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 __meminitdata
  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 void show_mem(void)
  90 {
  91         long i, total = 0, reserved = 0;
  92         long shared = 0, cached = 0;
  93         struct page *page;
  94         pg_data_t *pgdat;
  95
  96         printk(KERN_INFO "Mem-info:\n");
  97         show_free_areas();
  98         for_each_online_pgdat(pgdat) {
  99                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
 100                         /*
 101                          * This loop can take a while with 256 GB and
 102                          * 4k pages so defer the NMI watchdog:
 103                          */
 104                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
 105                                 touch_nmi_watchdog();
 106
 107                         if (!pfn_valid(pgdat->node_start_pfn + i))
 108                                 continue;
 109
 110                         page = pfn_to_page(pgdat->node_start_pfn + i);
 111                         total++;
 112                         if (PageReserved(page))
 113                                 reserved++;
 114                         else if (PageSwapCache(page))
 115                                 cached++;
 116                         else if (page_count(page))
 117                                 shared += page_count(page) - 1;
 118                 }
 119         }
 120         printk(KERN_INFO "%lu pages of RAM\n",          total);
 121         printk(KERN_INFO "%lu reserved pages\n",        reserved);
 122         printk(KERN_INFO "%lu pages shared\n",          shared);
 123         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 124 }
 125
 126 int after_bootmem;
 127
 128 static __init void *spp_getpage(void)
 129 {
 130         void *ptr;
 131
 132         if (after_bootmem)
 133                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 134         else
 135                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 136
 137         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 138                 panic("set_pte_phys: cannot allocate page data %s\n",
 139                         after_bootmem ? "after bootmem" : "");
 140         }
 141
 142         pr_debug("spp_getpage %p\n", ptr);
 143
 144         return ptr;
 145 }
 146
 147 void
 148 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 149 {
 150         pud_t *pud;
 151         pmd_t *pmd;
 152         pte_t *pte;
 153
 154         pud = pud_page + pud_index(vaddr);
 155         if (pud_none(*pud)) {
 156                 pmd = (pmd_t *) spp_getpage();
 157                 pud_populate(&init_mm, pud, pmd);
 158                 if (pmd != pmd_offset(pud, 0)) {
 159                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 160                                 pmd, pmd_offset(pud, 0));
 161                         return;
 162                 }
 163         }
 164         pmd = pmd_offset(pud, vaddr);
 165         if (pmd_none(*pmd)) {
 166                 pte = (pte_t *) spp_getpage();
 167                 pmd_populate_kernel(&init_mm, pmd, pte);
 168                 if (pte != pte_offset_kernel(pmd, 0)) {
 169                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 170                         return;
 171                 }
 172         }
 173
 174         pte = pte_offset_kernel(pmd, vaddr);
 175         if (!pte_none(*pte) && pte_val(new_pte) &&
 176             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 177                 pte_ERROR(*pte);
 178         set_pte(pte, new_pte);
 179
 180         /*
 181          * It's enough to flush this one mapping.
 182          * (PGE mappings get flushed as well)
 183          */
 184         __flush_tlb_one(vaddr);
 185 }
 186
 187 void
 188 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 189 {
 190         pgd_t *pgd;
 191         pud_t *pud_page;
 192
 193         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 194
 195         pgd = pgd_offset_k(vaddr);
 196         if (pgd_none(*pgd)) {
 197                 printk(KERN_ERR
 198                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 199                 return;
 200         }
 201         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 202         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 203 }
 204
 205 /*
 206  * Create large page table mappings for a range of physical addresses.
 207  */
 208 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 209                                                 pgprot_t prot)
 210 {
 211         pgd_t *pgd;
 212         pud_t *pud;
 213         pmd_t *pmd;
 214
 215         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 216         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 217                 pgd = pgd_offset_k((unsigned long)__va(phys));
 218                 if (pgd_none(*pgd)) {
 219                         pud = (pud_t *) spp_getpage();
 220                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 221                                                 _PAGE_USER));
 222                 }
 223                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 224                 if (pud_none(*pud)) {
 225                         pmd = (pmd_t *) spp_getpage();
 226                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 227                                                 _PAGE_USER));
 228                 }
 229                 pmd = pmd_offset(pud, phys);
 230                 BUG_ON(!pmd_none(*pmd));
 231                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 232         }
 233 }
 234
 235 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 236 {
 237         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 238 }
 239
 240 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 241 {
 242         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 243 }
 244
 245 /*
 246  * The head.S code sets up the kernel high mapping:
 247  *
 248  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 249  *
 250  * phys_addr holds the negative offset to the kernel, which is added
 251  * to the compile time generated pmds. This results in invalid pmds up
 252  * to the point where we hit the physaddr 0 mapping.
