arch/x86/mm/init_64.c

   1 /*
   2  *  linux/arch/x86_64/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
   6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
   7  */
   8
   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/swap.h>
  19 #include <linux/smp.h>
  20 #include <linux/init.h>
  21 #include <linux/initrd.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/proc_fs.h>
  25 #include <linux/pci.h>
  26 #include <linux/pfn.h>
  27 #include <linux/poison.h>
  28 #include <linux/dma-mapping.h>
  29 #include <linux/module.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/nmi.h>
  32
  33 #include <asm/processor.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36 #include <asm/pgtable.h>
  37 #include <asm/pgalloc.h>
  38 #include <asm/dma.h>
  39 #include <asm/fixmap.h>
  40 #include <asm/e820.h>
  41 #include <asm/apic.h>
  42 #include <asm/tlb.h>
  43 #include <asm/mmu_context.h>
  44 #include <asm/proto.h>
  45 #include <asm/smp.h>
  46 #include <asm/sections.h>
  47 #include <asm/kdebug.h>
  48 #include <asm/numa.h>
  49 #include <asm/cacheflush.h>
  50
  51 /*
  52  * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
  53  * The direct mapping extends to max_pfn_mapped, so that we can directly access
  54  * apertures, ACPI and other tables without having to play with fixmaps.
  55  */
  56 unsigned long max_low_pfn_mapped;
  57 unsigned long max_pfn_mapped;
  58
  59 static unsigned long dma_reserve __initdata;
  60
  61 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  62
  63 int direct_gbpages
  64 #ifdef CONFIG_DIRECT_GBPAGES
  65                                 = 1
  66 #endif
  67 ;
  68
  69 static int __init parse_direct_gbpages_off(char *arg)
  70 {
  71         direct_gbpages = 0;
  72         return 0;
  73 }
  74 early_param("nogbpages", parse_direct_gbpages_off);
  75
  76 static int __init parse_direct_gbpages_on(char *arg)
  77 {
  78         direct_gbpages = 1;
  79         return 0;
  80 }
  81 early_param("gbpages", parse_direct_gbpages_on);
  82
  83 /*
  84  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  85  * physical space so we can cache the place of the first one and move
  86  * around without checking the pgd every time.
  87  */
  88
  89 int after_bootmem;
  90
  91 unsigned long __supported_pte_mask __read_mostly = ~0UL;
  92 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  93
  94 static int do_not_nx __cpuinitdata;
  95
  96 /*
  97  * noexec=on|off
  98  * Control non-executable mappings for 64-bit processes.
  99  *
 100  * on   Enable (default)
 101  * off  Disable
 102  */
 103 static int __init nonx_setup(char *str)
 104 {
 105         if (!str)
 106                 return -EINVAL;
 107         if (!strncmp(str, "on", 2)) {
 108                 __supported_pte_mask |= _PAGE_NX;
 109                 do_not_nx = 0;
 110         } else if (!strncmp(str, "off", 3)) {
 111                 do_not_nx = 1;
 112                 __supported_pte_mask &= ~_PAGE_NX;
 113         }
 114         return 0;
 115 }
 116 early_param("noexec", nonx_setup);
 117
 118 void __cpuinit check_efer(void)
 119 {
 120         unsigned long efer;
 121
 122         rdmsrl(MSR_EFER, efer);
 123         if (!(efer & EFER_NX) || do_not_nx)
 124                 __supported_pte_mask &= ~_PAGE_NX;
 125 }
 126
 127 int force_personality32;
 128
 129 /*
 130  * noexec32=on|off
 131  * Control non executable heap for 32bit processes.
