git.oblomov.eu Git - linux-2.6/blob - arch/x86/mm/init_64.c

   1 /*
   2  *  linux/arch/x86_64/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
   6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
   7  */
   8
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/string.h>
  14 #include <linux/types.h>
  15 #include <linux/ptrace.h>
  16 #include <linux/mman.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19 #include <linux/smp.h>
  20 #include <linux/init.h>
  21 #include <linux/initrd.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/proc_fs.h>
  25 #include <linux/pci.h>
  26 #include <linux/pfn.h>
  27 #include <linux/poison.h>
  28 #include <linux/dma-mapping.h>
  29 #include <linux/module.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/nmi.h>
  32
  33 #include <asm/processor.h>
  34 #include <asm/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 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 172 {
 173         if (pgd_none(*pgd)) {
 174                 pud_t *pud = (pud_t *)spp_getpage();
 175                 pgd_populate(&init_mm, pgd, pud);
 176                 if (pud != pud_offset(pgd, 0))
 177                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 178                                pud, pud_offset(pgd, 0));
 179         }
 180         return pud_offset(pgd, vaddr);
 181 }
 182
 183 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 184 {
 185         if (pud_none(*pud)) {
 186                 pmd_t *pmd = (pmd_t *) spp_getpage();
 187                 pud_populate(&init_mm, pud, pmd);
 188                 if (pmd != pmd_offset(pud, 0))
 189                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 190                                pmd, pmd_offset(pud, 0));
 191         }
 192         return pmd_offset(pud, vaddr);
 193 }
 194
 195 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 196 {
 197         if (pmd_none(*pmd)) {
 198                 pte_t *pte = (pte_t *) spp_getpage();
 199                 pmd_populate_kernel(&init_mm, pmd, pte);
 200                 if (pte != pte_offset_kernel(pmd, 0))
 201                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 202         }
 203         return pte_offset_kernel(pmd, vaddr);
 204 }
 205
 206 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 207 {
 208         pud_t *pud;
 209         pmd_t *pmd;
 210         pte_t *pte;
 211
 212         pud = pud_page + pud_index(vaddr);
 213         pmd = fill_pmd(pud, vaddr);
 214         pte = fill_pte(pmd, vaddr);
 215
 216         set_pte(pte, new_pte);
 217
 218         /*
 219          * It's enough to flush this one mapping.
 220          * (PGE mappings get flushed as well)
 221          */
 222         __flush_tlb_one(vaddr);
 223 }
 224
 225 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 226 {
 227         pgd_t *pgd;
 228         pud_t *pud_page;
 229
 230         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 231
 232         pgd = pgd_offset_k(vaddr);
 233         if (pgd_none(*pgd)) {
 234                 printk(KERN_ERR
 235                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 236                 return;
 237         }
 238         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 239         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 240 }
 241
 242 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 243 {
 244         pgd_t *pgd;
 245         pud_t *pud;
 246
 247         pgd = pgd_offset_k(vaddr);
 248         pud = fill_pud(pgd, vaddr);
 249         return fill_pmd(pud, vaddr);
 250 }
 251
 252 pte_t * __init populate_extra_pte(unsigned long vaddr)
 253 {
 254         pmd_t *pmd;
 255
 256         pmd = populate_extra_pmd(vaddr);
 257         return fill_pte(pmd, vaddr);
 258 }
 259
 260 /*
 261  * Create large page table mappings for a range of physical addresses.
 262  */
 263 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 264                                                 pgprot_t prot)
 265 {
 266         pgd_t *pgd;
 267         pud_t *pud;
 268         pmd_t *pmd;
 269
 270         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 271         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 272                 pgd = pgd_offset_k((unsigned long)__va(phys));
 273                 if (pgd_none(*pgd)) {
 274                         pud = (pud_t *) spp_getpage();
 275                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 276                                                 _PAGE_USER));
 277                 }
 278                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 279                 if (pud_none(*pud)) {
 280                         pmd = (pmd_t *) spp_getpage();
 281                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 282                                                 _PAGE_USER));
 283                 }
 284                 pmd = pmd_offset(pud, phys);
 285                 BUG_ON(!pmd_none(*pmd));
 286                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 287         }
 288 }
 289
 290 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 291 {
 292         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 293 }
 294
 295 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 296 {
 297         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 298 }
 299
 300 /*
 301  * The head.S code sets up the kernel high mapping:
 302  *
 303  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 304  *
 305  * phys_addr holds the negative offset to the kernel, which is added
 306  * to the compile time generated pmds. This results in invalid pmds up
 307  * to the point where we hit the physaddr 0 mapping.
