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 #include <asm/init.h>
  52
  53 static unsigned long dma_reserve __initdata;
  54
  55 static int __init parse_direct_gbpages_off(char *arg)
  56 {
  57         direct_gbpages = 0;
  58         return 0;
  59 }
  60 early_param("nogbpages", parse_direct_gbpages_off);
  61
  62 static int __init parse_direct_gbpages_on(char *arg)
  63 {
  64         direct_gbpages = 1;
  65         return 0;
  66 }
  67 early_param("gbpages", parse_direct_gbpages_on);
  68
  69 /*
  70  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  71  * physical space so we can cache the place of the first one and move
  72  * around without checking the pgd every time.
  73  */
  74
  75 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
  76 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  77
  78 int force_personality32;
  79
  80 /*
  81  * noexec32=on|off
  82  * Control non executable heap for 32bit processes.
  83  * To control the stack too use noexec=off
  84  *
  85  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
  86  * off  PROT_READ implies PROT_EXEC
  87  */
  88 static int __init nonx32_setup(char *str)
  89 {
  90         if (!strcmp(str, "on"))
  91                 force_personality32 &= ~READ_IMPLIES_EXEC;
  92         else if (!strcmp(str, "off"))
  93                 force_personality32 |= READ_IMPLIES_EXEC;
  94         return 1;
  95 }
  96 __setup("noexec32=", nonx32_setup);
  97
  98 /*
  99  * NOTE: This function is marked __ref because it calls __init function
 100  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 101  */
 102 static __ref void *spp_getpage(void)
 103 {
 104         void *ptr;
 105
 106         if (after_bootmem)
 107                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 108         else
 109                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 110
 111         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 112                 panic("set_pte_phys: cannot allocate page data %s\n",
 113                         after_bootmem ? "after bootmem" : "");
 114         }
 115
 116         pr_debug("spp_getpage %p\n", ptr);
 117
 118         return ptr;
 119 }
 120
 121 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 122 {
 123         if (pgd_none(*pgd)) {
 124                 pud_t *pud = (pud_t *)spp_getpage();
 125                 pgd_populate(&init_mm, pgd, pud);
 126                 if (pud != pud_offset(pgd, 0))
 127                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 128                                pud, pud_offset(pgd, 0));
 129         }
 130         return pud_offset(pgd, vaddr);
 131 }
 132
 133 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 134 {
 135         if (pud_none(*pud)) {
 136                 pmd_t *pmd = (pmd_t *) spp_getpage();
 137                 pud_populate(&init_mm, pud, pmd);
 138                 if (pmd != pmd_offset(pud, 0))
 139                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 140                                pmd, pmd_offset(pud, 0));
 141         }
 142         return pmd_offset(pud, vaddr);
 143 }
 144
 145 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 146 {
 147         if (pmd_none(*pmd)) {
 148                 pte_t *pte = (pte_t *) spp_getpage();
 149                 pmd_populate_kernel(&init_mm, pmd, pte);
 150                 if (pte != pte_offset_kernel(pmd, 0))
 151                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 152         }
 153         return pte_offset_kernel(pmd, vaddr);
 154 }
 155
 156 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 157 {
 158         pud_t *pud;
 159         pmd_t *pmd;
 160         pte_t *pte;
 161
 162         pud = pud_page + pud_index(vaddr);
 163         pmd = fill_pmd(pud, vaddr);
 164         pte = fill_pte(pmd, vaddr);
 165
 166         set_pte(pte, new_pte);
 167
 168         /*
 169          * It's enough to flush this one mapping.
