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/bios_ebda.h>
  35 #include <asm/system.h>
  36 #include <asm/uaccess.h>
  37 #include <asm/pgtable.h>
  38 #include <asm/pgalloc.h>
  39 #include <asm/dma.h>
  40 #include <asm/fixmap.h>
  41 #include <asm/e820.h>
  42 #include <asm/apic.h>
  43 #include <asm/tlb.h>
  44 #include <asm/mmu_context.h>
  45 #include <asm/proto.h>
  46 #include <asm/smp.h>
  47 #include <asm/sections.h>
  48 #include <asm/kdebug.h>
  49 #include <asm/numa.h>
  50 #include <asm/cacheflush.h>
  51
  52 /*
  53  * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
  54  * The direct mapping extends to max_pfn_mapped, so that we can directly access
  55  * apertures, ACPI and other tables without having to play with fixmaps.
  56  */
  57 unsigned long max_low_pfn_mapped;
  58 unsigned long max_pfn_mapped;
  59
  60 static unsigned long dma_reserve __initdata;
  61
  62 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  63
  64 int direct_gbpages
  65 #ifdef CONFIG_DIRECT_GBPAGES
  66                                 = 1
  67 #endif
  68 ;
  69
  70 static int __init parse_direct_gbpages_off(char *arg)
  71 {
  72         direct_gbpages = 0;
  73         return 0;
  74 }
  75 early_param("nogbpages", parse_direct_gbpages_off);
  76
  77 static int __init parse_direct_gbpages_on(char *arg)
  78 {
  79         direct_gbpages = 1;
  80         return 0;
  81 }
  82 early_param("gbpages", parse_direct_gbpages_on);
  83
  84 /*
  85  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  86  * physical space so we can cache the place of the first one and move
  87  * around without checking the pgd every time.
  88  */
  89
  90 int after_bootmem;
  91
  92 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
  93 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  94
  95 static int do_not_nx __cpuinitdata;
  96
  97 /*
  98  * noexec=on|off
  99  * Control non-executable mappings for 64-bit processes.
 100  *
 101  * on   Enable (default)
 102  * off  Disable
 103  */
 104 static int __init nonx_setup(char *str)
 105 {
 106         if (!str)
 107                 return -EINVAL;
 108         if (!strncmp(str, "on", 2)) {
 109                 __supported_pte_mask |= _PAGE_NX;
 110                 do_not_nx = 0;
 111         } else if (!strncmp(str, "off", 3)) {
 112                 do_not_nx = 1;
 113                 __supported_pte_mask &= ~_PAGE_NX;
 114         }
 115         return 0;
 116 }
 117 early_param("noexec", nonx_setup);
 118
 119 void __cpuinit check_efer(void)
 120 {
 121         unsigned long efer;
 122
 123         rdmsrl(MSR_EFER, efer);
 124         if (!(efer & EFER_NX) || do_not_nx)
 125                 __supported_pte_mask &= ~_PAGE_NX;
 126 }
 127
 128 int force_personality32;
 129
 130 /*
 131  * noexec32=on|off
 132  * Control non executable heap for 32bit processes.
