arch/x86/mm/init_64.c

   1 /*
   2  *  linux/arch/x86_64/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
   6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
   7  */
   8
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/string.h>
  14 #include <linux/types.h>
  15 #include <linux/ptrace.h>
  16 #include <linux/mman.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19 #include <linux/smp.h>
  20 #include <linux/init.h>
  21 #include <linux/initrd.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/proc_fs.h>
  25 #include <linux/pci.h>
  26 #include <linux/pfn.h>
  27 #include <linux/poison.h>
  28 #include <linux/dma-mapping.h>
  29 #include <linux/module.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/nmi.h>
  32
  33 #include <asm/processor.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36 #include <asm/pgtable.h>
  37 #include <asm/pgalloc.h>
  38 #include <asm/dma.h>
  39 #include <asm/fixmap.h>
  40 #include <asm/e820.h>
  41 #include <asm/apic.h>
  42 #include <asm/tlb.h>
  43 #include <asm/mmu_context.h>
  44 #include <asm/proto.h>
  45 #include <asm/smp.h>
  46 #include <asm/sections.h>
  47 #include <asm/kdebug.h>
  48 #include <asm/numa.h>
  49 #include <asm/cacheflush.h>
  50
  51 /*
  52  * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
  53  * The direct mapping extends to max_pfn_mapped, so that we can directly access
  54  * apertures, ACPI and other tables without having to play with fixmaps.
  55  */
  56 unsigned long max_low_pfn_mapped;
  57 unsigned long max_pfn_mapped;
  58
  59 static unsigned long dma_reserve __initdata;
  60
  61 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  62
  63 int direct_gbpages __meminitdata
  64 #ifdef CONFIG_DIRECT_GBPAGES
  65                                 = 1
  66 #endif
  67 ;
  68
  69 static int __init parse_direct_gbpages_off(char *arg)
  70 {
  71         direct_gbpages = 0;
  72         return 0;
  73 }
  74 early_param("nogbpages", parse_direct_gbpages_off);
  75
  76 static int __init parse_direct_gbpages_on(char *arg)
  77 {
  78         direct_gbpages = 1;
  79         return 0;
  80 }
  81 early_param("gbpages", parse_direct_gbpages_on);
  82
  83 /*
  84  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  85  * physical space so we can cache the place of the first one and move
  86  * around without checking the pgd every time.
  87  */
  88
  89 void show_mem(void)
  90 {
  91         long i, total = 0, reserved = 0;
  92         long shared = 0, cached = 0;
  93         struct page *page;
  94         pg_data_t *pgdat;
  95
  96         printk(KERN_INFO "Mem-info:\n");
  97         show_free_areas();
  98         for_each_online_pgdat(pgdat) {
  99                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
 100                         /*
 101                          * This loop can take a while with 256 GB and
 102                          * 4k pages so defer the NMI watchdog:
 103                          */
 104                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
 105                                 touch_nmi_watchdog();
 106
 107                         if (!pfn_valid(pgdat->node_start_pfn + i))
 108                                 continue;
 109
 110                         page = pfn_to_page(pgdat->node_start_pfn + i);
 111                         total++;
 112                         if (PageReserved(page))
 113                                 reserved++;
 114                         else if (PageSwapCache(page))
 115                                 cached++;
 116                         else if (page_count(page))
 117                                 shared += page_count(page) - 1;
 118                 }
 119         }
 120         printk(KERN_INFO "%lu pages of RAM\n",          total);
 121         printk(KERN_INFO "%lu reserved pages\n",        reserved);
 122         printk(KERN_INFO "%lu pages shared\n",          shared);
 123         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 124 }
 125
 126 int after_bootmem;
 127
 128 static __init void *spp_getpage(void)
 129 {
 130         void *ptr;
 131
 132         if (after_bootmem)
 133                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 134         else
 135                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 136
 137         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 138                 panic("set_pte_phys: cannot allocate page data %s\n",
 139                         after_bootmem ? "after bootmem" : "");
 140         }
 141
 142         pr_debug("spp_getpage %p\n", ptr);
 143
 144         return ptr;
 145 }
 146
 147 void
 148 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 149 {
 150         pud_t *pud;
 151         pmd_t *pmd;
 152         pte_t *pte;
 153
 154         pud = pud_page + pud_index(vaddr);
 155         if (pud_none(*pud)) {
 156                 pmd = (pmd_t *) spp_getpage();
 157                 pud_populate(&init_mm, pud, pmd);
 158                 if (pmd != pmd_offset(pud, 0)) {
 159                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 160                                 pmd, pmd_offset(pud, 0));
 161                         return;
 162                 }
 163         }
 164         pmd = pmd_offset(pud, vaddr);
 165         if (pmd_none(*pmd)) {
 166                 pte = (pte_t *) spp_getpage();
 167                 pmd_populate_kernel(&init_mm, pmd, pte);
 168                 if (pte != pte_offset_kernel(pmd, 0)) {
 169                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 170                         return;
 171                 }
 172         }
 173
 174         pte = pte_offset_kernel(pmd, vaddr);
 175         if (!pte_none(*pte) && pte_val(new_pte) &&
 176             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 177                 pte_ERROR(*pte);
 178         set_pte(pte, new_pte);
 179
 180         /*
 181          * It's enough to flush this one mapping.