 170          * (PGE mappings get flushed as well)
 171          */
 172         __flush_tlb_one(vaddr);
 173 }
 174
 175 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 176 {
 177         pgd_t *pgd;
 178         pud_t *pud_page;
 179
 180         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 181
 182         pgd = pgd_offset_k(vaddr);
 183         if (pgd_none(*pgd)) {
 184                 printk(KERN_ERR
 185                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 186                 return;
 187         }
 188         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 189         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 190 }
 191
 192 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 193 {
 194         pgd_t *pgd;
 195         pud_t *pud;
 196
 197         pgd = pgd_offset_k(vaddr);
 198         pud = fill_pud(pgd, vaddr);
 199         return fill_pmd(pud, vaddr);
 200 }
 201
 202 pte_t * __init populate_extra_pte(unsigned long vaddr)
 203 {
 204         pmd_t *pmd;
 205
 206         pmd = populate_extra_pmd(vaddr);
 207         return fill_pte(pmd, vaddr);
 208 }
 209
 210 /*
 211  * Create large page table mappings for a range of physical addresses.
 212  */
 213 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 214                                                 pgprot_t prot)
 215 {
 216         pgd_t *pgd;
 217         pud_t *pud;
 218         pmd_t *pmd;
 219
 220         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 221         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 222                 pgd = pgd_offset_k((unsigned long)__va(phys));
 223                 if (pgd_none(*pgd)) {
 224                         pud = (pud_t *) spp_getpage();
 225                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 226                                                 _PAGE_USER));
 227                 }
 228                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 229                 if (pud_none(*pud)) {
 230                         pmd = (pmd_t *) spp_getpage();
 231                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 232                                                 _PAGE_USER));
 233                 }
 234                 pmd = pmd_offset(pud, phys);
 235                 BUG_ON(!pmd_none(*pmd));
 236                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 237         }
 238 }
 239
 240 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 241 {
 242         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 243 }
 244
 245 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 246 {
 247         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 248 }
 249
 250 /*
 251  * The head.S code sets up the kernel high mapping:
 252  *
 253  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 254  *
 255  * phys_addr holds the negative offset to the kernel, which is added
 256  * to the compile time generated pmds. This results in invalid pmds up
 257  * to the point where we hit the physaddr 0 mapping.
 258  *
 259  * We limit the mappings to the region from _text to _end.  _end is
 260  * rounded up to the 2MB boundary. This catches the invalid pmds as
 261  * well, as they are located before _text:
 262  */
 263 void __init cleanup_highmap(void)
 264 {
 265         unsigned long vaddr = __START_KERNEL_map;
 266         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
 267         pmd_t *pmd = level2_kernel_pgt;
 268         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 269
 270         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 271                 if (pmd_none(*pmd))
 272                         continue;
 273                 if (vaddr < (unsigned long) _text || vaddr > end)
 274                         set_pmd(pmd, __pmd(0));
 275         }
 276 }
 277
 278 static __ref void *alloc_low_page(unsigned long *phys)
 279 {
 280         unsigned long pfn = e820_table_end++;
 281         void *adr;
 282
 283         if (after_bootmem) {
 284                 adr = (void *)get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 285                 *phys = __pa(adr);
 286
 287                 return adr;
 288         }
 289
 290         if (pfn >= e820_table_top)
 291                 panic("alloc_low_page: ran out of memory");
 292
 293         adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
 294         memset(adr, 0, PAGE_SIZE);
 295         *phys  = pfn * PAGE_SIZE;
 296         return adr;
 297 }
 298
 299 static __ref void unmap_low_page(void *adr)
 300 {
 301         if (after_bootmem)
 302                 return;
 303
 304         early_iounmap(adr, PAGE_SIZE);
 305 }
 306
 307 static unsigned long __meminit
 308 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 309               pgprot_t prot)
 310 {
 311         unsigned pages = 0;
 312         unsigned long last_map_addr = end;
 313         int i;
 314
 315         pte_t *pte = pte_page + pte_index(addr);
 316
 317         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 318
 319                 if (addr >= end) {
 320                         if (!after_bootmem) {
 321                                 for(; i < PTRS_PER_PTE; i++, pte++)
 322                                         set_pte(pte, __pte(0));
 323                         }
 324                         break;
 325                 }
 326
 327                 /*
 328                  * We will re-use the existing mapping.
 329                  * Xen for example has some special requirements, like mapping
 330                  * pagetable pages as RO. So assume someone who pre-setup
 331                  * these mappings are more intelligent.