 133  * To control the stack too use noexec=off
 134  *
 135  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 136  * off  PROT_READ implies PROT_EXEC
 137  */
 138 static int __init nonx32_setup(char *str)
 139 {
 140         if (!strcmp(str, "on"))
 141                 force_personality32 &= ~READ_IMPLIES_EXEC;
 142         else if (!strcmp(str, "off"))
 143                 force_personality32 |= READ_IMPLIES_EXEC;
 144         return 1;
 145 }
 146 __setup("noexec32=", nonx32_setup);
 147
 148 /*
 149  * NOTE: This function is marked __ref because it calls __init function
 150  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 151  */
 152 static __ref void *spp_getpage(void)
 153 {
 154         void *ptr;
 155
 156         if (after_bootmem)
 157                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 158         else
 159                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 160
 161         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 162                 panic("set_pte_phys: cannot allocate page data %s\n",
 163                         after_bootmem ? "after bootmem" : "");
 164         }
 165
 166         pr_debug("spp_getpage %p\n", ptr);
 167
 168         return ptr;
 169 }
 170
 171 void
 172 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 173 {
 174         pud_t *pud;
 175         pmd_t *pmd;
 176         pte_t *pte;
 177
 178         pud = pud_page + pud_index(vaddr);
 179         if (pud_none(*pud)) {
 180                 pmd = (pmd_t *) spp_getpage();
 181                 pud_populate(&init_mm, pud, pmd);
 182                 if (pmd != pmd_offset(pud, 0)) {
 183                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 184                                 pmd, pmd_offset(pud, 0));
 185                         return;
 186                 }
 187         }
 188         pmd = pmd_offset(pud, vaddr);
 189         if (pmd_none(*pmd)) {
 190                 pte = (pte_t *) spp_getpage();
 191                 pmd_populate_kernel(&init_mm, pmd, pte);
 192                 if (pte != pte_offset_kernel(pmd, 0)) {
 193                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 194                         return;
 195                 }
 196         }
 197
 198         pte = pte_offset_kernel(pmd, vaddr);
 199         set_pte(pte, new_pte);
 200
 201         /*
 202          * It's enough to flush this one mapping.
 203          * (PGE mappings get flushed as well)
 204          */
 205         __flush_tlb_one(vaddr);
 206 }
 207
 208 void
 209 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 210 {
 211         pgd_t *pgd;
 212         pud_t *pud_page;
 213
 214         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 215
 216         pgd = pgd_offset_k(vaddr);
 217         if (pgd_none(*pgd)) {
 218                 printk(KERN_ERR
 219                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 220                 return;
 221         }
 222         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 223         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 224 }
 225
 226 /*
 227  * Create large page table mappings for a range of physical addresses.
 228  */
 229 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 230                                                 pgprot_t prot)
 231 {
 232         pgd_t *pgd;
 233         pud_t *pud;
 234         pmd_t *pmd;
 235
 236         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 237         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 238                 pgd = pgd_offset_k((unsigned long)__va(phys));
 239                 if (pgd_none(*pgd)) {
 240                         pud = (pud_t *) spp_getpage();
 241                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 242                                                 _PAGE_USER));
 243                 }
 244                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 245                 if (pud_none(*pud)) {
 246                         pmd = (pmd_t *) spp_getpage();
 247                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 248                                                 _PAGE_USER));
 249                 }
 250                 pmd = pmd_offset(pud, phys);
 251                 BUG_ON(!pmd_none(*pmd));
 252                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 253         }
 254 }
 255
 256 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 257 {
 258         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 259 }
 260
 261 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 262 {
 263         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 264 }
 265
 266 /*
 267  * The head.S code sets up the kernel high mapping:
 268  *
 269  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 270  *
 271  * phys_addr holds the negative offset to the kernel, which is added
 272  * to the compile time generated pmds. This results in invalid pmds up
 273  * to the point where we hit the physaddr 0 mapping.
