arch/x86/mm/init_32.c

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