 332                  */
 333                 if (pte_val(*pte)) {
 334                         pages++;
 335                         continue;
 336                 }
 337
 338                 if (0)
 339                         printk("   pte=%p addr=%lx pte=%016lx\n",
 340                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 341                 pages++;
 342                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 343                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 344         }
 345
 346         update_page_count(PG_LEVEL_4K, pages);
 347
 348         return last_map_addr;
 349 }
 350
 351 static unsigned long __meminit
 352 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end,
 353                 pgprot_t prot)
 354 {
 355         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 356
 357         return phys_pte_init(pte, address, end, prot);
 358 }
 359
 360 static unsigned long __meminit
 361 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 362               unsigned long page_size_mask, pgprot_t prot)
 363 {
 364         unsigned long pages = 0;
 365         unsigned long last_map_addr = end;
 366
 367         int i = pmd_index(address);
 368
 369         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 370                 unsigned long pte_phys;
 371                 pmd_t *pmd = pmd_page + pmd_index(address);
 372                 pte_t *pte;
 373                 pgprot_t new_prot = prot;
 374
 375                 if (address >= end) {
 376                         if (!after_bootmem) {
 377                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 378                                         set_pmd(pmd, __pmd(0));
 379                         }
 380                         break;
 381                 }
 382
 383                 if (pmd_val(*pmd)) {
 384                         if (!pmd_large(*pmd)) {
 385                                 spin_lock(&init_mm.page_table_lock);
 386                                 last_map_addr = phys_pte_update(pmd, address,
 387                                                                 end, prot);
 388                                 spin_unlock(&init_mm.page_table_lock);
 389                                 continue;
 390                         }
 391                         /*
 392                          * If we are ok with PG_LEVEL_2M mapping, then we will
 393                          * use the existing mapping,
 394                          *
 395                          * Otherwise, we will split the large page mapping but
 396                          * use the same existing protection bits except for
 397                          * large page, so that we don't violate Intel's TLB
 398                          * Application note (317080) which says, while changing
 399                          * the page sizes, new and old translations should
 400                          * not differ with respect to page frame and
 401                          * attributes.
 402                          */
 403                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 404                                 pages++;
 405                                 continue;
 406                         }
 407                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 408                 }
 409
 410                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 411                         pages++;
 412                         spin_lock(&init_mm.page_table_lock);
 413                         set_pte((pte_t *)pmd,
 414                                 pfn_pte(address >> PAGE_SHIFT,
 415                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 416                         spin_unlock(&init_mm.page_table_lock);
 417                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 418                         continue;
 419                 }
 420
 421                 pte = alloc_low_page(&pte_phys);
 422                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 423                 unmap_low_page(pte);
 424
 425                 spin_lock(&init_mm.page_table_lock);
 426                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 427                 spin_unlock(&init_mm.page_table_lock);
 428         }
 429         update_page_count(PG_LEVEL_2M, pages);
 430         return last_map_addr;
 431 }
 432
 433 static unsigned long __meminit
 434 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 435                 unsigned long page_size_mask, pgprot_t prot)
 436 {
 437         pmd_t *pmd = pmd_offset(pud, 0);
 438         unsigned long last_map_addr;
 439
 440         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
 441         __flush_tlb_all();
 442         return last_map_addr;
 443 }
 444
 445 static unsigned long __meminit
 446 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 447                          unsigned long page_size_mask)
 448 {
 449         unsigned long pages = 0;
 450         unsigned long last_map_addr = end;
 451         int i = pud_index(addr);
 452
 453         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 454                 unsigned long pmd_phys;
 455                 pud_t *pud = pud_page + pud_index(addr);
 456                 pmd_t *pmd;
 457                 pgprot_t prot = PAGE_KERNEL;
 458
 459                 if (addr >= end)
 460                         break;
 461
 462                 if (!after_bootmem &&
 463                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 464                         set_pud(pud, __pud(0));
 465                         continue;
 466                 }
 467
 468                 if (pud_val(*pud)) {
 469                         if (!pud_large(*pud)) {
 470                                 last_map_addr = phys_pmd_update(pud, addr, end,
 471                                                          page_size_mask, prot);
 472                                 continue;
 473                         }
 474                         /*
 475                          * If we are ok with PG_LEVEL_1G mapping, then we will
 476                          * use the existing mapping.