 274  *
 275  * We limit the mappings to the region from _text to _end.  _end is
 276  * rounded up to the 2MB boundary. This catches the invalid pmds as
 277  * well, as they are located before _text:
 278  */
 279 void __init cleanup_highmap(void)
 280 {
 281         unsigned long vaddr = __START_KERNEL_map;
 282         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
 283         pmd_t *pmd = level2_kernel_pgt;
 284         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 285
 286         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 287                 if (pmd_none(*pmd))
 288                         continue;
 289                 if (vaddr < (unsigned long) _text || vaddr > end)
 290                         set_pmd(pmd, __pmd(0));
 291         }
 292 }
 293
 294 static unsigned long __initdata table_start;
 295 static unsigned long __meminitdata table_end;
 296 static unsigned long __meminitdata table_top;
 297
 298 static __ref void *alloc_low_page(unsigned long *phys)
 299 {
 300         unsigned long pfn = table_end++;
 301         void *adr;
 302
 303         if (after_bootmem) {
 304                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 305                 *phys = __pa(adr);
 306
 307                 return adr;
 308         }
 309
 310         if (pfn >= table_top)
 311                 panic("alloc_low_page: ran out of memory");
 312
 313         adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
 314         memset(adr, 0, PAGE_SIZE);
 315         *phys  = pfn * PAGE_SIZE;
 316         return adr;
 317 }
 318
 319 static __ref void unmap_low_page(void *adr)
 320 {
 321         if (after_bootmem)
 322                 return;
 323
 324         early_iounmap(adr, PAGE_SIZE);
 325 }
 326
 327 static unsigned long __meminit
 328 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 329               pgprot_t prot)
 330 {
 331         unsigned pages = 0;
 332         unsigned long last_map_addr = end;
 333         int i;
 334
 335         pte_t *pte = pte_page + pte_index(addr);
 336
 337         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 338
 339                 if (addr >= end) {
 340                         if (!after_bootmem) {
 341                                 for(; i < PTRS_PER_PTE; i++, pte++)
 342                                         set_pte(pte, __pte(0));
 343                         }
 344                         break;
 345                 }
 346
 347                 /*
 348                  * We will re-use the existing mapping.
 349                  * Xen for example has some special requirements, like mapping
 350                  * pagetable pages as RO. So assume someone who pre-setup
 351                  * these mappings are more intelligent.
 352                  */
 353                 if (pte_val(*pte)) {
 354                         pages++;
 355                         continue;
 356                 }
 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                                 pages++;
 425                                 continue;
 426                         }
 427                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 428                 }
 429
 430                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 431                         pages++;
 432                         spin_lock(&init_mm.page_table_lock);
 433                         set_pte((pte_t *)pmd,
 434                                 pfn_pte(address >> PAGE_SHIFT,
 435                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 436                         spin_unlock(&init_mm.page_table_lock);
 437                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 438                         continue;
 439                 }
 440
 441                 pte = alloc_low_page(&pte_phys);
 442                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 443                 unmap_low_page(pte);
 444
 445                 spin_lock(&init_mm.page_table_lock);
 446                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 447                 spin_unlock(&init_mm.page_table_lock);
 448         }
 449         update_page_count(PG_LEVEL_2M, pages);
 450         return last_map_addr;
 451 }
 452
 453 static unsigned long __meminit
 454 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 455                 unsigned long page_size_mask, pgprot_t prot)
 456 {
 457         pmd_t *pmd = pmd_offset(pud, 0);
 458         unsigned long last_map_addr;
 459
 460         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
 461         __flush_tlb_all();
 462         return last_map_addr;
 463 }
 464
 465 static unsigned long __meminit
 466 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 467                          unsigned long page_size_mask)
 468 {
 469         unsigned long pages = 0;
 470         unsigned long last_map_addr = end;
 471         int i = pud_index(addr);
 472
 473         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 474                 unsigned long pmd_phys;
 475                 pud_t *pud = pud_page + pud_index(addr);
 476                 pmd_t *pmd;
 477                 pgprot_t prot = PAGE_KERNEL;
 478
 479                 if (addr >= end)
 480                         break;
 481
 482                 if (!after_bootmem &&
 483                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 484                         set_pud(pud, __pud(0));
 485                         continue;
 486                 }
 487
 488                 if (pud_val(*pud)) {
 489                         if (!pud_large(*pud)) {
 490                                 last_map_addr = phys_pmd_update(pud, addr, end,
 491                                                          page_size_mask, prot);
 492                                 continue;
 493                         }
 494                         /*
 495                          * If we are ok with PG_LEVEL_1G mapping, then we will
 496                          * use the existing mapping.
 497                          *
 498                          * Otherwise, we will split the gbpage mapping but use
 499                          * the same existing protection  bits except for large
 500                          * page, so that we don't violate Intel's TLB
 501                          * Application note (317080) which says, while changing
 502                          * the page sizes, new and old translations should
 503                          * not differ with respect to page frame and
 504                          * attributes.
