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 void __meminit
 306 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end)
 307 {
 308         unsigned pages = 0;
 309         int i;
 310         pte_t *pte = pte_page + pte_index(addr);
 311
 312         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 313
 314                 if (addr >= end) {
 315                         if (!after_bootmem) {
 316                                 for(; i < PTRS_PER_PTE; i++, pte++)
 317                                         set_pte(pte, __pte(0));
 318                         }
 319                         break;
 320                 }
 321
 322                 if (pte_val(*pte))
 323                         continue;
 324
 325                 if (0)
 326                         printk("   pte=%p addr=%lx pte=%016lx\n",
 327                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 328                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL));
 329                 pages++;
 330         }
 331         update_page_count(PG_LEVEL_4K, pages);
 332 }
 333
 334 static void __meminit
 335 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end)
 336 {
 337         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 338
 339         phys_pte_init(pte, address, end);
 340 }
 341
 342 static unsigned long __meminit
 343 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 344                          unsigned long page_size_mask)
 345 {
 346         unsigned long pages = 0;
 347
 348         int i = pmd_index(address);
 349
 350         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 351                 unsigned long pte_phys;
 352                 pmd_t *pmd = pmd_page + pmd_index(address);
 353                 pte_t *pte;
 354
 355                 if (address >= end) {
 356                         if (!after_bootmem) {
 357                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 358                                         set_pmd(pmd, __pmd(0));
 359                         }
 360                         break;
 361                 }
 362
 363                 if (pmd_val(*pmd)) {
 364                         if (!pmd_large(*pmd))
 365                                 phys_pte_update(pmd, address, end);
 366                         continue;
 367                 }
 368
 369                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 370                         pages++;
 371                         set_pte((pte_t *)pmd,
 372                                 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 373                         continue;
 374                 }
 375
 376                 pte = alloc_low_page(&pte_phys);
 377                 phys_pte_init(pte, address, end);
 378                 unmap_low_page(pte);
 379
 380                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 381         }
 382         update_page_count(PG_LEVEL_2M, pages);
 383         return address;
 384 }
 385
 386 static unsigned long __meminit
 387 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 388                          unsigned long page_size_mask)
 389 {
 390         pmd_t *pmd = pmd_offset(pud, 0);
 391         unsigned long last_map_addr;
 392
 393         spin_lock(&init_mm.page_table_lock);
 394         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask);
 395         spin_unlock(&init_mm.page_table_lock);
 396         __flush_tlb_all();
 397         return last_map_addr;
 398 }
 399
 400 static unsigned long __meminit
 401 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 402                          unsigned long page_size_mask)
 403 {
 404         unsigned long pages = 0;
 405         unsigned long last_map_addr = end;
 406         int i = pud_index(addr);
 407
 408         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 409                 unsigned long pmd_phys;
 410                 pud_t *pud = pud_page + pud_index(addr);
 411                 pmd_t *pmd;
 412
 413                 if (addr >= end)
 414                         break;
 415
 416                 if (!after_bootmem &&
 417                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 418                         set_pud(pud, __pud(0));
 419                         continue;
 420                 }
 421
 422                 if (pud_val(*pud)) {
 423                         if (!pud_large(*pud))
 424                                 last_map_addr = phys_pmd_update(pud, addr, end,
 425                                                          page_size_mask);
 426                         continue;
 427                 }
 428
 429                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 430                         pages++;
 431                         set_pte((pte_t *)pud,
 432                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 433                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 434                         continue;
 435                 }
 436
 437                 pmd = alloc_low_page(&pmd_phys);
 438
 439                 spin_lock(&init_mm.page_table_lock);
 440                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask);
 441                 unmap_low_page(pmd);
 442                 pud_populate(&init_mm, pud, __va(pmd_phys));
 443                 spin_unlock(&init_mm.page_table_lock);
 444
 445         }
 446         __flush_tlb_all();
 447         update_page_count(PG_LEVEL_1G, pages);
 448
 449         return last_map_addr;
 450 }
 451
 452 static unsigned long __meminit
 453 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 454                  unsigned long page_size_mask)
 455 {
 456         pud_t *pud;
 457
 458         pud = (pud_t *)pgd_page_vaddr(*pgd);
 459
 460         return phys_pud_init(pud, addr, end, page_size_mask);
 461 }
 462
 463 static void __init find_early_table_space(unsigned long end)
 464 {
 465         unsigned long puds, tables, start;
 466
 467         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 468         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 469         if (!direct_gbpages) {
 470                 unsigned long pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 471                 tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 472         }
 473         if (!cpu_has_pse) {
 474                 unsigned long ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 475                 tables += round_up(ptes * sizeof(pte_t), PAGE_SIZE);
 476         }
 477
 478         /*
 479          * RED-PEN putting page tables only on node 0 could
 480          * cause a hotspot and fill up ZONE_DMA. The page tables
 481          * need roughly 0.5KB per GB.