 477                          *
 478                          * Otherwise, we will split the gbpage mapping but use
 479                          * the same existing protection  bits except for large
 480                          * page, so that we don't violate Intel's TLB
 481                          * Application note (317080) which says, while changing
 482                          * the page sizes, new and old translations should
 483                          * not differ with respect to page frame and
 484                          * attributes.
 485                          */
 486                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 487                                 pages++;
 488                                 continue;
 489                         }
 490                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 491                 }
 492
 493                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 494                         pages++;
 495                         spin_lock(&init_mm.page_table_lock);
 496                         set_pte((pte_t *)pud,
 497                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 498                         spin_unlock(&init_mm.page_table_lock);
 499                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 500                         continue;
 501                 }
 502
 503                 pmd = alloc_low_page(&pmd_phys);
 504                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 505                                               prot);
 506                 unmap_low_page(pmd);
 507
 508                 spin_lock(&init_mm.page_table_lock);
 509                 pud_populate(&init_mm, pud, __va(pmd_phys));
 510                 spin_unlock(&init_mm.page_table_lock);
 511         }
 512         __flush_tlb_all();
 513
 514         update_page_count(PG_LEVEL_1G, pages);
 515
 516         return last_map_addr;
 517 }
 518
 519 static unsigned long __meminit
 520 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 521                  unsigned long page_size_mask)
 522 {
 523         pud_t *pud;
 524
 525         pud = (pud_t *)pgd_page_vaddr(*pgd);
 526
 527         return phys_pud_init(pud, addr, end, page_size_mask);
 528 }
 529
 530 unsigned long __meminit
 531 kernel_physical_mapping_init(unsigned long start,
 532                              unsigned long end,
 533                              unsigned long page_size_mask)
 534 {
 535
 536         unsigned long next, last_map_addr = end;
 537
 538         start = (unsigned long)__va(start);
 539         end = (unsigned long)__va(end);
 540
 541         for (; start < end; start = next) {
 542                 pgd_t *pgd = pgd_offset_k(start);
 543                 unsigned long pud_phys;
 544                 pud_t *pud;
 545
 546                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 547                 if (next > end)
 548                         next = end;
 549
 550                 if (pgd_val(*pgd)) {
 551                         last_map_addr = phys_pud_update(pgd, __pa(start),
 552                                                  __pa(end), page_size_mask);
 553                         continue;
 554                 }
 555
 556                 pud = alloc_low_page(&pud_phys);
 557                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 558                                                  page_size_mask);
 559                 unmap_low_page(pud);
 560
 561                 spin_lock(&init_mm.page_table_lock);
 562                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 563                 spin_unlock(&init_mm.page_table_lock);
 564         }
 565         __flush_tlb_all();
 566
 567         return last_map_addr;
 568 }
 569
 570 #ifndef CONFIG_NUMA
 571 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn,
 572                                 int acpi, int k8)
 573 {
 574         unsigned long bootmap_size, bootmap;
 575
 576         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 577         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 578                                  PAGE_SIZE);
 579         if (bootmap == -1L)
 580                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 581         /* don't touch min_low_pfn */
 582         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 583                                          0, end_pfn);
 584         e820_register_active_regions(0, start_pfn, end_pfn);
 585         free_bootmem_with_active_regions(0, end_pfn);
 586         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 587         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 588 }
 589 #endif
 590
 591 void __init paging_init(void)
 592 {
 593         unsigned long max_zone_pfns[MAX_NR_ZONES];
 594
 595         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 596         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 597         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 598         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 599
 600         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 601         sparse_init();
 602
 603         /*
 604          * clear the default setting with node 0
 605          * note: don't use nodes_clear here, that is really clearing when
 606          *       numa support is not compiled in, and later node_set_state
 607          *       will not set it back.
