git.oblomov.eu Git - linux-2.6/blob - 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_pfn_mapped;
  57
  58 static unsigned long dma_reserve __initdata;
  59
  60 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  61
  62 int direct_gbpages __meminitdata
  63 #ifdef CONFIG_DIRECT_GBPAGES
  64                                 = 1
  65 #endif
  66 ;
  67
  68 static int __init parse_direct_gbpages_off(char *arg)
  69 {
  70         direct_gbpages = 0;
  71         return 0;
  72 }
  73 early_param("nogbpages", parse_direct_gbpages_off);
  74
  75 static int __init parse_direct_gbpages_on(char *arg)
  76 {
  77         direct_gbpages = 1;
  78         return 0;
  79 }
  80 early_param("gbpages", parse_direct_gbpages_on);
  81
  82 /*
  83  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  84  * physical space so we can cache the place of the first one and move
  85  * around without checking the pgd every time.
  86  */
  87
  88 void show_mem(void)
  89 {
  90         long i, total = 0, reserved = 0;
  91         long shared = 0, cached = 0;
  92         struct page *page;
  93         pg_data_t *pgdat;
  94
  95         printk(KERN_INFO "Mem-info:\n");
  96         show_free_areas();
  97         for_each_online_pgdat(pgdat) {
  98                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
  99                         /*
 100                          * This loop can take a while with 256 GB and
 101                          * 4k pages so defer the NMI watchdog:
 102                          */
 103                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
 104                                 touch_nmi_watchdog();
 105
 106                         if (!pfn_valid(pgdat->node_start_pfn + i))
 107                                 continue;
 108
 109                         page = pfn_to_page(pgdat->node_start_pfn + i);
 110                         total++;
 111                         if (PageReserved(page))
 112                                 reserved++;
 113                         else if (PageSwapCache(page))
 114                                 cached++;
 115                         else if (page_count(page))
 116                                 shared += page_count(page) - 1;
 117                 }
 118         }
 119         printk(KERN_INFO "%lu pages of RAM\n",          total);
 120         printk(KERN_INFO "%lu reserved pages\n",        reserved);
 121         printk(KERN_INFO "%lu pages shared\n",          shared);
 122         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 123 }
 124
 125 int after_bootmem;
 126
 127 static __init void *spp_getpage(void)
 128 {
 129         void *ptr;
 130
 131         if (after_bootmem)
 132                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 133         else
 134                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 135
 136         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 137                 panic("set_pte_phys: cannot allocate page data %s\n",
 138                         after_bootmem ? "after bootmem" : "");
 139         }
 140
 141         pr_debug("spp_getpage %p\n", ptr);
 142
 143         return ptr;
 144 }
 145
 146 void
 147 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 148 {
 149         pud_t *pud;
 150         pmd_t *pmd;
 151         pte_t *pte;
 152
 153         pud = pud_page + pud_index(vaddr);
 154         if (pud_none(*pud)) {
 155                 pmd = (pmd_t *) spp_getpage();
 156                 pud_populate(&init_mm, pud, pmd);
 157                 if (pmd != pmd_offset(pud, 0)) {
 158                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 159                                 pmd, pmd_offset(pud, 0));
 160                         return;
 161                 }
 162         }
 163         pmd = pmd_offset(pud, vaddr);
 164         if (pmd_none(*pmd)) {
 165                 pte = (pte_t *) spp_getpage();
 166                 pmd_populate_kernel(&init_mm, pmd, pte);
 167                 if (pte != pte_offset_kernel(pmd, 0)) {
 168                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 169                         return;
 170                 }
 171         }
 172
 173         pte = pte_offset_kernel(pmd, vaddr);
 174         if (!pte_none(*pte) && pte_val(new_pte) &&
 175             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 176                 pte_ERROR(*pte);
 177         set_pte(pte, new_pte);
 178
 179         /*
 180          * It's enough to flush this one mapping.
 181          * (PGE mappings get flushed as well)
 182          */
 183         __flush_tlb_one(vaddr);
 184 }
 185
 186 void
 187 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 188 {
 189         pgd_t *pgd;
 190         pud_t *pud_page;
 191
 192         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 193
 194         pgd = pgd_offset_k(vaddr);
 195         if (pgd_none(*pgd)) {
 196                 printk(KERN_ERR
 197                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 198                 return;
 199         }
 200         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 201         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 202 }
 203
 204 /*
 205  * Create large page table mappings for a range of physical addresses.