 482          */
 483         start = 0x8000;
 484         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 485         if (table_start == -1UL)
 486                 panic("Cannot find space for the kernel page tables");
 487
 488         table_start >>= PAGE_SHIFT;
 489         table_end = table_start;
 490         table_top = table_start + (tables >> PAGE_SHIFT);
 491
 492         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 493                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 494 }
 495
 496 static void __init init_gbpages(void)
 497 {
 498         if (direct_gbpages && cpu_has_gbpages)
 499                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 500         else
 501                 direct_gbpages = 0;
 502 }
 503
 504 #ifdef CONFIG_MEMTEST
 505
 506 static void __init memtest(unsigned long start_phys, unsigned long size,
 507                                  unsigned pattern)
 508 {
 509         unsigned long i;
 510         unsigned long *start;
 511         unsigned long start_bad;
 512         unsigned long last_bad;
 513         unsigned long val;
 514         unsigned long start_phys_aligned;
 515         unsigned long count;
 516         unsigned long incr;
 517
 518         switch (pattern) {
 519         case 0:
 520                 val = 0UL;
 521                 break;
 522         case 1:
 523                 val = -1UL;
 524                 break;
 525         case 2:
 526                 val = 0x5555555555555555UL;
 527                 break;
 528         case 3:
 529                 val = 0xaaaaaaaaaaaaaaaaUL;
 530                 break;
 531         default:
 532                 return;
 533         }
 534
 535         incr = sizeof(unsigned long);
 536         start_phys_aligned = ALIGN(start_phys, incr);
 537         count = (size - (start_phys_aligned - start_phys))/incr;
 538         start = __va(start_phys_aligned);
 539         start_bad = 0;
 540         last_bad = 0;
 541
 542         for (i = 0; i < count; i++)
 543                 start[i] = val;
 544         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 545                 if (*start != val) {
 546                         if (start_phys_aligned == last_bad + incr) {
 547                                 last_bad += incr;
 548                         } else {
 549                                 if (start_bad) {
 550                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 551                                                 val, start_bad, last_bad + incr);
 552                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 553                                 }
 554                                 start_bad = last_bad = start_phys_aligned;
 555                         }
 556                 }
 557         }
 558         if (start_bad) {
 559                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 560                         val, start_bad, last_bad + incr);
 561                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 562         }
 563
 564 }
 565
 566 /* default is disabled */
 567 static int memtest_pattern __initdata;
 568
 569 static int __init parse_memtest(char *arg)
 570 {
 571         if (arg)
 572                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 573         return 0;
 574 }
 575
 576 early_param("memtest", parse_memtest);
 577
 578 static void __init early_memtest(unsigned long start, unsigned long end)
 579 {
 580         u64 t_start, t_size;
 581         unsigned pattern;
 582
 583         if (!memtest_pattern)
 584                 return;
 585
 586         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 587         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 588                 t_start = start;
 589                 t_size = 0;
 590                 while (t_start < end) {
 591                         t_start = find_e820_area_size(t_start, &t_size, 1);
 592
 593                         /* done ? */
 594                         if (t_start >= end)
 595                                 break;
 596                         if (t_start + t_size > end)
 597                                 t_size = end - t_start;
 598
 599                         printk(KERN_CONT "\n  %016llx - %016llx pattern %d",
 600                                 (unsigned long long)t_start,
 601                                 (unsigned long long)t_start + t_size, pattern);
 602
 603                         memtest(t_start, t_size, pattern);
 604
 605                         t_start += t_size;
 606                 }
 607         }
 608         printk(KERN_CONT "\n");
 609 }
 610 #else
 611 static void __init early_memtest(unsigned long start, unsigned long end)
 612 {
 613 }
 614 #endif
 615
 616 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 617                                                 unsigned long end,