 608          */
 609         node_clear_state(0, N_NORMAL_MEMORY);
 610
 611         free_area_init_nodes(max_zone_pfns);
 612 }
 613
 614 /*
 615  * Memory hotplug specific functions
 616  */
 617 #ifdef CONFIG_MEMORY_HOTPLUG
 618 /*
 619  * Memory is added always to NORMAL zone. This means you will never get
 620  * additional DMA/DMA32 memory.
 621  */
 622 int arch_add_memory(int nid, u64 start, u64 size)
 623 {
 624         struct pglist_data *pgdat = NODE_DATA(nid);
 625         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 626         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 627         unsigned long nr_pages = size >> PAGE_SHIFT;
 628         int ret;
 629
 630         last_mapped_pfn = init_memory_mapping(start, start + size);
 631         if (last_mapped_pfn > max_pfn_mapped)
 632                 max_pfn_mapped = last_mapped_pfn;
 633
 634         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 635         WARN_ON_ONCE(ret);
 636
 637         return ret;
 638 }
 639 EXPORT_SYMBOL_GPL(arch_add_memory);
 640
 641 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 642 int memory_add_physaddr_to_nid(u64 start)
 643 {
 644         return 0;
 645 }
 646 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 647 #endif
 648
 649 #endif /* CONFIG_MEMORY_HOTPLUG */
 650
 651 static struct kcore_list kcore_vsyscall;
 652
 653 void __init mem_init(void)
 654 {
 655         long codesize, reservedpages, datasize, initsize;
 656         unsigned long absent_pages;
 657
 658         pci_iommu_alloc();
 659
 660         /* clear_bss() already clear the empty_zero_page */
 661
 662         reservedpages = 0;
 663
 664         /* this will put all low memory onto the freelists */
 665 #ifdef CONFIG_NUMA
 666         totalram_pages = numa_free_all_bootmem();
 667 #else
 668         totalram_pages = free_all_bootmem();
 669 #endif
 670
 671         absent_pages = absent_pages_in_range(0, max_pfn);
 672         reservedpages = max_pfn - totalram_pages - absent_pages;
 673         after_bootmem = 1;
 674
 675         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 676         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 677         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 678
 679         /* Register memory areas for /proc/kcore */
 680         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 681                          VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
 682
 683         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 684                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 685                 nr_free_pages() << (PAGE_SHIFT-10),
 686                 max_pfn << (PAGE_SHIFT-10),
 687                 codesize >> 10,
 688                 absent_pages << (PAGE_SHIFT-10),
 689                 reservedpages << (PAGE_SHIFT-10),
 690                 datasize >> 10,
 691                 initsize >> 10);
 692 }
 693
 694 #ifdef CONFIG_DEBUG_RODATA
 695 const int rodata_test_data = 0xC3;
 696 EXPORT_SYMBOL_GPL(rodata_test_data);
 697
 698 int kernel_set_to_readonly;
 699
 700 void set_kernel_text_rw(void)
 701 {
 702         unsigned long start = PFN_ALIGN(_text);
 703         unsigned long end = PFN_ALIGN(__stop___ex_table);
 704
 705         if (!kernel_set_to_readonly)
 706                 return;
 707
 708         pr_debug("Set kernel text: %lx - %lx for read write\n",
 709                  start, end);
 710
 711         /*
 712          * Make the kernel identity mapping for text RW. Kernel text
 713          * mapping will always be RO. Refer to the comment in
 714          * static_protections() in pageattr.c
 715          */
 716         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
 717 }
 718
 719 void set_kernel_text_ro(void)
 720 {
 721         unsigned long start = PFN_ALIGN(_text);
 722         unsigned long end = PFN_ALIGN(__stop___ex_table);
 723
 724         if (!kernel_set_to_readonly)
 725                 return;
 726
 727         pr_debug("Set kernel text: %lx - %lx for read only\n",
 728                  start, end);
 729
 730         /*
 731          * Set the kernel identity mapping for text RO.
