  * switch to using exceptions got rid of all that.
 948  */
 949 static void __init test_wp_bit(void)
 950 {
 951         printk(KERN_INFO
 952   "Checking if this processor honours the WP bit even in supervisor mode...");
 953
 954         /* Any page-aligned address will do, the test is non-destructive */
 955         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
 956         boot_cpu_data.wp_works_ok = do_test_wp_bit();
 957         clear_fixmap(FIX_WP_TEST);
 958
 959         if (!boot_cpu_data.wp_works_ok) {
 960                 printk(KERN_CONT "No.\n");
 961 #ifdef CONFIG_X86_WP_WORKS_OK
 962                 panic(
 963   "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
 964 #endif
 965         } else {
 966                 printk(KERN_CONT "Ok.\n");
 967         }
 968 }
 969
 970 static struct kcore_list kcore_mem, kcore_vmalloc;
 971
 972 void __init mem_init(void)
 973 {
 974         int codesize, reservedpages, datasize, initsize;
 975         int tmp;
 976
 977         pci_iommu_alloc();
 978
 979 #ifdef CONFIG_FLATMEM
 980         BUG_ON(!mem_map);
 981 #endif
 982         /* this will put all low memory onto the freelists */
 983         totalram_pages += free_all_bootmem();
 984
 985         reservedpages = 0;
 986         for (tmp = 0; tmp < max_low_pfn; tmp++)
 987                 /*
 988                  * Only count reserved RAM pages:
 989                  */
 990                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
 991                         reservedpages++;
 992
 993         set_highmem_pages_init();
 994
 995         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 996         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 997         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 998
 999         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
1000         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
1001                    VMALLOC_END-VMALLOC_START);
1002
1003         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
1004                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
1005                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
1006                 num_physpages << (PAGE_SHIFT-10),
1007                 codesize >> 10,
1008                 reservedpages << (PAGE_SHIFT-10),
1009                 datasize >> 10,
1010                 initsize >> 10,
1011                 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
1012                );
1013
1014         printk(KERN_INFO "virtual kernel memory layout:\n"
1015                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1016 #ifdef CONFIG_HIGHMEM
1017                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1018 #endif
1019                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1020                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1021                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1022                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1023                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
1024                 FIXADDR_START, FIXADDR_TOP,
1025                 (FIXADDR_TOP - FIXADDR_START) >> 10,
1026
1027 #ifdef CONFIG_HIGHMEM
1028                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
1029                 (LAST_PKMAP*PAGE_SIZE) >> 10,
1030 #endif
1031
1032                 VMALLOC_START, VMALLOC_END,
1033                 (VMALLOC_END - VMALLOC_START) >> 20,
1034
1035                 (unsigned long)__va(0), (unsigned long)high_memory,
1036                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
1037
1038                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
1039                 ((unsigned long)&__init_end -
1040                  (unsigned long)&__init_begin) >> 10,
1041
1042                 (unsigned long)&_etext, (unsigned long)&_edata,
1043                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
1044
1045                 (unsigned long)&_text, (unsigned long)&_etext,
1046                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
1047
1048         /*
1049          * Check boundaries twice: Some fundamental inconsistencies can
1050          * be detected at build time already.
1051          */
1052 #define __FIXADDR_TOP (-PAGE_SIZE)
1053 #ifdef CONFIG_HIGHMEM
1054         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
1055         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
1056 #endif
1057 #define high_memory (-128UL << 20)
1058         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
1059 #undef high_memory
1060 #undef __FIXADDR_TOP
1061
1062 #ifdef CONFIG_HIGHMEM
1063         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
1064         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
1065 #endif
1066         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
1067         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
1068
1069         if (boot_cpu_data.wp_works_ok < 0)
1070                 test_wp_bit();
1071
1072         save_pg_dir();
1073         zap_low_mappings();
1074 }
1075
1076 #ifdef CONFIG_MEMORY_HOTPLUG
1077 int arch_add_memory(int nid, u64 start, u64 size)
1078 {
1079         struct pglist_data *pgdata = NODE_DATA(nid);
1080         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
1081         unsigned long start_pfn = start >> PAGE_SHIFT;
1082         unsigned long nr_pages = size >> PAGE_SHIFT;
1083
1084         return __add_pages(nid, zone, start_pfn, nr_pages);
1085 }
1086 #endif
1087
1088 /*
1089  * This function cannot be __init, since exceptions don't work in that
1090  * section.  Put this after the callers, so that it cannot be inlined.
