err.no 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  * PFN of last memory page.
  53  */
  54 unsigned long end_pfn;
  55
  56 /*
  57  * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
  58  * The direct mapping extends to max_pfn_mapped, so that we can directly access
  59  * apertures, ACPI and other tables without having to play with fixmaps.
  60  */
  61 unsigned long max_pfn_mapped;
  62
  63 static unsigned long dma_reserve __initdata;
  64
  65 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  66
  67 int direct_gbpages __meminitdata
  68 #ifdef CONFIG_DIRECT_GBPAGES
  69                                 = 1
  70 #endif
  71 ;
  72
  73 static int __init parse_direct_gbpages_off(char *arg)
  74 {
  75         direct_gbpages = 0;
  76         return 0;
  77 }
  78 early_param("nogbpages", parse_direct_gbpages_off);
  79
  80 static int __init parse_direct_gbpages_on(char *arg)
  81 {
  82         direct_gbpages = 1;
  83         return 0;
  84 }
  85 early_param("gbpages", parse_direct_gbpages_on);
  86
  87 /*
  88  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  89  * physical space so we can cache the place of the first one and move
  90  * around without checking the pgd every time.
  91  */
  92
  93 void show_mem(void)
  94 {
  95         long i, total = 0, reserved = 0;
  96         long shared = 0, cached = 0;
  97         struct page *page;
  98         pg_data_t *pgdat;
  99
 100         printk(KERN_INFO "Mem-info:\n");
 101         show_free_areas();
 102         for_each_online_pgdat(pgdat) {
 103                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
 104                         /*
 105                          * This loop can take a while with 256 GB and
 106                          * 4k pages so defer the NMI watchdog:
 107                          */
 108                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
 109                                 touch_nmi_watchdog();
 110
 111                         if (!pfn_valid(pgdat->node_start_pfn + i))
 112                                 continue;
 113
 114                         page = pfn_to_page(pgdat->node_start_pfn + i);
 115                         total++;
 116                         if (PageReserved(page))
 117                                 reserved++;
 118                         else if (PageSwapCache(page))
 119                                 cached++;
 120                         else if (page_count(page))
 121                                 shared += page_count(page) - 1;
 122                 }
 123         }
 124         printk(KERN_INFO "%lu pages of RAM\n",          total);
 125         printk(KERN_INFO "%lu reserved pages\n",        reserved);
 126         printk(KERN_INFO "%lu pages shared\n",          shared);
 127         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 128 }
 129
 130 int after_bootmem;
 131
 132 static __init void *spp_getpage(void)
 133 {
 134         void *ptr;
 135
 136         if (after_bootmem)
 137                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 138         else
 139                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 140
 141         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 142                 panic("set_pte_phys: cannot allocate page data %s\n",
 143                         after_bootmem ? "after bootmem" : "");
 144         }
 145
 146         pr_debug("spp_getpage %p\n", ptr);
 147
 148         return ptr;
 149 }
 150
 151 void
 152 set_pte_vaddr(unsigned long vaddr, pte_t new_pte)
 153 {
 154         pgd_t *pgd;
 155         pud_t *pud;
 156         pmd_t *pmd;
 157         pte_t *pte;
 158
 159         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(new_pte));
 160
 161         pgd = pgd_offset_k(vaddr);
 162         if (pgd_none(*pgd)) {
 163                 printk(KERN_ERR
 164                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 165                 return;
 166         }
 167         pud = pud_offset(pgd, vaddr);
 168         if (pud_none(*pud)) {
 169                 pmd = (pmd_t *) spp_getpage();
 170                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
 171                 if (pmd != pmd_offset(pud, 0)) {
 172                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 173                                 pmd, pmd_offset(pud, 0));
 174                         return;
 175                 }
 176         }
 177         pmd = pmd_offset(pud, vaddr);
 178         if (pmd_none(*pmd)) {
 179                 pte = (pte_t *) spp_getpage();
 180                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
 181                 if (pte != pte_offset_kernel(pmd, 0)) {
 182                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 183                         return;
 184                 }
 185         }
 186
 187         pte = pte_offset_kernel(pmd, vaddr);
 188         if (!pte_none(*pte) && pte_val(new_pte) &&
 189             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 190                 pte_ERROR(*pte);
 191         set_pte(pte, new_pte);
 192
 193         /*
 194          * It's enough to flush this one mapping.
