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
  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 int after_bootmem;
  90
  91 unsigned long __supported_pte_mask __read_mostly = ~0UL;
  92 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  93
  94 static int do_not_nx __cpuinitdata;
  95
  96 /*
  97  * noexec=on|off
  98  * Control non-executable mappings for 64-bit processes.
  99  *
 100  * on   Enable (default)
 101  * off  Disable
 102  */
 103 static int __init nonx_setup(char *str)
 104 {
 105         if (!str)
 106                 return -EINVAL;
 107         if (!strncmp(str, "on", 2)) {
 108                 __supported_pte_mask |= _PAGE_NX;
 109                 do_not_nx = 0;
 110         } else if (!strncmp(str, "off", 3)) {
 111                 do_not_nx = 1;
 112                 __supported_pte_mask &= ~_PAGE_NX;
 113         }
 114         return 0;
 115 }
 116 early_param("noexec", nonx_setup);
 117
 118 void __cpuinit check_efer(void)
 119 {
 120         unsigned long efer;
 121
 122         rdmsrl(MSR_EFER, efer);
 123         if (!(efer & EFER_NX) || do_not_nx)
 124                 __supported_pte_mask &= ~_PAGE_NX;
 125 }
 126
 127 int force_personality32;
 128
 129 /*
 130  * noexec32=on|off
 131  * Control non executable heap for 32bit processes.
 132  * To control the stack too use noexec=off
 133  *
 134  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 135  * off  PROT_READ implies PROT_EXEC
 136  */
 137 static int __init nonx32_setup(char *str)
 138 {
 139         if (!strcmp(str, "on"))
 140                 force_personality32 &= ~READ_IMPLIES_EXEC;
 141         else if (!strcmp(str, "off"))
 142                 force_personality32 |= READ_IMPLIES_EXEC;
 143         return 1;
 144 }
 145 __setup("noexec32=", nonx32_setup);
 146
 147 /*
 148  * NOTE: This function is marked __ref because it calls __init function
 149  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 150  */
 151 static __ref void *spp_getpage(void)
 152 {
 153         void *ptr;
 154
 155         if (after_bootmem)
 156                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 157         else
 158                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 159
 160         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 161                 panic("set_pte_phys: cannot allocate page data %s\n",
 162                         after_bootmem ? "after bootmem" : "");
 163         }
 164
 165         pr_debug("spp_getpage %p\n", ptr);
 166
 167         return ptr;
 168 }
 169
 170 void
 171 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 172 {
 173         pud_t *pud;
 174         pmd_t *pmd;
 175         pte_t *pte;
 176
 177         pud = pud_page + pud_index(vaddr);
 178         if (pud_none(*pud)) {
 179                 pmd = (pmd_t *) spp_getpage();
 180                 pud_populate(&init_mm, pud, pmd);
 181                 if (pmd != pmd_offset(pud, 0)) {
 182                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 183                                 pmd, pmd_offset(pud, 0));
 184                         return;
 185                 }
 186         }
 187         pmd = pmd_offset(pud, vaddr);
 188         if (pmd_none(*pmd)) {
 189                 pte = (pte_t *) spp_getpage();
 190                 pmd_populate_kernel(&init_mm, pmd, pte);
 191                 if (pte != pte_offset_kernel(pmd, 0)) {
 192                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 193                         return;
 194                 }
 195         }
 196
 197         pte = pte_offset_kernel(pmd, vaddr);
 198         if (!pte_none(*pte) && pte_val(new_pte) &&
 199             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 200                 pte_ERROR(*pte);
 201         set_pte(pte, new_pte);
 202
 203         /*
 204          * It's enough to flush this one mapping.
 205          * (PGE mappings get flushed as well)
 206          */
 207         __flush_tlb_one(vaddr);
 208 }
 209
 210 void
 211 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 212 {
 213         pgd_t *pgd;
 214         pud_t *pud_page;
 215
 216         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 217
 218         pgd = pgd_offset_k(vaddr);
 219         if (pgd_none(*pgd)) {
 220                 printk(KERN_ERR
 221                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 222                 return;
 223         }
 224         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 225         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 226 }
 227
 228 /*
 229  * Create large page table mappings for a range of physical addresses.