 253  *
 254  * We limit the mappings to the region from _text to _end.  _end is
 255  * rounded up to the 2MB boundary. This catches the invalid pmds as
 256  * well, as they are located before _text:
 257  */
 258 void __init cleanup_highmap(void)
 259 {
 260         unsigned long vaddr = __START_KERNEL_map;
 261         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 262         pmd_t *pmd = level2_kernel_pgt;
 263         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 264
 265         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 266                 if (pmd_none(*pmd))
 267                         continue;
 268                 if (vaddr < (unsigned long) _text || vaddr > end)
 269                         set_pmd(pmd, __pmd(0));
 270         }
 271 }
 272
 273 static unsigned long __initdata table_start;
 274 static unsigned long __meminitdata table_end;
 275 static unsigned long __meminitdata table_top;
 276
 277 static __meminit void *alloc_low_page(unsigned long *phys)
 278 {
 279         unsigned long pfn = table_end++;
 280         void *adr;
 281
 282         if (after_bootmem) {
 283                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 284                 *phys = __pa(adr);
 285
 286                 return adr;
 287         }
 288
 289         if (pfn >= table_top)
 290                 panic("alloc_low_page: ran out of memory");
 291
 292         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 293         memset(adr, 0, PAGE_SIZE);
 294         *phys  = pfn * PAGE_SIZE;
 295         return adr;
 296 }
 297
 298 static __meminit void unmap_low_page(void *adr)
 299 {
 300         if (after_bootmem)
 301                 return;
 302
 303         early_iounmap(adr, PAGE_SIZE);
 304 }
 305
 306 static unsigned long __meminit
 307 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end)
 308 {
 309         unsigned pages = 0;
 310         unsigned long last_map_addr = end;
 311         int i;
 312
 313         pte_t *pte = pte_page + pte_index(addr);
 314
 315         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 316
 317                 if (addr >= end) {
 318                         if (!after_bootmem) {
 319                                 for(; i < PTRS_PER_PTE; i++, pte++)
 320                                         set_pte(pte, __pte(0));
 321                         }
 322                         break;
 323                 }
 324
 325                 if (pte_val(*pte))
 326                         continue;
 327
 328                 if (0)
 329                         printk("   pte=%p addr=%lx pte=%016lx\n",
 330                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 331                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL));
 332                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 333                 pages++;
 334         }
 335         update_page_count(PG_LEVEL_4K, pages);
 336
 337         return last_map_addr;
 338 }
 339
 340 static unsigned long __meminit
 341 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end)
 342 {
 343         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 344
 345         return phys_pte_init(pte, address, end);
 346 }
 347
 348 static unsigned long __meminit
 349 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 350                          unsigned long page_size_mask)
 351 {
 352         unsigned long pages = 0;
 353         unsigned long last_map_addr = end;
 354
 355         int i = pmd_index(address);
 356
 357         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 358                 unsigned long pte_phys;
 359                 pmd_t *pmd = pmd_page + pmd_index(address);
 360                 pte_t *pte;
 361
 362                 if (address >= end) {
 363                         if (!after_bootmem) {
 364                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 365                                         set_pmd(pmd, __pmd(0));
 366                         }
 367                         break;
 368                 }
 369
 370                 if (pmd_val(*pmd)) {
 371                         if (!pmd_large(*pmd))
 372                                 last_map_addr = phys_pte_update(pmd, address,
 373                                                                  end);
 374                         continue;
 375                 }
 376
 377                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 378                         pages++;
 379                         set_pte((pte_t *)pmd,