 618                                                 unsigned long page_size_mask)
 619 {
 620
 621         unsigned long next, last_map_addr = end;
 622
 623         start = (unsigned long)__va(start);
 624         end = (unsigned long)__va(end);
 625
 626         for (; start < end; start = next) {
 627                 pgd_t *pgd = pgd_offset_k(start);
 628                 unsigned long pud_phys;
 629                 pud_t *pud;
 630
 631                 next = start + PGDIR_SIZE;
 632                 if (next > end)
 633                         next = end;
 634
 635                 if (pgd_val(*pgd)) {
 636                         last_map_addr = phys_pud_update(pgd, __pa(start),
 637                                                  __pa(end), page_size_mask);
 638                         continue;
 639                 }
 640
 641                 if (after_bootmem)
 642                         pud = pud_offset(pgd, start & PGDIR_MASK);
 643                 else
 644                         pud = alloc_low_page(&pud_phys);
 645
 646                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 647                                                  page_size_mask);
 648                 unmap_low_page(pud);
 649                 pgd_populate(&init_mm, pgd_offset_k(start),
 650                              __va(pud_phys));
 651         }
 652
 653         return last_map_addr;
 654 }
 655 /*
 656  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 657  * This runs before bootmem is initialized and gets pages directly from
 658  * the physical memory. To access them they are temporarily mapped.
 659  */
 660 unsigned long __init_refok init_memory_mapping(unsigned long start,
 661                                                unsigned long end)
 662 {
 663         unsigned long last_map_addr;
 664         unsigned long page_size_mask = 0;
 665
 666         printk(KERN_INFO "init_memory_mapping\n");
 667
 668         /*
 669          * Find space for the kernel direct mapping tables.
 670          *
 671          * Later we should allocate these tables in the local node of the
 672          * memory mapped. Unfortunately this is done currently before the
 673          * nodes are discovered.
 674          */
 675         if (!after_bootmem) {
 676                 init_gbpages();
 677                 find_early_table_space(end);
 678         }
 679
 680         if (direct_gbpages)
 681                 page_size_mask |= 1 << PG_LEVEL_1G;
 682         if (cpu_has_pse)
 683                 page_size_mask |= 1 << PG_LEVEL_2M;
 684
 685         last_map_addr = kernel_physical_mapping_init(start, end,
 686                                                          page_size_mask);
 687
 688         if (!after_bootmem)
 689                 mmu_cr4_features = read_cr4();
 690         __flush_tlb_all();
 691
 692         if (!after_bootmem && table_end > table_start)
 693                 reserve_early(table_start << PAGE_SHIFT,
 694                                  table_end << PAGE_SHIFT, "PGTABLE");
 695
 696         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 697                          last_map_addr, end);
 698
 699         if (!after_bootmem)
 700                 early_memtest(start, end);
 701
 702         return last_map_addr >> PAGE_SHIFT;
 703 }
 704
 705 #ifndef CONFIG_NUMA
 706 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 707 {
 708         unsigned long bootmap_size, bootmap;
 709
 710         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 711         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 712                                  PAGE_SIZE);
 713         if (bootmap == -1L)
 714                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 715         /* don't touch min_low_pfn */
 716         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 717                                          0, end_pfn);
 718         e820_register_active_regions(0, start_pfn, end_pfn);
 719         free_bootmem_with_active_regions(0, end_pfn);
 720         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 721         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 722 }
 723
 724 void __init paging_init(void)
 725 {
 726         unsigned long max_zone_pfns[MAX_NR_ZONES];
 727
 728         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 729         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 730         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 731         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 732
 733         memory_present(0, 0, max_pfn);
 734         sparse_init();
 735         free_area_init_nodes(max_zone_pfns);
 736 }
 737 #endif
 738
 739 /*
 740  * Memory hotplug specific functions
 741  */
 742 #ifdef CONFIG_MEMORY_HOTPLUG
 743 /*
 744  * Memory is added always to NORMAL zone. This means you will never get
 745  * additional DMA/DMA32 memory.