 195          * (PGE mappings get flushed as well)
 196          */
 197         __flush_tlb_one(vaddr);
 198 }
 199
 200 /*
 201  * The head.S code sets up the kernel high mapping:
 202  *
 203  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 204  *
 205  * phys_addr holds the negative offset to the kernel, which is added
 206  * to the compile time generated pmds. This results in invalid pmds up
 207  * to the point where we hit the physaddr 0 mapping.
 208  *
 209  * We limit the mappings to the region from _text to _end.  _end is
 210  * rounded up to the 2MB boundary. This catches the invalid pmds as
 211  * well, as they are located before _text:
 212  */
 213 void __init cleanup_highmap(void)
 214 {
 215         unsigned long vaddr = __START_KERNEL_map;
 216         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 217         pmd_t *pmd = level2_kernel_pgt;
 218         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 219
 220         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 221                 if (pmd_none(*pmd))
 222                         continue;
 223                 if (vaddr < (unsigned long) _text || vaddr > end)
 224                         set_pmd(pmd, __pmd(0));
 225         }
 226 }
 227
 228 static unsigned long __initdata table_start;
 229 static unsigned long __meminitdata table_end;
 230
 231 static __meminit void *alloc_low_page(unsigned long *phys)
 232 {
 233         unsigned long pfn = table_end++;
 234         void *adr;
 235
 236         if (after_bootmem) {
 237                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 238                 *phys = __pa(adr);
 239
 240                 return adr;
 241         }
 242
 243         if (pfn >= end_pfn)
 244                 panic("alloc_low_page: ran out of memory");
 245
 246         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 247         memset(adr, 0, PAGE_SIZE);
 248         *phys  = pfn * PAGE_SIZE;
 249         return adr;
 250 }
 251
 252 static __meminit void unmap_low_page(void *adr)
 253 {
 254         if (after_bootmem)
 255                 return;
 256
 257         early_iounmap(adr, PAGE_SIZE);
 258 }
 259
 260 /* Must run before zap_low_mappings */
 261 __meminit void *early_ioremap(unsigned long addr, unsigned long size)
 262 {
 263         pmd_t *pmd, *last_pmd;
 264         unsigned long vaddr;
 265         int i, pmds;
 266
 267         pmds = ((addr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 268         vaddr = __START_KERNEL_map;
 269         pmd = level2_kernel_pgt;
 270         last_pmd = level2_kernel_pgt + PTRS_PER_PMD - 1;
 271
 272         for (; pmd <= last_pmd; pmd++, vaddr += PMD_SIZE) {
 273                 for (i = 0; i < pmds; i++) {
 274                         if (pmd_present(pmd[i]))
 275                                 goto continue_outer_loop;
 276                 }
 277                 vaddr += addr & ~PMD_MASK;
 278                 addr &= PMD_MASK;
 279
 280                 for (i = 0; i < pmds; i++, addr += PMD_SIZE)
 281                         set_pmd(pmd+i, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
 282                 __flush_tlb_all();
 283
 284                 return (void *)vaddr;
 285 continue_outer_loop:
 286                 ;
 287         }
 288         printk(KERN_ERR "early_ioremap(0x%lx, %lu) failed\n", addr, size);
 289
 290         return NULL;
 291 }
 292
 293 /*
 294  * To avoid virtual aliases later:
 295  */
 296 __meminit void early_iounmap(void *addr, unsigned long size)
 297 {
 298         unsigned long vaddr;
 299         pmd_t *pmd;
 300         int i, pmds;
 301
 302         vaddr = (unsigned long)addr;
 303         pmds = ((vaddr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 304         pmd = level2_kernel_pgt + pmd_index(vaddr);
 305
 306         for (i = 0; i < pmds; i++)
 307                 pmd_clear(pmd + i);
 308
 309         __flush_tlb_all();
 310 }
 311