 206  */
 207 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 208                                                 pgprot_t prot)
 209 {
 210         pgd_t *pgd;
 211         pud_t *pud;
 212         pmd_t *pmd;
 213
 214         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 215         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 216                 pgd = pgd_offset_k((unsigned long)__va(phys));
 217                 if (pgd_none(*pgd)) {
 218                         pud = (pud_t *) spp_getpage();
 219                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 220                                                 _PAGE_USER));
 221                 }
 222                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 223                 if (pud_none(*pud)) {
 224                         pmd = (pmd_t *) spp_getpage();
 225                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 226                                                 _PAGE_USER));
 227                 }
 228                 pmd = pmd_offset(pud, phys);
 229                 BUG_ON(!pmd_none(*pmd));
 230                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 231         }
 232 }
 233
 234 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 235 {
 236         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 237 }
 238
 239 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 240 {
 241         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 242 }
 243
 244 /*
 245  * The head.S code sets up the kernel high mapping:
 246  *
 247  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 248  *
 249  * phys_addr holds the negative offset to the kernel, which is added
 250  * to the compile time generated pmds. This results in invalid pmds up
 251  * to the point where we hit the physaddr 0 mapping.
 252  *
 253  * We limit the mappings to the region from _text to _end.  _end is
 254  * rounded up to the 2MB boundary. This catches the invalid pmds as
 255  * well, as they are located before _text:
 256  */
 257 void __init cleanup_highmap(void)
 258 {
 259         unsigned long vaddr = __START_KERNEL_map;
 260         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 261         pmd_t *pmd = level2_kernel_pgt;
 262         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 263
 264         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 265                 if (pmd_none(*pmd))
 266                         continue;
 267                 if (vaddr < (unsigned long) _text || vaddr > end)
 268                         set_pmd(pmd, __pmd(0));
 269         }
 270 }
 271
 272 static unsigned long __initdata table_start;
 273 static unsigned long __meminitdata table_end;
 274 static unsigned long __meminitdata table_top;
 275
 276 static __meminit void *alloc_low_page(unsigned long *phys)
 277 {
 278         unsigned long pfn = table_end++;
 279         void *adr;
 280
 281         if (after_bootmem) {
 282                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 283                 *phys = __pa(adr);
 284
 285                 return adr;
 286         }
 287
 288         if (pfn >= table_top)
 289                 panic("alloc_low_page: ran out of memory");
 290
 291         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 292         memset(adr, 0, PAGE_SIZE);
 293         *phys  = pfn * PAGE_SIZE;
 294         return adr;
 295 }
 296
 297 static __meminit void unmap_low_page(void *adr)
 298 {
 299         if (after_bootmem)
 300                 return;
 301
 302         early_iounmap(adr, PAGE_SIZE);
 303 }
 304
 305 static unsigned long __meminit
 306 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end)
 307 {
 308         unsigned pages = 0;
 309         unsigned long last_map_addr = end;
 310         int i;
 311
 312         pte_t *pte = pte_page + pte_index(addr);
 313
 314         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 315
 316                 if (addr >= end) {
 317                         if (!after_bootmem) {
 318                                 for(; i < PTRS_PER_PTE; i++, pte++)
 319                                         set_pte(pte, __pte(0));
 320                         }
 321                         break;
 322                 }
 323
 324                 if (pte_val(*pte))
 325                         continue;
 326
 327                 if (0)
 328                         printk("   pte=%p addr=%lx pte=%016lx\n",
 329                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 330                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL));
 331                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 332                 pages++;
 333         }
 334         update_page_count(PG_LEVEL_4K, pages);
 335
 336         return last_map_addr;
 337 }
 338
 339 static unsigned long __meminit
 340 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end)
 341 {
 342         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 343
 344         return phys_pte_init(pte, address, end);
 345 }
 346
 347 static unsigned long __meminit
 348 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 349                          unsigned long page_size_mask)
 350 {
 351         unsigned long pages = 0;
 352         unsigned long last_map_addr = end;
 353
 354         int i = pmd_index(address);
 355
 356         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 357                 unsigned long pte_phys;
 358                 pmd_t *pmd = pmd_page + pmd_index(address);