 182          * (PGE mappings get flushed as well)
 183          */
 184         __flush_tlb_one(vaddr);
 185 }
 186
 187 void
 188 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 189 {
 190         pgd_t *pgd;
 191         pud_t *pud_page;
 192
 193         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 194
 195         pgd = pgd_offset_k(vaddr);
 196         if (pgd_none(*pgd)) {
 197                 printk(KERN_ERR
 198                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 199                 return;
 200         }
 201         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 202         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 203 }
 204
 205 /*
 206  * Create large page table mappings for a range of physical addresses.
 207  */
 208 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 209                                                 pgprot_t prot)
 210 {
 211         pgd_t *pgd;
 212         pud_t *pud;
 213         pmd_t *pmd;
 214
 215         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 216         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 217                 pgd = pgd_offset_k((unsigned long)__va(phys));
 218                 if (pgd_none(*pgd)) {
 219                         pud = (pud_t *) spp_getpage();
 220                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 221                                                 _PAGE_USER));
 222                 }
 223                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 224                 if (pud_none(*pud)) {
 225                         pmd = (pmd_t *) spp_getpage();
 226                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 227                                                 _PAGE_USER));
 228                 }
 229                 pmd = pmd_offset(pud, phys);
 230                 BUG_ON(!pmd_none(*pmd));
 231                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 232         }
 233 }
 234
 235 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 236 {
 237         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 238 }
 239
 240 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 241 {
 242         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 243 }
 244
 245 /*
 246  * The head.S code sets up the kernel high mapping:
 247  *
 248  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 249  *
 250  * phys_addr holds the negative offset to the kernel, which is added
 251  * to the compile time generated pmds. This results in invalid pmds up
 252  * to the point where we hit the physaddr 0 mapping.
 253  *
 254  * We limit the mappings to the region from _text to _end.  _end is
 255  * rounded up to the 2MB boundary. This catches the invalid pmds as
 256  * well, as they are located before _text:
 257  */
 258 void __init cleanup_highmap(void)
 259 {
 260         unsigned long vaddr = __START_KERNEL_map;
 261         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 262         pmd_t *pmd = level2_kernel_pgt;
 263         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 264
 265         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 266                 if (pmd_none(*pmd))
 267                         continue;
 268                 if (vaddr < (unsigned long) _text || vaddr > end)
 269                         set_pmd(pmd, __pmd(0));
 270         }
 271 }
 272
 273 static unsigned long __initdata table_start;
 274 static unsigned long __meminitdata table_end;
 275 static unsigned long __meminitdata table_top;
 276
 277 static __meminit void *alloc_low_page(unsigned long *phys)
 278 {
 279         unsigned long pfn = table_end++;
 280         void *adr;
 281
 282         if (after_bootmem) {
 283                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 284                 *phys = __pa(adr);
 285
 286                 return adr;
 287         }
 288
 289         if (pfn >= table_top)
 290                 panic("alloc_low_page: ran out of memory");
 291
 292         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 293         memset(adr, 0, PAGE_SIZE);
 294         *phys  = pfn * PAGE_SIZE;
 295         return adr;
 296 }
 297
 298 static __meminit void unmap_low_page(void *adr)
 299 {
 300         if (after_bootmem)
 301                 return;
 302
 303         early_iounmap(adr, PAGE_SIZE);
 304 }
 305
 306 static unsigned long __meminit
 307 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end)
 308 {
 309         unsigned pages = 0;
 310         unsigned long last_map_addr = end;
 311         int i;
 312
 313         pte_t *pte = pte_page + pte_index(addr);
 314
 315         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 316
 317                 if (addr >= end) {
 318                         if (!after_bootmem) {
 319                                 for(; i < PTRS_PER_PTE; i++, pte++)
 320                                         set_pte(pte, __pte(0));
 321                         }
 322                         break;
 323                 }
 324
 325                 if (pte_val(*pte))
 326                         continue;
 327
 328                 if (0)
 329                         printk("   pte=%p addr=%lx pte=%016lx\n",
 330                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 331                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL));
 332                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 333                 pages++;
 334         }
 335         update_page_count(PG_LEVEL_4K, pages);
 336
 337         return last_map_addr;
 338 }
 339
 340 static unsigned long __meminit
 341 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end)
 342 {
 343         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 344
 345         return phys_pte_init(pte, address, end);
 346 }
 347
 348 static unsigned long __meminit
