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 static int __init parse_direct_gbpages_off(char *arg)
  65 {
  66         direct_gbpages = 0;
  67         return 0;
  68 }
  69 early_param("nogbpages", parse_direct_gbpages_off);
  70
  71 static int __init parse_direct_gbpages_on(char *arg)
  72 {
  73         direct_gbpages = 1;
  74         return 0;
  75 }
  76 early_param("gbpages", parse_direct_gbpages_on);
  77
  78 /*
  79  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  80  * physical space so we can cache the place of the first one and move
  81  * around without checking the pgd every time.
  82  */
  83
  84 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
  85 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  86
  87 static int do_not_nx __cpuinitdata;
  88
  89 /*
  90  * noexec=on|off
  91  * Control non-executable mappings for 64-bit processes.
  92  *
  93  * on   Enable (default)
  94  * off  Disable
  95  */
  96 static int __init nonx_setup(char *str)
  97 {
  98         if (!str)
  99                 return -EINVAL;
 100         if (!strncmp(str, "on", 2)) {
 101                 __supported_pte_mask |= _PAGE_NX;
 102                 do_not_nx = 0;
 103         } else if (!strncmp(str, "off", 3)) {
 104                 do_not_nx = 1;
 105                 __supported_pte_mask &= ~_PAGE_NX;
 106         }
 107         return 0;
 108 }
 109 early_param("noexec", nonx_setup);
 110
 111 void __cpuinit check_efer(void)
 112 {
 113         unsigned long efer;
 114
 115         rdmsrl(MSR_EFER, efer);
 116         if (!(efer & EFER_NX) || do_not_nx)
 117                 __supported_pte_mask &= ~_PAGE_NX;
 118 }
 119
 120 int force_personality32;
 121
 122 /*
 123  * noexec32=on|off
 124  * Control non executable heap for 32bit processes.
 125  * To control the stack too use noexec=off
 126  *
 127  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 128  * off  PROT_READ implies PROT_EXEC
 129  */
 130 static int __init nonx32_setup(char *str)
 131 {
 132         if (!strcmp(str, "on"))
 133                 force_personality32 &= ~READ_IMPLIES_EXEC;
 134         else if (!strcmp(str, "off"))
 135                 force_personality32 |= READ_IMPLIES_EXEC;
 136         return 1;
 137 }
 138 __setup("noexec32=", nonx32_setup);
 139
 140 /*
 141  * NOTE: This function is marked __ref because it calls __init function
 142  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 143  */
 144 static __ref void *spp_getpage(void)
 145 {
 146         void *ptr;
 147
 148         if (after_bootmem)
 149                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 150         else
 151                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 152
 153         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 154                 panic("set_pte_phys: cannot allocate page data %s\n",
 155                         after_bootmem ? "after bootmem" : "");
 156         }
 157
 158         pr_debug("spp_getpage %p\n", ptr);
 159
 160         return ptr;
 161 }
 162
 163 void
 164 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 165 {
 166         pud_t *pud;
 167         pmd_t *pmd;
 168         pte_t *pte;
 169
 170         pud = pud_page + pud_index(vaddr);
 171         if (pud_none(*pud)) {
 172                 pmd = (pmd_t *) spp_getpage();
 173                 pud_populate(&init_mm, pud, pmd);
 174                 if (pmd != pmd_offset(pud, 0)) {
 175                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 176                                 pmd, pmd_offset(pud, 0));
 177                         return;
 178                 }
 179         }
 180         pmd = pmd_offset(pud, vaddr);
 181         if (pmd_none(*pmd)) {
 182                 pte = (pte_t *) spp_getpage();
 183                 pmd_populate_kernel(&init_mm, pmd, pte);
 184                 if (pte != pte_offset_kernel(pmd, 0)) {
 185                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 186                         return;
 187                 }
 188         }
 189
 190         pte = pte_offset_kernel(pmd, vaddr);
 191         set_pte(pte, new_pte);
 192
 193         /*
 194          * It's enough to flush this one mapping.
 195          * (PGE mappings get flushed as well)
 196          */
 197         __flush_tlb_one(vaddr);
 198 }
 199
 200 void
 201 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 202 {
 203         pgd_t *pgd;
 204         pud_t *pud_page;
 205
 206         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 207
 208         pgd = pgd_offset_k(vaddr);
 209         if (pgd_none(*pgd)) {
 210                 printk(KERN_ERR
 211                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 212                 return;
 213         }
 214         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 215         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 216 }
 217
 218 /*
 219  * Create large page table mappings for a range of physical addresses.
