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@ucw.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/memblock.h>
  25 #include <linux/proc_fs.h>
  26 #include <linux/pci.h>
  27 #include <linux/pfn.h>
  28 #include <linux/poison.h>
  29 #include <linux/dma-mapping.h>
  30 #include <linux/module.h>
  31 #include <linux/memory.h>
  32 #include <linux/memory_hotplug.h>
  33 #include <linux/nmi.h>
  34 #include <linux/gfp.h>
  35
  36 #include <asm/processor.h>
  37 #include <asm/bios_ebda.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/pgalloc.h>
  41 #include <asm/dma.h>
  42 #include <asm/fixmap.h>
  43 #include <asm/e820.h>
  44 #include <asm/apic.h>
  45 #include <asm/tlb.h>
  46 #include <asm/mmu_context.h>
  47 #include <asm/proto.h>
  48 #include <asm/smp.h>
  49 #include <asm/sections.h>
  50 #include <asm/kdebug.h>
  51 #include <asm/numa.h>
  52 #include <asm/cacheflush.h>
  53 #include <asm/init.h>
  54 #include <asm/uv/uv.h>
  55 #include <asm/setup.h>
  56
  57 #include "mm_internal.h"
  58
  59 static void ident_pmd_init(unsigned long pmd_flag, pmd_t *pmd_page,
  60                            unsigned long addr, unsigned long end)
  61 {
  62         addr &= PMD_MASK;
  63         for (; addr < end; addr += PMD_SIZE) {
  64                 pmd_t *pmd = pmd_page + pmd_index(addr);
  65
  66                 if (!pmd_present(*pmd))
  67                         set_pmd(pmd, __pmd(addr | pmd_flag));
  68         }
  69 }
  70 static int ident_pud_init(struct x86_mapping_info *info, pud_t *pud_page,
  71                           unsigned long addr, unsigned long end)
  72 {
  73         unsigned long next;
  74
  75         for (; addr < end; addr = next) {
  76                 pud_t *pud = pud_page + pud_index(addr);
  77                 pmd_t *pmd;
  78
  79                 next = (addr & PUD_MASK) + PUD_SIZE;
  80                 if (next > end)
  81                         next = end;
  82
  83                 if (pud_present(*pud)) {
  84                         pmd = pmd_offset(pud, 0);
  85                         ident_pmd_init(info->pmd_flag, pmd, addr, next);
  86                         continue;
  87                 }
  88                 pmd = (pmd_t *)info->alloc_pgt_page(info->context);
  89                 if (!pmd)
  90                         return -ENOMEM;
  91                 ident_pmd_init(info->pmd_flag, pmd, addr, next);
  92                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
  93         }
  94
  95         return 0;
  96 }
  97
  98 int kernel_ident_mapping_init(struct x86_mapping_info *info, pgd_t *pgd_page,
  99                               unsigned long addr, unsigned long end)
 100 {
 101         unsigned long next;
 102         int result;
 103         int off = info->kernel_mapping ? pgd_index(__PAGE_OFFSET) : 0;
 104
 105         for (; addr < end; addr = next) {
 106                 pgd_t *pgd = pgd_page + pgd_index(addr) + off;
 107                 pud_t *pud;
 108
 109                 next = (addr & PGDIR_MASK) + PGDIR_SIZE;
 110                 if (next > end)
 111                         next = end;
 112
 113                 if (pgd_present(*pgd)) {
 114                         pud = pud_offset(pgd, 0);
 115                         result = ident_pud_init(info, pud, addr, next);
 116                         if (result)
 117                                 return result;
 118                         continue;
 119                 }
 120
 121                 pud = (pud_t *)info->alloc_pgt_page(info->context);
 122                 if (!pud)
 123                         return -ENOMEM;
 124                 result = ident_pud_init(info, pud, addr, next);
 125                 if (result)
 126                         return result;
 127                 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE));
 128         }
 129
 130         return 0;
 131 }
 132
 133 static int __init parse_direct_gbpages_off(char *arg)
 134 {
 135         direct_gbpages = 0;
 136         return 0;
 137 }
 138 early_param("nogbpages", parse_direct_gbpages_off);
 139
 140 static int __init parse_direct_gbpages_on(char *arg)
 141 {
 142         direct_gbpages = 1;
 143         return 0;
 144 }
 145 early_param("gbpages", parse_direct_gbpages_on);
 146
 147 /*
 148  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
 149  * physical space so we can cache the place of the first one and move
 150  * around without checking the pgd every time.
 151  */
 152
 153 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
 154 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 155
 156 int force_personality32;
 157
 158 /*
 159  * noexec32=on|off
 160  * Control non executable heap for 32bit processes.