 132  * To control the stack too use noexec=off
 133  *
 134  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 135  * off  PROT_READ implies PROT_EXEC
 136  */
 137 static int __init nonx32_setup(char *str)
 138 {
 139         if (!strcmp(str, "on"))
 140                 force_personality32 &= ~READ_IMPLIES_EXEC;
 141         else if (!strcmp(str, "off"))
 142                 force_personality32 |= READ_IMPLIES_EXEC;
 143         return 1;
 144 }
 145 __setup("noexec32=", nonx32_setup);
 146
 147 /*
 148  * NOTE: This function is marked __ref because it calls __init function
 149  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 150  */
 151 static __ref void *spp_getpage(void)
 152 {
 153         void *ptr;
 154
 155         if (after_bootmem)
 156                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 157         else
 158                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 159
 160         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 161                 panic("set_pte_phys: cannot allocate page data %s\n",
 162                         after_bootmem ? "after bootmem" : "");
 163         }
 164
 165         pr_debug("spp_getpage %p\n", ptr);
 166
 167         return ptr;
 168 }
 169
 170 void
 171 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 172 {
 173         pud_t *pud;
 174         pmd_t *pmd;
 175         pte_t *pte;
 176
 177         pud = pud_page + pud_index(vaddr);
 178         if (pud_none(*pud)) {
 179                 pmd = (pmd_t *) spp_getpage();
 180                 pud_populate(&init_mm, pud, pmd);
 181                 if (pmd != pmd_offset(pud, 0)) {
 182                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 183                                 pmd, pmd_offset(pud, 0));
 184                         return;
 185                 }
 186         }
 187         pmd = pmd_offset(pud, vaddr);
 188         if (pmd_none(*pmd)) {
 189                 pte = (pte_t *) spp_getpage();
 190                 pmd_populate_kernel(&init_mm, pmd, pte);
 191                 if (pte != pte_offset_kernel(pmd, 0)) {
 192                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 193                         return;
 194                 }
 195         }
 196
 197         pte = pte_offset_kernel(pmd, vaddr);
 198         if (!pte_none(*pte) && pte_val(new_pte) &&
 199             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 200                 pte_ERROR(*pte);
 201         set_pte(pte, new_pte);
 202
 203         /*
 204          * It's enough to flush this one mapping.
 205          * (PGE mappings get flushed as well)
 206          */
 207         __flush_tlb_one(vaddr);
 208 }
 209
 210 void
 211 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 212 {
 213         pgd_t *pgd;
 214         pud_t *pud_page;
 215
 216         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 217
 218         pgd = pgd_offset_k(vaddr);
 219         if (pgd_none(*pgd)) {
 220                 printk(KERN_ERR
 221                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 222                 return;
 223         }
 224         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 225         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 226 }
 227
 228 /*
 229  * Create large page table mappings for a range of physical addresses.
 230  */
 231 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 232                                                 pgprot_t prot)
 233 {
 234         pgd_t *pgd;
 235         pud_t *pud;
 236         pmd_t *pmd;
 237
 238         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 239         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 240                 pgd = pgd_offset_k((unsigned long)__va(phys));
 241                 if (pgd_none(*pgd)) {
 242                         pud = (pud_t *) spp_getpage();
 243                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 244                                                 _PAGE_USER));
 245                 }
 246                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 247                 if (pud_none(*pud)) {
 248                         pmd = (pmd_t *) spp_getpage();
 249                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 250                                                 _PAGE_USER));
 251                 }
 252                 pmd = pmd_offset(pud, phys);
 253                 BUG_ON(!pmd_none(*pmd));
 254                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 255         }
 256 }
 257
 258 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 259 {
 260         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 261 }
 262
 263 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 264 {
 265         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 266 }
 267
 268 /*
 269  * The head.S code sets up the kernel high mapping:
 270  *
 271  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 272  *
 273  * phys_addr holds the negative offset to the kernel, which is added
 274  * to the compile time generated pmds. This results in invalid pmds up
 275  * to the point where we hit the physaddr 0 mapping.