 230  */
 231 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 232                                                 pgprot_t prot)
 233 {
 234         pgd_t *pgd;
 235         pud_t *pud;
 236         pmd_t *pmd;
 237
 238         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 239         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 240                 pgd = pgd_offset_k((unsigned long)__va(phys));
 241                 if (pgd_none(*pgd)) {
 242                         pud = (pud_t *) spp_getpage();
 243                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 244                                                 _PAGE_USER));
 245                 }
 246                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 247                 if (pud_none(*pud)) {
 248                         pmd = (pmd_t *) spp_getpage();
 249                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 250                                                 _PAGE_USER));
 251                 }
 252                 pmd = pmd_offset(pud, phys);
 253                 BUG_ON(!pmd_none(*pmd));
 254                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 255         }
 256 }
 257
 258 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 259 {
 260         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 261 }
 262
 263 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 264 {
 265         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 266 }
 267
 268 /*
 269  * The head.S code sets up the kernel high mapping:
 270  *
 271  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 272  *
 273  * phys_addr holds the negative offset to the kernel, which is added
 274  * to the compile time generated pmds. This results in invalid pmds up
 275  * to the point where we hit the physaddr 0 mapping.
 276  *
 277  * We limit the mappings to the region from _text to _end.  _end is
 278  * rounded up to the 2MB boundary. This catches the invalid pmds as
 279  * well, as they are located before _text:
 280  */
 281 void __init cleanup_highmap(void)
 282 {
 283         unsigned long vaddr = __START_KERNEL_map;
 284         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
 285         pmd_t *pmd = level2_kernel_pgt;
 286         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 287
 288         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 289                 if (pmd_none(*pmd))
 290                         continue;
 291                 if (vaddr < (unsigned long) _text || vaddr > end)
 292                         set_pmd(pmd, __pmd(0));
 293         }
 294 }
 295
 296 static unsigned long __initdata table_start;
 297 static unsigned long __meminitdata table_end;
 298 static unsigned long __meminitdata table_top;
 299
 300 static __ref void *alloc_low_page(unsigned long *phys)
 301 {
 302         unsigned long pfn = table_end++;
 303         void *adr;
 304
 305         if (after_bootmem) {
 306                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 307                 *phys = __pa(adr);
 308
 309                 return adr;
 310         }
 311
 312         if (pfn >= table_top)
 313                 panic("alloc_low_page: ran out of memory");
 314
 315         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 316         memset(adr, 0, PAGE_SIZE);
 317         *phys  = pfn * PAGE_SIZE;
 318         return adr;
 319 }
 320
 321 static __ref void unmap_low_page(void *adr)
 322 {
 323         if (after_bootmem)
 324                 return;
 325
 326         early_iounmap(adr, PAGE_SIZE);
 327 }
 328
 329 static unsigned long __meminit
 330 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 331               pgprot_t prot)
 332 {
 333         unsigned pages = 0;
 334         unsigned long last_map_addr = end;
 335         int i;
 336
 337         pte_t *pte = pte_page + pte_index(addr);
 338
 339         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 340
 341                 if (addr >= end) {
 342                         if (!after_bootmem) {
 343                                 for(; i < PTRS_PER_PTE; i++, pte++)
 344                                         set_pte(pte, __pte(0));
 345                         }
 346                         break;
 347                 }
 348
 349                 /*
 350                  * We will re-use the existing mapping.
 351                  * Xen for example has some special requirements, like mapping
 352                  * pagetable pages as RO. So assume someone who pre-setup
 353                  * these mappings are more intelligent.