 220  */
 221 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 222                                                 pgprot_t prot)
 223 {
 224         pgd_t *pgd;
 225         pud_t *pud;
 226         pmd_t *pmd;
 227
 228         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 229         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 230                 pgd = pgd_offset_k((unsigned long)__va(phys));
 231                 if (pgd_none(*pgd)) {
 232                         pud = (pud_t *) spp_getpage();
 233                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 234                                                 _PAGE_USER));
 235                 }
 236                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 237                 if (pud_none(*pud)) {
 238                         pmd = (pmd_t *) spp_getpage();
 239                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 240                                                 _PAGE_USER));
 241                 }
 242                 pmd = pmd_offset(pud, phys);
 243                 BUG_ON(!pmd_none(*pmd));
 244                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 245         }
 246 }
 247
 248 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 249 {
 250         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 251 }
 252
 253 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 254 {
 255         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 256 }
 257
 258 /*
 259  * The head.S code sets up the kernel high mapping:
 260  *
 261  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 262  *
 263  * phys_addr holds the negative offset to the kernel, which is added
 264  * to the compile time generated pmds. This results in invalid pmds up
 265  * to the point where we hit the physaddr 0 mapping.
 266  *
 267  * We limit the mappings to the region from _text to _end.  _end is
 268  * rounded up to the 2MB boundary. This catches the invalid pmds as
 269  * well, as they are located before _text:
 270  */
 271 void __init cleanup_highmap(void)
 272 {
 273         unsigned long vaddr = __START_KERNEL_map;
 274         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
 275         pmd_t *pmd = level2_kernel_pgt;
 276         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 277
 278         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 279                 if (pmd_none(*pmd))
 280                         continue;
 281                 if (vaddr < (unsigned long) _text || vaddr > end)
 282                         set_pmd(pmd, __pmd(0));
 283         }
 284 }
 285
 286 extern unsigned long __initdata e820_table_start;
 287 extern unsigned long __meminitdata e820_table_end;
 288 extern unsigned long __meminitdata e820_table_top;
 289
 290 static __ref void *alloc_low_page(unsigned long *phys)
 291 {
 292         unsigned long pfn = e820_table_end++;
 293         void *adr;
 294
 295         if (after_bootmem) {
 296                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 297                 *phys = __pa(adr);
 298
 299                 return adr;
 300         }
 301
 302         if (pfn >= e820_table_top)
 303                 panic("alloc_low_page: ran out of memory");
 304
 305         adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
 306         memset(adr, 0, PAGE_SIZE);
 307         *phys  = pfn * PAGE_SIZE;
 308         return adr;
 309 }
 310
 311 static __ref void unmap_low_page(void *adr)
 312 {
 313         if (after_bootmem)
 314                 return;
 315
 316         early_iounmap(adr, PAGE_SIZE);
 317 }
 318
 319 static unsigned long __meminit
 320 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 321               pgprot_t prot)
 322 {
 323         unsigned pages = 0;
 324         unsigned long last_map_addr = end;
 325         int i;
 326
 327         pte_t *pte = pte_page + pte_index(addr);
 328
 329         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 330
 331                 if (addr >= end) {
 332                         if (!after_bootmem) {
 333                                 for(; i < PTRS_PER_PTE; i++, pte++)
 334                                         set_pte(pte, __pte(0));
 335                         }
 336                         break;
 337                 }
 338
 339                 /*
 340                  * We will re-use the existing mapping.
 341                  * Xen for example has some special requirements, like mapping
 342                  * pagetable pages as RO. So assume someone who pre-setup
 343                  * these mappings are more intelligent.