 505                          */
 506                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 507                                 pages++;
 508                                 continue;
 509                         }
 510                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 511                 }
 512
 513                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 514                         pages++;
 515                         spin_lock(&init_mm.page_table_lock);
 516                         set_pte((pte_t *)pud,
 517                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 518                         spin_unlock(&init_mm.page_table_lock);
 519                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 520                         continue;
 521                 }
 522
 523                 pmd = alloc_low_page(&pmd_phys);
 524                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 525                                               prot);
 526                 unmap_low_page(pmd);
 527
 528                 spin_lock(&init_mm.page_table_lock);
 529                 pud_populate(&init_mm, pud, __va(pmd_phys));
 530                 spin_unlock(&init_mm.page_table_lock);
 531         }
 532         __flush_tlb_all();
 533
 534         update_page_count(PG_LEVEL_1G, pages);
 535
 536         return last_map_addr;
 537 }
 538
 539 static unsigned long __meminit
 540 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 541                  unsigned long page_size_mask)
 542 {
 543         pud_t *pud;
 544
 545         pud = (pud_t *)pgd_page_vaddr(*pgd);
 546
 547         return phys_pud_init(pud, addr, end, page_size_mask);
 548 }
 549
 550 static void __init find_early_table_space(unsigned long end, int use_pse,
 551                                           int use_gbpages)
 552 {
 553         unsigned long puds, pmds, ptes, tables, start;
 554
 555         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 556         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
 557
 558         if (use_gbpages) {
 559                 unsigned long extra;
 560
 561                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 562                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 563         } else
 564                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 565
 566         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 567
 568         if (use_pse) {
 569                 unsigned long extra;
 570
 571                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 572                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 573         } else
 574                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 575
 576         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 577
 578         /*
 579          * RED-PEN putting page tables only on node 0 could
 580          * cause a hotspot and fill up ZONE_DMA. The page tables
 581          * need roughly 0.5KB per GB.
 582          */
 583         start = 0x8000;
 584         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 585         if (table_start == -1UL)
 586                 panic("Cannot find space for the kernel page tables");
 587
 588         table_start >>= PAGE_SHIFT;
 589         table_end = table_start;
 590         table_top = table_start + (tables >> PAGE_SHIFT);
 591
 592         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 593                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 594 }
 595
 596 static void __init init_gbpages(void)
 597 {
 598         if (direct_gbpages && cpu_has_gbpages)
 599                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 600         else
 601                 direct_gbpages = 0;
 602 }
 603
 604 static unsigned long __meminit kernel_physical_mapping_init(unsigned long start,
 605                                                 unsigned long end,
 606                                                 unsigned long page_size_mask)
 607 {
 608
 609         unsigned long next, last_map_addr = end;
 610
 611         start = (unsigned long)__va(start);
 612         end = (unsigned long)__va(end);
 613
 614         for (; start < end; start = next) {
 615                 pgd_t *pgd = pgd_offset_k(start);
 616                 unsigned long pud_phys;
 617                 pud_t *pud;
 618
 619                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 620                 if (next > end)
 621                         next = end;
 622
 623                 if (pgd_val(*pgd)) {
 624                         last_map_addr = phys_pud_update(pgd, __pa(start),
 625                                                  __pa(end), page_size_mask);
 626                         continue;
 627                 }
 628
 629                 pud = alloc_low_page(&pud_phys);
 630                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 631                                                  page_size_mask);
 632                 unmap_low_page(pud);
 633
 634                 spin_lock(&init_mm.page_table_lock);
 635                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 636                 spin_unlock(&init_mm.page_table_lock);
 637         }
 638         __flush_tlb_all();
 639
 640         return last_map_addr;
 641 }
 642
 643 struct map_range {
 644         unsigned long start;
 645         unsigned long end;
 646         unsigned page_size_mask;
 647 };
 648
 649 #define NR_RANGE_MR 5
 650
 651 static int save_mr(struct map_range *mr, int nr_range,
 652                    unsigned long start_pfn, unsigned long end_pfn,
 653                    unsigned long page_size_mask)
 654 {
 655         if (start_pfn < end_pfn) {
 656                 if (nr_range >= NR_RANGE_MR)
 657                         panic("run out of range for init_memory_mapping\n");
 658                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 659                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 660                 mr[nr_range].page_size_mask = page_size_mask;
 661                 nr_range++;
 662         }
 663
 664         return nr_range;
 665 }
 666
 667 /*
 668  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 669  * This runs before bootmem is initialized and gets pages directly from
 670  * the physical memory. To access them they are temporarily mapped.