 359                 pte_t *pte;
 360
 361                 if (address >= end) {
 362                         if (!after_bootmem) {
 363                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 364                                         set_pmd(pmd, __pmd(0));
 365                         }
 366                         break;
 367                 }
 368
 369                 if (pmd_val(*pmd)) {
 370                         if (!pmd_large(*pmd))
 371                                 last_map_addr = phys_pte_update(pmd, address,
 372                                                                  end);
 373                         continue;
 374                 }
 375
 376                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 377                         pages++;
 378                         set_pte((pte_t *)pmd,
 379                                 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 380                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 381                         continue;
 382                 }
 383
 384                 pte = alloc_low_page(&pte_phys);
 385                 last_map_addr = phys_pte_init(pte, address, end);
 386                 unmap_low_page(pte);
 387
 388                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 389         }
 390         update_page_count(PG_LEVEL_2M, pages);
 391         return last_map_addr;
 392 }
 393
 394 static unsigned long __meminit
 395 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 396                          unsigned long page_size_mask)
 397 {
 398         pmd_t *pmd = pmd_offset(pud, 0);
 399         unsigned long last_map_addr;
 400
 401         spin_lock(&init_mm.page_table_lock);
 402         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask);
 403         spin_unlock(&init_mm.page_table_lock);
 404         __flush_tlb_all();
 405         return last_map_addr;
 406 }
 407
 408 static unsigned long __meminit
 409 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 410                          unsigned long page_size_mask)
 411 {
 412         unsigned long pages = 0;
 413         unsigned long last_map_addr = end;
 414         int i = pud_index(addr);
 415
 416         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 417                 unsigned long pmd_phys;
 418                 pud_t *pud = pud_page + pud_index(addr);
 419                 pmd_t *pmd;
 420
 421                 if (addr >= end)
 422                         break;
 423
 424                 if (!after_bootmem &&
 425                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 426                         set_pud(pud, __pud(0));
 427                         continue;
 428                 }
 429
 430                 if (pud_val(*pud)) {
 431                         if (!pud_large(*pud))
 432                                 last_map_addr = phys_pmd_update(pud, addr, end,
 433                                                          page_size_mask);
 434                         continue;
 435                 }
 436
 437                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 438                         pages++;
 439                         set_pte((pte_t *)pud,
 440                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 441                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 442                         continue;
 443                 }
 444
 445                 pmd = alloc_low_page(&pmd_phys);
 446
 447                 spin_lock(&init_mm.page_table_lock);
 448                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask);
 449                 unmap_low_page(pmd);
 450                 pud_populate(&init_mm, pud, __va(pmd_phys));
 451                 spin_unlock(&init_mm.page_table_lock);
 452
 453         }
 454         __flush_tlb_all();
 455         update_page_count(PG_LEVEL_1G, pages);
 456
 457         return last_map_addr;
 458 }
 459
 460 static unsigned long __meminit
 461 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 462                  unsigned long page_size_mask)
 463 {
 464         pud_t *pud;
 465
 466         pud = (pud_t *)pgd_page_vaddr(*pgd);
 467
 468         return phys_pud_init(pud, addr, end, page_size_mask);
 469 }
 470
 471 static void __init find_early_table_space(unsigned long end)
 472 {
 473         unsigned long puds, pmds, ptes, tables, start;
 474
 475         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 476         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 477         if (direct_gbpages) {
 478                 unsigned long extra;
 479                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 480                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 481         } else
 482                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 483         tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 484
 485         if (cpu_has_pse) {
 486                 unsigned long extra;
 487                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 488                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 489         } else
 490                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 491         tables += round_up(ptes * sizeof(pte_t), PAGE_SIZE);
 492
 493         /*
 494          * RED-PEN putting page tables only on node 0 could
 495          * cause a hotspot and fill up ZONE_DMA. The page tables
 496          * need roughly 0.5KB per GB.