 308  *
 309  * We limit the mappings to the region from _text to _end.  _end is
 310  * rounded up to the 2MB boundary. This catches the invalid pmds as
 311  * well, as they are located before _text:
 312  */
 313 void __init cleanup_highmap(void)
 314 {
 315         unsigned long vaddr = __START_KERNEL_map;
 316         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
 317         pmd_t *pmd = level2_kernel_pgt;
 318         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 319
 320         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 321                 if (pmd_none(*pmd))
 322                         continue;
 323                 if (vaddr < (unsigned long) _text || vaddr > end)
 324                         set_pmd(pmd, __pmd(0));
 325         }
 326 }
 327
 328 static unsigned long __initdata table_start;
 329 static unsigned long __meminitdata table_end;
 330 static unsigned long __meminitdata table_top;
 331
 332 static __ref void *alloc_low_page(unsigned long *phys)
 333 {
 334         unsigned long pfn = table_end++;
 335         void *adr;
 336
 337         if (after_bootmem) {
 338                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 339                 *phys = __pa(adr);
 340
 341                 return adr;
 342         }
 343
 344         if (pfn >= table_top)
 345                 panic("alloc_low_page: ran out of memory");
 346
 347         adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
 348         memset(adr, 0, PAGE_SIZE);
 349         *phys  = pfn * PAGE_SIZE;
 350         return adr;
 351 }
 352
 353 static __ref void unmap_low_page(void *adr)
 354 {
 355         if (after_bootmem)
 356                 return;
 357
 358         early_iounmap(adr, PAGE_SIZE);
 359 }
 360
 361 static unsigned long __meminit
 362 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 363               pgprot_t prot)
 364 {
 365         unsigned pages = 0;
 366         unsigned long last_map_addr = end;
 367         int i;
 368
 369         pte_t *pte = pte_page + pte_index(addr);
 370
 371         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 372
 373                 if (addr >= end) {
 374                         if (!after_bootmem) {
 375                                 for(; i < PTRS_PER_PTE; i++, pte++)
 376                                         set_pte(pte, __pte(0));
 377                         }
 378                         break;
 379                 }
 380
 381                 /*
 382                  * We will re-use the existing mapping.
 383                  * Xen for example has some special requirements, like mapping
 384                  * pagetable pages as RO. So assume someone who pre-setup
 385                  * these mappings are more intelligent.
 386                  */
 387                 if (pte_val(*pte)) {
 388                         pages++;
 389                         continue;
 390                 }
 391
 392                 if (0)
 393                         printk("   pte=%p addr=%lx pte=%016lx\n",
 394                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 395                 pages++;
 396                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 397                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 398         }
 399
 400         update_page_count(PG_LEVEL_4K, pages);
 401
 402         return last_map_addr;
 403 }
 404
 405 static unsigned long __meminit
 406 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end,
 407                 pgprot_t prot)
 408 {
 409         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 410
 411         return phys_pte_init(pte, address, end, prot);
 412 }
 413
 414 static unsigned long __meminit
 415 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 416               unsigned long page_size_mask, pgprot_t prot)
 417 {
 418         unsigned long pages = 0;
 419         unsigned long last_map_addr = end;
 420
 421         int i = pmd_index(address);
 422
 423         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 424                 unsigned long pte_phys;
 425                 pmd_t *pmd = pmd_page + pmd_index(address);
 426                 pte_t *pte;
 427                 pgprot_t new_prot = prot;
 428
 429                 if (address >= end) {
 430                         if (!after_bootmem) {
 431                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 432                                         set_pmd(pmd, __pmd(0));
 433                         }
 434                         break;
 435                 }
 436
 437                 if (pmd_val(*pmd)) {
 438                         if (!pmd_large(*pmd)) {
 439                                 spin_lock(&init_mm.page_table_lock);
 440                                 last_map_addr = phys_pte_update(pmd, address,
 441                                                                 end, prot);
 442                                 spin_unlock(&init_mm.page_table_lock);
 443                                 continue;
 444                         }
 445                         /*
 446                          * If we are ok with PG_LEVEL_2M mapping, then we will
 447                          * use the existing mapping,
 448                          *
 449                          * Otherwise, we will split the large page mapping but
 450                          * use the same existing protection bits except for
 451                          * large page, so that we don't violate Intel's TLB
 452                          * Application note (317080) which says, while changing
 453                          * the page sizes, new and old translations should
 454                          * not differ with respect to page frame and
 455                          * attributes.