 354                  */
 355                 if (pte_val(*pte))
 356                         continue;
 357
 358                 if (0)
 359                         printk("   pte=%p addr=%lx pte=%016lx\n",
 360                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 361                 pages++;
 362                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 363                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 364         }
 365
 366         update_page_count(PG_LEVEL_4K, pages);
 367
 368         return last_map_addr;
 369 }
 370
 371 static unsigned long __meminit
 372 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end,
 373                 pgprot_t prot)
 374 {
 375         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 376
 377         return phys_pte_init(pte, address, end, prot);
 378 }
 379
 380 static unsigned long __meminit
 381 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 382               unsigned long page_size_mask, pgprot_t prot)
 383 {
 384         unsigned long pages = 0;
 385         unsigned long last_map_addr = end;
 386
 387         int i = pmd_index(address);
 388
 389         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 390                 unsigned long pte_phys;
 391                 pmd_t *pmd = pmd_page + pmd_index(address);
 392                 pte_t *pte;
 393                 pgprot_t new_prot = prot;
 394
 395                 if (address >= end) {
 396                         if (!after_bootmem) {
 397                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 398                                         set_pmd(pmd, __pmd(0));
 399                         }
 400                         break;
 401                 }
 402
 403                 if (pmd_val(*pmd)) {
 404                         if (!pmd_large(*pmd)) {
 405                                 spin_lock(&init_mm.page_table_lock);
 406                                 last_map_addr = phys_pte_update(pmd, address,
 407                                                                 end, prot);
 408                                 spin_unlock(&init_mm.page_table_lock);
 409                                 continue;
 410                         }
 411                         /*
 412                          * If we are ok with PG_LEVEL_2M mapping, then we will
 413                          * use the existing mapping,
 414                          *
 415                          * Otherwise, we will split the large page mapping but
 416                          * use the same existing protection bits except for
 417                          * large page, so that we don't violate Intel's TLB
 418                          * Application note (317080) which says, while changing
 419                          * the page sizes, new and old translations should
 420                          * not differ with respect to page frame and
 421                          * attributes.
 422                          */
 423                         if (page_size_mask & (1 << PG_LEVEL_2M))
 424                                 continue;
 425                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 426                 }
 427
 428                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 429                         pages++;
 430                         spin_lock(&init_mm.page_table_lock);
 431                         set_pte((pte_t *)pmd,
 432                                 pfn_pte(address >> PAGE_SHIFT,
 433                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 434                         spin_unlock(&init_mm.page_table_lock);
 435                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 436                         continue;
 437                 }
 438
 439                 pte = alloc_low_page(&pte_phys);
 440                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 441                 unmap_low_page(pte);
 442
 443                 spin_lock(&init_mm.page_table_lock);
 444                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 445                 spin_unlock(&init_mm.page_table_lock);
 446         }
 447         update_page_count(PG_LEVEL_2M, pages);
 448         return last_map_addr;
 449 }
 450
 451 static unsigned long __meminit
 452 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 453                 unsigned long page_size_mask, pgprot_t prot)
 454 {
 455         pmd_t *pmd = pmd_offset(pud, 0);
 456         unsigned long last_map_addr;
 457
 458         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
 459         __flush_tlb_all();
 460         return last_map_addr;
 461 }
 462
 463 static unsigned long __meminit
 464 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 465                          unsigned long page_size_mask)
 466 {
 467         unsigned long pages = 0;
 468         unsigned long last_map_addr = end;
 469         int i = pud_index(addr);
 470
 471         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 472                 unsigned long pmd_phys;
 473                 pud_t *pud = pud_page + pud_index(addr);
 474                 pmd_t *pmd;
 475                 pgprot_t prot = PAGE_KERNEL;
 476
 477                 if (addr >= end)
 478                         break;
 479
 480                 if (!after_bootmem &&
 481                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 482                         set_pud(pud, __pud(0));
 483                         continue;
 484                 }
 485
 486                 if (pud_val(*pud)) {
 487                         if (!pud_large(*pud)) {
 488                                 last_map_addr = phys_pmd_update(pud, addr, end,
 489                                                          page_size_mask, prot);
 490                                 continue;
 491                         }
 492                         /*
 493                          * If we are ok with PG_LEVEL_1G mapping, then we will
 494                          * use the existing mapping.
 495                          *
 496                          * Otherwise, we will split the gbpage mapping but use
 497                          * the same existing protection  bits except for large
 498                          * page, so that we don't violate Intel's TLB
 499                          * Application note (317080) which says, while changing
 500                          * the page sizes, new and old translations should
 501                          * not differ with respect to page frame and
 502                          * attributes.