 276  *
 277  * We limit the mappings to the region from _text to _end.  _end is
 278  * rounded up to the 2MB boundary. This catches the invalid pmds as
 279  * well, as they are located before _text:
 280  */
 281 void __init cleanup_highmap(void)
 282 {
 283         unsigned long vaddr = __START_KERNEL_map;
 284         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 285         pmd_t *pmd = level2_kernel_pgt;
 286         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 287
 288         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 289                 if (pmd_none(*pmd))
 290                         continue;
 291                 if (vaddr < (unsigned long) _text || vaddr > end)
 292                         set_pmd(pmd, __pmd(0));
 293         }
 294 }
 295
 296 static unsigned long __initdata table_start;
 297 static unsigned long __meminitdata table_end;
 298 static unsigned long __meminitdata table_top;
 299
 300 static __ref void *alloc_low_page(unsigned long *phys)
 301 {
 302         unsigned long pfn = table_end++;
 303         void *adr;
 304
 305         if (after_bootmem) {
 306                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 307                 *phys = __pa(adr);
 308
 309                 return adr;
 310         }
 311
 312         if (pfn >= table_top)
 313                 panic("alloc_low_page: ran out of memory");
 314
 315         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 316         memset(adr, 0, PAGE_SIZE);
 317         *phys  = pfn * PAGE_SIZE;
 318         return adr;
 319 }
 320
 321 static __ref void unmap_low_page(void *adr)
 322 {
 323         if (after_bootmem)
 324                 return;
 325
 326         early_iounmap(adr, PAGE_SIZE);
 327 }
 328
 329 static unsigned long __meminit
 330 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end)
 331 {
 332         unsigned pages = 0;
 333         unsigned long last_map_addr = end;
 334         int i;
 335
 336         pte_t *pte = pte_page + pte_index(addr);
 337
 338         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 339
 340                 if (addr >= end) {
 341                         if (!after_bootmem) {
 342                                 for(; i < PTRS_PER_PTE; i++, pte++)
 343                                         set_pte(pte, __pte(0));
 344                         }
 345                         break;
 346                 }
 347
 348                 if (pte_val(*pte))
 349                         continue;
 350
 351                 if (0)
 352                         printk("   pte=%p addr=%lx pte=%016lx\n",
 353                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 354                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL));
 355                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 356                 pages++;
 357         }
 358         update_page_count(PG_LEVEL_4K, pages);
 359
 360         return last_map_addr;
 361 }
 362
 363 static unsigned long __meminit
 364 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end)
 365 {
 366         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 367
 368         return phys_pte_init(pte, address, end);
 369 }
 370
 371 static unsigned long __meminit
 372 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 373                          unsigned long page_size_mask)
 374 {
 375         unsigned long pages = 0;
 376         unsigned long last_map_addr = end;
 377         unsigned long start = address;
 378
 379         int i = pmd_index(address);
 380
 381         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 382                 unsigned long pte_phys;
 383                 pmd_t *pmd = pmd_page + pmd_index(address);
 384                 pte_t *pte;
 385
 386                 if (address >= end) {
 387                         if (!after_bootmem) {
 388                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 389                                         set_pmd(pmd, __pmd(0));
 390                         }
 391                         break;
 392                 }
 393
 394                 if (pmd_val(*pmd)) {
 395                         if (!pmd_large(*pmd)) {
 396                                 spin_lock(&init_mm.page_table_lock);
 397                                 last_map_addr = phys_pte_update(pmd, address,
 398                                                                 end);
 399                                 spin_unlock(&init_mm.page_table_lock);
 400                         }
 401                         /* Count entries we're using from level2_ident_pgt */
 402                         if (start == 0)
 403                                 pages++;
 404                         continue;
 405                 }
 406
 407                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 408                         pages++;
 409                         spin_lock(&init_mm.page_table_lock);
 410                         set_pte((pte_t *)pmd,
 411                                 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 412                         spin_unlock(&init_mm.page_table_lock);
 413                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 414                         continue;
 415                 }
 416
 417                 pte = alloc_low_page(&pte_phys);
 418                 last_map_addr = phys_pte_init(pte, address, end);
 419                 unmap_low_page(pte);
 420
 421                 spin_lock(&init_mm.page_table_lock);
 422                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 423                 spin_unlock(&init_mm.page_table_lock);
 424         }
 425         update_page_count(PG_LEVEL_2M, pages);
 426         return last_map_addr;
 427 }
 428
 429 static unsigned long __meminit
 430 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 431                          unsigned long page_size_mask)
 432 {
 433         pmd_t *pmd = pmd_offset(pud, 0);
 434         unsigned long last_map_addr;
 435
 436         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask);
 437         __flush_tlb_all();
 438         return last_map_addr;
 439 }
 440
 441 static unsigned long __meminit
 442 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 443                          unsigned long page_size_mask)
 444 {
 445         unsigned long pages = 0;
 446         unsigned long last_map_addr = end;
 447         int i = pud_index(addr);
 448
 449         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 450                 unsigned long pmd_phys;