 161  * To control the stack too use noexec=off
 162  *
 163  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 164  * off  PROT_READ implies PROT_EXEC
 165  */
 166 static int __init nonx32_setup(char *str)
 167 {
 168         if (!strcmp(str, "on"))
 169                 force_personality32 &= ~READ_IMPLIES_EXEC;
 170         else if (!strcmp(str, "off"))
 171                 force_personality32 |= READ_IMPLIES_EXEC;
 172         return 1;
 173 }
 174 __setup("noexec32=", nonx32_setup);
 175
 176 /*
 177  * When memory was added/removed make sure all the processes MM have
 178  * suitable PGD entries in the local PGD level page.
 179  */
 180 void sync_global_pgds(unsigned long start, unsigned long end)
 181 {
 182         unsigned long address;
 183
 184         for (address = start; address <= end; address += PGDIR_SIZE) {
 185                 const pgd_t *pgd_ref = pgd_offset_k(address);
 186                 struct page *page;
 187
 188                 if (pgd_none(*pgd_ref))
 189                         continue;
 190
 191                 spin_lock(&pgd_lock);
 192                 list_for_each_entry(page, &pgd_list, lru) {
 193                         pgd_t *pgd;
 194                         spinlock_t *pgt_lock;
 195
 196                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 197                         /* the pgt_lock only for Xen */
 198                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 199                         spin_lock(pgt_lock);
 200
 201                         if (pgd_none(*pgd))
 202                                 set_pgd(pgd, *pgd_ref);
 203                         else
 204                                 BUG_ON(pgd_page_vaddr(*pgd)
 205                                        != pgd_page_vaddr(*pgd_ref));
 206
 207                         spin_unlock(pgt_lock);
 208                 }
 209                 spin_unlock(&pgd_lock);
 210         }
 211 }
 212
 213 /*
 214  * NOTE: This function is marked __ref because it calls __init function
 215  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 216  */
 217 static __ref void *spp_getpage(void)
 218 {
 219         void *ptr;
 220
 221         if (after_bootmem)
 222                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 223         else
 224                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 225
 226         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 227                 panic("set_pte_phys: cannot allocate page data %s\n",
 228                         after_bootmem ? "after bootmem" : "");
 229         }
 230
 231         pr_debug("spp_getpage %p\n", ptr);
 232
 233         return ptr;
 234 }
 235
 236 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 237 {
 238         if (pgd_none(*pgd)) {
 239                 pud_t *pud = (pud_t *)spp_getpage();
 240                 pgd_populate(&init_mm, pgd, pud);
 241                 if (pud != pud_offset(pgd, 0))
 242                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 243                                pud, pud_offset(pgd, 0));
 244         }
 245         return pud_offset(pgd, vaddr);
 246 }
 247
 248 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 249 {
 250         if (pud_none(*pud)) {
 251                 pmd_t *pmd = (pmd_t *) spp_getpage();
 252                 pud_populate(&init_mm, pud, pmd);
 253                 if (pmd != pmd_offset(pud, 0))
 254                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 255                                pmd, pmd_offset(pud, 0));
 256         }
 257         return pmd_offset(pud, vaddr);
 258 }
 259
 260 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 261 {
 262         if (pmd_none(*pmd)) {
 263                 pte_t *pte = (pte_t *) spp_getpage();
 264                 pmd_populate_kernel(&init_mm, pmd, pte);
 265                 if (pte != pte_offset_kernel(pmd, 0))
 266                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 267         }
 268         return pte_offset_kernel(pmd, vaddr);
 269 }
 270
 271 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 272 {
 273         pud_t *pud;
 274         pmd_t *pmd;
 275         pte_t *pte;
 276
 277         pud = pud_page + pud_index(vaddr);
 278         pmd = fill_pmd(pud, vaddr);
 279         pte = fill_pte(pmd, vaddr);
 280
 281         set_pte(pte, new_pte);
 282
 283         /*
 284          * It's enough to flush this one mapping.
 285          * (PGE mappings get flushed as well)
 286          */
 287         __flush_tlb_one(vaddr);
 288 }
 289
 290 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 291 {
 292         pgd_t *pgd;
 293         pud_t *pud_page;
 294
 295         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 296
 297         pgd = pgd_offset_k(vaddr);
 298         if (pgd_none(*pgd)) {
 299                 printk(KERN_ERR
 300                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 301                 return;
 302         }
 303         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 304         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 305 }
 306
 307 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 308 {
 309         pgd_t *pgd;
 310         pud_t *pud;
 311
 312         pgd = pgd_offset_k(vaddr);
 313         pud = fill_pud(pgd, vaddr);
 314         return fill_pmd(pud, vaddr);
 315 }
 316
 317 pte_t * __init populate_extra_pte(unsigned long vaddr)
 318 {
 319         pmd_t *pmd;
 320
 321         pmd = populate_extra_pmd(vaddr);
 322         return fill_pte(pmd, vaddr);
 323 }
 324
 325 /*
 326  * Create large page table mappings for a range of physical addresses.