 671  */
 672 unsigned long __init_refok init_memory_mapping(unsigned long start,
 673                                                unsigned long end)
 674 {
 675         unsigned long last_map_addr = 0;
 676         unsigned long page_size_mask = 0;
 677         unsigned long start_pfn, end_pfn;
 678         unsigned long pos;
 679
 680         struct map_range mr[NR_RANGE_MR];
 681         int nr_range, i;
 682         int use_pse, use_gbpages;
 683
 684         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 685
 686         if (!after_bootmem)
 687                 init_gbpages();
 688
 689 #ifdef CONFIG_DEBUG_PAGEALLOC
 690         /*
 691          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 692          * This will simplify cpa(), which otherwise needs to support splitting
 693          * large pages into small in interrupt context, etc.
 694          */
 695         use_pse = use_gbpages = 0;
 696 #else
 697         use_pse = cpu_has_pse;
 698         use_gbpages = direct_gbpages;
 699 #endif
 700
 701         if (use_gbpages)
 702                 page_size_mask |= 1 << PG_LEVEL_1G;
 703         if (use_pse)
 704                 page_size_mask |= 1 << PG_LEVEL_2M;
 705
 706         memset(mr, 0, sizeof(mr));
 707         nr_range = 0;
 708
 709         /* head if not big page alignment ? */
 710         start_pfn = start >> PAGE_SHIFT;
 711         pos = start_pfn << PAGE_SHIFT;
 712         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 713                         << (PMD_SHIFT - PAGE_SHIFT);
 714         if (end_pfn > (end >> PAGE_SHIFT))
 715                 end_pfn = end >> PAGE_SHIFT;
 716         if (start_pfn < end_pfn) {
 717                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 718                 pos = end_pfn << PAGE_SHIFT;
 719         }
 720
 721         /* big page (2M) range */
 722         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 723                          << (PMD_SHIFT - PAGE_SHIFT);
 724         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 725                          << (PUD_SHIFT - PAGE_SHIFT);
 726         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 727                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 728         if (start_pfn < end_pfn) {
 729                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 730                                 page_size_mask & (1<<PG_LEVEL_2M));
 731                 pos = end_pfn << PAGE_SHIFT;
 732         }
 733
 734         /* big page (1G) range */
 735         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 736                          << (PUD_SHIFT - PAGE_SHIFT);
 737         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 738         if (start_pfn < end_pfn) {
 739                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 740                                 page_size_mask &
 741                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 742                 pos = end_pfn << PAGE_SHIFT;
 743         }
 744
 745         /* tail is not big page (1G) alignment */
 746         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 747                          << (PMD_SHIFT - PAGE_SHIFT);
 748         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 749         if (start_pfn < end_pfn) {
 750                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 751                                 page_size_mask & (1<<PG_LEVEL_2M));
 752                 pos = end_pfn << PAGE_SHIFT;
 753         }
 754
 755         /* tail is not big page (2M) alignment */
 756         start_pfn = pos>>PAGE_SHIFT;
 757         end_pfn = end>>PAGE_SHIFT;
 758         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 759
 760         /* try to merge same page size and continuous */
 761         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 762                 unsigned long old_start;
 763                 if (mr[i].end != mr[i+1].start ||
 764                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 765                         continue;
 766                 /* move it */
 767                 old_start = mr[i].start;
 768                 memmove(&mr[i], &mr[i+1],
 769                         (nr_range - 1 - i) * sizeof(struct map_range));
 770                 mr[i--].start = old_start;
 771                 nr_range--;
 772         }
 773
 774         for (i = 0; i < nr_range; i++)
 775                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 776                                 mr[i].start, mr[i].end,
 777                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 778                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 779
 780         /*
 781          * Find space for the kernel direct mapping tables.