 344                  */
 345                 if (pte_val(*pte)) {
 346                         pages++;
 347                         continue;
 348                 }
 349
 350                 if (0)
 351                         printk("   pte=%p addr=%lx pte=%016lx\n",
 352                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 353                 pages++;
 354                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 355                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 356         }
 357
 358         update_page_count(PG_LEVEL_4K, pages);
 359
 360         return last_map_addr;
 361 }
 362
 363 static unsigned long __meminit
 364 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end,
 365                 pgprot_t prot)
 366 {
 367         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 368
 369         return phys_pte_init(pte, address, end, prot);
 370 }
 371
 372 static unsigned long __meminit
 373 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 374               unsigned long page_size_mask, pgprot_t prot)
 375 {
 376         unsigned long pages = 0;
 377         unsigned long last_map_addr = end;
 378
 379         int i = pmd_index(address);
 380
 381         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 382                 unsigned long pte_phys;
 383                 pmd_t *pmd = pmd_page + pmd_index(address);
 384                 pte_t *pte;
 385                 pgprot_t new_prot = prot;
 386
 387                 if (address >= end) {
 388                         if (!after_bootmem) {
 389                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 390                                         set_pmd(pmd, __pmd(0));
 391                         }
 392                         break;
 393                 }
 394
 395                 if (pmd_val(*pmd)) {
 396                         if (!pmd_large(*pmd)) {
 397                                 spin_lock(&init_mm.page_table_lock);
 398                                 last_map_addr = phys_pte_update(pmd, address,
 399                                                                 end, prot);
 400                                 spin_unlock(&init_mm.page_table_lock);
 401                                 continue;
 402                         }
 403                         /*
 404                          * If we are ok with PG_LEVEL_2M mapping, then we will
 405                          * use the existing mapping,
 406                          *
 407                          * Otherwise, we will split the large page mapping but
 408                          * use the same existing protection bits except for
 409                          * large page, so that we don't violate Intel's TLB
 410                          * Application note (317080) which says, while changing
 411                          * the page sizes, new and old translations should
 412                          * not differ with respect to page frame and
 413                          * attributes.
 414                          */
 415                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 416                                 pages++;
 417                                 continue;
 418                         }
 419                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 420                 }
 421
 422                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 423                         pages++;
 424                         spin_lock(&init_mm.page_table_lock);
 425                         set_pte((pte_t *)pmd,
 426                                 pfn_pte(address >> PAGE_SHIFT,
 427                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 428                         spin_unlock(&init_mm.page_table_lock);
 429                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 430                         continue;
 431                 }
 432
 433                 pte = alloc_low_page(&pte_phys);
 434                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 435                 unmap_low_page(pte);
 436
 437                 spin_lock(&init_mm.page_table_lock);
 438                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 439                 spin_unlock(&init_mm.page_table_lock);
 440         }
 441         update_page_count(PG_LEVEL_2M, pages);
 442         return last_map_addr;
 443 }
 444
 445 static unsigned long __meminit
 446 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 447                 unsigned long page_size_mask, pgprot_t prot)
 448 {
 449         pmd_t *pmd = pmd_offset(pud, 0);
 450         unsigned long last_map_addr;
 451
 452         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
 453         __flush_tlb_all();
 454         return last_map_addr;
 455 }
 456
 457 static unsigned long __meminit
 458 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 459                          unsigned long page_size_mask)
 460 {
 461         unsigned long pages = 0;
 462         unsigned long last_map_addr = end;
 463         int i = pud_index(addr);
 464
 465         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 466                 unsigned long pmd_phys;
 467                 pud_t *pud = pud_page + pud_index(addr);
 468                 pmd_t *pmd;
 469                 pgprot_t prot = PAGE_KERNEL;
 470
 471                 if (addr >= end)
 472                         break;
 473
 474                 if (!after_bootmem &&
 475                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 476                         set_pud(pud, __pud(0));
 477                         continue;
 478                 }
 479
 480                 if (pud_val(*pud)) {
 481                         if (!pud_large(*pud)) {
 482                                 last_map_addr = phys_pmd_update(pud, addr, end,
 483                                                          page_size_mask, prot);
 484                                 continue;
 485                         }
 486                         /*
 487                          * If we are ok with PG_LEVEL_1G mapping, then we will
 488                          * use the existing mapping.
 489                          *
 490                          * Otherwise, we will split the gbpage mapping but use
 491                          * the same existing protection  bits except for large
 492                          * page, so that we don't violate Intel's TLB
 493                          * Application note (317080) which says, while changing
 494                          * the page sizes, new and old translations should
 495                          * not differ with respect to page frame and
 496                          * attributes.