 380                                 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 381                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 382                         continue;
 383                 }
 384
 385                 pte = alloc_low_page(&pte_phys);
 386                 last_map_addr = phys_pte_init(pte, address, end);
 387                 unmap_low_page(pte);
 388
 389                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 390         }
 391         update_page_count(PG_LEVEL_2M, pages);
 392         return last_map_addr;
 393 }
 394
 395 static unsigned long __meminit
 396 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 397                          unsigned long page_size_mask)
 398 {
 399         pmd_t *pmd = pmd_offset(pud, 0);
 400         unsigned long last_map_addr;
 401
 402         spin_lock(&init_mm.page_table_lock);
 403         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask);
 404         spin_unlock(&init_mm.page_table_lock);
 405         __flush_tlb_all();
 406         return last_map_addr;
 407 }
 408
 409 static unsigned long __meminit
 410 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 411                          unsigned long page_size_mask)
 412 {
 413         unsigned long pages = 0;
 414         unsigned long last_map_addr = end;
 415         int i = pud_index(addr);
 416
 417         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 418                 unsigned long pmd_phys;
 419                 pud_t *pud = pud_page + pud_index(addr);
 420                 pmd_t *pmd;
 421
 422                 if (addr >= end)
 423                         break;
 424
 425                 if (!after_bootmem &&
 426                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 427                         set_pud(pud, __pud(0));
 428                         continue;
 429                 }
 430
 431                 if (pud_val(*pud)) {
 432                         if (!pud_large(*pud))
 433                                 last_map_addr = phys_pmd_update(pud, addr, end,
 434                                                          page_size_mask);
 435                         continue;
 436                 }
 437
 438                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 439                         pages++;
 440                         set_pte((pte_t *)pud,
 441                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 442                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 443                         continue;
 444                 }
 445
 446                 pmd = alloc_low_page(&pmd_phys);
 447
 448                 spin_lock(&init_mm.page_table_lock);
 449                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask);
 450                 unmap_low_page(pmd);
 451                 pud_populate(&init_mm, pud, __va(pmd_phys));
 452                 spin_unlock(&init_mm.page_table_lock);
 453
 454         }
 455         __flush_tlb_all();
 456         update_page_count(PG_LEVEL_1G, pages);
 457
 458         return last_map_addr;
 459 }
 460
 461 static unsigned long __meminit
 462 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 463                  unsigned long page_size_mask)
 464 {
 465         pud_t *pud;
 466
 467         pud = (pud_t *)pgd_page_vaddr(*pgd);
 468
 469         return phys_pud_init(pud, addr, end, page_size_mask);
 470 }
 471
 472 static void __init find_early_table_space(unsigned long end)
 473 {
 474         unsigned long puds, pmds, ptes, tables, start;
 475
 476         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 477         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 478         if (direct_gbpages) {
 479                 unsigned long extra;
 480                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 481                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 482         } else
 483                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 484         tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 485
 486         if (cpu_has_pse) {
 487                 unsigned long extra;
 488                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 489                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 490         } else
 491                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 492         tables += round_up(ptes * sizeof(pte_t), PAGE_SIZE);
 493
 494         /*
 495          * RED-PEN putting page tables only on node 0 could
 496          * cause a hotspot and fill up ZONE_DMA. The page tables
 497          * need roughly 0.5KB per GB.