 503                          */
 504                         if (page_size_mask & (1 << PG_LEVEL_1G))
 505                                 continue;
 506                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 507                 }
 508
 509                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 510                         pages++;
 511                         spin_lock(&init_mm.page_table_lock);
 512                         set_pte((pte_t *)pud,
 513                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 514                         spin_unlock(&init_mm.page_table_lock);
 515                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 516                         continue;
 517                 }
 518
 519                 pmd = alloc_low_page(&pmd_phys);
 520                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 521                                               prot);
 522                 unmap_low_page(pmd);
 523
 524                 spin_lock(&init_mm.page_table_lock);
 525                 pud_populate(&init_mm, pud, __va(pmd_phys));
 526                 spin_unlock(&init_mm.page_table_lock);
 527         }
 528         __flush_tlb_all();
 529
 530         update_page_count(PG_LEVEL_1G, pages);
 531
 532         return last_map_addr;
 533 }
 534
 535 static unsigned long __meminit
 536 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 537                  unsigned long page_size_mask)
 538 {
 539         pud_t *pud;
 540
 541         pud = (pud_t *)pgd_page_vaddr(*pgd);
 542
 543         return phys_pud_init(pud, addr, end, page_size_mask);
 544 }
 545
 546 static void __init find_early_table_space(unsigned long end, int use_pse,
 547                                           int use_gbpages)
 548 {
 549         unsigned long puds, pmds, ptes, tables, start;
 550
 551         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 552         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
 553         if (use_gbpages) {
 554                 unsigned long extra;
 555                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 556                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 557         } else
 558                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 559         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 560
 561         if (use_pse) {
 562                 unsigned long extra;
 563                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 564                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 565         } else
 566                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 567         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 568
 569         /*
 570          * RED-PEN putting page tables only on node 0 could
 571          * cause a hotspot and fill up ZONE_DMA. The page tables
 572          * need roughly 0.5KB per GB.
 573          */
 574         start = 0x8000;
 575         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 576         if (table_start == -1UL)
 577                 panic("Cannot find space for the kernel page tables");
 578
 579         table_start >>= PAGE_SHIFT;
 580         table_end = table_start;
 581         table_top = table_start + (tables >> PAGE_SHIFT);
 582
 583         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 584                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 585 }
 586
 587 static void __init init_gbpages(void)
 588 {
 589         if (direct_gbpages && cpu_has_gbpages)
 590                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 591         else
 592                 direct_gbpages = 0;
 593 }
 594
 595 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 596                                                 unsigned long end,
 597                                                 unsigned long page_size_mask)
 598 {
 599
 600         unsigned long next, last_map_addr = end;
 601
 602         start = (unsigned long)__va(start);
 603         end = (unsigned long)__va(end);
 604
 605         for (; start < end; start = next) {
 606                 pgd_t *pgd = pgd_offset_k(start);
 607                 unsigned long pud_phys;
 608                 pud_t *pud;
 609
 610                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 611                 if (next > end)
 612                         next = end;
 613
 614                 if (pgd_val(*pgd)) {
 615                         last_map_addr = phys_pud_update(pgd, __pa(start),
 616                                                  __pa(end), page_size_mask);
 617                         continue;
 618                 }
 619
 620                 pud = alloc_low_page(&pud_phys);
 621                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 622                                                  page_size_mask);
 623                 unmap_low_page(pud);
 624
 625                 spin_lock(&init_mm.page_table_lock);
 626                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 627                 spin_unlock(&init_mm.page_table_lock);
 628         }
 629         __flush_tlb_all();
 630
 631         return last_map_addr;
 632 }
 633
 634 struct map_range {
 635         unsigned long start;
 636         unsigned long end;
 637         unsigned page_size_mask;
 638 };
 639
 640 #define NR_RANGE_MR 5
 641
 642 static int save_mr(struct map_range *mr, int nr_range,
 643                    unsigned long start_pfn, unsigned long end_pfn,
 644                    unsigned long page_size_mask)
 645 {
 646
 647         if (start_pfn < end_pfn) {
 648                 if (nr_range >= NR_RANGE_MR)
 649                         panic("run out of range for init_memory_mapping\n");
 650                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 651                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 652                 mr[nr_range].page_size_mask = page_size_mask;
 653                 nr_range++;
 654         }
 655
 656         return nr_range;
 657 }
 658
 659 /*
 660  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 661  * This runs before bootmem is initialized and gets pages directly from
 662  * the physical memory. To access them they are temporarily mapped.