 456                          */
 457                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 458                                 pages++;
 459                                 continue;
 460                         }
 461                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 462                 }
 463
 464                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 465                         pages++;
 466                         spin_lock(&init_mm.page_table_lock);
 467                         set_pte((pte_t *)pmd,
 468                                 pfn_pte(address >> PAGE_SHIFT,
 469                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 470                         spin_unlock(&init_mm.page_table_lock);
 471                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 472                         continue;
 473                 }
 474
 475                 pte = alloc_low_page(&pte_phys);
 476                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 477                 unmap_low_page(pte);
 478
 479                 spin_lock(&init_mm.page_table_lock);
 480                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 481                 spin_unlock(&init_mm.page_table_lock);
 482         }
 483         update_page_count(PG_LEVEL_2M, pages);
 484         return last_map_addr;
 485 }
 486
 487 static unsigned long __meminit
 488 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 489                 unsigned long page_size_mask, pgprot_t prot)
 490 {
 491         pmd_t *pmd = pmd_offset(pud, 0);
 492         unsigned long last_map_addr;
 493
 494         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
 495         __flush_tlb_all();
 496         return last_map_addr;
 497 }
 498
 499 static unsigned long __meminit
 500 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 501                          unsigned long page_size_mask)
 502 {
 503         unsigned long pages = 0;
 504         unsigned long last_map_addr = end;
 505         int i = pud_index(addr);
 506
 507         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 508                 unsigned long pmd_phys;
 509                 pud_t *pud = pud_page + pud_index(addr);
 510                 pmd_t *pmd;
 511                 pgprot_t prot = PAGE_KERNEL;
 512
 513                 if (addr >= end)
 514                         break;
 515
 516                 if (!after_bootmem &&
 517                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 518                         set_pud(pud, __pud(0));
 519                         continue;
 520                 }
 521
 522                 if (pud_val(*pud)) {
 523                         if (!pud_large(*pud)) {
 524                                 last_map_addr = phys_pmd_update(pud, addr, end,
 525                                                          page_size_mask, prot);
 526                                 continue;
 527                         }
 528                         /*
 529                          * If we are ok with PG_LEVEL_1G mapping, then we will
 530                          * use the existing mapping.
 531                          *
 532                          * Otherwise, we will split the gbpage mapping but use
 533                          * the same existing protection  bits except for large
 534                          * page, so that we don't violate Intel's TLB
 535                          * Application note (317080) which says, while changing
 536                          * the page sizes, new and old translations should
 537                          * not differ with respect to page frame and
 538                          * attributes.