 349 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 350                          unsigned long page_size_mask)
 351 {
 352         unsigned long pages = 0;
 353         unsigned long last_map_addr = end;
 354
 355         int i = pmd_index(address);
 356
 357         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 358                 unsigned long pte_phys;
 359                 pmd_t *pmd = pmd_page + pmd_index(address);
 360                 pte_t *pte;
 361
 362                 if (address >= end) {
 363                         if (!after_bootmem) {
 364                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 365                                         set_pmd(pmd, __pmd(0));
 366                         }
 367                         break;
 368                 }
 369
 370                 if (pmd_val(*pmd)) {
 371                         if (!pmd_large(*pmd))
 372                                 last_map_addr = phys_pte_update(pmd, address,
 373                                                                  end);
 374                         continue;
 375                 }
 376
 377                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 378                         pages++;
 379                         set_pte((pte_t *)pmd,
 380                                 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 381                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 382                         continue;
 383                 }
 384
 385                 pte = alloc_low_page(&pte_phys);
 386                 last_map_addr = phys_pte_init(pte, address, end);
 387                 unmap_low_page(pte);
 388
 389                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 390         }
 391         update_page_count(PG_LEVEL_2M, pages);
 392         return last_map_addr;
 393 }
 394
 395 static unsigned long __meminit
 396 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 397                          unsigned long page_size_mask)
 398 {
 399         pmd_t *pmd = pmd_offset(pud, 0);
 400         unsigned long last_map_addr;
 401
 402         spin_lock(&init_mm.page_table_lock);
 403         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask);
 404         spin_unlock(&init_mm.page_table_lock);
 405         __flush_tlb_all();
 406         return last_map_addr;
 407 }
 408
 409 static unsigned long __meminit
 410 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 411                          unsigned long page_size_mask)
 412 {
 413         unsigned long pages = 0;
 414         unsigned long last_map_addr = end;
 415         int i = pud_index(addr);
 416
 417         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 418                 unsigned long pmd_phys;
 419                 pud_t *pud = pud_page + pud_index(addr);
 420                 pmd_t *pmd;
 421
 422                 if (addr >= end)
 423                         break;
 424
 425                 if (!after_bootmem &&
 426                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 427                         set_pud(pud, __pud(0));
 428                         continue;
 429                 }
 430
 431                 if (pud_val(*pud)) {
 432                         if (!pud_large(*pud))
 433                                 last_map_addr = phys_pmd_update(pud, addr, end,
 434                                                          page_size_mask);
 435                         continue;
 436                 }
 437
 438                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 439                         pages++;
 440                         set_pte((pte_t *)pud,
 441                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 442                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 443                         continue;
 444                 }
 445
 446                 pmd = alloc_low_page(&pmd_phys);
 447
 448                 spin_lock(&init_mm.page_table_lock);
 449                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask);
 450                 unmap_low_page(pmd);
 451                 pud_populate(&init_mm, pud, __va(pmd_phys));
 452                 spin_unlock(&init_mm.page_table_lock);
 453
 454         }
 455         __flush_tlb_all();
 456         update_page_count(PG_LEVEL_1G, pages);
 457
 458         return last_map_addr;
 459 }
 460
 461 static unsigned long __meminit
 462 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 463                  unsigned long page_size_mask)
 464 {
 465         pud_t *pud;
 466
 467         pud = (pud_t *)pgd_page_vaddr(*pgd);
 468
 469         return phys_pud_init(pud, addr, end, page_size_mask);
 470 }
 471
 472 static void __init find_early_table_space(unsigned long end)
 473 {
 474         unsigned long puds, pmds, ptes, tables, start;
 475
 476         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 477         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 478         if (direct_gbpages) {
 479                 unsigned long extra;
 480                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 481                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 482         } else
 483                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 484         tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 485
 486         if (cpu_has_pse) {
 487                 unsigned long extra;
 488                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 489                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 490         } else
 491                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 492         tables += round_up(ptes * sizeof(pte_t), PAGE_SIZE);
 493
 494         /*
 495          * RED-PEN putting page tables only on node 0 could
 496          * cause a hotspot and fill up ZONE_DMA. The page tables
 497          * need roughly 0.5KB per GB.