 498          */
 499         start = 0x8000;
 500         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 501         if (table_start == -1UL)
 502                 panic("Cannot find space for the kernel page tables");
 503
 504         table_start >>= PAGE_SHIFT;
 505         table_end = table_start;
 506         table_top = table_start + (tables >> PAGE_SHIFT);
 507
 508         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 509                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 510 }
 511
 512 static void __init init_gbpages(void)
 513 {
 514         if (direct_gbpages && cpu_has_gbpages)
 515                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 516         else
 517                 direct_gbpages = 0;
 518 }
 519
 520 #ifdef CONFIG_MEMTEST
 521
 522 static void __init memtest(unsigned long start_phys, unsigned long size,
 523                                  unsigned pattern)
 524 {
 525         unsigned long i;
 526         unsigned long *start;
 527         unsigned long start_bad;
 528         unsigned long last_bad;
 529         unsigned long val;
 530         unsigned long start_phys_aligned;
 531         unsigned long count;
 532         unsigned long incr;
 533
 534         switch (pattern) {
 535         case 0:
 536                 val = 0UL;
 537                 break;
 538         case 1:
 539                 val = -1UL;
 540                 break;
 541         case 2:
 542                 val = 0x5555555555555555UL;
 543                 break;
 544         case 3:
 545                 val = 0xaaaaaaaaaaaaaaaaUL;
 546                 break;
 547         default:
 548                 return;
 549         }
 550
 551         incr = sizeof(unsigned long);
 552         start_phys_aligned = ALIGN(start_phys, incr);
 553         count = (size - (start_phys_aligned - start_phys))/incr;
 554         start = __va(start_phys_aligned);
 555         start_bad = 0;
 556         last_bad = 0;
 557
 558         for (i = 0; i < count; i++)
 559                 start[i] = val;
 560         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 561                 if (*start != val) {
 562                         if (start_phys_aligned == last_bad + incr) {
 563                                 last_bad += incr;
 564                         } else {
 565                                 if (start_bad) {
 566                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 567                                                 val, start_bad, last_bad + incr);
 568                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 569                                 }
 570                                 start_bad = last_bad = start_phys_aligned;
 571                         }
 572                 }
 573         }
 574         if (start_bad) {
 575                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 576                         val, start_bad, last_bad + incr);
 577                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 578         }
 579
 580 }
 581
 582 /* default is disabled */
 583 static int memtest_pattern __initdata;
 584
 585 static int __init parse_memtest(char *arg)
 586 {
 587         if (arg)
 588                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 589         return 0;
 590 }
 591
 592 early_param("memtest", parse_memtest);
 593
 594 static void __init early_memtest(unsigned long start, unsigned long end)
 595 {
 596         u64 t_start, t_size;
 597         unsigned pattern;
 598
 599         if (!memtest_pattern)
 600                 return;
 601
 602         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 603         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 604                 t_start = start;
 605                 t_size = 0;
 606                 while (t_start < end) {
 607                         t_start = find_e820_area_size(t_start, &t_size, 1);
 608
 609                         /* done ? */
 610                         if (t_start >= end)
 611                                 break;
 612                         if (t_start + t_size > end)
 613                                 t_size = end - t_start;
 614
 615                         printk(KERN_CONT "\n  %016llx - %016llx pattern %d",
 616                                 (unsigned long long)t_start,
 617                                 (unsigned long long)t_start + t_size, pattern);
 618
 619                         memtest(t_start, t_size, pattern);
 620
 621                         t_start += t_size;
 622                 }
 623         }
 624         printk(KERN_CONT "\n");
 625 }
 626 #else
 627 static void __init early_memtest(unsigned long start, unsigned long end)
 628 {
 629 }
 630 #endif
 631
 632 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 633                                                 unsigned long end,
 634                                                 unsigned long page_size_mask)
 635 {
 636
 637         unsigned long next, last_map_addr = end;
 638
 639         start = (unsigned long)__va(start);
 640         end = (unsigned long)__va(end);
 641
 642         for (; start < end; start = next) {
 643                 pgd_t *pgd = pgd_offset_k(start);
 644                 unsigned long pud_phys;
 645                 pud_t *pud;
 646
 647                 next = start + PGDIR_SIZE;
 648                 if (next > end)
 649                         next = end;
 650
 651                 if (pgd_val(*pgd)) {
 652                         last_map_addr = phys_pud_update(pgd, __pa(start),
 653                                                  __pa(end), page_size_mask);
 654                         continue;
 655                 }
 656
 657                 if (after_bootmem)
 658                         pud = pud_offset(pgd, start & PGDIR_MASK);
 659                 else
 660                         pud = alloc_low_page(&pud_phys);
 661
 662                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 663                                                  page_size_mask);
 664                 unmap_low_page(pud);
 665                 pgd_populate(&init_mm, pgd_offset_k(start),
 666                              __va(pud_phys));
 667         }
 668
 669         return last_map_addr;
 670 }
 671
 672 struct map_range {
 673         unsigned long start;
 674         unsigned long end;
 675         unsigned page_size_mask;
 676 };
 677
 678 #define NR_RANGE_MR 5
 679
 680 static int save_mr(struct map_range *mr, int nr_range,
 681                    unsigned long start_pfn, unsigned long end_pfn,
 682                    unsigned long page_size_mask)
 683 {
 684
 685         if (start_pfn < end_pfn) {
 686                 if (nr_range >= NR_RANGE_MR)
 687                         panic("run out of range for init_memory_mapping\n");
 688                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 689                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 690                 mr[nr_range].page_size_mask = page_size_mask;
 691                 nr_range++;
 692         }
 693
 694         return nr_range;
 695 }
 696
 697 /*
 698  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 699  * This runs before bootmem is initialized and gets pages directly from
 700  * the physical memory. To access them they are temporarily mapped.