 663  */
 664 unsigned long __init_refok init_memory_mapping(unsigned long start,
 665                                                unsigned long end)
 666 {
 667         unsigned long last_map_addr = 0;
 668         unsigned long page_size_mask = 0;
 669         unsigned long start_pfn, end_pfn;
 670
 671         struct map_range mr[NR_RANGE_MR];
 672         int nr_range, i;
 673         int use_pse, use_gbpages;
 674
 675         printk(KERN_INFO "init_memory_mapping\n");
 676
 677         /*
 678          * Find space for the kernel direct mapping tables.
 679          *
 680          * Later we should allocate these tables in the local node of the
 681          * memory mapped. Unfortunately this is done currently before the
 682          * nodes are discovered.
 683          */
 684         if (!after_bootmem)
 685                 init_gbpages();
 686
 687 #ifdef CONFIG_DEBUG_PAGEALLOC
 688         /*
 689          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 690          * This will simplify cpa(), which otherwise needs to support splitting
 691          * large pages into small in interrupt context, etc.
 692          */
 693         use_pse = use_gbpages = 0;
 694 #else
 695         use_pse = cpu_has_pse;
 696         use_gbpages = direct_gbpages;
 697 #endif
 698
 699         if (use_gbpages)
 700                 page_size_mask |= 1 << PG_LEVEL_1G;
 701         if (use_pse)
 702                 page_size_mask |= 1 << PG_LEVEL_2M;
 703
 704         memset(mr, 0, sizeof(mr));
 705         nr_range = 0;
 706
 707         /* head if not big page alignment ?*/
 708         start_pfn = start >> PAGE_SHIFT;
 709         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 710                         << (PMD_SHIFT - PAGE_SHIFT);
 711         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 712
 713         /* big page (2M) range*/
 714         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 715                          << (PMD_SHIFT - PAGE_SHIFT);
 716         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 717                          << (PUD_SHIFT - PAGE_SHIFT);
 718         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 719                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 720         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 721                         page_size_mask & (1<<PG_LEVEL_2M));
 722
 723         /* big page (1G) range */
 724         start_pfn = end_pfn;
 725         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 726         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 727                                 page_size_mask &
 728                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 729
 730         /* tail is not big page (1G) alignment */
 731         start_pfn = end_pfn;
 732         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 733         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 734                         page_size_mask & (1<<PG_LEVEL_2M));
 735
 736         /* tail is not big page (2M) alignment */
 737         start_pfn = end_pfn;
 738         end_pfn = end>>PAGE_SHIFT;
 739         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 740
 741         /* try to merge same page size and continuous */
 742         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 743                 unsigned long old_start;
 744                 if (mr[i].end != mr[i+1].start ||
 745                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 746                         continue;
 747                 /* move it */
 748                 old_start = mr[i].start;
 749                 memmove(&mr[i], &mr[i+1],
 750                          (nr_range - 1 - i) * sizeof (struct map_range));
 751                 mr[i].start = old_start;
 752                 nr_range--;
 753         }
 754
 755         for (i = 0; i < nr_range; i++)
 756                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 757                                 mr[i].start, mr[i].end,
 758                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 759                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 760
 761         if (!after_bootmem)
 762                 find_early_table_space(end, use_pse, use_gbpages);
 763
 764         for (i = 0; i < nr_range; i++)
 765                 last_map_addr = kernel_physical_mapping_init(
 766                                         mr[i].start, mr[i].end,
 767                                         mr[i].page_size_mask);
 768
 769         if (!after_bootmem)
 770                 mmu_cr4_features = read_cr4();
 771         __flush_tlb_all();
 772
 773         if (!after_bootmem && table_end > table_start)
 774                 reserve_early(table_start << PAGE_SHIFT,
 775                                  table_end << PAGE_SHIFT, "PGTABLE");
 776
 777         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 778                          last_map_addr, end);
 779
 780         if (!after_bootmem)
 781                 early_memtest(start, end);
 782
 783         return last_map_addr >> PAGE_SHIFT;
 784 }
 785
 786 #ifndef CONFIG_NUMA
 787 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 788 {
 789         unsigned long bootmap_size, bootmap;
 790
 791         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 792         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 793                                  PAGE_SIZE);
 794         if (bootmap == -1L)
 795                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 796         /* don't touch min_low_pfn */
 797         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 798                                          0, end_pfn);
 799         e820_register_active_regions(0, start_pfn, end_pfn);
 800         free_bootmem_with_active_regions(0, end_pfn);
 801         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 802         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 803 }
 804
 805 void __init paging_init(void)
 806 {
 807         unsigned long max_zone_pfns[MAX_NR_ZONES];
 808
 809         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 810         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 811         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 812         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 813
 814         memory_present(0, 0, max_pfn);
 815         sparse_init();
 816         free_area_init_nodes(max_zone_pfns);
 817 }
 818 #endif
 819
 820 /*
 821  * Memory hotplug specific functions
 822  */
 823 #ifdef CONFIG_MEMORY_HOTPLUG
 824 /*
 825  * Memory is added always to NORMAL zone. This means you will never get
 826  * additional DMA/DMA32 memory.