 312 static unsigned long __meminit
 313 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
 314 {
 315         unsigned long pages = 0;
 316
 317         int i = pmd_index(address);
 318
 319         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 320                 pmd_t *pmd = pmd_page + pmd_index(address);
 321
 322                 if (address >= end) {
 323                         if (!after_bootmem) {
 324                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 325                                         set_pmd(pmd, __pmd(0));
 326                         }
 327                         break;
 328                 }
 329
 330                 if (pmd_val(*pmd))
 331                         continue;
 332
 333                 pages++;
 334                 set_pte((pte_t *)pmd,
 335                         pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 336         }
 337         update_page_count(PG_LEVEL_2M, pages);
 338         return address;
 339 }
 340
 341 static unsigned long __meminit
 342 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
 343 {
 344         pmd_t *pmd = pmd_offset(pud, 0);
 345         unsigned long last_map_addr;
 346
 347         spin_lock(&init_mm.page_table_lock);
 348         last_map_addr = phys_pmd_init(pmd, address, end);
 349         spin_unlock(&init_mm.page_table_lock);
 350         __flush_tlb_all();
 351         return last_map_addr;
 352 }
 353
 354 static unsigned long __meminit
 355 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
 356 {
 357         unsigned long pages = 0;
 358         unsigned long last_map_addr = end;
 359         int i = pud_index(addr);
 360
 361         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 362                 unsigned long pmd_phys;
 363                 pud_t *pud = pud_page + pud_index(addr);
 364                 pmd_t *pmd;
 365
 366                 if (addr >= end)
 367                         break;
 368
 369                 if (!after_bootmem &&
 370                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 371                         set_pud(pud, __pud(0));
 372                         continue;
 373                 }
 374
 375                 if (pud_val(*pud)) {
 376                         if (!pud_large(*pud))
 377                                 last_map_addr = phys_pmd_update(pud, addr, end);
 378                         continue;
 379                 }
 380
 381                 if (direct_gbpages) {
 382                         pages++;
 383                         set_pte((pte_t *)pud,
 384                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 385                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 386                         continue;
 387                 }
 388
 389                 pmd = alloc_low_page(&pmd_phys);
 390
 391                 spin_lock(&init_mm.page_table_lock);
 392                 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
 393                 last_map_addr = phys_pmd_init(pmd, addr, end);
 394                 spin_unlock(&init_mm.page_table_lock);
 395
 396                 unmap_low_page(pmd);
 397         }
 398         __flush_tlb_all();
 399         update_page_count(PG_LEVEL_1G, pages);
 400
 401         return last_map_addr >> PAGE_SHIFT;
 402 }
 403
 404 static void __init find_early_table_space(unsigned long end)
 405 {
 406         unsigned long puds, pmds, tables, start;
 407
 408         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 409         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 410         if (!direct_gbpages) {
 411                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 412                 tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 413         }
 414
 415         /*
 416          * RED-PEN putting page tables only on node 0 could
 417          * cause a hotspot and fill up ZONE_DMA. The page tables
 418          * need roughly 0.5KB per GB.