 451                 pud_t *pud = pud_page + pud_index(addr);
 452                 pmd_t *pmd;
 453
 454                 if (addr >= end)
 455                         break;
 456
 457                 if (!after_bootmem &&
 458                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 459                         set_pud(pud, __pud(0));
 460                         continue;
 461                 }
 462
 463                 if (pud_val(*pud)) {
 464                         if (!pud_large(*pud))
 465                                 last_map_addr = phys_pmd_update(pud, addr, end,
 466                                                          page_size_mask);
 467                         continue;
 468                 }
 469
 470                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 471                         pages++;
 472                         spin_lock(&init_mm.page_table_lock);
 473                         set_pte((pte_t *)pud,
 474                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 475                         spin_unlock(&init_mm.page_table_lock);
 476                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 477                         continue;
 478                 }
 479
 480                 pmd = alloc_low_page(&pmd_phys);
 481                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask);
 482                 unmap_low_page(pmd);
 483
 484                 spin_lock(&init_mm.page_table_lock);
 485                 pud_populate(&init_mm, pud, __va(pmd_phys));
 486                 spin_unlock(&init_mm.page_table_lock);
 487         }
 488         __flush_tlb_all();
 489         update_page_count(PG_LEVEL_1G, pages);
 490
 491         return last_map_addr;
 492 }
 493
 494 static unsigned long __meminit
 495 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 496                  unsigned long page_size_mask)
 497 {
 498         pud_t *pud;
 499
 500         pud = (pud_t *)pgd_page_vaddr(*pgd);
 501
 502         return phys_pud_init(pud, addr, end, page_size_mask);
 503 }
 504
 505 static void __init find_early_table_space(unsigned long end)
 506 {
 507         unsigned long puds, pmds, ptes, tables, start;
 508
 509         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 510         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 511         if (direct_gbpages) {
 512                 unsigned long extra;
 513                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 514                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 515         } else
 516                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 517         tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 518
 519         if (cpu_has_pse) {
 520                 unsigned long extra;
 521                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 522                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 523         } else
 524                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 525         tables += round_up(ptes * sizeof(pte_t), PAGE_SIZE);
 526
 527         /*
 528          * RED-PEN putting page tables only on node 0 could
 529          * cause a hotspot and fill up ZONE_DMA. The page tables
 530          * need roughly 0.5KB per GB.
 531          */
 532         start = 0x8000;
 533         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 534         if (table_start == -1UL)
 535                 panic("Cannot find space for the kernel page tables");
 536
 537         table_start >>= PAGE_SHIFT;
 538         table_end = table_start;
 539         table_top = table_start + (tables >> PAGE_SHIFT);
 540
 541         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 542                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 543 }
 544
 545 static void __init init_gbpages(void)
 546 {
 547         if (direct_gbpages && cpu_has_gbpages)
 548                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 549         else
 550                 direct_gbpages = 0;
 551 }
 552
 553 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 554                                                 unsigned long end,
 555                                                 unsigned long page_size_mask)
 556 {
 557
 558         unsigned long next, last_map_addr = end;
 559
 560         start = (unsigned long)__va(start);
 561         end = (unsigned long)__va(end);
 562
 563         for (; start < end; start = next) {
 564                 pgd_t *pgd = pgd_offset_k(start);
 565                 unsigned long pud_phys;
 566                 pud_t *pud;
 567
 568                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 569                 if (next > end)
 570                         next = end;
 571
 572                 if (pgd_val(*pgd)) {
 573                         last_map_addr = phys_pud_update(pgd, __pa(start),
 574                                                  __pa(end), page_size_mask);
 575                         continue;
 576                 }
 577
 578                 pud = alloc_low_page(&pud_phys);
 579                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 580                                                  page_size_mask);
 581                 unmap_low_page(pud);
 582
 583                 spin_lock(&init_mm.page_table_lock);
 584                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 585                 spin_unlock(&init_mm.page_table_lock);
 586         }
 587
 588         return last_map_addr;
 589 }
 590
 591 struct map_range {
 592         unsigned long start;
 593         unsigned long end;
 594         unsigned page_size_mask;
 595 };
 596
 597 #define NR_RANGE_MR 5
 598
 599 static int save_mr(struct map_range *mr, int nr_range,
 600                    unsigned long start_pfn, unsigned long end_pfn,
 601                    unsigned long page_size_mask)
 602 {
 603
 604         if (start_pfn < end_pfn) {
 605                 if (nr_range >= NR_RANGE_MR)
 606                         panic("run out of range for init_memory_mapping\n");
 607                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 608                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 609                 mr[nr_range].page_size_mask = page_size_mask;
 610                 nr_range++;
 611         }
 612
 613         return nr_range;
 614 }
 615
 616 /*
 617  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 618  * This runs before bootmem is initialized and gets pages directly from
 619  * the physical memory. To access them they are temporarily mapped.