 827  */
 828 int arch_add_memory(int nid, u64 start, u64 size)
 829 {
 830         struct pglist_data *pgdat = NODE_DATA(nid);
 831         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 832         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 833         unsigned long nr_pages = size >> PAGE_SHIFT;
 834         int ret;
 835
 836         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 837         if (last_mapped_pfn > max_pfn_mapped)
 838                 max_pfn_mapped = last_mapped_pfn;
 839
 840         ret = __add_pages(zone, start_pfn, nr_pages);
 841         WARN_ON(1);
 842
 843         return ret;
 844 }
 845 EXPORT_SYMBOL_GPL(arch_add_memory);
 846
 847 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 848 int memory_add_physaddr_to_nid(u64 start)
 849 {
 850         return 0;
 851 }
 852 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 853 #endif
 854
 855 #endif /* CONFIG_MEMORY_HOTPLUG */
 856
 857 /*
 858  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 859  * is valid. The argument is a physical page number.
 860  *
 861  *
 862  * On x86, access has to be given to the first megabyte of ram because that area
 863  * contains bios code and data regions used by X and dosemu and similar apps.
 864  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 865  * mmio resources as well as potential bios/acpi data regions.
 866  */
 867 int devmem_is_allowed(unsigned long pagenr)
 868 {
 869         if (pagenr <= 256)
 870                 return 1;
 871         if (!page_is_ram(pagenr))
 872                 return 1;
 873         return 0;
 874 }
 875
 876
 877 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 878                          kcore_modules, kcore_vsyscall;
 879
 880 void __init mem_init(void)
 881 {
 882         long codesize, reservedpages, datasize, initsize;
 883
 884         pci_iommu_alloc();
 885
 886         /* clear_bss() already clear the empty_zero_page */
 887
 888         reservedpages = 0;
 889
 890         /* this will put all low memory onto the freelists */
 891 #ifdef CONFIG_NUMA
 892         totalram_pages = numa_free_all_bootmem();
 893 #else
 894         totalram_pages = free_all_bootmem();
 895 #endif
 896         reservedpages = max_pfn - totalram_pages -
 897                                         absent_pages_in_range(0, max_pfn);
 898         after_bootmem = 1;
 899
 900         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 901         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 902         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 903
 904         /* Register memory areas for /proc/kcore */
 905         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 906         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 907                    VMALLOC_END-VMALLOC_START);
 908         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 909         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 910         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 911                                  VSYSCALL_END - VSYSCALL_START);
 912
 913         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 914                                 "%ldk reserved, %ldk data, %ldk init)\n",
 915                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 916                 max_pfn << (PAGE_SHIFT-10),
 917                 codesize >> 10,
 918                 reservedpages << (PAGE_SHIFT-10),
 919                 datasize >> 10,
 920                 initsize >> 10);
 921 }
 922
 923 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 924 {
 925         unsigned long addr = begin;
 926
 927         if (addr >= end)
 928                 return;
 929
 930         /*
 931          * If debugging page accesses then do not free this memory but
 932          * mark them not present - any buggy init-section access will
 933          * create a kernel page fault:
 934          */
 935 #ifdef CONFIG_DEBUG_PAGEALLOC
 936         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 937                 begin, PAGE_ALIGN(end));
 938         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 939 #else
 940         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 941
 942         for (; addr < end; addr += PAGE_SIZE) {
 943                 ClearPageReserved(virt_to_page(addr));
 944                 init_page_count(virt_to_page(addr));
 945                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 946                         POISON_FREE_INITMEM, PAGE_SIZE);
 947                 free_page(addr);
 948                 totalram_pages++;
