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