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