 419          */
 420         start = 0x8000;
 421         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 422         if (table_start == -1UL)
 423                 panic("Cannot find space for the kernel page tables");
 424
 425         table_start >>= PAGE_SHIFT;
 426         table_end = table_start;
 427
 428         early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
 429                 end, table_start << PAGE_SHIFT,
 430                 (table_start << PAGE_SHIFT) + tables);
 431 }
 432
 433 static void __init init_gbpages(void)
 434 {
 435         if (direct_gbpages && cpu_has_gbpages)
 436                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 437         else
 438                 direct_gbpages = 0;
 439 }
 440
 441 #ifdef CONFIG_MEMTEST
 442
 443 static void __init memtest(unsigned long start_phys, unsigned long size,
 444                                  unsigned pattern)
 445 {
 446         unsigned long i;
 447         unsigned long *start;
 448         unsigned long start_bad;
 449         unsigned long last_bad;
 450         unsigned long val;
 451         unsigned long start_phys_aligned;
 452         unsigned long count;
 453         unsigned long incr;
 454
 455         switch (pattern) {
 456         case 0:
 457                 val = 0UL;
 458                 break;
 459         case 1:
 460                 val = -1UL;
 461                 break;
 462         case 2:
 463                 val = 0x5555555555555555UL;
 464                 break;
 465         case 3:
 466                 val = 0xaaaaaaaaaaaaaaaaUL;
 467                 break;
 468         default:
 469                 return;
 470         }
 471
 472         incr = sizeof(unsigned long);
 473         start_phys_aligned = ALIGN(start_phys, incr);
 474         count = (size - (start_phys_aligned - start_phys))/incr;
 475         start = __va(start_phys_aligned);
 476         start_bad = 0;
 477         last_bad = 0;
 478
 479         for (i = 0; i < count; i++)
 480                 start[i] = val;
 481         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 482                 if (*start != val) {
 483                         if (start_phys_aligned == last_bad + incr) {
 484                                 last_bad += incr;
 485                         } else {
 486                                 if (start_bad) {
 487                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 488                                                 val, start_bad, last_bad + incr);
 489                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 490                                 }
 491                                 start_bad = last_bad = start_phys_aligned;
 492                         }
 493                 }
 494         }
 495         if (start_bad) {
 496                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 497                         val, start_bad, last_bad + incr);
 498                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 499         }
 500
 501 }
 502
 503 /* default is disabled */
 504 static int memtest_pattern __initdata;
 505
 506 static int __init parse_memtest(char *arg)
 507 {
 508         if (arg)
 509                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 510         return 0;
 511 }
 512
 513 early_param("memtest", parse_memtest);
 514
 515 static void __init early_memtest(unsigned long start, unsigned long end)
 516 {
 517         u64 t_start, t_size;
 518         unsigned pattern;
 519
 520         if (!memtest_pattern)
 521                 return;
 522
 523         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 524         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 525                 t_start = start;
 526                 t_size = 0;
 527                 while (t_start < end) {
 528                         t_start = find_e820_area_size(t_start, &t_size, 1);
 529
 530                         /* done ? */
 531                         if (t_start >= end)
 532                                 break;
 533                         if (t_start + t_size > end)
 534                                 t_size = end - t_start;
 535
 536                         printk(KERN_CONT "\n  %016llx - %016llx pattern %d",
 537                                 (unsigned long long)t_start,
 538                                 (unsigned long long)t_start + t_size, pattern);
 539
 540                         memtest(t_start, t_size, pattern);
 541
 542                         t_start += t_size;
 543                 }
 544         }
 545         printk(KERN_CONT "\n");
 546 }
 547 #else
 548 static void __init early_memtest(unsigned long start, unsigned long end)
 549 {
 550 }
 551 #endif
 552
 553 /*
 554  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 555  * This runs before bootmem is initialized and gets pages directly from
 556  * the physical memory. To access them they are temporarily mapped.
 557  */
 558 unsigned long __init_refok init_memory_mapping(unsigned long start, unsigned long end)
 559 {
 560         unsigned long next, last_map_addr = end;
 561         unsigned long start_phys = start, end_phys = end;
 562
 563         printk(KERN_INFO "init_memory_mapping\n");
 564
 565         /*
 566          * Find space for the kernel direct mapping tables.
 567          *
 568          * Later we should allocate these tables in the local node of the
 569          * memory mapped. Unfortunately this is done currently before the
 570          * nodes are discovered.