 701  */
 702 unsigned long __init_refok init_memory_mapping(unsigned long start,
 703                                                unsigned long end)
 704 {
 705         unsigned long last_map_addr = 0;
 706         unsigned long page_size_mask = 0;
 707         unsigned long start_pfn, end_pfn;
 708
 709         struct map_range mr[NR_RANGE_MR];
 710         int nr_range, i;
 711
 712         printk(KERN_INFO "init_memory_mapping\n");
 713
 714         /*
 715          * Find space for the kernel direct mapping tables.
 716          *
 717          * Later we should allocate these tables in the local node of the
 718          * memory mapped. Unfortunately this is done currently before the
 719          * nodes are discovered.
 720          */
 721         if (!after_bootmem)
 722                 init_gbpages();
 723
 724         if (direct_gbpages)
 725                 page_size_mask |= 1 << PG_LEVEL_1G;
 726         if (cpu_has_pse)
 727                 page_size_mask |= 1 << PG_LEVEL_2M;
 728
 729         memset(mr, 0, sizeof(mr));
 730         nr_range = 0;
 731
 732         /* head if not big page alignment ?*/
 733         start_pfn = start >> PAGE_SHIFT;
 734         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 735                         << (PMD_SHIFT - PAGE_SHIFT);
 736         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 737
 738         /* big page (2M) range*/
 739         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 740                          << (PMD_SHIFT - PAGE_SHIFT);
 741         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 742                          << (PUD_SHIFT - PAGE_SHIFT);
 743         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 744                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 745         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 746                         page_size_mask & (1<<PG_LEVEL_2M));
 747
 748         /* big page (1G) range */
 749         start_pfn = end_pfn;
 750         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 751         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 752                                 page_size_mask &
 753                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 754
 755         /* tail is not big page (1G) alignment */
 756         start_pfn = end_pfn;
 757         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 758         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 759                         page_size_mask & (1<<PG_LEVEL_2M));
 760
 761         /* tail is not big page (2M) alignment */
 762         start_pfn = end_pfn;
 763         end_pfn = end>>PAGE_SHIFT;
 764         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 765
 766         /* try to merge same page size and continuous */
 767         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 768                 unsigned long old_start;
 769                 if (mr[i].end != mr[i+1].start ||
 770                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 771                         continue;
 772                 /* move it */
 773                 old_start = mr[i].start;
 774                 memmove(&mr[i], &mr[i+1],
 775                          (nr_range - 1 - i) * sizeof (struct map_range));
 776                 mr[i].start = old_start;
 777                 nr_range--;
 778         }
 779
 780         for (i = 0; i < nr_range; i++)
 781                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 782                                 mr[i].start, mr[i].end,
 783                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 784                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 785
 786         if (!after_bootmem)
 787                 find_early_table_space(end);
 788
 789         for (i = 0; i < nr_range; i++)
 790                 last_map_addr = kernel_physical_mapping_init(
 791                                         mr[i].start, mr[i].end,
 792                                         mr[i].page_size_mask);
 793
 794         if (!after_bootmem)
 795                 mmu_cr4_features = read_cr4();
 796         __flush_tlb_all();
 797
 798         if (!after_bootmem && table_end > table_start)
 799                 reserve_early(table_start << PAGE_SHIFT,
 800                                  table_end << PAGE_SHIFT, "PGTABLE");
 801
 802         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 803                          last_map_addr, end);
 804
 805         if (!after_bootmem)
 806                 early_memtest(start, end);
 807
 808         return last_map_addr >> PAGE_SHIFT;
 809 }
 810
 811 #ifndef CONFIG_NUMA
 812 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 813 {
 814         unsigned long bootmap_size, bootmap;
 815
 816         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 817         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 818                                  PAGE_SIZE);