 497          */
 498         start = 0x8000;
 499         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 500         if (table_start == -1UL)
 501                 panic("Cannot find space for the kernel page tables");
 502
 503         table_start >>= PAGE_SHIFT;
 504         table_end = table_start;
 505         table_top = table_start + (tables >> PAGE_SHIFT);
 506
 507         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 508                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 509 }
 510
 511 static void __init init_gbpages(void)
 512 {
 513         if (direct_gbpages && cpu_has_gbpages)
 514                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 515         else
 516                 direct_gbpages = 0;
 517 }
 518
 519 #ifdef CONFIG_MEMTEST
 520
 521 static void __init memtest(unsigned long start_phys, unsigned long size,
 522                                  unsigned pattern)
 523 {
 524         unsigned long i;
 525         unsigned long *start;
 526         unsigned long start_bad;
 527         unsigned long last_bad;
 528         unsigned long val;
 529         unsigned long start_phys_aligned;
 530         unsigned long count;
 531         unsigned long incr;
 532
 533         switch (pattern) {
 534         case 0:
 535                 val = 0UL;
 536                 break;
 537         case 1:
 538                 val = -1UL;
 539                 break;
 540         case 2:
 541                 val = 0x5555555555555555UL;
 542                 break;
 543         case 3:
 544                 val = 0xaaaaaaaaaaaaaaaaUL;
 545                 break;
 546         default:
 547                 return;
 548         }
 549
 550         incr = sizeof(unsigned long);
 551         start_phys_aligned = ALIGN(start_phys, incr);
 552         count = (size - (start_phys_aligned - start_phys))/incr;
 553         start = __va(start_phys_aligned);
 554         start_bad = 0;
 555         last_bad = 0;
 556
 557         for (i = 0; i < count; i++)
 558                 start[i] = val;
 559         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 560                 if (*start != val) {
 561                         if (start_phys_aligned == last_bad + incr) {
 562                                 last_bad += incr;
 563                         } else {
 564                                 if (start_bad) {
 565                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 566                                                 val, start_bad, last_bad + incr);
 567                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 568                                 }
 569                                 start_bad = last_bad = start_phys_aligned;
 570                         }
 571                 }
 572         }
 573         if (start_bad) {
 574                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 575                         val, start_bad, last_bad + incr);
 576                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 577         }
 578
 579 }
 580
 581 /* default is disabled */
 582 static int memtest_pattern __initdata;
 583
 584 static int __init parse_memtest(char *arg)
 585 {
 586         if (arg)
 587                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 588         return 0;
 589 }
 590
 591 early_param("memtest", parse_memtest);
 592
 593 static void __init early_memtest(unsigned long start, unsigned long end)
 594 {
 595         u64 t_start, t_size;
 596         unsigned pattern;
 597
 598         if (!memtest_pattern)
 599                 return;
 600
 601         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 602         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 603                 t_start = start;
 604                 t_size = 0;
 605                 while (t_start < end) {
 606                         t_start = find_e820_area_size(t_start, &t_size, 1);
 607
 608                         /* done ? */
 609                         if (t_start >= end)
 610                                 break;
 611                         if (t_start + t_size > end)
 612                                 t_size = end - t_start;
 613
 614                         printk(KERN_CONT "\n  %016llx - %016llx pattern %d",
 615                                 (unsigned long long)t_start,
 616                                 (unsigned long long)t_start + t_size, pattern);
 617
 618                         memtest(t_start, t_size, pattern);
 619
 620                         t_start += t_size;
 621                 }
 622         }
 623         printk(KERN_CONT "\n");
 624 }
 625 #else
 626 static void __init early_memtest(unsigned long start, unsigned long end)
 627 {
 628 }
 629 #endif
 630
 631 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 632                                                 unsigned long end,
 633                                                 unsigned long page_size_mask)
 634 {
 635
 636         unsigned long next, last_map_addr = end;
 637
 638         start = (unsigned long)__va(start);
 639         end = (unsigned long)__va(end);
 640
 641         for (; start < end; start = next) {
 642                 pgd_t *pgd = pgd_offset_k(start);
 643                 unsigned long pud_phys;
 644                 pud_t *pud;
 645