 782          *
 783          * Later we should allocate these tables in the local node of the
 784          * memory mapped. Unfortunately this is done currently before the
 785          * nodes are discovered.
 786          */
 787         if (!after_bootmem)
 788                 find_early_table_space(end, use_pse, use_gbpages);
 789
 790         for (i = 0; i < nr_range; i++)
 791                 last_map_addr = kernel_physical_mapping_init(
 792                                         mr[i].start, mr[i].end,
 793                                         mr[i].page_size_mask);
 794
 795         if (!after_bootmem)
 796                 mmu_cr4_features = read_cr4();
 797         __flush_tlb_all();
 798
 799         if (!after_bootmem && table_end > table_start)
 800                 reserve_early(table_start << PAGE_SHIFT,
 801                                  table_end << PAGE_SHIFT, "PGTABLE");
 802
 803         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 804                          last_map_addr, end);
 805
 806         if (!after_bootmem)
 807                 early_memtest(start, end);
 808
 809         return last_map_addr >> PAGE_SHIFT;
 810 }
 811
 812 #ifndef CONFIG_NUMA
 813 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 814 {
 815         unsigned long bootmap_size, bootmap;
 816
 817         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 818         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 819                                  PAGE_SIZE);
 820         if (bootmap == -1L)
 821                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 822         /* don't touch min_low_pfn */
 823         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 824                                          0, end_pfn);
 825         e820_register_active_regions(0, start_pfn, end_pfn);
 826         free_bootmem_with_active_regions(0, end_pfn);
 827         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 828         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 829 }
 830
 831 void __init paging_init(void)
 832 {
 833         unsigned long max_zone_pfns[MAX_NR_ZONES];
 834
 835         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 836         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 837         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 838         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 839
 840         memory_present(0, 0, max_pfn);
 841         sparse_init();
 842         free_area_init_nodes(max_zone_pfns);
 843 }
 844 #endif
 845
 846 /*
 847  * Memory hotplug specific functions
 848  */
 849 #ifdef CONFIG_MEMORY_HOTPLUG
 850 /*
 851  * Memory is added always to NORMAL zone. This means you will never get
 852  * additional DMA/DMA32 memory.
 853  */
 854 int arch_add_memory(int nid, u64 start, u64 size)
 855 {
 856         struct pglist_data *pgdat = NODE_DATA(nid);
 857         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 858         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 859         unsigned long nr_pages = size >> PAGE_SHIFT;
 860         int ret;
 861
 862         last_mapped_pfn = init_memory_mapping(start, start + size);
 863         if (last_mapped_pfn > max_pfn_mapped)
 864                 max_pfn_mapped = last_mapped_pfn;
 865
 866         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 867         WARN_ON_ONCE(ret);
 868
 869         return ret;
 870 }
 871 EXPORT_SYMBOL_GPL(arch_add_memory);
 872
 873 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 874 int memory_add_physaddr_to_nid(u64 start)
 875 {
 876         return 0;
 877 }
 878 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 879 #endif
 880
 881 #endif /* CONFIG_MEMORY_HOTPLUG */
 882
 883 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 884                          kcore_modules, kcore_vsyscall;
 885
 886 void __init mem_init(void)
 887 {
 888         long codesize, reservedpages, datasize, initsize;
 889         unsigned long absent_pages;
 890
 891         pci_iommu_alloc();
 892
 893         /* clear_bss() already clear the empty_zero_page */
 894
 895         reservedpages = 0;
 896
 897         /* this will put all low memory onto the freelists */
 898 #ifdef CONFIG_NUMA
 899         totalram_pages = numa_free_all_bootmem();
 900 #else
 901         totalram_pages = free_all_bootmem();
 902 #endif
 903
 904         absent_pages = absent_pages_in_range(0, max_pfn);
 905         reservedpages = max_pfn - totalram_pages - absent_pages;
 906         after_bootmem = 1;
 907
 908         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 909         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 910         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 911
 912         /* Register memory areas for /proc/kcore */