 539                          */
 540                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 541                                 pages++;
 542                                 continue;
 543                         }
 544                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 545                 }
 546
 547                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 548                         pages++;
 549                         spin_lock(&init_mm.page_table_lock);
 550                         set_pte((pte_t *)pud,
 551                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 552                         spin_unlock(&init_mm.page_table_lock);
 553                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 554                         continue;
 555                 }
 556
 557                 pmd = alloc_low_page(&pmd_phys);
 558                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 559                                               prot);
 560                 unmap_low_page(pmd);
 561
 562                 spin_lock(&init_mm.page_table_lock);
 563                 pud_populate(&init_mm, pud, __va(pmd_phys));
 564                 spin_unlock(&init_mm.page_table_lock);
 565         }
 566         __flush_tlb_all();
 567
 568         update_page_count(PG_LEVEL_1G, pages);
 569
 570         return last_map_addr;
 571 }
 572
 573 static unsigned long __meminit
 574 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 575                  unsigned long page_size_mask)
 576 {
 577         pud_t *pud;
 578
 579         pud = (pud_t *)pgd_page_vaddr(*pgd);
 580
 581         return phys_pud_init(pud, addr, end, page_size_mask);
 582 }
 583
 584 static void __init find_early_table_space(unsigned long end, int use_pse,
 585                                           int use_gbpages)
 586 {
 587         unsigned long puds, pmds, ptes, tables, start;
 588
 589         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 590         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
 591         if (use_gbpages) {
 592                 unsigned long extra;
 593                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 594                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 595         } else
 596                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 597         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 598
 599         if (use_pse) {
 600                 unsigned long extra;
 601                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 602                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 603         } else
 604                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 605         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 606
 607         /*
 608          * RED-PEN putting page tables only on node 0 could
 609          * cause a hotspot and fill up ZONE_DMA. The page tables
 610          * need roughly 0.5KB per GB.
 611          */
 612         start = 0x8000;
 613         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 614         if (table_start == -1UL)
 615                 panic("Cannot find space for the kernel page tables");
 616
 617         table_start >>= PAGE_SHIFT;
 618         table_end = table_start;
 619         table_top = table_start + (tables >> PAGE_SHIFT);
 620
 621         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 622                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 623 }
 624
 625 static void __init init_gbpages(void)
 626 {
 627         if (direct_gbpages && cpu_has_gbpages)
 628                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 629         else
 630                 direct_gbpages = 0;
 631 }
 632
 633 static unsigned long __meminit kernel_physical_mapping_init(unsigned long start,
 634                                                 unsigned long end,
 635                                                 unsigned long page_size_mask)
 636 {
 637
 638         unsigned long next, last_map_addr = end;
 639
 640         start = (unsigned long)__va(start);
 641         end = (unsigned long)__va(end);
 642
 643         for (; start < end; start = next) {
 644                 pgd_t *pgd = pgd_offset_k(start);
 645                 unsigned long pud_phys;
 646                 pud_t *pud;
 647
 648                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 649                 if (next > end)
 650                         next = end;
 651
 652                 if (pgd_val(*pgd)) {
 653                         last_map_addr = phys_pud_update(pgd, __pa(start),
 654                                                  __pa(end), page_size_mask);
 655                         continue;
 656                 }
 657
 658                 pud = alloc_low_page(&pud_phys);
 659                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 660                                                  page_size_mask);
 661                 unmap_low_page(pud);
 662
 663                 spin_lock(&init_mm.page_table_lock);
 664                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 665                 spin_unlock(&init_mm.page_table_lock);
 666         }
 667         __flush_tlb_all();
 668
 669         return last_map_addr;
 670 }
 671
 672 struct map_range {
 673         unsigned long start;
 674         unsigned long end;
 675         unsigned page_size_mask;
 676 };
 677
 678 #define NR_RANGE_MR 5
 679
 680 static int save_mr(struct map_range *mr, int nr_range,
 681                    unsigned long start_pfn, unsigned long end_pfn,
 682                    unsigned long page_size_mask)
 683 {
 684
 685         if (start_pfn < end_pfn) {
 686                 if (nr_range >= NR_RANGE_MR)
 687                         panic("run out of range for init_memory_mapping\n");
 688                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 689                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 690                 mr[nr_range].page_size_mask = page_size_mask;
 691                 nr_range++;
 692         }
 693
 694         return nr_range;
 695 }
 696
 697 /*
 698  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 699  * This runs before bootmem is initialized and gets pages directly from
 700  * the physical memory. To access them they are temporarily mapped.