 819         if (bootmap == -1L)
 820                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 821         /* don't touch min_low_pfn */
 822         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 823                                          0, end_pfn);
 824         e820_register_active_regions(0, start_pfn, end_pfn);
 825         free_bootmem_with_active_regions(0, end_pfn);
 826         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 827         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 828 }
 829
 830 void __init paging_init(void)
 831 {
 832         unsigned long max_zone_pfns[MAX_NR_ZONES];
 833
 834         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 835         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 836         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 837         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 838
 839         memory_present(0, 0, max_pfn);
 840         sparse_init();
 841         free_area_init_nodes(max_zone_pfns);
 842 }
 843 #endif
 844
 845 /*
 846  * Memory hotplug specific functions
 847  */
 848 #ifdef CONFIG_MEMORY_HOTPLUG
 849 /*
 850  * Memory is added always to NORMAL zone. This means you will never get
 851  * additional DMA/DMA32 memory.
 852  */
 853 int arch_add_memory(int nid, u64 start, u64 size)
 854 {
 855         struct pglist_data *pgdat = NODE_DATA(nid);
 856         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 857         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 858         unsigned long nr_pages = size >> PAGE_SHIFT;
 859         int ret;
 860
 861         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 862         if (last_mapped_pfn > max_pfn_mapped)
 863                 max_pfn_mapped = last_mapped_pfn;
 864
 865         ret = __add_pages(zone, start_pfn, nr_pages);
 866         WARN_ON(1);
 867
 868         return ret;
 869 }
 870 EXPORT_SYMBOL_GPL(arch_add_memory);
 871
 872 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 873 int memory_add_physaddr_to_nid(u64 start)
 874 {
 875         return 0;
 876 }
 877 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 878 #endif
 879
 880 #endif /* CONFIG_MEMORY_HOTPLUG */
 881
 882 /*
 883  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 884  * is valid. The argument is a physical page number.
 885  *
 886  *
 887  * On x86, access has to be given to the first megabyte of ram because that area
 888  * contains bios code and data regions used by X and dosemu and similar apps.
 889  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 890  * mmio resources as well as potential bios/acpi data regions.
 891  */
 892 int devmem_is_allowed(unsigned long pagenr)
 893 {
 894         if (pagenr <= 256)
 895                 return 1;
 896         if (!page_is_ram(pagenr))
 897                 return 1;
 898         return 0;
 899 }
 900
 901
 902 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 903                          kcore_modules, kcore_vsyscall;
 904
 905 void __init mem_init(void)
 906 {
 907         long codesize, reservedpages, datasize, initsize;
 908
 909         pci_iommu_alloc();
 910
 911         /* clear_bss() already clear the empty_zero_page */
 912
 913         reservedpages = 0;
 914
 915         /* this will put all low memory onto the freelists */
 916 #ifdef CONFIG_NUMA
 917         totalram_pages = numa_free_all_bootmem();
 918 #else
 919         totalram_pages = free_all_bootmem();
 920 #endif
 921         reservedpages = max_pfn - totalram_pages -
 922                                         absent_pages_in_range(0, max_pfn);
 923         after_bootmem = 1;
 924
 925         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 926         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 927         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 928
 929         /* Register memory areas for /proc/kcore */
 930         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 931         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 932                    VMALLOC_END-VMALLOC_START);
 933         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 934         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 935         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 936                                  VSYSCALL_END - VSYSCALL_START);
 937
 938         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 939                                 "%ldk reserved, %ldk data, %ldk init)\n",
 940                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 941                 max_pfn << (PAGE_SHIFT-10),
 942                 codesize >> 10,
 943                 reservedpages << (PAGE_SHIFT-10),
 944                 datasize >> 10,
 945                 initsize >> 10);
 946
 947         cpa_init();
 948 }
 949
 950 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 951 {
 952         unsigned long addr = begin;
 953