 646                 next = start + PGDIR_SIZE;
 647                 if (next > end)
 648                         next = end;
 649
 650                 if (pgd_val(*pgd)) {
 651                         last_map_addr = phys_pud_update(pgd, __pa(start),
 652                                                  __pa(end), page_size_mask);
 653                         continue;
 654                 }
 655
 656                 if (after_bootmem)
 657                         pud = pud_offset(pgd, start & PGDIR_MASK);
 658                 else
 659                         pud = alloc_low_page(&pud_phys);
 660
 661                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 662                                                  page_size_mask);
 663                 unmap_low_page(pud);
 664                 pgd_populate(&init_mm, pgd_offset_k(start),
 665                              __va(pud_phys));
 666         }
 667
 668         return last_map_addr;
 669 }
 670
 671 struct map_range {
 672         unsigned long start;
 673         unsigned long end;
 674         unsigned page_size_mask;
 675 };
 676
 677 #define NR_RANGE_MR 5
 678
 679 static int save_mr(struct map_range *mr, int nr_range,
 680                    unsigned long start_pfn, unsigned long end_pfn,
 681                    unsigned long page_size_mask)
 682 {
 683
 684         if (start_pfn < end_pfn) {
 685                 if (nr_range >= NR_RANGE_MR)
 686                         panic("run out of range for init_memory_mapping\n");
 687                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 688                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 689                 mr[nr_range].page_size_mask = page_size_mask;
 690                 nr_range++;
 691         }
 692
 693         return nr_range;
 694 }
 695
 696 /*
 697  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 698  * This runs before bootmem is initialized and gets pages directly from
 699  * the physical memory. To access them they are temporarily mapped.
 700  */
 701 unsigned long __init_refok init_memory_mapping(unsigned long start,
 702                                                unsigned long end)
 703 {
 704         unsigned long last_map_addr = 0;
 705         unsigned long page_size_mask = 0;
 706         unsigned long start_pfn, end_pfn;
 707
 708         struct map_range mr[NR_RANGE_MR];
 709         int nr_range, i;
 710
 711         printk(KERN_INFO "init_memory_mapping\n");
 712
 713         /*
 714          * Find space for the kernel direct mapping tables.
 715          *
 716          * Later we should allocate these tables in the local node of the
 717          * memory mapped. Unfortunately this is done currently before the
 718          * nodes are discovered.
 719          */
 720         if (!after_bootmem)
 721                 init_gbpages();
 722
 723         if (direct_gbpages)
 724                 page_size_mask |= 1 << PG_LEVEL_1G;
 725         if (cpu_has_pse)
 726                 page_size_mask |= 1 << PG_LEVEL_2M;
 727
 728         memset(mr, 0, sizeof(mr));
 729         nr_range = 0;
 730
 731         /* head if not big page alignment ?*/
 732         start_pfn = start >> PAGE_SHIFT;
 733         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 734                         << (PMD_SHIFT - PAGE_SHIFT);
 735         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 736
 737         /* big page (2M) range*/
 738         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 739                          << (PMD_SHIFT - PAGE_SHIFT);
 740         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 741                          << (PUD_SHIFT - PAGE_SHIFT);
 742         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 743                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 744         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 745                         page_size_mask & (1<<PG_LEVEL_2M));
 746
 747         /* big page (1G) range */
 748         start_pfn = end_pfn;
 749         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 750         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 751                                 page_size_mask &
 752                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 753
 754         /* tail is not big page (1G) alignment */
 755         start_pfn = end_pfn;
 756         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 757         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 758                         page_size_mask & (1<<PG_LEVEL_2M));
 759
 760         /* tail is not big page (2M) alignment */
 761         start_pfn = end_pfn;
 762         end_pfn = end>>PAGE_SHIFT;
 763         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 764
 765         for (i = 0; i < nr_range; i++)
 766                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 767                                 mr[i].start, mr[i].end,
 768                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 769                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 770
 771         if (!after_bootmem)
 772                 find_early_table_space(end);
 773
 774         for (i = 0; i < nr_range; i++)
 775                 last_map_addr = kernel_physical_mapping_init(