 949         }
 950 #endif
 951 }
 952
 953 void free_initmem(void)
 954 {
 955         free_init_pages("unused kernel memory",
 956                         (unsigned long)(&__init_begin),
 957                         (unsigned long)(&__init_end));
 958 }
 959
 960 #ifdef CONFIG_DEBUG_RODATA
 961 const int rodata_test_data = 0xC3;
 962 EXPORT_SYMBOL_GPL(rodata_test_data);
 963
 964 void mark_rodata_ro(void)
 965 {
 966         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 967         unsigned long rodata_start =
 968                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 969
 970 #ifdef CONFIG_DYNAMIC_FTRACE
 971         /* Dynamic tracing modifies the kernel text section */
 972         start = rodata_start;
 973 #endif
 974
 975         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 976                (end - start) >> 10);
 977         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 978
 979         /*
 980          * The rodata section (but not the kernel text!) should also be
 981          * not-executable.
 982          */
 983         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 984
 985         rodata_test();
 986
 987 #ifdef CONFIG_CPA_DEBUG
 988         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 989         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 990
 991         printk(KERN_INFO "Testing CPA: again\n");
 992         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 993 #endif
 994 }
 995
 996 #endif
 997
 998 #ifdef CONFIG_BLK_DEV_INITRD
 999 void free_initrd_mem(unsigned long start, unsigned long end)
1000 {
1001         free_init_pages("initrd memory", start, end);
1002 }
1003 #endif
1004
1005 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1006                                    int flags)
1007 {
1008 #ifdef CONFIG_NUMA
1009         int nid, next_nid;
1010         int ret;
1011 #endif
1012         unsigned long pfn = phys >> PAGE_SHIFT;
1013
1014         if (pfn >= max_pfn) {
1015                 /*
1016                  * This can happen with kdump kernels when accessing
1017                  * firmware tables:
1018                  */
1019                 if (pfn < max_pfn_mapped)
1020                         return -EFAULT;
1021
1022                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
1023                                 phys, len);
1024                 return -EFAULT;
1025         }
1026
1027         /* Should check here against the e820 map to avoid double free */
1028 #ifdef CONFIG_NUMA
1029         nid = phys_to_nid(phys);
1030         next_nid = phys_to_nid(phys + len - 1);
1031         if (nid == next_nid)
1032                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
1033         else
1034                 ret = reserve_bootmem(phys, len, flags);
1035
1036         if (ret != 0)
1037                 return ret;
1038
1039 #else
1040         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1041 #endif
1042
1043         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1044                 dma_reserve += len / PAGE_SIZE;
1045                 set_dma_reserve(dma_reserve);
1046         }
1047
1048         return 0;
1049 }
1050
1051 int kern_addr_valid(unsigned long addr)
1052 {
1053         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1054         pgd_t *pgd;
1055         pud_t *pud;
1056         pmd_t *pmd;
1057         pte_t *pte;
1058
1059         if (above != 0 && above != -1UL)
1060                 return 0;
1061
1062         pgd = pgd_offset_k(addr);
1063         if (pgd_none(*pgd))
1064                 return 0;
1065
1066         pud = pud_offset(pgd, addr);
1067         if (pud_none(*pud))
1068                 return 0;
1069
1070         pmd = pmd_offset(pud, addr);
1071         if (pmd_none(*pmd))
1072                 return 0;
1073
1074         if (pmd_large(*pmd))
1075                 return pfn_valid(pmd_pfn(*pmd));
1076
1077         pte = pte_offset_kernel(pmd, addr);
1078         if (pte_none(*pte))
1079                 return 0;
1080
1081         return pfn_valid(pte_pfn(*pte));
1082 }
1083
1084 /*
1085  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1086  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1087  * not need special handling anymore:
1088  */
1089 static struct vm_area_struct gate_vma = {
1090         .vm_start       = VSYSCALL_START,
1091         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1092         .vm_page_prot   = PAGE_READONLY_EXEC,
1093         .vm_flags       = VM_READ | VM_EXEC
1094 };
1095