 327  */
 328 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 329                                                 pgprot_t prot)
 330 {
 331         pgd_t *pgd;
 332         pud_t *pud;
 333         pmd_t *pmd;
 334
 335         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 336         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 337                 pgd = pgd_offset_k((unsigned long)__va(phys));
 338                 if (pgd_none(*pgd)) {
 339                         pud = (pud_t *) spp_getpage();
 340                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 341                                                 _PAGE_USER));
 342                 }
 343                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 344                 if (pud_none(*pud)) {
 345                         pmd = (pmd_t *) spp_getpage();
 346                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 347                                                 _PAGE_USER));
 348                 }
 349                 pmd = pmd_offset(pud, phys);
 350                 BUG_ON(!pmd_none(*pmd));
 351                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 352         }
 353 }
 354
 355 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 356 {
 357         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 358 }
 359
 360 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 361 {
 362         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 363 }
 364
 365 /*
 366  * The head.S code sets up the kernel high mapping:
 367  *
 368  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 369  *
 370  * phys_addr holds the negative offset to the kernel, which is added
 371  * to the compile time generated pmds. This results in invalid pmds up
 372  * to the point where we hit the physaddr 0 mapping.
 373  *
 374  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 375  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 376  * well, as they are located before _text:
 377  */
 378 void __init cleanup_highmap(void)
 379 {
 380         unsigned long vaddr = __START_KERNEL_map;
 381         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 382         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 383         pmd_t *pmd = level2_kernel_pgt;
 384
 385         /*
 386          * Native path, max_pfn_mapped is not set yet.
 387          * Xen has valid max_pfn_mapped set in
 388          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 389          */
 390         if (max_pfn_mapped)
 391                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 392
 393         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 394                 if (pmd_none(*pmd))
 395                         continue;
 396                 if (vaddr < (unsigned long) _text || vaddr > end)
 397                         set_pmd(pmd, __pmd(0));
 398         }
 399 }
 400
 401 static unsigned long __meminit
 402 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 403               pgprot_t prot)
 404 {
 405         unsigned long pages = 0, next;
 406         unsigned long last_map_addr = end;
 407         int i;
 408
 409         pte_t *pte = pte_page + pte_index(addr);
 410
 411         for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) {
 412                 next = (addr & PAGE_MASK) + PAGE_SIZE;
 413                 if (addr >= end) {
 414                         if (!after_bootmem &&
 415                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) &&
 416                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN))
 417                                 set_pte(pte, __pte(0));
 418                         continue;
 419                 }
 420
 421                 /*
 422                  * We will re-use the existing mapping.
 423                  * Xen for example has some special requirements, like mapping
 424                  * pagetable pages as RO. So assume someone who pre-setup
 425                  * these mappings are more intelligent.
 426                  */
 427                 if (pte_val(*pte)) {
 428                         if (!after_bootmem)
 429                                 pages++;
 430                         continue;
 431                 }
 432
 433                 if (0)
 434                         printk("   pte=%p addr=%lx pte=%016lx\n",
 435                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 436                 pages++;
 437                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 438                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 439         }
 440
 441         update_page_count(PG_LEVEL_4K, pages);
 442
 443         return last_map_addr;
 444 }
 445
 446 static unsigned long __meminit
 447 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 448               unsigned long page_size_mask, pgprot_t prot)
 449 {
 450         unsigned long pages = 0, next;
 451         unsigned long last_map_addr = end;
 452
 453         int i = pmd_index(address);
 454
 455         for (; i < PTRS_PER_PMD; i++, address = next) {
 456                 pmd_t *pmd = pmd_page + pmd_index(address);
 457                 pte_t *pte;
 458                 pgprot_t new_prot = prot;
 459
 460                 next = (address & PMD_MASK) + PMD_SIZE;
 461                 if (address >= end) {
 462                         if (!after_bootmem &&
 463                             !e820_any_mapped(address & PMD_MASK, next, E820_RAM) &&
 464                             !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN))
 465                                 set_pmd(pmd, __pmd(0));
 466                         continue;
 467                 }
 468
 469                 if (pmd_val(*pmd)) {
 470                         if (!pmd_large(*pmd)) {
 471                                 spin_lock(&init_mm.page_table_lock);
 472                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 473                                 last_map_addr = phys_pte_init(pte, address,
 474                                                                 end, prot);
 475                                 spin_unlock(&init_mm.page_table_lock);
 476                                 continue;
 477                         }
 478                         /*
 479                          * If we are ok with PG_LEVEL_2M mapping, then we will
 480                          * use the existing mapping,
 481                          *
 482                          * Otherwise, we will split the large page mapping but
 483                          * use the same existing protection bits except for
 484                          * large page, so that we don't violate Intel's TLB
 485                          * Application note (317080) which says, while changing
 486                          * the page sizes, new and old translations should
 487                          * not differ with respect to page frame and
 488                          * attributes.