 571          */
 572         if (!after_bootmem) {
 573                 init_gbpages();
 574                 find_early_table_space(end);
 575         }
 576
 577         start = (unsigned long)__va(start);
 578         end = (unsigned long)__va(end);
 579
 580         for (; start < end; start = next) {
 581                 pgd_t *pgd = pgd_offset_k(start);
 582                 unsigned long pud_phys;
 583                 pud_t *pud;
 584
 585                 if (after_bootmem)
 586                         pud = pud_offset(pgd, start & PGDIR_MASK);
 587                 else
 588                         pud = alloc_low_page(&pud_phys);
 589
 590                 next = start + PGDIR_SIZE;
 591                 if (next > end)
 592                         next = end;
 593                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next));
 594                 if (!after_bootmem)
 595                         set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
 596                 unmap_low_page(pud);
 597         }
 598
 599         if (!after_bootmem)
 600                 mmu_cr4_features = read_cr4();
 601         __flush_tlb_all();
 602
 603         if (!after_bootmem)
 604                 reserve_early(table_start << PAGE_SHIFT,
 605                                  table_end << PAGE_SHIFT, "PGTABLE");
 606
 607         if (!after_bootmem)
 608                 early_memtest(start_phys, end_phys);
 609
 610         return last_map_addr;
 611 }
 612
 613 #ifndef CONFIG_NUMA
 614 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 615 {
 616         unsigned long bootmap_size, bootmap;
 617
 618         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 619         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 620                                  PAGE_SIZE);
 621         if (bootmap == -1L)
 622                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 623         /* don't touch min_low_pfn */
 624         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 625                                          0, end_pfn);
 626         e820_register_active_regions(0, start_pfn, end_pfn);
 627         free_bootmem_with_active_regions(0, end_pfn);
 628         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 629         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 630 }
 631
 632 void __init paging_init(void)
 633 {
 634         unsigned long max_zone_pfns[MAX_NR_ZONES];
 635
 636         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 637         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 638         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 639         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 640
 641         memory_present(0, 0, end_pfn);
 642         sparse_init();
 643         free_area_init_nodes(max_zone_pfns);
 644 }
 645 #endif
 646
 647 /*
 648  * Memory hotplug specific functions
 649  */
 650 #ifdef CONFIG_MEMORY_HOTPLUG
 651 /*
 652  * Memory is added always to NORMAL zone. This means you will never get
 653  * additional DMA/DMA32 memory.
 654  */
 655 int arch_add_memory(int nid, u64 start, u64 size)
 656 {
 657         struct pglist_data *pgdat = NODE_DATA(nid);
 658         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 659         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 660         unsigned long nr_pages = size >> PAGE_SHIFT;
 661         int ret;
 662
 663         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 664         if (last_mapped_pfn > max_pfn_mapped)
 665                 max_pfn_mapped = last_mapped_pfn;
 666
 667         ret = __add_pages(zone, start_pfn, nr_pages);
 668         WARN_ON(1);
 669
 670         return ret;
 671 }
 672 EXPORT_SYMBOL_GPL(arch_add_memory);
 673
 674 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 675 int memory_add_physaddr_to_nid(u64 start)
 676 {
 677         return 0;
 678 }
 679 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 680 #endif
 681
 682 #endif /* CONFIG_MEMORY_HOTPLUG */
 683
 684 /*
 685  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 686  * is valid. The argument is a physical page number.
 687  *
 688  *
 689  * On x86, access has to be given to the first megabyte of ram because that area
 690  * contains bios code and data regions used by X and dosemu and similar apps.
 691  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 692  * mmio resources as well as potential bios/acpi data regions.