 701  */
 702 unsigned long __init_refok init_memory_mapping(unsigned long start,
 703                                                unsigned long end)
 704 {
 705         unsigned long last_map_addr = 0;
 706         unsigned long page_size_mask = 0;
 707         unsigned long start_pfn, end_pfn;
 708         unsigned long pos;
 709
 710         struct map_range mr[NR_RANGE_MR];
 711         int nr_range, i;
 712         int use_pse, use_gbpages;
 713
 714         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 715
 716         /*
 717          * Find space for the kernel direct mapping tables.
 718          *
 719          * Later we should allocate these tables in the local node of the
 720          * memory mapped. Unfortunately this is done currently before the
 721          * nodes are discovered.
 722          */
 723         if (!after_bootmem)
 724                 init_gbpages();
 725
 726 #ifdef CONFIG_DEBUG_PAGEALLOC
 727         /*
 728          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 729          * This will simplify cpa(), which otherwise needs to support splitting
 730          * large pages into small in interrupt context, etc.
 731          */
 732         use_pse = use_gbpages = 0;
 733 #else
 734         use_pse = cpu_has_pse;
 735         use_gbpages = direct_gbpages;
 736 #endif
 737
 738         if (use_gbpages)
 739                 page_size_mask |= 1 << PG_LEVEL_1G;
 740         if (use_pse)
 741                 page_size_mask |= 1 << PG_LEVEL_2M;
 742
 743         memset(mr, 0, sizeof(mr));
 744         nr_range = 0;
 745
 746         /* head if not big page alignment ?*/
 747         start_pfn = start >> PAGE_SHIFT;
 748         pos = start_pfn << PAGE_SHIFT;
 749         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 750                         << (PMD_SHIFT - PAGE_SHIFT);
 751         if (end_pfn > (end >> PAGE_SHIFT))
 752                 end_pfn = end >> PAGE_SHIFT;
 753         if (start_pfn < end_pfn) {
 754                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 755                 pos = end_pfn << PAGE_SHIFT;
 756         }
 757
 758         /* big page (2M) range*/
 759         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 760                          << (PMD_SHIFT - PAGE_SHIFT);
 761         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 762                          << (PUD_SHIFT - PAGE_SHIFT);
 763         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 764                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 765         if (start_pfn < end_pfn) {
 766                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 767                                 page_size_mask & (1<<PG_LEVEL_2M));
 768                 pos = end_pfn << PAGE_SHIFT;
 769         }
 770
 771         /* big page (1G) range */
 772         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 773                          << (PUD_SHIFT - PAGE_SHIFT);
 774         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 775         if (start_pfn < end_pfn) {
 776                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 777                                 page_size_mask &
 778                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 779                 pos = end_pfn << PAGE_SHIFT;
 780         }
 781
 782         /* tail is not big page (1G) alignment */
 783         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 784                          << (PMD_SHIFT - PAGE_SHIFT);
 785         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 786         if (start_pfn < end_pfn) {
 787                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 788                                 page_size_mask & (1<<PG_LEVEL_2M));
 789                 pos = end_pfn << PAGE_SHIFT;
 790         }
 791
 792         /* tail is not big page (2M) alignment */
 793         start_pfn = pos>>PAGE_SHIFT;
 794         end_pfn = end>>PAGE_SHIFT;
 795         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 796
 797         /* try to merge same page size and continuous */
 798         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 799                 unsigned long old_start;
 800                 if (mr[i].end != mr[i+1].start ||
 801                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 802                         continue;
 803                 /* move it */
 804                 old_start = mr[i].start;