 620  */
 621 unsigned long __init_refok init_memory_mapping(unsigned long start,
 622                                                unsigned long end)
 623 {
 624         unsigned long last_map_addr = 0;
 625         unsigned long page_size_mask = 0;
 626         unsigned long start_pfn, end_pfn;
 627
 628         struct map_range mr[NR_RANGE_MR];
 629         int nr_range, i;
 630
 631         printk(KERN_INFO "init_memory_mapping\n");
 632
 633         /*
 634          * Find space for the kernel direct mapping tables.
 635          *
 636          * Later we should allocate these tables in the local node of the
 637          * memory mapped. Unfortunately this is done currently before the
 638          * nodes are discovered.
 639          */
 640         if (!after_bootmem)
 641                 init_gbpages();
 642
 643         if (direct_gbpages)
 644                 page_size_mask |= 1 << PG_LEVEL_1G;
 645         if (cpu_has_pse)
 646                 page_size_mask |= 1 << PG_LEVEL_2M;
 647
 648         memset(mr, 0, sizeof(mr));
 649         nr_range = 0;
 650
 651         /* head if not big page alignment ?*/
 652         start_pfn = start >> PAGE_SHIFT;
 653         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 654                         << (PMD_SHIFT - PAGE_SHIFT);
 655         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 656
 657         /* big page (2M) range*/
 658         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 659                          << (PMD_SHIFT - PAGE_SHIFT);
 660         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 661                          << (PUD_SHIFT - PAGE_SHIFT);
 662         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 663                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 664         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 665                         page_size_mask & (1<<PG_LEVEL_2M));
 666
 667         /* big page (1G) range */
 668         start_pfn = end_pfn;
 669         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 670         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 671                                 page_size_mask &
 672                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 673
 674         /* tail is not big page (1G) alignment */
 675         start_pfn = end_pfn;
 676         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 677         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 678                         page_size_mask & (1<<PG_LEVEL_2M));
 679
 680         /* tail is not big page (2M) alignment */
 681         start_pfn = end_pfn;
 682         end_pfn = end>>PAGE_SHIFT;
 683         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 684
 685         /* try to merge same page size and continuous */
 686         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 687                 unsigned long old_start;
 688                 if (mr[i].end != mr[i+1].start ||
 689                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 690                         continue;
 691                 /* move it */
 692                 old_start = mr[i].start;
 693                 memmove(&mr[i], &mr[i+1],
 694                          (nr_range - 1 - i) * sizeof (struct map_range));
 695                 mr[i].start = old_start;
 696                 nr_range--;
 697         }
 698
 699         for (i = 0; i < nr_range; i++)
 700                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 701                                 mr[i].start, mr[i].end,
 702                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 703                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 704
 705         if (!after_bootmem)
 706                 find_early_table_space(end);
 707
 708         for (i = 0; i < nr_range; i++)
 709                 last_map_addr = kernel_physical_mapping_init(
 710                                         mr[i].start, mr[i].end,
 711                                         mr[i].page_size_mask);
 712
 713         if (!after_bootmem)
 714                 mmu_cr4_features = read_cr4();
 715         __flush_tlb_all();
 716
 717         if (!after_bootmem && table_end > table_start)
 718                 reserve_early(table_start << PAGE_SHIFT,
 719                                  table_end << PAGE_SHIFT, "PGTABLE");
 720
 721         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 722                          last_map_addr, end);
 723
 724         if (!after_bootmem)
 725                 early_memtest(start, end);
 726
 727         return last_map_addr >> PAGE_SHIFT;
 728 }
 729
 730 #ifndef CONFIG_NUMA
 731 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 732 {
 733         unsigned long bootmap_size, bootmap;
 734
 735         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 736         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 737                                  PAGE_SIZE);
 738         if (bootmap == -1L)