 498          */
 499         start = 0x8000;
 500         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 501         if (table_start == -1UL)
 502                 panic("Cannot find space for the kernel page tables");
 503
 504         table_start >>= PAGE_SHIFT;
 505         table_end = table_start;
 506         table_top = table_start + (tables >> PAGE_SHIFT);
 507
 508         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 509                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 510 }
 511
 512 static void __init init_gbpages(void)
 513 {
 514         if (direct_gbpages && cpu_has_gbpages)
 515                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 516         else
 517                 direct_gbpages = 0;
 518 }
 519
 520 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 521                                                 unsigned long end,
 522                                                 unsigned long page_size_mask)
 523 {
 524
 525         unsigned long next, last_map_addr = end;
 526
 527         start = (unsigned long)__va(start);
 528         end = (unsigned long)__va(end);
 529
 530         for (; start < end; start = next) {
 531                 pgd_t *pgd = pgd_offset_k(start);
 532                 unsigned long pud_phys;
 533                 pud_t *pud;
 534
 535                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 536                 if (next > end)
 537                         next = end;
 538
 539                 if (pgd_val(*pgd)) {
 540                         last_map_addr = phys_pud_update(pgd, __pa(start),
 541                                                  __pa(end), page_size_mask);
 542                         continue;
 543                 }
 544
 545                 if (after_bootmem)
 546                         pud = pud_offset(pgd, start & PGDIR_MASK);
 547                 else
 548                         pud = alloc_low_page(&pud_phys);
 549
 550                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 551                                                  page_size_mask);
 552                 unmap_low_page(pud);
 553                 pgd_populate(&init_mm, pgd_offset_k(start),
 554                              __va(pud_phys));
 555         }
 556
 557         return last_map_addr;
 558 }
 559
 560 struct map_range {
 561         unsigned long start;
 562         unsigned long end;
 563         unsigned page_size_mask;
 564 };
 565
 566 #define NR_RANGE_MR 5
 567
 568 static int save_mr(struct map_range *mr, int nr_range,
 569                    unsigned long start_pfn, unsigned long end_pfn,
 570                    unsigned long page_size_mask)
 571 {
 572
 573         if (start_pfn < end_pfn) {
 574                 if (nr_range >= NR_RANGE_MR)
 575                         panic("run out of range for init_memory_mapping\n");
 576                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 577                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 578                 mr[nr_range].page_size_mask = page_size_mask;
 579                 nr_range++;
 580         }
 581
 582         return nr_range;
 583 }
 584
 585 /*
 586  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 587  * This runs before bootmem is initialized and gets pages directly from
 588  * the physical memory. To access them they are temporarily mapped.
 589  */
 590 unsigned long __init_refok init_memory_mapping(unsigned long start,
 591                                                unsigned long end)
 592 {
 593         unsigned long last_map_addr = 0;
 594         unsigned long page_size_mask = 0;
 595         unsigned long start_pfn, end_pfn;
 596
 597         struct map_range mr[NR_RANGE_MR];
 598         int nr_range, i;
 599
 600         printk(KERN_INFO "init_memory_mapping\n");
 601
 602         /*
 603          * Find space for the kernel direct mapping tables.
 604          *
 605          * Later we should allocate these tables in the local node of the
 606          * memory mapped. Unfortunately this is done currently before the
 607          * nodes are discovered.