 913         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 914         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 915                    VMALLOC_END-VMALLOC_START);
 916         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 917         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 918         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 919                                  VSYSCALL_END - VSYSCALL_START);
 920
 921         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 922                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 923                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 924                 max_pfn << (PAGE_SHIFT-10),
 925                 codesize >> 10,
 926                 absent_pages << (PAGE_SHIFT-10),
 927                 reservedpages << (PAGE_SHIFT-10),
 928                 datasize >> 10,
 929                 initsize >> 10);
 930 }
 931
 932 #ifdef CONFIG_DEBUG_RODATA
 933 const int rodata_test_data = 0xC3;
 934 EXPORT_SYMBOL_GPL(rodata_test_data);
 935
 936 void mark_rodata_ro(void)
 937 {
 938         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 939         unsigned long rodata_start =
 940                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 941
 942 #ifdef CONFIG_DYNAMIC_FTRACE
 943         /* Dynamic tracing modifies the kernel text section */
 944         start = rodata_start;
 945 #endif
 946
 947         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 948                (end - start) >> 10);
 949         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 950
 951         /*
 952          * The rodata section (but not the kernel text!) should also be
 953          * not-executable.
 954          */
 955         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 956
 957         rodata_test();
 958
 959 #ifdef CONFIG_CPA_DEBUG
 960         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 961         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 962
 963         printk(KERN_INFO "Testing CPA: again\n");
 964         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 965 #endif
 966 }
 967
 968 #endif
 969
 970 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 971                                    int flags)
 972 {
 973 #ifdef CONFIG_NUMA
 974         int nid, next_nid;
 975         int ret;
 976 #endif
 977         unsigned long pfn = phys >> PAGE_SHIFT;
 978
 979         if (pfn >= max_pfn) {
 980                 /*
 981                  * This can happen with kdump kernels when accessing
 982                  * firmware tables:
 983                  */
 984                 if (pfn < max_pfn_mapped)
 985                         return -EFAULT;
 986
 987                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 988                                 phys, len);
 989                 return -EFAULT;
 990         }
 991
 992         /* Should check here against the e820 map to avoid double free */
 993 #ifdef CONFIG_NUMA
 994         nid = phys_to_nid(phys);
 995         next_nid = phys_to_nid(phys + len - 1);
 996         if (nid == next_nid)
 997                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 998         else
 999                 ret = reserve_bootmem(phys, len, flags);
1000
1001         if (ret != 0)
1002                 return ret;
1003
1004 #else
1005         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1006 #endif
1007
1008         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1009                 dma_reserve += len / PAGE_SIZE;
1010                 set_dma_reserve(dma_reserve);
1011         }
1012
1013         return 0;
1014 }
1015
1016 int kern_addr_valid(unsigned long addr)
1017 {
1018         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1019         pgd_t *pgd;
1020         pud_t *pud;
1021         pmd_t *pmd;
1022         pte_t *pte;
1023
1024         if (above != 0 && above != -1UL)
1025                 return 0;
1026
1027         pgd = pgd_offset_k(addr);
1028         if (pgd_none(*pgd))
1029                 return 0;
1030
1031         pud = pud_offset(pgd, addr);
1032         if (pud_none(*pud))
1033                 return 0;
1034
1035         pmd = pmd_offset(pud, addr);
1036         if (pmd_none(*pmd))
1037                 return 0;
1038
1039         if (pmd_large(*pmd))
1040                 return pfn_valid(pmd_pfn(*pmd));
1041
1042         pte = pte_offset_kernel(pmd, addr);
1043         if (pte_none(*pte))
1044                 return 0;
1045
1046         return pfn_valid(pte_pfn(*pte));
1047 }
1048
1049 /*
1050  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1051  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1052  * not need special handling anymore:
1053  */
1054 static struct vm_area_struct gate_vma = {
1055         .vm_start       = VSYSCALL_START,