 805                 memmove(&mr[i], &mr[i+1],
 806                          (nr_range - 1 - i) * sizeof (struct map_range));
 807                 mr[i--].start = old_start;
 808                 nr_range--;
 809         }
 810
 811         for (i = 0; i < nr_range; i++)
 812                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 813                                 mr[i].start, mr[i].end,
 814                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 815                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 816
 817         if (!after_bootmem)
 818                 find_early_table_space(end, use_pse, use_gbpages);
 819
 820         for (i = 0; i < nr_range; i++)
 821                 last_map_addr = kernel_physical_mapping_init(
 822                                         mr[i].start, mr[i].end,
 823                                         mr[i].page_size_mask);
 824
 825         if (!after_bootmem)
 826                 mmu_cr4_features = read_cr4();
 827         __flush_tlb_all();
 828
 829         if (!after_bootmem && table_end > table_start)
 830                 reserve_early(table_start << PAGE_SHIFT,
 831                                  table_end << PAGE_SHIFT, "PGTABLE");
 832
 833         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 834                          last_map_addr, end);
 835
 836         if (!after_bootmem)
 837                 early_memtest(start, end);
 838
 839         return last_map_addr >> PAGE_SHIFT;
 840 }
 841
 842 #ifndef CONFIG_NUMA
 843 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 844 {
 845         unsigned long bootmap_size, bootmap;
 846
 847         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 848         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 849                                  PAGE_SIZE);
 850         if (bootmap == -1L)
 851                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 852         /* don't touch min_low_pfn */
 853         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 854                                          0, end_pfn);
 855         e820_register_active_regions(0, start_pfn, end_pfn);
 856         free_bootmem_with_active_regions(0, end_pfn);
 857         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 858         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 859 }
 860
 861 void __init paging_init(void)
 862 {
 863         unsigned long max_zone_pfns[MAX_NR_ZONES];
 864
 865         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 866         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 867         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 868         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 869
 870         memory_present(0, 0, max_pfn);
 871         sparse_init();
 872         free_area_init_nodes(max_zone_pfns);
 873 }
 874 #endif
 875
 876 /*
 877  * Memory hotplug specific functions
 878  */
 879 #ifdef CONFIG_MEMORY_HOTPLUG
 880 /*
 881  * Memory is added always to NORMAL zone. This means you will never get
 882  * additional DMA/DMA32 memory.
 883  */
 884 int arch_add_memory(int nid, u64 start, u64 size)
 885 {
 886         struct pglist_data *pgdat = NODE_DATA(nid);
 887         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 888         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 889         unsigned long nr_pages = size >> PAGE_SHIFT;
 890         int ret;
 891
 892         last_mapped_pfn = init_memory_mapping(start, start + size);
 893         if (last_mapped_pfn > max_pfn_mapped)
 894                 max_pfn_mapped = last_mapped_pfn;
 895
 896         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 897         WARN_ON_ONCE(ret);
 898
 899         return ret;
 900 }
 901 EXPORT_SYMBOL_GPL(arch_add_memory);
 902
 903 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 904 int memory_add_physaddr_to_nid(u64 start)
 905 {
 906         return 0;
 907 }
 908 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 909 #endif
 910
 911 #endif /* CONFIG_MEMORY_HOTPLUG */
 912
 913 /*
 914  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 915  * is valid. The argument is a physical page number.
 916  *
 917  *
 918  * On x86, access has to be given to the first megabyte of ram because that area
 919  * contains bios code and data regions used by X and dosemu and similar apps.
 920  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 921  * mmio resources as well as potential bios/acpi data regions.