1096 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1097 {
1098 #ifdef CONFIG_IA32_EMULATION
1099         if (test_tsk_thread_flag(tsk, TIF_IA32))
1100                 return NULL;
1101 #endif
1102         return &gate_vma;
1103 }
1104
1105 int in_gate_area(struct task_struct *task, unsigned long addr)
1106 {
1107         struct vm_area_struct *vma = get_gate_vma(task);
1108
1109         if (!vma)
1110                 return 0;
1111
1112         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1113 }
1114
1115 /*
1116  * Use this when you have no reliable task/vma, typically from interrupt
1117  * context. It is less reliable than using the task's vma and may give
1118  * false positives:
1119  */
1120 int in_gate_area_no_task(unsigned long addr)
1121 {
1122         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1123 }
1124
1125 const char *arch_vma_name(struct vm_area_struct *vma)
1126 {
1127         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1128                 return "[vdso]";
1129         if (vma == &gate_vma)
1130                 return "[vsyscall]";
1131         return NULL;
1132 }
1133
1134 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1135 /*
1136  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1137  */
1138 static long __meminitdata addr_start, addr_end;
1139 static void __meminitdata *p_start, *p_end;
1140 static int __meminitdata node_start;
1141
1142 int __meminit
1143 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1144 {
1145         unsigned long addr = (unsigned long)start_page;
1146         unsigned long end = (unsigned long)(start_page + size);
1147         unsigned long next;
1148         pgd_t *pgd;
1149         pud_t *pud;
1150         pmd_t *pmd;
1151
1152         for (; addr < end; addr = next) {
1153                 void *p = NULL;
1154
1155                 pgd = vmemmap_pgd_populate(addr, node);
1156                 if (!pgd)
1157                         return -ENOMEM;
1158
1159                 pud = vmemmap_pud_populate(pgd, addr, node);
1160                 if (!pud)
1161                         return -ENOMEM;
1162
1163                 if (!cpu_has_pse) {
1164                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1165                         pmd = vmemmap_pmd_populate(pud, addr, node);
1166
1167                         if (!pmd)
1168                                 return -ENOMEM;
1169
1170                         p = vmemmap_pte_populate(pmd, addr, node);
1171
1172                         if (!p)
1173                                 return -ENOMEM;
1174
1175                         addr_end = addr + PAGE_SIZE;
1176                         p_end = p + PAGE_SIZE;
1177                 } else {
1178                         next = pmd_addr_end(addr, end);
1179
1180                         pmd = pmd_offset(pud, addr);
1181                         if (pmd_none(*pmd)) {
1182                                 pte_t entry;
1183
1184                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1185                                 if (!p)
1186                                         return -ENOMEM;
1187
1188                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1189                                                 PAGE_KERNEL_LARGE);
1190                                 set_pmd(pmd, __pmd(pte_val(entry)));
1191
1192                                 /* check to see if we have contiguous blocks */
1193                                 if (p_end != p || node_start != node) {
1194                                         if (p_start)
1195                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1196                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1197                                         addr_start = addr;
1198                                         node_start = node;
1199                                         p_start = p;
1200                                 }
1201
1202                                 addr_end = addr + PMD_SIZE;
1203                                 p_end = p + PMD_SIZE;
1204                         } else
1205                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1206                 }
1207
1208         }
1209         return 0;
1210 }
1211
1212 void __meminit vmemmap_populate_print_last(void)
1213 {
1214         if (p_start) {
1215                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1216                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1217                 p_start = NULL;
1218                 p_end = NULL;
1219                 node_start = 0;
1220         }
1221 }
1222 #endif