 693  */
 694 int devmem_is_allowed(unsigned long pagenr)
 695 {
 696         if (pagenr <= 256)
 697                 return 1;
 698         if (!page_is_ram(pagenr))
 699                 return 1;
 700         return 0;
 701 }
 702
 703
 704 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 705                          kcore_modules, kcore_vsyscall;
 706
 707 void __init mem_init(void)
 708 {
 709         long codesize, reservedpages, datasize, initsize;
 710
 711         pci_iommu_alloc();
 712
 713         /* clear_bss() already clear the empty_zero_page */
 714
 715         reservedpages = 0;
 716
 717         /* this will put all low memory onto the freelists */
 718 #ifdef CONFIG_NUMA
 719         totalram_pages = numa_free_all_bootmem();
 720 #else
 721         totalram_pages = free_all_bootmem();
 722 #endif
 723         reservedpages = end_pfn - totalram_pages -
 724                                         absent_pages_in_range(0, end_pfn);
 725         after_bootmem = 1;
 726
 727         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 728         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 729         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 730
 731         /* Register memory areas for /proc/kcore */
 732         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 733         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 734                    VMALLOC_END-VMALLOC_START);
 735         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 736         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 737         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 738                                  VSYSCALL_END - VSYSCALL_START);
 739
 740         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 741                                 "%ldk reserved, %ldk data, %ldk init)\n",
 742                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 743                 end_pfn << (PAGE_SHIFT-10),
 744                 codesize >> 10,
 745                 reservedpages << (PAGE_SHIFT-10),
 746                 datasize >> 10,
 747                 initsize >> 10);
 748
 749         cpa_init();
 750 }
 751
 752 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 753 {
 754         unsigned long addr = begin;
 755
 756         if (addr >= end)
 757                 return;
 758
 759         /*
 760          * If debugging page accesses then do not free this memory but
 761          * mark them not present - any buggy init-section access will
 762          * create a kernel page fault:
 763          */
 764 #ifdef CONFIG_DEBUG_PAGEALLOC
 765         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 766                 begin, PAGE_ALIGN(end));
 767         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 768 #else
 769         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 770
 771         for (; addr < end; addr += PAGE_SIZE) {
 772                 ClearPageReserved(virt_to_page(addr));
 773                 init_page_count(virt_to_page(addr));
 774                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 775                         POISON_FREE_INITMEM, PAGE_SIZE);
 776                 free_page(addr);
 777                 totalram_pages++;
 778         }
 779 #endif
 780 }
 781
 782 void free_initmem(void)
 783 {
 784         free_init_pages("unused kernel memory",
 785                         (unsigned long)(&__init_begin),
 786                         (unsigned long)(&__init_end));
 787 }
 788
 789 #ifdef CONFIG_DEBUG_RODATA
 790 const int rodata_test_data = 0xC3;
 791 EXPORT_SYMBOL_GPL(rodata_test_data);
 792
 793 void mark_rodata_ro(void)
 794 {
 795         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 796
 797         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 798                (end - start) >> 10);
 799         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 800
 801         /*
 802          * The rodata section (but not the kernel text!) should also be
 803          * not-executable.
 804          */
 805         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 806         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 807
 808         rodata_test();
 809
 810 #ifdef CONFIG_CPA_DEBUG
 811         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 812         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 813
 814         printk(KERN_INFO "Testing CPA: again\n");
 815         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 816 #endif
 817 }
 818
 819 #endif
 820
 821 #ifdef CONFIG_BLK_DEV_INITRD
 822 void free_initrd_mem(unsigned long start, unsigned long end)
 823 {
 824         free_init_pages("initrd memory", start, end);
 825 }
 826 #endif
 827
 828 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 829                                    int flags)
 830 {
 831 #ifdef CONFIG_NUMA
 832         int nid, next_nid;
 833 #endif
 834         unsigned long pfn = phys >> PAGE_SHIFT;
 835         int ret;
 836
 837         if (pfn >= end_pfn) {
 838                 /*
 839                  * This can happen with kdump kernels when accessing
 840                  * firmware tables:
 841                  */
 842                 if (pfn < max_pfn_mapped)
 843                         return -EFAULT;
 844
 845                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
 846                                 phys, len);
 847                 return -EFAULT;
 848         }
 849
 850         /* Should check here against the e820 map to avoid double free */
 851 #ifdef CONFIG_NUMA
 852         nid = phys_to_nid(phys);
 853         next_nid = phys_to_nid(phys + len - 1);
 854         if (nid == next_nid)
 855                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 856         else