 922  */
 923 int devmem_is_allowed(unsigned long pagenr)
 924 {
 925         if (pagenr <= 256)
 926                 return 1;
 927         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 928                 return 0;
 929         if (!page_is_ram(pagenr))
 930                 return 1;
 931         return 0;
 932 }
 933
 934
 935 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 936                          kcore_modules, kcore_vsyscall;
 937
 938 void __init mem_init(void)
 939 {
 940         long codesize, reservedpages, datasize, initsize;
 941         unsigned long absent_pages;
 942
 943         pci_iommu_alloc();
 944
 945         /* clear_bss() already clear the empty_zero_page */
 946
 947         reservedpages = 0;
 948
 949         /* this will put all low memory onto the freelists */
 950 #ifdef CONFIG_NUMA
 951         totalram_pages = numa_free_all_bootmem();
 952 #else
 953         totalram_pages = free_all_bootmem();
 954 #endif
 955
 956         absent_pages = absent_pages_in_range(0, max_pfn);
 957         reservedpages = max_pfn - totalram_pages - absent_pages;
 958         after_bootmem = 1;
 959
 960         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 961         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 962         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 963
 964         /* Register memory areas for /proc/kcore */
 965         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 966         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 967                    VMALLOC_END-VMALLOC_START);
 968         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 969         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 970         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 971                                  VSYSCALL_END - VSYSCALL_START);
 972
 973         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 974                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 975                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 976                 max_pfn << (PAGE_SHIFT-10),
 977                 codesize >> 10,
 978                 absent_pages << (PAGE_SHIFT-10),
 979                 reservedpages << (PAGE_SHIFT-10),
 980                 datasize >> 10,
 981                 initsize >> 10);
 982 }
 983
 984 #ifdef CONFIG_DEBUG_RODATA
 985 const int rodata_test_data = 0xC3;
 986 EXPORT_SYMBOL_GPL(rodata_test_data);
 987
 988 void mark_rodata_ro(void)
 989 {
 990         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 991         unsigned long rodata_start =
 992                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 993
 994 #ifdef CONFIG_DYNAMIC_FTRACE
 995         /* Dynamic tracing modifies the kernel text section */
 996         start = rodata_start;
 997 #endif
 998
 999         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1000                (end - start) >> 10);
1001         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1002
1003         /*
1004          * The rodata section (but not the kernel text!) should also be
1005          * not-executable.
1006          */
1007         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
1008
1009         rodata_test();
1010
1011 #ifdef CONFIG_CPA_DEBUG
1012         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1013         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1014
1015         printk(KERN_INFO "Testing CPA: again\n");
1016         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1017 #endif
1018 }
1019
1020 #endif
1021
1022 #ifdef CONFIG_BLK_DEV_INITRD
1023 void free_initrd_mem(unsigned long start, unsigned long end)
1024 {
1025         free_init_pages("initrd memory", start, end);
1026 }
1027 #endif
1028
1029 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1030                                    int flags)
1031 {
1032 #ifdef CONFIG_NUMA
1033         int nid, next_nid;
1034         int ret;
1035 #endif
1036         unsigned long pfn = phys >> PAGE_SHIFT;
1037
1038         if (pfn >= max_pfn) {
1039                 /*
1040                  * This can happen with kdump kernels when accessing
1041                  * firmware tables:
1042                  */
1043                 if (pfn < max_pfn_mapped)
1044                         return -EFAULT;
1045
1046                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
1047                                 phys, len);
1048                 return -EFAULT;
1049         }
1050
1051         /* Should check here against the e820 map to avoid double free */
1052 #ifdef CONFIG_NUMA
1053         nid = phys_to_nid(phys);
1054         next_nid = phys_to_nid(phys + len - 1);
1055         if (nid == next_nid)
1056                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
1057         else
1058                 ret = reserve_bootmem(phys, len, flags);
1059
1060         if (ret != 0)
1061                 return ret;
1062
1063 #else
1064         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1065 #endif
1066
1067         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1068                 dma_reserve += len / PAGE_SIZE;
1069                 set_dma_reserve(dma_reserve);
1070         }
1071
1072         return 0;
1073 }
1074
1075 int kern_addr_valid(unsigned long addr)
1076 {
1077         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1078         pgd_t *pgd;
1079         pud_t *pud;
1080         pmd_t *pmd;
1081         pte_t *pte;
1082
1083         if (above != 0 && above != -1UL)
1084                 return 0;
1085
1086         pgd = pgd_offset_k(addr);