 739                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 740         /* don't touch min_low_pfn */
 741         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 742                                          0, end_pfn);
 743         e820_register_active_regions(0, start_pfn, end_pfn);
 744         free_bootmem_with_active_regions(0, end_pfn);
 745         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 746         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 747 }
 748
 749 void __init paging_init(void)
 750 {
 751         unsigned long max_zone_pfns[MAX_NR_ZONES];
 752
 753         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 754         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 755         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 756         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 757
 758         memory_present(0, 0, max_pfn);
 759         sparse_init();
 760         free_area_init_nodes(max_zone_pfns);
 761 }
 762 #endif
 763
 764 /*
 765  * Memory hotplug specific functions
 766  */
 767 #ifdef CONFIG_MEMORY_HOTPLUG
 768 /*
 769  * Memory is added always to NORMAL zone. This means you will never get
 770  * additional DMA/DMA32 memory.
 771  */
 772 int arch_add_memory(int nid, u64 start, u64 size)
 773 {
 774         struct pglist_data *pgdat = NODE_DATA(nid);
 775         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 776         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 777         unsigned long nr_pages = size >> PAGE_SHIFT;
 778         int ret;
 779
 780         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 781         if (last_mapped_pfn > max_pfn_mapped)
 782                 max_pfn_mapped = last_mapped_pfn;
 783
 784         ret = __add_pages(zone, start_pfn, nr_pages);
 785         WARN_ON(1);
 786
 787         return ret;
 788 }
 789 EXPORT_SYMBOL_GPL(arch_add_memory);
 790
 791 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 792 int memory_add_physaddr_to_nid(u64 start)
 793 {
 794         return 0;
 795 }
 796 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 797 #endif
 798
 799 #endif /* CONFIG_MEMORY_HOTPLUG */
 800
 801 /*
 802  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 803  * is valid. The argument is a physical page number.
 804  *
 805  *
 806  * On x86, access has to be given to the first megabyte of ram because that area
 807  * contains bios code and data regions used by X and dosemu and similar apps.
 808  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 809  * mmio resources as well as potential bios/acpi data regions.
 810  */
 811 int devmem_is_allowed(unsigned long pagenr)
 812 {
 813         if (pagenr <= 256)
 814                 return 1;
 815         if (!page_is_ram(pagenr))
 816                 return 1;
 817         return 0;
 818 }
 819
 820
 821 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 822                          kcore_modules, kcore_vsyscall;
 823
 824 void __init mem_init(void)
 825 {
 826         long codesize, reservedpages, datasize, initsize;
 827
 828         pci_iommu_alloc();
 829
 830         /* clear_bss() already clear the empty_zero_page */
 831
 832         reservedpages = 0;
 833
 834         /* this will put all low memory onto the freelists */
 835 #ifdef CONFIG_NUMA
 836         totalram_pages = numa_free_all_bootmem();
 837 #else
 838         totalram_pages = free_all_bootmem();
 839 #endif
 840         reservedpages = max_pfn - totalram_pages -
 841                                         absent_pages_in_range(0, max_pfn);
 842         after_bootmem = 1;
 843
 844         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 845         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 846         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 847
 848         /* Register memory areas for /proc/kcore */
 849         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 850         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 851                    VMALLOC_END-VMALLOC_START);
 852         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 853         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 854         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 855                                  VSYSCALL_END - VSYSCALL_START);
 856
 857         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 858                                 "%ldk reserved, %ldk data, %ldk init)\n",
 859                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 860                 max_pfn << (PAGE_SHIFT-10),
 861                 codesize >> 10,
 862                 reservedpages << (PAGE_SHIFT-10),
 863                 datasize >> 10,
 864                 initsize >> 10);
 865
 866         cpa_init();
 867 }
 868
 869 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 870 {
 871         unsigned long addr = begin;
 872
 873         if (addr >= end)
 874                 return;
 875
 876         /*