 489                          */
 490                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 491                                 if (!after_bootmem)
 492                                         pages++;
 493                                 last_map_addr = next;
 494                                 continue;
 495                         }
 496                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 497                 }
 498
 499                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 500                         pages++;
 501                         spin_lock(&init_mm.page_table_lock);
 502                         set_pte((pte_t *)pmd,
 503                                 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT,
 504                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 505                         spin_unlock(&init_mm.page_table_lock);
 506                         last_map_addr = next;
 507                         continue;
 508                 }
 509
 510                 pte = alloc_low_page();
 511                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 512
 513                 spin_lock(&init_mm.page_table_lock);
 514                 pmd_populate_kernel(&init_mm, pmd, pte);
 515                 spin_unlock(&init_mm.page_table_lock);
 516         }
 517         update_page_count(PG_LEVEL_2M, pages);
 518         return last_map_addr;
 519 }
 520
 521 static unsigned long __meminit
 522 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 523                          unsigned long page_size_mask)
 524 {
 525         unsigned long pages = 0, next;
 526         unsigned long last_map_addr = end;
 527         int i = pud_index(addr);
 528
 529         for (; i < PTRS_PER_PUD; i++, addr = next) {
 530                 pud_t *pud = pud_page + pud_index(addr);
 531                 pmd_t *pmd;
 532                 pgprot_t prot = PAGE_KERNEL;
 533
 534                 next = (addr & PUD_MASK) + PUD_SIZE;
 535                 if (addr >= end) {
 536                         if (!after_bootmem &&
 537                             !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) &&
 538                             !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN))
 539                                 set_pud(pud, __pud(0));
 540                         continue;
 541                 }
 542
 543                 if (pud_val(*pud)) {
 544                         if (!pud_large(*pud)) {
 545                                 pmd = pmd_offset(pud, 0);
 546                                 last_map_addr = phys_pmd_init(pmd, addr, end,
 547                                                          page_size_mask, prot);
 548                                 __flush_tlb_all();
 549                                 continue;
 550                         }
 551                         /*
 552                          * If we are ok with PG_LEVEL_1G mapping, then we will
 553                          * use the existing mapping.
 554                          *
 555                          * Otherwise, we will split the gbpage mapping but use
 556                          * the same existing protection  bits except for large
 557                          * page, so that we don't violate Intel's TLB
 558                          * Application note (317080) which says, while changing
 559                          * the page sizes, new and old translations should
 560                          * not differ with respect to page frame and
 561                          * attributes.
 562                          */
 563                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 564                                 if (!after_bootmem)
 565                                         pages++;
 566                                 last_map_addr = next;
 567                                 continue;
 568                         }
 569                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 570                 }
 571
 572                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 573                         pages++;
 574                         spin_lock(&init_mm.page_table_lock);
 575                         set_pte((pte_t *)pud,
 576                                 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT,
 577                                         PAGE_KERNEL_LARGE));
 578                         spin_unlock(&init_mm.page_table_lock);
 579                         last_map_addr = next;
 580                         continue;
 581                 }
 582
 583                 pmd = alloc_low_page();
 584                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 585                                               prot);
 586
 587                 spin_lock(&init_mm.page_table_lock);
 588                 pud_populate(&init_mm, pud, pmd);
 589                 spin_unlock(&init_mm.page_table_lock);
 590         }
 591         __flush_tlb_all();
 592
 593         update_page_count(PG_LEVEL_1G, pages);
 594
 595         return last_map_addr;
 596 }
 597
 598 unsigned long __meminit
 599 kernel_physical_mapping_init(unsigned long start,
 600                              unsigned long end,
 601                              unsigned long page_size_mask)
 602 {
 603         bool pgd_changed = false;
 604         unsigned long next, last_map_addr = end;
 605         unsigned long addr;
 606
 607         start = (unsigned long)__va(start);
 608         end = (unsigned long)__va(end);
 609         addr = start;
 610
 611         for (; start < end; start = next) {
 612                 pgd_t *pgd = pgd_offset_k(start);
 613                 pud_t *pud;
 614
 615                 next = (start & PGDIR_MASK) + PGDIR_SIZE;
 616
 617                 if (pgd_val(*pgd)) {
 618                         pud = (pud_t *)pgd_page_vaddr(*pgd);
 619                         last_map_addr = phys_pud_init(pud, __pa(start),
 620                                                  __pa(end), page_size_mask);
 621                         continue;
 622                 }
 623
 624                 pud = alloc_low_page();
 625                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(end),
 626                                                  page_size_mask);
 627
 628                 spin_lock(&init_mm.page_table_lock);
 629                 pgd_populate(&init_mm, pgd, pud);
 630                 spin_unlock(&init_mm.page_table_lock);
 631                 pgd_changed = true;
 632         }
 633
 634         if (pgd_changed)
 635                 sync_global_pgds(addr, end - 1);
 636
 637         __flush_tlb_all();
 638
 639         return last_map_addr;
 640 }
 641
 642 #ifndef CONFIG_NUMA
 643 void __init initmem_init(void)
 644 {
 645         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
 646 }
 647 #endif
 648
 649 void __init paging_init(void)
 650 {
 651         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 652         sparse_init();
 653
 654         /*
 655          * clear the default setting with node 0
 656          * note: don't use nodes_clear here, that is really clearing when
 657          *       numa support is not compiled in, and later node_set_state
 658          *       will not set it back.