 776                                         mr[i].start, mr[i].end,
 777                                         mr[i].page_size_mask);
 778
 779         if (!after_bootmem)
 780                 mmu_cr4_features = read_cr4();
 781         __flush_tlb_all();
 782
 783         if (!after_bootmem && table_end > table_start)
 784                 reserve_early(table_start << PAGE_SHIFT,
 785                                  table_end << PAGE_SHIFT, "PGTABLE");
 786
 787         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 788                          last_map_addr, end);
 789
 790         if (!after_bootmem)
 791                 early_memtest(start, end);
 792
 793         return last_map_addr >> PAGE_SHIFT;
 794 }
 795
 796 #ifndef CONFIG_NUMA
 797 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 798 {
 799         unsigned long bootmap_size, bootmap;
 800
 801         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 802         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 803                                  PAGE_SIZE);
 804         if (bootmap == -1L)
 805                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 806         /* don't touch min_low_pfn */
 807         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 808                                          0, end_pfn);
 809         e820_register_active_regions(0, start_pfn, end_pfn);
 810         free_bootmem_with_active_regions(0, end_pfn);
 811         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 812         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 813 }
 814
 815 void __init paging_init(void)
 816 {
 817         unsigned long max_zone_pfns[MAX_NR_ZONES];
 818
 819         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 820         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 821         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 822         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 823
 824         memory_present(0, 0, max_pfn);
 825         sparse_init();
 826         free_area_init_nodes(max_zone_pfns);
 827 }
 828 #endif
 829
 830 /*
 831  * Memory hotplug specific functions
 832  */
 833 #ifdef CONFIG_MEMORY_HOTPLUG
 834 /*
 835  * Memory is added always to NORMAL zone. This means you will never get
 836  * additional DMA/DMA32 memory.
 837  */
 838 int arch_add_memory(int nid, u64 start, u64 size)
 839 {
 840         struct pglist_data *pgdat = NODE_DATA(nid);
 841         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 842         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 843         unsigned long nr_pages = size >> PAGE_SHIFT;
 844         int ret;
 845
 846         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 847         if (last_mapped_pfn > max_pfn_mapped)
 848                 max_pfn_mapped = last_mapped_pfn;
 849
 850         ret = __add_pages(zone, start_pfn, nr_pages);
 851         WARN_ON(1);
 852
 853         return ret;
 854 }
 855 EXPORT_SYMBOL_GPL(arch_add_memory);
 856
 857 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 858 int memory_add_physaddr_to_nid(u64 start)
 859 {
 860         return 0;
 861 }
 862 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 863 #endif
 864
 865 #endif /* CONFIG_MEMORY_HOTPLUG */
 866
 867 /*
 868  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 869  * is valid. The argument is a physical page number.
 870  *
 871  *
 872  * On x86, access has to be given to the first megabyte of ram because that area
 873  * contains bios code and data regions used by X and dosemu and similar apps.
 874  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 875  * mmio resources as well as potential bios/acpi data regions.
 876  */
 877 int devmem_is_allowed(unsigned long pagenr)
 878 {
 879         if (pagenr <= 256)
 880                 return 1;
 881         if (!page_is_ram(pagenr))
 882                 return 1;
 883         return 0;
 884 }
 885
 886
 887 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 888                          kcore_modules, kcore_vsyscall;
 889
 890 void __init mem_init(void)
 891 {
 892         long codesize, reservedpages, datasize, initsize;
 893
 894         pci_iommu_alloc();
 895
 896         /* clear_bss() already clear the empty_zero_page */
 897
 898         reservedpages = 0;
 899
 900         /* this will put all low memory onto the freelists */
 901 #ifdef CONFIG_NUMA
 902         totalram_pages = numa_free_all_bootmem();
 903 #else
 904         totalram_pages = free_all_bootmem();
 905 #endif
 906         reservedpages = max_pfn - totalram_pages -
 907                                         absent_pages_in_range(0, max_pfn);
 908         after_bootmem = 1;
 909
 910         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 911         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 912         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 913
 914         /* Register memory areas for /proc/kcore */
 915         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 916         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 917                    VMALLOC_END-VMALLOC_START);
 918         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 919         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 920         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 921                                  VSYSCALL_END - VSYSCALL_START);