 877          * If debugging page accesses then do not free this memory but
 878          * mark them not present - any buggy init-section access will
 879          * create a kernel page fault:
 880          */
 881 #ifdef CONFIG_DEBUG_PAGEALLOC
 882         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 883                 begin, PAGE_ALIGN(end));
 884         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 885 #else
 886         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 887
 888         for (; addr < end; addr += PAGE_SIZE) {
 889                 ClearPageReserved(virt_to_page(addr));
 890                 init_page_count(virt_to_page(addr));
 891                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 892                         POISON_FREE_INITMEM, PAGE_SIZE);
 893                 free_page(addr);
 894                 totalram_pages++;
 895         }
 896 #endif
 897 }
 898
 899 void free_initmem(void)
 900 {
 901         free_init_pages("unused kernel memory",
 902                         (unsigned long)(&__init_begin),
 903                         (unsigned long)(&__init_end));
 904 }
 905
 906 #ifdef CONFIG_DEBUG_RODATA
 907 const int rodata_test_data = 0xC3;
 908 EXPORT_SYMBOL_GPL(rodata_test_data);
 909
 910 void mark_rodata_ro(void)
 911 {
 912         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 913         unsigned long rodata_start =
 914                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 915
 916 #ifdef CONFIG_DYNAMIC_FTRACE
 917         /* Dynamic tracing modifies the kernel text section */
 918         start = rodata_start;
 919 #endif
 920
 921         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 922                (end - start) >> 10);
 923         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 924
 925         /*
 926          * The rodata section (but not the kernel text!) should also be
 927          * not-executable.
 928          */
 929         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 930
 931         rodata_test();
 932
 933 #ifdef CONFIG_CPA_DEBUG
 934         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 935         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 936
 937         printk(KERN_INFO "Testing CPA: again\n");
 938         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 939 #endif
 940 }
 941
 942 #endif
 943
 944 #ifdef CONFIG_BLK_DEV_INITRD
 945 void free_initrd_mem(unsigned long start, unsigned long end)
 946 {
 947         free_init_pages("initrd memory", start, end);
 948 }
 949 #endif
 950
 951 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 952                                    int flags)
 953 {
 954 #ifdef CONFIG_NUMA
 955         int nid, next_nid;
 956         int ret;
 957 #endif
 958         unsigned long pfn = phys >> PAGE_SHIFT;
 959
 960         if (pfn >= max_pfn) {
 961                 /*
 962                  * This can happen with kdump kernels when accessing
 963                  * firmware tables:
 964                  */
 965                 if (pfn < max_pfn_mapped)
 966                         return -EFAULT;
 967
 968                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 969                                 phys, len);
 970                 return -EFAULT;
 971         }
 972
 973         /* Should check here against the e820 map to avoid double free */
 974 #ifdef CONFIG_NUMA
 975         nid = phys_to_nid(phys);
 976         next_nid = phys_to_nid(phys + len - 1);
 977         if (nid == next_nid)
 978                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 979         else
 980                 ret = reserve_bootmem(phys, len, flags);
 981
 982         if (ret != 0)
 983                 return ret;
 984
 985 #else
 986         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 987 #endif
 988
 989         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 990                 dma_reserve += len / PAGE_SIZE;
 991                 set_dma_reserve(dma_reserve);
 992         }
 993
 994         return 0;
 995 }
 996
 997 int kern_addr_valid(unsigned long addr)
 998 {
 999         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1000         pgd_t *pgd;
1001         pud_t *pud;
1002         pmd_t *pmd;
1003         pte_t *pte;
1004
1005         if (above != 0 && above != -1UL)
1006                 return 0;
1007
1008         pgd = pgd_offset_k(addr);
1009         if (pgd_none(*pgd))
1010                 return 0;
1011
1012         pud = pud_offset(pgd, addr);
1013         if (pud_none(*pud))
1014                 return 0;
1015
1016         pmd = pmd_offset(pud, addr);
1017         if (pmd_none(*pmd))
1018                 return 0;
1019
1020         if (pmd_large(*pmd))
1021                 return pfn_valid(pmd_pfn(*pmd));
1022
1023         pte = pte_offset_kernel(pmd, addr);
1024         if (pte_none(*pte))
1025                 return 0;
1026
1027         return pfn_valid(pte_pfn(*pte));
1028 }
1029
1030 /*