 732          */
 733         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 734 }
 735
 736 void mark_rodata_ro(void)
 737 {
 738         unsigned long start = PFN_ALIGN(_text);
 739         unsigned long rodata_start =
 740                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 741         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
 742         unsigned long text_end = PAGE_ALIGN((unsigned long) &__stop___ex_table);
 743         unsigned long rodata_end = PAGE_ALIGN((unsigned long) &__end_rodata);
 744         unsigned long data_start = (unsigned long) &_sdata;
 745
 746         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 747                (end - start) >> 10);
 748         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 749
 750         kernel_set_to_readonly = 1;
 751
 752         /*
 753          * The rodata section (but not the kernel text!) should also be
 754          * not-executable.
 755          */
 756         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 757
 758         rodata_test();
 759
 760 #ifdef CONFIG_CPA_DEBUG
 761         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 762         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 763
 764         printk(KERN_INFO "Testing CPA: again\n");
 765         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 766 #endif
 767
 768         free_init_pages("unused kernel memory",
 769                         (unsigned long) page_address(virt_to_page(text_end)),
 770                         (unsigned long)
 771                                  page_address(virt_to_page(rodata_start)));
 772         free_init_pages("unused kernel memory",
 773                         (unsigned long) page_address(virt_to_page(rodata_end)),
 774                         (unsigned long) page_address(virt_to_page(data_start)));
 775 }
 776
 777 #endif
 778
 779 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 780                                    int flags)
 781 {
 782 #ifdef CONFIG_NUMA
 783         int nid, next_nid;
 784         int ret;
 785 #endif
 786         unsigned long pfn = phys >> PAGE_SHIFT;
 787
 788         if (pfn >= max_pfn) {
 789                 /*
 790                  * This can happen with kdump kernels when accessing
 791                  * firmware tables:
 792                  */
 793                 if (pfn < max_pfn_mapped)
 794                         return -EFAULT;
 795
 796                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 797                                 phys, len);
 798                 return -EFAULT;
 799         }
 800
 801         /* Should check here against the e820 map to avoid double free */
 802 #ifdef CONFIG_NUMA
 803         nid = phys_to_nid(phys);
 804         next_nid = phys_to_nid(phys + len - 1);
 805         if (nid == next_nid)
 806                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 807         else
 808                 ret = reserve_bootmem(phys, len, flags);
 809
 810         if (ret != 0)
 811                 return ret;
 812
 813 #else
 814         reserve_bootmem(phys, len, flags);
 815 #endif
 816
 817         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 818                 dma_reserve += len / PAGE_SIZE;
 819                 set_dma_reserve(dma_reserve);
 820         }
 821
 822         return 0;
 823 }
 824
 825 int kern_addr_valid(unsigned long addr)
 826 {
 827         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 828         pgd_t *pgd;
 829         pud_t *pud;
 830         pmd_t *pmd;
 831         pte_t *pte;
 832
 833         if (above != 0 && above != -1UL)
 834                 return 0;
 835
 836         pgd = pgd_offset_k(addr);
 837         if (pgd_none(*pgd))
 838                 return 0;
 839
 840         pud = pud_offset(pgd, addr);
 841         if (pud_none(*pud))
 842                 return 0;
 843
 844         pmd = pmd_offset(pud, addr);
 845         if (pmd_none(*pmd))
 846                 return 0;
 847
 848         if (pmd_large(*pmd))
 849                 return pfn_valid(pmd_pfn(*pmd));
 850
 851         pte = pte_offset_kernel(pmd, addr);
 852         if (pte_none(*pte))
 853                 return 0;
 854
 855         return pfn_valid(pte_pfn(*pte));
 856 }
 857
 858 /*
 859  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 860  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 861  * not need special handling anymore:
 862  */
 863 static struct vm_area_struct gate_vma = {
 864         .vm_start       = VSYSCALL_START,
 865         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 866         .vm_page_prot   = PAGE_READONLY_EXEC,
