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