 922
 923         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 924                                 "%ldk reserved, %ldk data, %ldk init)\n",
 925                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 926                 max_pfn << (PAGE_SHIFT-10),
 927                 codesize >> 10,
 928                 reservedpages << (PAGE_SHIFT-10),
 929                 datasize >> 10,
 930                 initsize >> 10);
 931
 932         cpa_init();
 933 }
 934
 935 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 936 {
 937         unsigned long addr = begin;
 938
 939         if (addr >= end)
 940                 return;
 941
 942         /*
 943          * If debugging page accesses then do not free this memory but
 944          * mark them not present - any buggy init-section access will
 945          * create a kernel page fault:
 946          */
 947 #ifdef CONFIG_DEBUG_PAGEALLOC
 948         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 949                 begin, PAGE_ALIGN(end));
 950         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 951 #else
 952         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 953
 954         for (; addr < end; addr += PAGE_SIZE) {
 955                 ClearPageReserved(virt_to_page(addr));
 956                 init_page_count(virt_to_page(addr));
 957                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 958                         POISON_FREE_INITMEM, PAGE_SIZE);
 959                 free_page(addr);
 960                 totalram_pages++;
 961         }
 962 #endif
 963 }
 964
 965 void free_initmem(void)
 966 {
 967         free_init_pages("unused kernel memory",
 968                         (unsigned long)(&__init_begin),
 969                         (unsigned long)(&__init_end));
 970 }
 971
 972 #ifdef CONFIG_DEBUG_RODATA
 973 const int rodata_test_data = 0xC3;
 974 EXPORT_SYMBOL_GPL(rodata_test_data);
 975
 976 void mark_rodata_ro(void)
 977 {
 978         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 979
 980         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 981                (end - start) >> 10);
 982         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 983
 984         /*
 985          * The rodata section (but not the kernel text!) should also be
 986          * not-executable.
 987          */
 988         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 989         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 990
 991         rodata_test();
 992
 993 #ifdef CONFIG_CPA_DEBUG
 994         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 995         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 996
 997         printk(KERN_INFO "Testing CPA: again\n");
 998         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 999 #endif
1000 }
1001
1002 #endif
1003
1004 #ifdef CONFIG_BLK_DEV_INITRD
1005 void free_initrd_mem(unsigned long start, unsigned long end)
1006 {
1007         free_init_pages("initrd memory", start, end);
1008 }
1009 #endif
1010
1011 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1012                                    int flags)
1013 {
1014 #ifdef CONFIG_NUMA
1015         int nid, next_nid;
1016         int ret;
1017 #endif
1018         unsigned long pfn = phys >> PAGE_SHIFT;
1019
1020         if (pfn >= max_pfn) {
1021                 /*
1022                  * This can happen with kdump kernels when accessing
1023                  * firmware tables:
1024                  */
1025                 if (pfn < max_pfn_mapped)
1026                         return -EFAULT;
1027
1028                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
1029                                 phys, len);
1030                 return -EFAULT;
1031         }
1032
1033         /* Should check here against the e820 map to avoid double free */
1034 #ifdef CONFIG_NUMA
1035         nid = phys_to_nid(phys);
1036         next_nid = phys_to_nid(phys + len - 1);
1037         if (nid == next_nid)
1038                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
1039         else
1040                 ret = reserve_bootmem(phys, len, flags);
1041
1042         if (ret != 0)
1043                 return ret;
1044
1045 #else
1046         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1047 #endif
1048
1049         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1050                 dma_reserve += len / PAGE_SIZE;
1051                 set_dma_reserve(dma_reserve);
1052         }
1053
1054         return 0;
1055 }
1056
1057 int kern_addr_valid(unsigned long addr)
1058 {
1059         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1060         pgd_t *pgd;
1061         pud_t *pud;
1062         pmd_t *pmd;
1063         pte_t *pte;
1064
1065         if (above != 0 && above != -1UL)
1066                 return 0;
1067
1068         pgd = pgd_offset_k(addr);
1069         if (pgd_none(*pgd))
1070                 return 0;
1071
1072         pud = pud_offset(pgd, addr);
1073         if (pud_none(*pud))
1074                 return 0;
1075
1076         pmd = pmd_offset(pud, addr);
1077         if (pmd_none(*pmd))
1078                 return 0;
1079
1080         if (pmd_large(*pmd))