 497                          */
 498                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 499                                 pages++;
 500                                 continue;
 501                         }
 502                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 503                 }
 504
 505                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 506                         pages++;
 507                         spin_lock(&init_mm.page_table_lock);
 508                         set_pte((pte_t *)pud,
 509                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 510                         spin_unlock(&init_mm.page_table_lock);
 511                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 512                         continue;
 513                 }
 514
 515                 pmd = alloc_low_page(&pmd_phys);
 516                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 517                                               prot);
 518                 unmap_low_page(pmd);
 519
 520                 spin_lock(&init_mm.page_table_lock);
 521                 pud_populate(&init_mm, pud, __va(pmd_phys));
 522                 spin_unlock(&init_mm.page_table_lock);
 523         }
 524         __flush_tlb_all();
 525
 526         update_page_count(PG_LEVEL_1G, pages);
 527
 528         return last_map_addr;
 529 }
 530
 531 static unsigned long __meminit
 532 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 533                  unsigned long page_size_mask)
 534 {
 535         pud_t *pud;
 536
 537         pud = (pud_t *)pgd_page_vaddr(*pgd);
 538
 539         return phys_pud_init(pud, addr, end, page_size_mask);
 540 }
 541
 542 unsigned long __meminit
 543 kernel_physical_mapping_init(unsigned long start,
 544                              unsigned long end,
 545                              unsigned long page_size_mask)
 546 {
 547
 548         unsigned long next, last_map_addr = end;
 549
 550         start = (unsigned long)__va(start);
 551         end = (unsigned long)__va(end);
 552
 553         for (; start < end; start = next) {
 554                 pgd_t *pgd = pgd_offset_k(start);
 555                 unsigned long pud_phys;
 556                 pud_t *pud;
 557
 558                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 559                 if (next > end)
 560                         next = end;
 561
 562                 if (pgd_val(*pgd)) {
 563                         last_map_addr = phys_pud_update(pgd, __pa(start),
 564                                                  __pa(end), page_size_mask);
 565                         continue;
 566                 }
 567
 568                 pud = alloc_low_page(&pud_phys);
 569                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 570                                                  page_size_mask);
 571                 unmap_low_page(pud);
 572
 573                 spin_lock(&init_mm.page_table_lock);
 574                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 575                 spin_unlock(&init_mm.page_table_lock);
 576         }
 577         __flush_tlb_all();
 578
 579         return last_map_addr;
 580 }
 581
 582 #ifndef CONFIG_NUMA
 583 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 584 {
 585         unsigned long bootmap_size, bootmap;
 586
 587         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 588         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 589                                  PAGE_SIZE);
 590         if (bootmap == -1L)
 591                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 592         /* don't touch min_low_pfn */
 593         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 594                                          0, end_pfn);
 595         e820_register_active_regions(0, start_pfn, end_pfn);
 596         free_bootmem_with_active_regions(0, end_pfn);
 597         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 598         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 599 }
 600
 601 void __init paging_init(void)
 602 {
 603         unsigned long max_zone_pfns[MAX_NR_ZONES];
 604
 605         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 606         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 607         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 608         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 609
 610         memory_present(0, 0, max_pfn);
 611         sparse_init();
 612         free_area_init_nodes(max_zone_pfns);
 613 }
 614 #endif
 615
 616 /*
 617  * Memory hotplug specific functions
 618  */
 619 #ifdef CONFIG_MEMORY_HOTPLUG
 620 /*
 621  * Memory is added always to NORMAL zone. This means you will never get
 622  * additional DMA/DMA32 memory.