1087         if (pgd_none(*pgd))
1088                 return 0;
1089
1090         pud = pud_offset(pgd, addr);
1091         if (pud_none(*pud))
1092                 return 0;
1093
1094         pmd = pmd_offset(pud, addr);
1095         if (pmd_none(*pmd))
1096                 return 0;
1097
1098         if (pmd_large(*pmd))
1099                 return pfn_valid(pmd_pfn(*pmd));
1100
1101         pte = pte_offset_kernel(pmd, addr);
1102         if (pte_none(*pte))
1103                 return 0;
1104
1105         return pfn_valid(pte_pfn(*pte));
1106 }
1107
1108 /*
1109  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1110  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1111  * not need special handling anymore:
1112  */
1113 static struct vm_area_struct gate_vma = {
1114         .vm_start       = VSYSCALL_START,
1115         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1116         .vm_page_prot   = PAGE_READONLY_EXEC,
1117         .vm_flags       = VM_READ | VM_EXEC
1118 };
1119
1120 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1121 {
1122 #ifdef CONFIG_IA32_EMULATION
1123         if (test_tsk_thread_flag(tsk, TIF_IA32))
1124                 return NULL;
1125 #endif
1126         return &gate_vma;
1127 }
1128
1129 int in_gate_area(struct task_struct *task, unsigned long addr)
1130 {
1131         struct vm_area_struct *vma = get_gate_vma(task);
1132
1133         if (!vma)
1134                 return 0;
1135
1136         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1137 }
1138
1139 /*
1140  * Use this when you have no reliable task/vma, typically from interrupt
1141  * context. It is less reliable than using the task's vma and may give
1142  * false positives:
1143  */
1144 int in_gate_area_no_task(unsigned long addr)
1145 {
1146         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1147 }
1148
1149 const char *arch_vma_name(struct vm_area_struct *vma)
1150 {
1151         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1152                 return "[vdso]";
1153         if (vma == &gate_vma)
1154                 return "[vsyscall]";
1155         return NULL;
1156 }
1157
1158 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1159 /*
1160  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1161  */
1162 static long __meminitdata addr_start, addr_end;
1163 static void __meminitdata *p_start, *p_end;
1164 static int __meminitdata node_start;
1165
1166 int __meminit
1167 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1168 {
1169         unsigned long addr = (unsigned long)start_page;
1170         unsigned long end = (unsigned long)(start_page + size);
1171         unsigned long next;
1172         pgd_t *pgd;
1173         pud_t *pud;
1174         pmd_t *pmd;
1175
1176         for (; addr < end; addr = next) {
1177                 void *p = NULL;
1178
1179                 pgd = vmemmap_pgd_populate(addr, node);
1180                 if (!pgd)
1181                         return -ENOMEM;
1182
1183                 pud = vmemmap_pud_populate(pgd, addr, node);
1184                 if (!pud)
1185                         return -ENOMEM;
1186
1187                 if (!cpu_has_pse) {
1188                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1189                         pmd = vmemmap_pmd_populate(pud, addr, node);
1190
1191                         if (!pmd)
1192                                 return -ENOMEM;
1193
1194                         p = vmemmap_pte_populate(pmd, addr, node);
1195
1196                         if (!p)
1197                                 return -ENOMEM;
1198
1199                         addr_end = addr + PAGE_SIZE;
1200                         p_end = p + PAGE_SIZE;
1201                 } else {
1202                         next = pmd_addr_end(addr, end);
1203
1204                         pmd = pmd_offset(pud, addr);
1205                         if (pmd_none(*pmd)) {
1206                                 pte_t entry;
1207
1208                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1209                                 if (!p)
1210                                         return -ENOMEM;
1211
1212                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1213                                                 PAGE_KERNEL_LARGE);
1214                                 set_pmd(pmd, __pmd(pte_val(entry)));
1215
1216                                 /* check to see if we have contiguous blocks */
1217                                 if (p_end != p || node_start != node) {
1218                                         if (p_start)
1219                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1220                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1221                                         addr_start = addr;
1222                                         node_start = node;
1223                                         p_start = p;
1224                                 }
1225
1226                                 addr_end = addr + PMD_SIZE;
1227                                 p_end = p + PMD_SIZE;
1228                         } else
1229                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1230                 }
1231
1232         }
1233         return 0;
1234 }
1235
1236 void __meminit vmemmap_populate_print_last(void)
1237 {
1238         if (p_start) {
1239                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1240                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1241                 p_start = NULL;
1242                 p_end = NULL;
1243                 node_start = 0;
1244         }
1245 }
1246 #endif