 608          */
 609         if (!after_bootmem)
 610                 init_gbpages();
 611
 612         if (direct_gbpages)
 613                 page_size_mask |= 1 << PG_LEVEL_1G;
 614         if (cpu_has_pse)
 615                 page_size_mask |= 1 << PG_LEVEL_2M;
 616
 617         memset(mr, 0, sizeof(mr));
 618         nr_range = 0;
 619
 620         /* head if not big page alignment ?*/
 621         start_pfn = start >> PAGE_SHIFT;
 622         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 623                         << (PMD_SHIFT - PAGE_SHIFT);
 624         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 625
 626         /* big page (2M) range*/
 627         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 628                          << (PMD_SHIFT - PAGE_SHIFT);
 629         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 630                          << (PUD_SHIFT - PAGE_SHIFT);
 631         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 632                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 633         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 634                         page_size_mask & (1<<PG_LEVEL_2M));
 635
 636         /* big page (1G) range */
 637         start_pfn = end_pfn;
 638         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 639         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 640                                 page_size_mask &
 641                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 642
 643         /* tail is not big page (1G) alignment */
 644         start_pfn = end_pfn;
 645         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 646         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 647                         page_size_mask & (1<<PG_LEVEL_2M));
 648
 649         /* tail is not big page (2M) alignment */
 650         start_pfn = end_pfn;
 651         end_pfn = end>>PAGE_SHIFT;
 652         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 653
 654         /* try to merge same page size and continuous */
 655         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 656                 unsigned long old_start;
 657                 if (mr[i].end != mr[i+1].start ||
 658                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 659                         continue;
 660                 /* move it */
 661                 old_start = mr[i].start;
 662                 memmove(&mr[i], &mr[i+1],
 663                          (nr_range - 1 - i) * sizeof (struct map_range));
 664                 mr[i].start = old_start;
 665                 nr_range--;
 666         }
 667
 668         for (i = 0; i < nr_range; i++)
 669                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 670                                 mr[i].start, mr[i].end,
 671                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 672                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 673
 674         if (!after_bootmem)
 675                 find_early_table_space(end);
 676
 677         for (i = 0; i < nr_range; i++)
 678                 last_map_addr = kernel_physical_mapping_init(
 679                                         mr[i].start, mr[i].end,
 680                                         mr[i].page_size_mask);
 681
 682         if (!after_bootmem)
 683                 mmu_cr4_features = read_cr4();
 684         __flush_tlb_all();
 685
 686         if (!after_bootmem && table_end > table_start)
 687                 reserve_early(table_start << PAGE_SHIFT,
 688                                  table_end << PAGE_SHIFT, "PGTABLE");
 689
 690         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 691                          last_map_addr, end);
 692
 693         if (!after_bootmem)
 694                 early_memtest(start, end);
 695
 696         return last_map_addr >> PAGE_SHIFT;
 697 }
 698
 699 #ifndef CONFIG_NUMA
 700 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 701 {
 702         unsigned long bootmap_size, bootmap;
 703
 704         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 705         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 706                                  PAGE_SIZE);
 707         if (bootmap == -1L)
 708                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 709         /* don't touch min_low_pfn */
 710         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 711                                          0, end_pfn);
 712         e820_register_active_regions(0, start_pfn, end_pfn);
 713         free_bootmem_with_active_regions(0, end_pfn);
 714         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 715         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 716 }
 717
 718 void __init paging_init(void)
 719 {
 720         unsigned long max_zone_pfns[MAX_NR_ZONES];
 721
 722         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 723         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 724         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 725         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 726
 727         memory_present(0, 0, max_pfn);
 728         sparse_init();
 729         free_area_init_nodes(max_zone_pfns);
 730 }
 731 #endif
 732
 733 /*
 734  * Memory hotplug specific functions
 735  */
 736 #ifdef CONFIG_MEMORY_HOTPLUG
 737 /*
 738  * Memory is added always to NORMAL zone. This means you will never get
 739  * additional DMA/DMA32 memory.