 659          */
 660         node_clear_state(0, N_MEMORY);
 661         if (N_MEMORY != N_NORMAL_MEMORY)
 662                 node_clear_state(0, N_NORMAL_MEMORY);
 663
 664         zone_sizes_init();
 665 }
 666
 667 /*
 668  * Memory hotplug specific functions
 669  */
 670 #ifdef CONFIG_MEMORY_HOTPLUG
 671 /*
 672  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 673  * updating.
 674  */
 675 static void  update_end_of_memory_vars(u64 start, u64 size)
 676 {
 677         unsigned long end_pfn = PFN_UP(start + size);
 678
 679         if (end_pfn > max_pfn) {
 680                 max_pfn = end_pfn;
 681                 max_low_pfn = end_pfn;
 682                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 683         }
 684 }
 685
 686 /*
 687  * Memory is added always to NORMAL zone. This means you will never get
 688  * additional DMA/DMA32 memory.
 689  */
 690 int arch_add_memory(int nid, u64 start, u64 size)
 691 {
 692         struct pglist_data *pgdat = NODE_DATA(nid);
 693         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 694         unsigned long start_pfn = start >> PAGE_SHIFT;
 695         unsigned long nr_pages = size >> PAGE_SHIFT;
 696         int ret;
 697
 698         init_memory_mapping(start, start + size);
 699
 700         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 701         WARN_ON_ONCE(ret);
 702
 703         /* update max_pfn, max_low_pfn and high_memory */
 704         update_end_of_memory_vars(start, size);
 705
 706         return ret;
 707 }
 708 EXPORT_SYMBOL_GPL(arch_add_memory);
 709
 710 #define PAGE_INUSE 0xFD
 711
 712 static void __meminit free_pagetable(struct page *page, int order)
 713 {
 714         struct zone *zone;
 715         bool bootmem = false;
 716         unsigned long magic;
 717         unsigned int nr_pages = 1 << order;
 718
 719         /* bootmem page has reserved flag */
 720         if (PageReserved(page)) {
 721                 __ClearPageReserved(page);
 722                 bootmem = true;
 723
 724                 magic = (unsigned long)page->lru.next;
 725                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 726                         while (nr_pages--)
 727                                 put_page_bootmem(page++);
 728                 } else
 729                         __free_pages_bootmem(page, order);
 730         } else
 731                 free_pages((unsigned long)page_address(page), order);
 732
 733         /*
 734          * SECTION_INFO pages and MIX_SECTION_INFO pages
 735          * are all allocated by bootmem.
 736          */
 737         if (bootmem) {
 738                 zone = page_zone(page);
 739                 zone_span_writelock(zone);
 740                 zone->present_pages += nr_pages;
 741                 zone_span_writeunlock(zone);
 742                 totalram_pages += nr_pages;
 743         }
 744 }
 745
 746 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 747 {
 748         pte_t *pte;
 749         int i;
 750
 751         for (i = 0; i < PTRS_PER_PTE; i++) {
 752                 pte = pte_start + i;
 753                 if (pte_val(*pte))
 754                         return;
 755         }
 756
 757         /* free a pte talbe */
 758         free_pagetable(pmd_page(*pmd), 0);
 759         spin_lock(&init_mm.page_table_lock);
 760         pmd_clear(pmd);
 761         spin_unlock(&init_mm.page_table_lock);
 762 }
 763
 764 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 765 {
 766         pmd_t *pmd;
 767         int i;
 768
 769         for (i = 0; i < PTRS_PER_PMD; i++) {
 770                 pmd = pmd_start + i;
 771                 if (pmd_val(*pmd))
 772                         return;
 773         }
 774
 775         /* free a pmd talbe */
 776         free_pagetable(pud_page(*pud), 0);
 777         spin_lock(&init_mm.page_table_lock);
 778         pud_clear(pud);
 779         spin_unlock(&init_mm.page_table_lock);
 780 }
 781
 782 /* Return true if pgd is changed, otherwise return false. */
 783 static bool __meminit free_pud_table(pud_t *pud_start, pgd_t *pgd)
 784 {
 785         pud_t *pud;
 786         int i;
 787
 788         for (i = 0; i < PTRS_PER_PUD; i++) {
 789                 pud = pud_start + i;
 790                 if (pud_val(*pud))
 791                         return false;
 792         }
 793
 794         /* free a pud table */
 795         free_pagetable(pgd_page(*pgd), 0);
 796         spin_lock(&init_mm.page_table_lock);
 797         pgd_clear(pgd);
 798         spin_unlock(&init_mm.page_table_lock);
 799
 800         return true;
 801 }
 802
 803 static void __meminit
 804 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 805                  bool direct)
 806 {
 807         unsigned long next, pages = 0;
 808         pte_t *pte;
 809         void *page_addr;
 810         phys_addr_t phys_addr;
 811
 812         pte = pte_start + pte_index(addr);
 813         for (; addr < end; addr = next, pte++) {
 814                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 815                 if (next > end)
 816                         next = end;
 817
 818                 if (!pte_present(*pte))
 819                         continue;
 820
 821                 /*
 822                  * We mapped [0,1G) memory as identity mapping when
 823                  * initializing, in arch/x86/kernel/head_64.S. These
 824                  * pagetables cannot be removed.