 857                 ret = reserve_bootmem(phys, len, flags);
 858
 859         if (ret != 0)
 860                 return ret;
 861
 862 #else
 863         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 864 #endif
 865
 866         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 867                 dma_reserve += len / PAGE_SIZE;
 868                 set_dma_reserve(dma_reserve);
 869         }
 870
 871         return 0;
 872 }
 873
 874 int kern_addr_valid(unsigned long addr)
 875 {
 876         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 877         pgd_t *pgd;
 878         pud_t *pud;
 879         pmd_t *pmd;
 880         pte_t *pte;
 881
 882         if (above != 0 && above != -1UL)
 883                 return 0;
 884
 885         pgd = pgd_offset_k(addr);
 886         if (pgd_none(*pgd))
 887                 return 0;
 888
 889         pud = pud_offset(pgd, addr);
 890         if (pud_none(*pud))
 891                 return 0;
 892
 893         pmd = pmd_offset(pud, addr);
 894         if (pmd_none(*pmd))
 895                 return 0;
 896
 897         if (pmd_large(*pmd))
 898                 return pfn_valid(pmd_pfn(*pmd));
 899
 900         pte = pte_offset_kernel(pmd, addr);
 901         if (pte_none(*pte))
 902                 return 0;
 903
 904         return pfn_valid(pte_pfn(*pte));
 905 }
 906
 907 /*
 908  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 909  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 910  * not need special handling anymore:
 911  */
 912 static struct vm_area_struct gate_vma = {
 913         .vm_start       = VSYSCALL_START,
 914         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 915         .vm_page_prot   = PAGE_READONLY_EXEC,
 916         .vm_flags       = VM_READ | VM_EXEC
 917 };
 918
 919 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 920 {
 921 #ifdef CONFIG_IA32_EMULATION
 922         if (test_tsk_thread_flag(tsk, TIF_IA32))
 923                 return NULL;
 924 #endif
 925         return &gate_vma;
 926 }
 927
 928 int in_gate_area(struct task_struct *task, unsigned long addr)
 929 {
 930         struct vm_area_struct *vma = get_gate_vma(task);
 931
 932         if (!vma)
 933                 return 0;
 934
 935         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 936 }
 937
 938 /*
 939  * Use this when you have no reliable task/vma, typically from interrupt
 940  * context. It is less reliable than using the task's vma and may give
 941  * false positives:
 942  */
 943 int in_gate_area_no_task(unsigned long addr)
 944 {
 945         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 946 }
 947
 948 const char *arch_vma_name(struct vm_area_struct *vma)
 949 {
 950         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 951                 return "[vdso]";
 952         if (vma == &gate_vma)
 953                 return "[vsyscall]";
 954         return NULL;
 955 }
 956
 957 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 958 /*
 959  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 960  */
 961 static long __meminitdata addr_start, addr_end;
 962 static void __meminitdata *p_start, *p_end;
 963 static int __meminitdata node_start;
 964
 965 int __meminit
 966 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 967 {
 968         unsigned long addr = (unsigned long)start_page;
 969         unsigned long end = (unsigned long)(start_page + size);
 970         unsigned long next;
 971         pgd_t *pgd;
 972         pud_t *pud;
 973         pmd_t *pmd;
 974
 975         for (; addr < end; addr = next) {
 976                 next = pmd_addr_end(addr, end);
 977
 978                 pgd = vmemmap_pgd_populate(addr, node);
 979                 if (!pgd)
 980                         return -ENOMEM;
 981
 982                 pud = vmemmap_pud_populate(pgd, addr, node);
 983                 if (!pud)
 984                         return -ENOMEM;
 985
 986                 pmd = pmd_offset(pud, addr);
 987                 if (pmd_none(*pmd)) {
 988                         pte_t entry;
 989                         void *p;
 990
 991                         p = vmemmap_alloc_block(PMD_SIZE, node);
 992                         if (!p)
 993                                 return -ENOMEM;
 994
 995                         entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 996                                                         PAGE_KERNEL_LARGE);
 997                         set_pmd(pmd, __pmd(pte_val(entry)));
 998
 999                         /* check to see if we have contiguous blocks */
1000                         if (p_end != p || node_start != node) {
1001                                 if (p_start)
1002                                         printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1003                                                 addr_start, addr_end-1, p_start, p_end-1, node_start);
1004                                 addr_start = addr;
1005                                 node_start = node;
1006                                 p_start = p;
1007                         }
1008                         addr_end = addr + PMD_SIZE;
1009                         p_end = p + PMD_SIZE;
1010                 } else {
1011                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1012                 }
1013         }
1014         return 0;
1015 }
1016
1017 void __meminit vmemmap_populate_print_last(void)
1018 {
1019         if (p_start) {
1020                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1021                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1022                 p_start = NULL;
1023                 p_end = NULL;
1024                 node_start = 0;
1025         }
1026 }
1027 #endif