 746  */
 747 int arch_add_memory(int nid, u64 start, u64 size)
 748 {
 749         struct pglist_data *pgdat = NODE_DATA(nid);
 750         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 751         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 752         unsigned long nr_pages = size >> PAGE_SHIFT;
 753         int ret;
 754
 755         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 756         if (last_mapped_pfn > max_pfn_mapped)
 757                 max_pfn_mapped = last_mapped_pfn;
 758
 759         ret = __add_pages(zone, start_pfn, nr_pages);
 760         WARN_ON(1);
 761
 762         return ret;
 763 }
 764 EXPORT_SYMBOL_GPL(arch_add_memory);
 765
 766 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 767 int memory_add_physaddr_to_nid(u64 start)
 768 {
 769         return 0;
 770 }
 771 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 772 #endif
 773
 774 #endif /* CONFIG_MEMORY_HOTPLUG */
 775
 776 /*
 777  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 778  * is valid. The argument is a physical page number.
 779  *
 780  *
 781  * On x86, access has to be given to the first megabyte of ram because that area
 782  * contains bios code and data regions used by X and dosemu and similar apps.
 783  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 784  * mmio resources as well as potential bios/acpi data regions.
 785  */
 786 int devmem_is_allowed(unsigned long pagenr)
 787 {
 788         if (pagenr <= 256)
 789                 return 1;
 790         if (!page_is_ram(pagenr))
 791                 return 1;
 792         return 0;
 793 }
 794
 795
 796 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 797                          kcore_modules, kcore_vsyscall;
 798
 799 void __init mem_init(void)
 800 {
 801         long codesize, reservedpages, datasize, initsize;
 802
 803         pci_iommu_alloc();
 804
 805         /* clear_bss() already clear the empty_zero_page */
 806
 807         reservedpages = 0;
 808
 809         /* this will put all low memory onto the freelists */
 810 #ifdef CONFIG_NUMA
 811         totalram_pages = numa_free_all_bootmem();
 812 #else
 813         totalram_pages = free_all_bootmem();
 814 #endif
 815         reservedpages = max_pfn - totalram_pages -
 816                                         absent_pages_in_range(0, max_pfn);
 817         after_bootmem = 1;
 818
 819         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 820         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 821         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 822
 823         /* Register memory areas for /proc/kcore */
 824         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 825         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 826                    VMALLOC_END-VMALLOC_START);
 827         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 828         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 829         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 830                                  VSYSCALL_END - VSYSCALL_START);
 831
 832         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 833                                 "%ldk reserved, %ldk data, %ldk init)\n",
 834                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 835                 max_pfn << (PAGE_SHIFT-10),
 836                 codesize >> 10,
 837                 reservedpages << (PAGE_SHIFT-10),
 838                 datasize >> 10,
 839                 initsize >> 10);
 840
 841         cpa_init();
 842 }
 843
 844 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 845 {
 846         unsigned long addr = begin;
 847
 848         if (addr >= end)
 849                 return;
 850
 851         /*
 852          * If debugging page accesses then do not free this memory but
 853          * mark them not present - any buggy init-section access will
 854          * create a kernel page fault:
 855          */
 856 #ifdef CONFIG_DEBUG_PAGEALLOC
 857         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 858                 begin, PAGE_ALIGN(end));
 859         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 860 #else
 861         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 862
 863         for (; addr < end; addr += PAGE_SIZE) {
 864                 ClearPageReserved(virt_to_page(addr));
 865                 init_page_count(virt_to_page(addr));
 866                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 867                         POISON_FREE_INITMEM, PAGE_SIZE);