 825                  */
 826                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 827                 if (phys_addr < (phys_addr_t)0x40000000)
 828                         return;
 829
 830                 if (IS_ALIGNED(addr, PAGE_SIZE) &&
 831                     IS_ALIGNED(next, PAGE_SIZE)) {
 832                         /*
 833                          * Do not free direct mapping pages since they were
 834                          * freed when offlining, or simplely not in use.
 835                          */
 836                         if (!direct)
 837                                 free_pagetable(pte_page(*pte), 0);
 838
 839                         spin_lock(&init_mm.page_table_lock);
 840                         pte_clear(&init_mm, addr, pte);
 841                         spin_unlock(&init_mm.page_table_lock);
 842
 843                         /* For non-direct mapping, pages means nothing. */
 844                         pages++;
 845                 } else {
 846                         /*
 847                          * If we are here, we are freeing vmemmap pages since
 848                          * direct mapped memory ranges to be freed are aligned.
 849                          *
 850                          * If we are not removing the whole page, it means
 851                          * other page structs in this page are being used and
 852                          * we canot remove them. So fill the unused page_structs
 853                          * with 0xFD, and remove the page when it is wholly
 854                          * filled with 0xFD.
 855                          */
 856                         memset((void *)addr, PAGE_INUSE, next - addr);
 857
 858                         page_addr = page_address(pte_page(*pte));
 859                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
 860                                 free_pagetable(pte_page(*pte), 0);
 861
 862                                 spin_lock(&init_mm.page_table_lock);
 863                                 pte_clear(&init_mm, addr, pte);
 864                                 spin_unlock(&init_mm.page_table_lock);
 865                         }
 866                 }
 867         }
 868
 869         /* Call free_pte_table() in remove_pmd_table(). */
 870         flush_tlb_all();
 871         if (direct)
 872                 update_page_count(PG_LEVEL_4K, -pages);
 873 }
 874
 875 static void __meminit
 876 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
 877                  bool direct)
 878 {
 879         unsigned long next, pages = 0;
 880         pte_t *pte_base;
 881         pmd_t *pmd;
 882         void *page_addr;
 883
 884         pmd = pmd_start + pmd_index(addr);
 885         for (; addr < end; addr = next, pmd++) {
 886                 next = pmd_addr_end(addr, end);
 887
 888                 if (!pmd_present(*pmd))
 889                         continue;
 890
 891                 if (pmd_large(*pmd)) {
 892                         if (IS_ALIGNED(addr, PMD_SIZE) &&
 893                             IS_ALIGNED(next, PMD_SIZE)) {
 894                                 if (!direct)
 895                                         free_pagetable(pmd_page(*pmd),
 896                                                        get_order(PMD_SIZE));
 897
 898                                 spin_lock(&init_mm.page_table_lock);
 899                                 pmd_clear(pmd);
 900                                 spin_unlock(&init_mm.page_table_lock);
 901                                 pages++;
 902                         } else {
 903                                 /* If here, we are freeing vmemmap pages. */
 904                                 memset((void *)addr, PAGE_INUSE, next - addr);
 905
 906                                 page_addr = page_address(pmd_page(*pmd));
 907                                 if (!memchr_inv(page_addr, PAGE_INUSE,
 908                                                 PMD_SIZE)) {
 909                                         free_pagetable(pmd_page(*pmd),
 910                                                        get_order(PMD_SIZE));
 911
 912                                         spin_lock(&init_mm.page_table_lock);
 913                                         pmd_clear(pmd);
 914                                         spin_unlock(&init_mm.page_table_lock);
 915                                 }
 916                         }
 917
 918                         continue;
 919                 }
 920
 921                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
 922                 remove_pte_table(pte_base, addr, next, direct);
 923                 free_pte_table(pte_base, pmd);
 924         }
 925
 926         /* Call free_pmd_table() in remove_pud_table(). */
 927         if (direct)
 928                 update_page_count(PG_LEVEL_2M, -pages);
 929 }
 930
 931 static void __meminit