 954         if (addr >= end)
 955                 return;
 956
 957         /*
 958          * If debugging page accesses then do not free this memory but
 959          * mark them not present - any buggy init-section access will
 960          * create a kernel page fault:
 961          */
 962 #ifdef CONFIG_DEBUG_PAGEALLOC
 963         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 964                 begin, PAGE_ALIGN(end));
 965         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 966 #else
 967         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 968
 969         for (; addr < end; addr += PAGE_SIZE) {
 970                 ClearPageReserved(virt_to_page(addr));
 971                 init_page_count(virt_to_page(addr));
 972                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 973                         POISON_FREE_INITMEM, PAGE_SIZE);
 974                 free_page(addr);
 975                 totalram_pages++;
 976         }
 977 #endif
 978 }
 979
 980 void free_initmem(void)
 981 {
 982         free_init_pages("unused kernel memory",
 983                         (unsigned long)(&__init_begin),
 984                         (unsigned long)(&__init_end));
 985 }
 986
 987 #ifdef CONFIG_DEBUG_RODATA
 988 const int rodata_test_data = 0xC3;
 989 EXPORT_SYMBOL_GPL(rodata_test_data);
 990
 991 void mark_rodata_ro(void)
 992 {
 993         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 994
 995         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 996                (end - start) >> 10);
 997         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 998
 999         /*
1000          * The rodata section (but not the kernel text!) should also be
1001          * not-executable.
1002          */
1003         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
1004         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
1005
1006         rodata_test();
1007
1008 #ifdef CONFIG_CPA_DEBUG
1009         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1010         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1011
1012         printk(KERN_INFO "Testing CPA: again\n");
1013         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1014 #endif
1015 }
1016
1017 #endif
1018
1019 #ifdef CONFIG_BLK_DEV_INITRD
1020 void free_initrd_mem(unsigned long start, unsigned long end)
1021 {
1022         free_init_pages("initrd memory", start, end);
1023 }
1024 #endif
1025
1026 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1027                                    int flags)
1028 {
1029 #ifdef CONFIG_NUMA
1030         int nid, next_nid;
1031         int ret;
1032 #endif
1033         unsigned long pfn = phys >> PAGE_SHIFT;
1034
1035         if (pfn >= max_pfn) {
1036                 /*
1037                  * This can happen with kdump kernels when accessing
1038                  * firmware tables:
1039                  */
1040                 if (pfn < max_pfn_mapped)
1041                         return -EFAULT;
1042
1043                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
1044                                 phys, len);
1045                 return -EFAULT;
1046         }
1047
1048         /* Should check here against the e820 map to avoid double free */
1049 #ifdef CONFIG_NUMA
1050         nid = phys_to_nid(phys);
1051         next_nid = phys_to_nid(phys + len - 1);
1052         if (nid == next_nid)
1053                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
1054         else
1055                 ret = reserve_bootmem(phys, len, flags);
1056
1057         if (ret != 0)
1058                 return ret;
1059
1060 #else
1061         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1062 #endif
1063
1064         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1065                 dma_reserve += len / PAGE_SIZE;
1066                 set_dma_reserve(dma_reserve);
1067         }
1068
1069         return 0;
1070 }
1071
1072 int kern_addr_valid(unsigned long addr)
1073 {
1074         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1075         pgd_t *pgd;
1076         pud_t *pud;
1077         pmd_t *pmd;
1078         pte_t *pte;
1079
1080         if (above != 0 && above != -1UL)
1081                 return 0;
1082
1083         pgd = pgd_offset_k(addr);
1084         if (pgd_none(*pgd))
1085                 return 0;
1086
1087         pud = pud_offset(pgd, addr);
1088         if (pud_none(*pud))
1089                 return 0;
1090
1091         pmd = pmd_offset(pud, addr);
1092         if (pmd_none(*pmd))
1093                 return 0;
1094
1095         if (pmd_large(*pmd))
1096                 return pfn_valid(pmd_pfn(*pmd));
1097
1098         pte = pte_offset_kernel(pmd, addr);
1099         if (pte_none(*pte))
1100                 return 0;
1101
1102         return pfn_valid(pte_pfn(*pte));
1103 }
1104
1105 /*
1106  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1107  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1108  * not need special handling anymore:
1109  */
1110 static struct vm_area_struct gate_vma = {
1111         .vm_start       = VSYSCALL_START,
1112         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1113         .vm_page_prot   = PAGE_READONLY_EXEC,
1114         .vm_flags       = VM_READ | VM_EXEC
1115 };
1116
1117 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1118 {
1119 #ifdef CONFIG_IA32_EMULATION
1120         if (test_tsk_thread_flag(tsk, TIF_IA32))
1121                 return NULL;
1122 #endif
1123         return &gate_vma;
1124 }
1125
1126 int in_gate_area(struct task_struct *task, unsigned long addr)
1127 {
1128         struct vm_area_struct *vma = get_gate_vma(task);
1129
1130         if (!vma)
1131                 return 0;
1132
1133         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1134 }
1135
1136 /*
1137  * Use this when you have no reliable task/vma, typically from interrupt
1138  * context. It is less reliable than using the task's vma and may give
1139  * false positives:
1140  */
1141 int in_gate_area_no_task(unsigned long addr)
1142 {
1143         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1144 }
1145
1146 const char *arch_vma_name(struct vm_area_struct *vma)
1147 {
1148         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1149                 return "[vdso]";
1150         if (vma == &gate_vma)
1151                 return "[vsyscall]";
1152         return NULL;
1153 }
1154
1155 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1156 /*
1157  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1158  */
1159 static long __meminitdata addr_start, addr_end;
1160 static void __meminitdata *p_start, *p_end;
1161 static int __meminitdata node_start;
1162
1163 int __meminit
1164 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1165 {
1166         unsigned long addr = (unsigned long)start_page;
1167         unsigned long end = (unsigned long)(start_page + size);
1168         unsigned long next;
1169         pgd_t *pgd;
1170         pud_t *pud;
1171         pmd_t *pmd;
1172
1173         for (; addr < end; addr = next) {
1174                 void *p = NULL;
1175
1176                 pgd = vmemmap_pgd_populate(addr, node);
1177                 if (!pgd)
1178                         return -ENOMEM;
1179
1180                 pud = vmemmap_pud_populate(pgd, addr, node);
1181                 if (!pud)
1182                         return -ENOMEM;
1183
1184                 if (!cpu_has_pse) {
1185                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1186                         pmd = vmemmap_pmd_populate(pud, addr, node);
1187
1188                         if (!pmd)
1189                                 return -ENOMEM;
1190
1191                         p = vmemmap_pte_populate(pmd, addr, node);
1192
1193                         if (!p)
1194                                 return -ENOMEM;
1195
1196                         addr_end = addr + PAGE_SIZE;
1197                         p_end = p + PAGE_SIZE;
1198                 } else {
1199                         next = pmd_addr_end(addr, end);
1200
1201                         pmd = pmd_offset(pud, addr);
1202                         if (pmd_none(*pmd)) {
1203                                 pte_t entry;
1204
1205                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1206                                 if (!p)
1207                                         return -ENOMEM;
1208
1209                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1210                                                 PAGE_KERNEL_LARGE);
1211                                 set_pmd(pmd, __pmd(pte_val(entry)));
1212
1213                                 /* check to see if we have contiguous blocks */
1214                                 if (p_end != p || node_start != node) {
1215                                         if (p_start)
1216                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1217                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1218                                         addr_start = addr;
1219                                         node_start = node;
1220                                         p_start = p;
1221                                 }
1222
1223                                 addr_end = addr + PMD_SIZE;
1224                                 p_end = p + PMD_SIZE;
1225                         } else
1226                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1227                 }
1228
1229         }
1230         return 0;
1231 }
1232
1233 void __meminit vmemmap_populate_print_last(void)
1234 {
1235         if (p_start) {
1236                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1237                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1238                 p_start = NULL;
1239                 p_end = NULL;
1240                 node_start = 0;
1241         }
1242 }
1243 #endif