 623  */
 624 int arch_add_memory(int nid, u64 start, u64 size)
 625 {
 626         struct pglist_data *pgdat = NODE_DATA(nid);
 627         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 628         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 629         unsigned long nr_pages = size >> PAGE_SHIFT;
 630         int ret;
 631
 632         last_mapped_pfn = init_memory_mapping(start, start + size);
 633         if (last_mapped_pfn > max_pfn_mapped)
 634                 max_pfn_mapped = last_mapped_pfn;
 635
 636         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 637         WARN_ON_ONCE(ret);
 638
 639         return ret;
 640 }
 641 EXPORT_SYMBOL_GPL(arch_add_memory);
 642
 643 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 644 int memory_add_physaddr_to_nid(u64 start)
 645 {
 646         return 0;
 647 }
 648 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 649 #endif
 650
 651 #endif /* CONFIG_MEMORY_HOTPLUG */
 652
 653 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 654                          kcore_modules, kcore_vsyscall;
 655
 656 void __init mem_init(void)
 657 {
 658         long codesize, reservedpages, datasize, initsize;
 659         unsigned long absent_pages;
 660
 661         pci_iommu_alloc();
 662
 663         /* clear_bss() already clear the empty_zero_page */
 664
 665         reservedpages = 0;
 666
 667         /* this will put all low memory onto the freelists */
 668 #ifdef CONFIG_NUMA
 669         totalram_pages = numa_free_all_bootmem();
 670 #else
 671         totalram_pages = free_all_bootmem();
 672 #endif
 673
 674         absent_pages = absent_pages_in_range(0, max_pfn);
 675         reservedpages = max_pfn - totalram_pages - absent_pages;
 676         after_bootmem = 1;
 677
 678         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 679         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 680         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 681
 682         /* Register memory areas for /proc/kcore */
 683         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 684         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 685                    VMALLOC_END-VMALLOC_START);
 686         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 687         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 688         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 689                                  VSYSCALL_END - VSYSCALL_START);
 690
 691         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 692                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 693                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 694                 max_pfn << (PAGE_SHIFT-10),
 695                 codesize >> 10,
 696                 absent_pages << (PAGE_SHIFT-10),
 697                 reservedpages << (PAGE_SHIFT-10),
 698                 datasize >> 10,
 699                 initsize >> 10);
 700 }
 701
 702 #ifdef CONFIG_DEBUG_RODATA
 703 const int rodata_test_data = 0xC3;
 704 EXPORT_SYMBOL_GPL(rodata_test_data);
 705
 706 void mark_rodata_ro(void)
 707 {
 708         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 709         unsigned long rodata_start =
 710                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 711
 712 #ifdef CONFIG_DYNAMIC_FTRACE
 713         /* Dynamic tracing modifies the kernel text section */
 714         start = rodata_start;
 715 #endif
 716
 717         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 718                (end - start) >> 10);
 719         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 720
 721         /*
 722          * The rodata section (but not the kernel text!) should also be
 723          * not-executable.
 724          */
 725         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 726
 727         rodata_test();
 728
 729 #ifdef CONFIG_CPA_DEBUG
 730         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 731         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 732
 733         printk(KERN_INFO "Testing CPA: again\n");
 734         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 735 #endif
 736 }
 737
 738 #endif
 739
 740 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 741                                    int flags)
 742 {
 743 #ifdef CONFIG_NUMA
 744         int nid, next_nid;
 745         int ret;
 746 #endif
 747         unsigned long pfn = phys >> PAGE_SHIFT;
 748
 749         if (pfn >= max_pfn) {
 750                 /*
 751                  * This can happen with kdump kernels when accessing
 752                  * firmware tables:
 753                  */
 754                 if (pfn < max_pfn_mapped)
 755                         return -EFAULT;
 756
 757                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 758                                 phys, len);
 759                 return -EFAULT;
 760         }
 761
 762         /* Should check here against the e820 map to avoid double free */
 763 #ifdef CONFIG_NUMA
 764         nid = phys_to_nid(phys);
 765         next_nid = phys_to_nid(phys + len - 1);
 766         if (nid == next_nid)
 767                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 768         else
 769                 ret = reserve_bootmem(phys, len, flags);
 770
 771         if (ret != 0)
 772                 return ret;
 773
 774 #else
 775         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 776 #endif
 777
 778         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 779                 dma_reserve += len / PAGE_SIZE;
 780                 set_dma_reserve(dma_reserve);
 781         }
 782
 783         return 0;
 784 }
 785
 786 int kern_addr_valid(unsigned long addr)
 787 {
 788         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 789         pgd_t *pgd;