 740  */
 741 int arch_add_memory(int nid, u64 start, u64 size)
 742 {
 743         struct pglist_data *pgdat = NODE_DATA(nid);
 744         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 745         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 746         unsigned long nr_pages = size >> PAGE_SHIFT;
 747         int ret;
 748
 749         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 750         if (last_mapped_pfn > max_pfn_mapped)
 751                 max_pfn_mapped = last_mapped_pfn;
 752
 753         ret = __add_pages(zone, start_pfn, nr_pages);
 754         WARN_ON(1);
 755
 756         return ret;
 757 }
 758 EXPORT_SYMBOL_GPL(arch_add_memory);
 759
 760 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 761 int memory_add_physaddr_to_nid(u64 start)
 762 {
 763         return 0;
 764 }
 765 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 766 #endif
 767
 768 #endif /* CONFIG_MEMORY_HOTPLUG */
 769
 770 /*
 771  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 772  * is valid. The argument is a physical page number.
 773  *
 774  *
 775  * On x86, access has to be given to the first megabyte of ram because that area
 776  * contains bios code and data regions used by X and dosemu and similar apps.
 777  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 778  * mmio resources as well as potential bios/acpi data regions.
 779  */
 780 int devmem_is_allowed(unsigned long pagenr)
 781 {
 782         if (pagenr <= 256)
 783                 return 1;
 784         if (!page_is_ram(pagenr))
 785                 return 1;
 786         return 0;
 787 }
 788
 789
 790 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 791                          kcore_modules, kcore_vsyscall;
 792
 793 void __init mem_init(void)
 794 {
 795         long codesize, reservedpages, datasize, initsize;
 796
 797         pci_iommu_alloc();
 798
 799         /* clear_bss() already clear the empty_zero_page */
 800
 801         reservedpages = 0;
 802
 803         /* this will put all low memory onto the freelists */
 804 #ifdef CONFIG_NUMA
 805         totalram_pages = numa_free_all_bootmem();
 806 #else
 807         totalram_pages = free_all_bootmem();
 808 #endif
 809         reservedpages = max_pfn - totalram_pages -
 810                                         absent_pages_in_range(0, max_pfn);
 811         after_bootmem = 1;
 812
 813         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 814         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 815         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 816
 817         /* Register memory areas for /proc/kcore */
 818         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 819         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 820                    VMALLOC_END-VMALLOC_START);
 821         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 822         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 823         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 824                                  VSYSCALL_END - VSYSCALL_START);
 825
 826         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 827                                 "%ldk reserved, %ldk data, %ldk init)\n",
 828                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 829                 max_pfn << (PAGE_SHIFT-10),
 830                 codesize >> 10,
 831                 reservedpages << (PAGE_SHIFT-10),
 832                 datasize >> 10,
 833                 initsize >> 10);
 834
 835         cpa_init();
 836 }
 837
 838 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 839 {
 840         unsigned long addr = begin;
 841
 842         if (addr >= end)
 843                 return;
 844
 845         /*
 846          * If debugging page accesses then do not free this memory but
 847          * mark them not present - any buggy init-section access will
 848          * create a kernel page fault:
 849          */
 850 #ifdef CONFIG_DEBUG_PAGEALLOC
 851         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 852                 begin, PAGE_ALIGN(end));
 853         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 854 #else
 855         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 856
 857         for (; addr < end; addr += PAGE_SIZE) {
 858                 ClearPageReserved(virt_to_page(addr));
 859                 init_page_count(virt_to_page(addr));
 860                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 861                         POISON_FREE_INITMEM, PAGE_SIZE);
 862                 free_page(addr);
 863                 totalram_pages++;
 864         }
 865 #endif
 866 }
 867
 868 void free_initmem(void)
 869 {
 870         free_init_pages("unused kernel memory",
 871                         (unsigned long)(&__init_begin),
 872                         (unsigned long)(&__init_end));
 873 }
 874
 875 #ifdef CONFIG_DEBUG_RODATA
 876 const int rodata_test_data = 0xC3;
 877 EXPORT_SYMBOL_GPL(rodata_test_data);
 878
 879 void mark_rodata_ro(void)
 880 {
 881         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 882         unsigned long rodata_start =
 883                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 884
 885 #ifdef CONFIG_DYNAMIC_FTRACE
 886         /* Dynamic tracing modifies the kernel text section */
 887         start = rodata_start;
 888 #endif
 889
 890         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 891                (end - start) >> 10);
 892         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 893
 894         /*
 895          * The rodata section (but not the kernel text!) should also be
 896          * not-executable.
 897          */
 898         set_memory_nx(rodata_start, (end - rodata_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