 868                 free_page(addr);
 869                 totalram_pages++;
 870         }
 871 #endif
 872 }
 873
 874 void free_initmem(void)
 875 {
 876         free_init_pages("unused kernel memory",
 877                         (unsigned long)(&__init_begin),
 878                         (unsigned long)(&__init_end));
 879 }
 880
 881 #ifdef CONFIG_DEBUG_RODATA
 882 const int rodata_test_data = 0xC3;
 883 EXPORT_SYMBOL_GPL(rodata_test_data);
 884
 885 void mark_rodata_ro(void)
 886 {
 887         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 888
 889         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 890                (end - start) >> 10);
 891         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 892
 893         /*
 894          * The rodata section (but not the kernel text!) should also be
 895          * not-executable.
 896          */
 897         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 898         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 899
 900         rodata_test();
 901
 902 #ifdef CONFIG_CPA_DEBUG
 903         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 904         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 905
 906         printk(KERN_INFO "Testing CPA: again\n");
 907         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 908 #endif
 909 }
 910
 911 #endif
 912
 913 #ifdef CONFIG_BLK_DEV_INITRD
 914 void free_initrd_mem(unsigned long start, unsigned long end)
 915 {
 916         free_init_pages("initrd memory", start, end);
 917 }
 918 #endif
 919
 920 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 921                                    int flags)
 922 {
 923 #ifdef CONFIG_NUMA
 924         int nid, next_nid;
 925         int ret;
 926 #endif
 927         unsigned long pfn = phys >> PAGE_SHIFT;
 928
 929         if (pfn >= max_pfn) {
 930                 /*
 931                  * This can happen with kdump kernels when accessing
 932                  * firmware tables:
 933                  */
 934                 if (pfn < max_pfn_mapped)
 935                         return -EFAULT;
 936
 937                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 938                                 phys, len);
 939                 return -EFAULT;
 940         }
 941
 942         /* Should check here against the e820 map to avoid double free */
 943 #ifdef CONFIG_NUMA
 944         nid = phys_to_nid(phys);
 945         next_nid = phys_to_nid(phys + len - 1);
 946         if (nid == next_nid)
 947                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 948         else
 949                 ret = reserve_bootmem(phys, len, flags);
 950
 951         if (ret != 0)
 952                 return ret;
 953
 954 #else
 955         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 956 #endif
 957
 958         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 959                 dma_reserve += len / PAGE_SIZE;
 960                 set_dma_reserve(dma_reserve);
 961         }
 962
 963         return 0;
 964 }
 965
 966 int kern_addr_valid(unsigned long addr)
 967 {
 968         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 969         pgd_t *pgd;
 970         pud_t *pud;
 971         pmd_t *pmd;
 972         pte_t *pte;
 973
 974         if (above != 0 && above != -1UL)
 975                 return 0;
 976
 977         pgd = pgd_offset_k(addr);
 978         if (pgd_none(*pgd))
 979                 return 0;
 980
 981         pud = pud_offset(pgd, addr);
 982         if (pud_none(*pud))
 983                 return 0;
 984
 985         pmd = pmd_offset(pud, addr);
 986         if (pmd_none(*pmd))
 987                 return 0;
 988
 989         if (pmd_large(*pmd))
 990                 return pfn_valid(pmd_pfn(*pmd));
 991
 992         pte = pte_offset_kernel(pmd, addr);
 993         if (pte_none(*pte))
 994                 return 0;
 995
 996         return pfn_valid(pte_pfn(*pte));
 997 }
 998
 999 /*
1000  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1001  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1002  * not need special handling anymore:
1003  */
1004 static struct vm_area_struct gate_vma = {
1005         .vm_start       = VSYSCALL_START,
1006         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1007         .vm_page_prot   = PAGE_READONLY_EXEC,
1008         .vm_flags       = VM_READ | VM_EXEC
1009 };
1010
1011 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1012 {
1013 #ifdef CONFIG_IA32_EMULATION