 790         pud_t *pud;
 791         pmd_t *pmd;
 792         pte_t *pte;
 793
 794         if (above != 0 && above != -1UL)
 795                 return 0;
 796
 797         pgd = pgd_offset_k(addr);
 798         if (pgd_none(*pgd))
 799                 return 0;
 800
 801         pud = pud_offset(pgd, addr);
 802         if (pud_none(*pud))
 803                 return 0;
 804
 805         pmd = pmd_offset(pud, addr);
 806         if (pmd_none(*pmd))
 807                 return 0;
 808
 809         if (pmd_large(*pmd))
 810                 return pfn_valid(pmd_pfn(*pmd));
 811
 812         pte = pte_offset_kernel(pmd, addr);
 813         if (pte_none(*pte))
 814                 return 0;
 815
 816         return pfn_valid(pte_pfn(*pte));
 817 }
 818
 819 /*
 820  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 821  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 822  * not need special handling anymore:
 823  */
 824 static struct vm_area_struct gate_vma = {
 825         .vm_start       = VSYSCALL_START,
 826         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 827         .vm_page_prot   = PAGE_READONLY_EXEC,
 828         .vm_flags       = VM_READ | VM_EXEC
 829 };
 830
 831 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 832 {
 833 #ifdef CONFIG_IA32_EMULATION
 834         if (test_tsk_thread_flag(tsk, TIF_IA32))
 835                 return NULL;
 836 #endif
 837         return &gate_vma;
 838 }
 839
 840 int in_gate_area(struct task_struct *task, unsigned long addr)
 841 {
 842         struct vm_area_struct *vma = get_gate_vma(task);
 843
 844         if (!vma)
 845                 return 0;
 846
 847         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 848 }
 849
 850 /*
 851  * Use this when you have no reliable task/vma, typically from interrupt
 852  * context. It is less reliable than using the task's vma and may give
 853  * false positives:
 854  */
 855 int in_gate_area_no_task(unsigned long addr)
 856 {
 857         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 858 }
 859
 860 const char *arch_vma_name(struct vm_area_struct *vma)
 861 {
 862         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 863                 return "[vdso]";
 864         if (vma == &gate_vma)
 865                 return "[vsyscall]";
 866         return NULL;
 867 }
 868
 869 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 870 /*
 871  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 872  */
 873 static long __meminitdata addr_start, addr_end;
 874 static void __meminitdata *p_start, *p_end;
 875 static int __meminitdata node_start;
 876
 877 int __meminit
 878 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 879 {
 880         unsigned long addr = (unsigned long)start_page;
 881         unsigned long end = (unsigned long)(start_page + size);
 882         unsigned long next;
 883         pgd_t *pgd;
 884         pud_t *pud;
 885         pmd_t *pmd;
 886
 887         for (; addr < end; addr = next) {
 888                 void *p = NULL;
 889
 890                 pgd = vmemmap_pgd_populate(addr, node);
 891                 if (!pgd)
 892                         return -ENOMEM;
 893
 894                 pud = vmemmap_pud_populate(pgd, addr, node);
 895                 if (!pud)
 896                         return -ENOMEM;
 897
 898                 if (!cpu_has_pse) {
 899                         next = (addr + PAGE_SIZE) & PAGE_MASK;
 900                         pmd = vmemmap_pmd_populate(pud, addr, node);
 901
 902                         if (!pmd)
 903                                 return -ENOMEM;
 904
 905                         p = vmemmap_pte_populate(pmd, addr, node);
 906
 907                         if (!p)
 908                                 return -ENOMEM;
 909
 910                         addr_end = addr + PAGE_SIZE;
 911                         p_end = p + PAGE_SIZE;
 912                 } else {
 913                         next = pmd_addr_end(addr, end);
 914
 915                         pmd = pmd_offset(pud, addr);
 916                         if (pmd_none(*pmd)) {
 917                                 pte_t entry;
 918
 919                                 p = vmemmap_alloc_block(PMD_SIZE, node);
 920                                 if (!p)
 921                                         return -ENOMEM;
 922
 923                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 924                                                 PAGE_KERNEL_LARGE);
 925                                 set_pmd(pmd, __pmd(pte_val(entry)));
 926
 927                                 /* check to see if we have contiguous blocks */
 928                                 if (p_end != p || node_start != node) {
 929                                         if (p_start)
 930                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 931                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
 932                                         addr_start = addr;
 933                                         node_start = node;
 934                                         p_start = p;
 935                                 }
 936
 937                                 addr_end = addr + PMD_SIZE;
 938                                 p_end = p + PMD_SIZE;
 939                         } else
 940                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
 941                 }
 942
 943         }
 944         return 0;
 945 }
 946
 947 void __meminit vmemmap_populate_print_last(void)
 948 {
 949         if (p_start) {
 950                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 951                         addr_start, addr_end-1, p_start, p_end-1, node_start);
 952                 p_start = NULL;
 953                 p_end = NULL;
 954                 node_start = 0;
 955         }
 956 }
 957 #endif