1091  */
1092 static noinline int do_test_wp_bit(void)
1093 {
1094         char tmp_reg;
1095         int flag;
1096
1097         __asm__ __volatile__(
1098                 "       movb %0, %1     \n"
1099                 "1:     movb %1, %0     \n"
1100                 "       xorl %2, %2     \n"
1101                 "2:                     \n"
1102                 _ASM_EXTABLE(1b,2b)
1103                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
1104                  "=q" (tmp_reg),
1105                  "=r" (flag)
1106                 :"2" (1)
1107                 :"memory");
1108
1109         return flag;
1110 }
1111
1112 #ifdef CONFIG_DEBUG_RODATA
1113 const int rodata_test_data = 0xC3;
1114 EXPORT_SYMBOL_GPL(rodata_test_data);
1115
1116 void mark_rodata_ro(void)
1117 {
1118         unsigned long start = PFN_ALIGN(_text);
1119         unsigned long size = PFN_ALIGN(_etext) - start;
1120
1121 #ifndef CONFIG_DYNAMIC_FTRACE
1122         /* Dynamic tracing modifies the kernel text section */
1123         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1124         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
1125                 size >> 10);
1126
1127 #ifdef CONFIG_CPA_DEBUG
1128         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
1129                 start, start+size);
1130         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
1131
1132         printk(KERN_INFO "Testing CPA: write protecting again\n");
1133         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
1134 #endif
1135 #endif /* CONFIG_DYNAMIC_FTRACE */
1136
1137         start += size;
1138         size = (unsigned long)__end_rodata - start;
1139         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1140         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1141                 size >> 10);
1142         rodata_test();
1143
1144 #ifdef CONFIG_CPA_DEBUG
1145         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
1146         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
1147
1148         printk(KERN_INFO "Testing CPA: write protecting again\n");
1149         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1150 #endif
1151 }
1152 #endif
1153
1154 void free_init_pages(char *what, unsigned long begin, unsigned long end)
1155 {
1156 #ifdef CONFIG_DEBUG_PAGEALLOC
1157         /*
1158          * If debugging page accesses then do not free this memory but
1159          * mark them not present - any buggy init-section access will
1160          * create a kernel page fault:
1161          */
1162         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
1163                 begin, PAGE_ALIGN(end));
1164         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
1165 #else
1166         unsigned long addr;
1167
1168         /*
1169          * We just marked the kernel text read only above, now that
1170          * we are going to free part of that, we need to make that
1171          * writeable first.
1172          */
1173         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
1174
1175         for (addr = begin; addr < end; addr += PAGE_SIZE) {
1176                 ClearPageReserved(virt_to_page(addr));
1177                 init_page_count(virt_to_page(addr));
1178                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
1179                 free_page(addr);
1180                 totalram_pages++;
1181         }
1182         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
1183 #endif
1184 }
1185
1186 void free_initmem(void)
1187 {
1188         free_init_pages("unused kernel memory",
1189                         (unsigned long)(&__init_begin),
1190                         (unsigned long)(&__init_end));
1191 }
1192
1193 #ifdef CONFIG_BLK_DEV_INITRD
1194 void free_initrd_mem(unsigned long start, unsigned long end)
1195 {
1196         free_init_pages("initrd memory", start, end);
1197 }
1198 #endif
1199
1200 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1201                                    int flags)
1202 {
1203         return reserve_bootmem(phys, len, flags);
1204 }