1081                 return pfn_valid(pmd_pfn(*pmd));
1082
1083         pte = pte_offset_kernel(pmd, addr);
1084         if (pte_none(*pte))
1085                 return 0;
1086
1087         return pfn_valid(pte_pfn(*pte));
1088 }
1089
1090 /*
1091  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1092  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1093  * not need special handling anymore:
1094  */
1095 static struct vm_area_struct gate_vma = {
1096         .vm_start       = VSYSCALL_START,
1097         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1098         .vm_page_prot   = PAGE_READONLY_EXEC,
1099         .vm_flags       = VM_READ | VM_EXEC
1100 };
1101
1102 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1103 {
1104 #ifdef CONFIG_IA32_EMULATION
1105         if (test_tsk_thread_flag(tsk, TIF_IA32))
1106                 return NULL;
1107 #endif
1108         return &gate_vma;
1109 }
1110
1111 int in_gate_area(struct task_struct *task, unsigned long addr)
1112 {
1113         struct vm_area_struct *vma = get_gate_vma(task);
1114
1115         if (!vma)
1116                 return 0;
1117
1118         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1119 }
1120
1121 /*
1122  * Use this when you have no reliable task/vma, typically from interrupt
1123  * context. It is less reliable than using the task's vma and may give
1124  * false positives:
1125  */
1126 int in_gate_area_no_task(unsigned long addr)
1127 {
1128         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1129 }
1130
1131 const char *arch_vma_name(struct vm_area_struct *vma)
1132 {
1133         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1134                 return "[vdso]";
1135         if (vma == &gate_vma)
1136                 return "[vsyscall]";
1137         return NULL;
1138 }
1139
1140 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1141 /*
1142  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1143  */
1144 static long __meminitdata addr_start, addr_end;
1145 static void __meminitdata *p_start, *p_end;
1146 static int __meminitdata node_start;
1147
1148 int __meminit
1149 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1150 {
1151         unsigned long addr = (unsigned long)start_page;
1152         unsigned long end = (unsigned long)(start_page + size);
1153         unsigned long next;
1154         pgd_t *pgd;
1155         pud_t *pud;
1156         pmd_t *pmd;
1157
1158         for (; addr < end; addr = next) {
1159                 void *p = NULL;
1160
1161                 pgd = vmemmap_pgd_populate(addr, node);
1162                 if (!pgd)
1163                         return -ENOMEM;
1164
1165                 pud = vmemmap_pud_populate(pgd, addr, node);
1166                 if (!pud)
1167                         return -ENOMEM;
1168
1169                 if (!cpu_has_pse) {
1170                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1171                         pmd = vmemmap_pmd_populate(pud, addr, node);
1172
1173                         if (!pmd)
1174                                 return -ENOMEM;
1175
1176                         p = vmemmap_pte_populate(pmd, addr, node);
1177
1178                         if (!p)
1179                                 return -ENOMEM;
1180
1181                         addr_end = addr + PAGE_SIZE;
1182                         p_end = p + PAGE_SIZE;
1183                 } else {
1184                         next = pmd_addr_end(addr, end);
1185
1186                         pmd = pmd_offset(pud, addr);
1187                         if (pmd_none(*pmd)) {
1188                                 pte_t entry;
1189
1190                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1191                                 if (!p)
1192                                         return -ENOMEM;
1193
1194                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1195                                                 PAGE_KERNEL_LARGE);
1196                                 set_pmd(pmd, __pmd(pte_val(entry)));
1197
1198                                 /* check to see if we have contiguous blocks */
1199                                 if (p_end != p || node_start != node) {
1200                                         if (p_start)
1201                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1202                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1203                                         addr_start = addr;
1204                                         node_start = node;
1205                                         p_start = p;
1206                                 }
1207
1208                                 addr_end = addr + PMD_SIZE;
1209                                 p_end = p + PMD_SIZE;
1210                         } else
1211                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1212                 }
1213
1214         }
1215         return 0;
1216 }
1217
1218 void __meminit vmemmap_populate_print_last(void)
1219 {
1220         if (p_start) {
1221                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1222                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1223                 p_start = NULL;
1224                 p_end = NULL;
1225                 node_start = 0;
1226         }
1227 }
1228 #endif