1014         if (test_tsk_thread_flag(tsk, TIF_IA32))
1015                 return NULL;
1016 #endif
1017         return &gate_vma;
1018 }
1019
1020 int in_gate_area(struct task_struct *task, unsigned long addr)
1021 {
1022         struct vm_area_struct *vma = get_gate_vma(task);
1023
1024         if (!vma)
1025                 return 0;
1026
1027         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1028 }
1029
1030 /*
1031  * Use this when you have no reliable task/vma, typically from interrupt
1032  * context. It is less reliable than using the task's vma and may give
1033  * false positives:
1034  */
1035 int in_gate_area_no_task(unsigned long addr)
1036 {
1037         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1038 }
1039
1040 const char *arch_vma_name(struct vm_area_struct *vma)
1041 {
1042         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1043                 return "[vdso]";
1044         if (vma == &gate_vma)
1045                 return "[vsyscall]";
1046         return NULL;
1047 }
1048
1049 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1050 /*
1051  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1052  */
1053 static long __meminitdata addr_start, addr_end;
1054 static void __meminitdata *p_start, *p_end;
1055 static int __meminitdata node_start;
1056
1057 int __meminit
1058 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1059 {
1060         unsigned long addr = (unsigned long)start_page;
1061         unsigned long end = (unsigned long)(start_page + size);
1062         unsigned long next;
1063         pgd_t *pgd;
1064         pud_t *pud;
1065         pmd_t *pmd;
1066
1067         for (; addr < end; addr = next) {
1068                 void *p = NULL;
1069
1070                 pgd = vmemmap_pgd_populate(addr, node);
1071                 if (!pgd)
1072                         return -ENOMEM;
1073
1074                 pud = vmemmap_pud_populate(pgd, addr, node);
1075                 if (!pud)
1076                         return -ENOMEM;
1077
1078                 if (!cpu_has_pse) {
1079                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1080                         pmd = vmemmap_pmd_populate(pud, addr, node);
1081
1082                         if (!pmd)
1083                                 return -ENOMEM;
1084
1085                         p = vmemmap_pte_populate(pmd, addr, node);
1086
1087                         if (!p)
1088                                 return -ENOMEM;
1089
1090                         addr_end = addr + PAGE_SIZE;
1091                         p_end = p + PAGE_SIZE;
1092                 } else {
1093                         next = pmd_addr_end(addr, end);
1094
1095                         pmd = pmd_offset(pud, addr);
1096                         if (pmd_none(*pmd)) {
1097                                 pte_t entry;
1098
1099                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1100                                 if (!p)
1101                                         return -ENOMEM;
1102
1103                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1104                                                 PAGE_KERNEL_LARGE);
1105                                 set_pmd(pmd, __pmd(pte_val(entry)));
1106
1107                                 /* check to see if we have contiguous blocks */
1108                                 if (p_end != p || node_start != node) {
1109                                         if (p_start)
1110                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1111                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1112                                         addr_start = addr;
1113                                         node_start = node;
1114                                         p_start = p;
1115                                 }
1116
1117                                 addr_end = addr + PMD_SIZE;
1118                                 p_end = p + PMD_SIZE;
1119                         } else
1120                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1121                 }
1122
1123         }
1124         return 0;
1125 }
1126
1127 void __meminit vmemmap_populate_print_last(void)
1128 {
1129         if (p_start) {
1130                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1131                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1132                 p_start = NULL;
1133                 p_end = NULL;
1134                 node_start = 0;
1135         }
1136 }
1137 #endif