 932 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
 933                  bool direct)
 934 {
 935         unsigned long next, pages = 0;
 936         pmd_t *pmd_base;
 937         pud_t *pud;
 938         void *page_addr;
 939
 940         pud = pud_start + pud_index(addr);
 941         for (; addr < end; addr = next, pud++) {
 942                 next = pud_addr_end(addr, end);
 943
 944                 if (!pud_present(*pud))
 945                         continue;
 946
 947                 if (pud_large(*pud)) {
 948                         if (IS_ALIGNED(addr, PUD_SIZE) &&
 949                             IS_ALIGNED(next, PUD_SIZE)) {
 950                                 if (!direct)
 951                                         free_pagetable(pud_page(*pud),
 952                                                        get_order(PUD_SIZE));
 953
 954                                 spin_lock(&init_mm.page_table_lock);
 955                                 pud_clear(pud);
 956                                 spin_unlock(&init_mm.page_table_lock);
 957                                 pages++;
 958                         } else {
 959                                 /* If here, we are freeing vmemmap pages. */
 960                                 memset((void *)addr, PAGE_INUSE, next - addr);
 961
 962                                 page_addr = page_address(pud_page(*pud));
 963                                 if (!memchr_inv(page_addr, PAGE_INUSE,
 964                                                 PUD_SIZE)) {
 965                                         free_pagetable(pud_page(*pud),
 966                                                        get_order(PUD_SIZE));
 967
 968                                         spin_lock(&init_mm.page_table_lock);
 969                                         pud_clear(pud);
 970                                         spin_unlock(&init_mm.page_table_lock);
 971                                 }
 972                         }
 973
 974                         continue;
 975                 }
 976
 977                 pmd_base = (pmd_t *)pud_page_vaddr(*pud);
 978                 remove_pmd_table(pmd_base, addr, next, direct);
 979                 free_pmd_table(pmd_base, pud);
 980         }
 981
 982         if (direct)
 983                 update_page_count(PG_LEVEL_1G, -pages);
 984 }
 985
 986 /* start and end are both virtual address. */
 987 static void __meminit
 988 remove_pagetable(unsigned long start, unsigned long end, bool direct)
 989 {
 990         unsigned long next;
 991         pgd_t *pgd;
 992         pud_t *pud;
 993         bool pgd_changed = false;
 994
 995         for (; start < end; start = next) {
 996                 next = pgd_addr_end(start, end);
 997
 998                 pgd = pgd_offset_k(start);
 999                 if (!pgd_present(*pgd))
1000                         continue;
1001
1002                 pud = (pud_t *)pgd_page_vaddr(*pgd);
1003                 remove_pud_table(pud, start, next, direct);
1004                 if (free_pud_table(pud, pgd))
1005                         pgd_changed = true;
1006         }
1007
1008         if (pgd_changed)
1009                 sync_global_pgds(start, end - 1);
1010
1011         flush_tlb_all();
1012 }
1013
1014 void __ref vmemmap_free(unsigned long start, unsigned long end)
1015 {
1016         remove_pagetable(start, end, false);
1017 }
1018
1019 static void __meminit
1020 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1021 {
1022         start = (unsigned long)__va(start);
1023         end = (unsigned long)__va(end);
1024
1025         remove_pagetable(start, end, true);
1026 }
1027
1028 #ifdef CONFIG_MEMORY_HOTREMOVE
1029 int __ref arch_remove_memory(u64 start, u64 size)
1030 {
1031         unsigned long start_pfn = start >> PAGE_SHIFT;
1032         unsigned long nr_pages = size >> PAGE_SHIFT;
1033         struct zone *zone;
1034         int ret;
1035
1036         zone = page_zone(pfn_to_page(start_pfn));
1037         kernel_physical_mapping_remove(start, start + size);
1038         ret = __remove_pages(zone, start_pfn, nr_pages);
1039         WARN_ON_ONCE(ret);
1040
1041         return ret;
1042 }
1043 #endif
1044 #endif /* CONFIG_MEMORY_HOTPLUG */
1045
1046 static struct kcore_list kcore_vsyscall;
1047
1048 static void __init register_page_bootmem_info(void)
1049 {
1050 #ifdef CONFIG_NUMA
1051         int i;
1052
1053         for_each_online_node(i)
1054                 register_page_bootmem_info_node(NODE_DATA(i));
1055 #endif
1056 }
1057
1058 void __init mem_init(void)
1059 {
1060         long codesize, reservedpages, datasize, initsize;
1061         unsigned long absent_pages;
1062
1063         pci_iommu_alloc();
1064
1065         /* clear_bss() already clear the empty_zero_page */
1066
1067         register_page_bootmem_info();
1068
1069         /* this will put all memory onto the freelists */
1070         totalram_pages = free_all_bootmem();