1056         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1057         .vm_page_prot   = PAGE_READONLY_EXEC,
1058         .vm_flags       = VM_READ | VM_EXEC
1059 };
1060
1061 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1062 {
1063 #ifdef CONFIG_IA32_EMULATION
1064         if (test_tsk_thread_flag(tsk, TIF_IA32))
1065                 return NULL;
1066 #endif
1067         return &gate_vma;
1068 }
1069
1070 int in_gate_area(struct task_struct *task, unsigned long addr)
1071 {
1072         struct vm_area_struct *vma = get_gate_vma(task);
1073
1074         if (!vma)
1075                 return 0;
1076
1077         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1078 }
1079
1080 /*
1081  * Use this when you have no reliable task/vma, typically from interrupt
1082  * context. It is less reliable than using the task's vma and may give
1083  * false positives:
1084  */
1085 int in_gate_area_no_task(unsigned long addr)
1086 {
1087         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1088 }
1089
1090 const char *arch_vma_name(struct vm_area_struct *vma)
1091 {
1092         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1093                 return "[vdso]";
1094         if (vma == &gate_vma)
1095                 return "[vsyscall]";
1096         return NULL;
1097 }
1098
1099 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1100 /*
1101  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1102  */
1103 static long __meminitdata addr_start, addr_end;
1104 static void __meminitdata *p_start, *p_end;
1105 static int __meminitdata node_start;
1106
1107 int __meminit
1108 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1109 {
1110         unsigned long addr = (unsigned long)start_page;
1111         unsigned long end = (unsigned long)(start_page + size);
1112         unsigned long next;
1113         pgd_t *pgd;
1114         pud_t *pud;
1115         pmd_t *pmd;
1116
1117         for (; addr < end; addr = next) {
1118                 void *p = NULL;
1119
1120                 pgd = vmemmap_pgd_populate(addr, node);
1121                 if (!pgd)
1122                         return -ENOMEM;
1123
1124                 pud = vmemmap_pud_populate(pgd, addr, node);
1125                 if (!pud)
1126                         return -ENOMEM;
1127
1128                 if (!cpu_has_pse) {
1129                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1130                         pmd = vmemmap_pmd_populate(pud, addr, node);
1131
1132                         if (!pmd)
1133                                 return -ENOMEM;
1134
1135                         p = vmemmap_pte_populate(pmd, addr, node);
1136
1137                         if (!p)
1138                                 return -ENOMEM;
1139
1140                         addr_end = addr + PAGE_SIZE;
1141                         p_end = p + PAGE_SIZE;
1142                 } else {
1143                         next = pmd_addr_end(addr, end);
1144
1145                         pmd = pmd_offset(pud, addr);
1146                         if (pmd_none(*pmd)) {
1147                                 pte_t entry;
1148
1149                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1150                                 if (!p)
1151                                         return -ENOMEM;
1152
1153                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1154                                                 PAGE_KERNEL_LARGE);
1155                                 set_pmd(pmd, __pmd(pte_val(entry)));
1156
1157                                 /* check to see if we have contiguous blocks */
1158                                 if (p_end != p || node_start != node) {
1159                                         if (p_start)
1160                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1161                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1162                                         addr_start = addr;
1163                                         node_start = node;
1164                                         p_start = p;
1165                                 }
1166
1167                                 addr_end = addr + PMD_SIZE;
1168                                 p_end = p + PMD_SIZE;
1169                         } else
1170                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1171                 }
1172
1173         }
1174         return 0;
1175 }
1176
1177 void __meminit vmemmap_populate_print_last(void)
1178 {
1179         if (p_start) {
1180                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1181                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1182                 p_start = NULL;
1183                 p_end = NULL;
1184                 node_start = 0;
1185         }
1186 }
1187 #endif