1031  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1032  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1033  * not need special handling anymore:
1034  */
1035 static struct vm_area_struct gate_vma = {
1036         .vm_start       = VSYSCALL_START,
1037         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1038         .vm_page_prot   = PAGE_READONLY_EXEC,
1039         .vm_flags       = VM_READ | VM_EXEC
1040 };
1041
1042 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1043 {
1044 #ifdef CONFIG_IA32_EMULATION
1045         if (test_tsk_thread_flag(tsk, TIF_IA32))
1046                 return NULL;
1047 #endif
1048         return &gate_vma;
1049 }
1050
1051 int in_gate_area(struct task_struct *task, unsigned long addr)
1052 {
1053         struct vm_area_struct *vma = get_gate_vma(task);
1054
1055         if (!vma)
1056                 return 0;
1057
1058         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1059 }
1060
1061 /*
1062  * Use this when you have no reliable task/vma, typically from interrupt
1063  * context. It is less reliable than using the task's vma and may give
1064  * false positives:
1065  */
1066 int in_gate_area_no_task(unsigned long addr)
1067 {
1068         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1069 }
1070
1071 const char *arch_vma_name(struct vm_area_struct *vma)
1072 {
1073         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1074                 return "[vdso]";
1075         if (vma == &gate_vma)
1076                 return "[vsyscall]";
1077         return NULL;
1078 }
1079
1080 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1081 /*
1082  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1083  */
1084 static long __meminitdata addr_start, addr_end;
1085 static void __meminitdata *p_start, *p_end;
1086 static int __meminitdata node_start;
1087
1088 int __meminit
1089 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1090 {
1091         unsigned long addr = (unsigned long)start_page;
1092         unsigned long end = (unsigned long)(start_page + size);
1093         unsigned long next;
1094         pgd_t *pgd;
1095         pud_t *pud;
1096         pmd_t *pmd;
1097
1098         for (; addr < end; addr = next) {
1099                 void *p = NULL;
1100
1101                 pgd = vmemmap_pgd_populate(addr, node);
1102                 if (!pgd)
1103                         return -ENOMEM;
1104
1105                 pud = vmemmap_pud_populate(pgd, addr, node);
1106                 if (!pud)
1107                         return -ENOMEM;
1108
1109                 if (!cpu_has_pse) {
1110                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1111                         pmd = vmemmap_pmd_populate(pud, addr, node);
1112
1113                         if (!pmd)
1114                                 return -ENOMEM;
1115
1116                         p = vmemmap_pte_populate(pmd, addr, node);
1117
1118                         if (!p)
1119                                 return -ENOMEM;
1120
1121                         addr_end = addr + PAGE_SIZE;
1122                         p_end = p + PAGE_SIZE;
1123                 } else {
1124                         next = pmd_addr_end(addr, end);
1125
1126                         pmd = pmd_offset(pud, addr);
1127                         if (pmd_none(*pmd)) {
1128                                 pte_t entry;
1129
1130                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1131                                 if (!p)
1132                                         return -ENOMEM;
1133
1134                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1135                                                 PAGE_KERNEL_LARGE);
1136                                 set_pmd(pmd, __pmd(pte_val(entry)));
1137
1138                                 /* check to see if we have contiguous blocks */
1139                                 if (p_end != p || node_start != node) {
1140                                         if (p_start)
1141                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1142                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1143                                         addr_start = addr;
1144                                         node_start = node;
1145                                         p_start = p;
1146                                 }
1147
1148                                 addr_end = addr + PMD_SIZE;
1149                                 p_end = p + PMD_SIZE;
1150                         } else
1151                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1152                 }
1153
1154         }
1155         return 0;
1156 }
1157
1158 void __meminit vmemmap_populate_print_last(void)
1159 {
1160         if (p_start) {
1161                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1162                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1163                 p_start = NULL;
1164                 p_end = NULL;
1165                 node_start = 0;
1166         }
1167 }
1168 #endif