 867         .vm_flags       = VM_READ | VM_EXEC
 868 };
 869
 870 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 871 {
 872 #ifdef CONFIG_IA32_EMULATION
 873         if (test_tsk_thread_flag(tsk, TIF_IA32))
 874                 return NULL;
 875 #endif
 876         return &gate_vma;
 877 }
 878
 879 int in_gate_area(struct task_struct *task, unsigned long addr)
 880 {
 881         struct vm_area_struct *vma = get_gate_vma(task);
 882
 883         if (!vma)
 884                 return 0;
 885
 886         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 887 }
 888
 889 /*
 890  * Use this when you have no reliable task/vma, typically from interrupt
 891  * context. It is less reliable than using the task's vma and may give
 892  * false positives:
 893  */
 894 int in_gate_area_no_task(unsigned long addr)
 895 {
 896         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 897 }
 898
 899 const char *arch_vma_name(struct vm_area_struct *vma)
 900 {
 901         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 902                 return "[vdso]";
 903         if (vma == &gate_vma)
 904                 return "[vsyscall]";
 905         return NULL;
 906 }
 907
 908 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 909 /*
 910  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 911  */
 912 static long __meminitdata addr_start, addr_end;
 913 static void __meminitdata *p_start, *p_end;
 914 static int __meminitdata node_start;
 915
 916 int __meminit
 917 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 918 {
 919         unsigned long addr = (unsigned long)start_page;
 920         unsigned long end = (unsigned long)(start_page + size);
 921         unsigned long next;
 922         pgd_t *pgd;
 923         pud_t *pud;
 924         pmd_t *pmd;
 925
 926         for (; addr < end; addr = next) {
 927                 void *p = NULL;
 928
 929                 pgd = vmemmap_pgd_populate(addr, node);
 930                 if (!pgd)
 931                         return -ENOMEM;
 932
 933                 pud = vmemmap_pud_populate(pgd, addr, node);
 934                 if (!pud)
 935                         return -ENOMEM;
 936
 937                 if (!cpu_has_pse) {
 938                         next = (addr + PAGE_SIZE) & PAGE_MASK;
 939                         pmd = vmemmap_pmd_populate(pud, addr, node);
 940
 941                         if (!pmd)
 942                                 return -ENOMEM;
 943
 944                         p = vmemmap_pte_populate(pmd, addr, node);
 945
 946                         if (!p)
 947                                 return -ENOMEM;
 948
 949                         addr_end = addr + PAGE_SIZE;
 950                         p_end = p + PAGE_SIZE;
 951                 } else {
 952                         next = pmd_addr_end(addr, end);
 953
 954                         pmd = pmd_offset(pud, addr);
 955                         if (pmd_none(*pmd)) {
 956                                 pte_t entry;
 957
 958                                 p = vmemmap_alloc_block(PMD_SIZE, node);
 959                                 if (!p)
 960                                         return -ENOMEM;
 961
 962                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 963                                                 PAGE_KERNEL_LARGE);
 964                                 set_pmd(pmd, __pmd(pte_val(entry)));
 965
 966                                 /* check to see if we have contiguous blocks */
 967                                 if (p_end != p || node_start != node) {
 968                                         if (p_start)
 969                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 970                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
 971                                         addr_start = addr;
 972                                         node_start = node;
 973                                         p_start = p;
 974                                 }
 975
 976                                 addr_end = addr + PMD_SIZE;
 977                                 p_end = p + PMD_SIZE;
 978                         } else
 979                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
 980                 }
 981
 982         }
 983         return 0;
 984 }
 985
 986 void __meminit vmemmap_populate_print_last(void)
 987 {
 988         if (p_start) {
 989                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 990                         addr_start, addr_end-1, p_start, p_end-1, node_start);
 991                 p_start = NULL;
 992                 p_end = NULL;
 993                 node_start = 0;
 994         }
 995 }
 996 #endif