1071
1072         absent_pages = absent_pages_in_range(0, max_pfn);
1073         reservedpages = max_pfn - totalram_pages - absent_pages;
1074         after_bootmem = 1;
1075
1076         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
1077         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
1078         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
1079
1080         /* Register memory areas for /proc/kcore */
1081         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
1082                          VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
1083
1084         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
1085                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
1086                 nr_free_pages() << (PAGE_SHIFT-10),
1087                 max_pfn << (PAGE_SHIFT-10),
1088                 codesize >> 10,
1089                 absent_pages << (PAGE_SHIFT-10),
1090                 reservedpages << (PAGE_SHIFT-10),
1091                 datasize >> 10,
1092                 initsize >> 10);
1093 }
1094
1095 #ifdef CONFIG_DEBUG_RODATA
1096 const int rodata_test_data = 0xC3;
1097 EXPORT_SYMBOL_GPL(rodata_test_data);
1098
1099 int kernel_set_to_readonly;
1100
1101 void set_kernel_text_rw(void)
1102 {
1103         unsigned long start = PFN_ALIGN(_text);
1104         unsigned long end = PFN_ALIGN(__stop___ex_table);
1105
1106         if (!kernel_set_to_readonly)
1107                 return;
1108
1109         pr_debug("Set kernel text: %lx - %lx for read write\n",
1110                  start, end);
1111
1112         /*
1113          * Make the kernel identity mapping for text RW. Kernel text
1114          * mapping will always be RO. Refer to the comment in
1115          * static_protections() in pageattr.c
1116          */
1117         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1118 }
1119
1120 void set_kernel_text_ro(void)
1121 {
1122         unsigned long start = PFN_ALIGN(_text);
1123         unsigned long end = PFN_ALIGN(__stop___ex_table);
1124
1125         if (!kernel_set_to_readonly)
1126                 return;
1127
1128         pr_debug("Set kernel text: %lx - %lx for read only\n",
1129                  start, end);
1130
1131         /*
1132          * Set the kernel identity mapping for text RO.
1133          */
1134         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1135 }
1136
1137 void mark_rodata_ro(void)
1138 {
1139         unsigned long start = PFN_ALIGN(_text);
1140         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1141         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1142         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1143         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1144         unsigned long all_end = PFN_ALIGN(&_end);
1145
1146         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1147                (end - start) >> 10);
1148         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1149
1150         kernel_set_to_readonly = 1;
1151
1152         /*
1153          * The rodata/data/bss/brk section (but not the kernel text!)
1154          * should also be not-executable.
1155          */
1156         set_memory_nx(rodata_start, (all_end - rodata_start) >> PAGE_SHIFT);
1157
1158         rodata_test();
1159
1160 #ifdef CONFIG_CPA_DEBUG
1161         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1162         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1163
1164         printk(KERN_INFO "Testing CPA: again\n");
1165         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1166 #endif
1167
1168         free_init_pages("unused kernel memory",
1169                         (unsigned long) __va(__pa_symbol(text_end)),
1170                         (unsigned long) __va(__pa_symbol(rodata_start)));
1171
1172         free_init_pages("unused kernel memory",
1173                         (unsigned long) __va(__pa_symbol(rodata_end)),
1174                         (unsigned long) __va(__pa_symbol(_sdata)));
1175 }
1176
1177 #endif
1178
1179 int kern_addr_valid(unsigned long addr)
1180 {
1181         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1182         pgd_t *pgd;
1183         pud_t *pud;
1184         pmd_t *pmd;
1185         pte_t *pte;
1186
1187         if (above != 0 && above != -1UL)
1188                 return 0;
1189
1190         pgd = pgd_offset_k(addr);
1191         if (pgd_none(*pgd))
1192                 return 0;
1193
1194         pud = pud_offset(pgd, addr);
1195         if (pud_none(*pud))
1196                 return 0;
1197
1198         if (pud_large(*pud))
1199                 return pfn_valid(pud_pfn(*pud));
1200
1201         pmd = pmd_offset(pud, addr);
1202         if (pmd_none(*pmd))
1203                 return 0;
1204
1205         if (pmd_large(*pmd))
1206                 return pfn_valid(pmd_pfn(*pmd));
1207
1208         pte = pte_offset_kernel(pmd, addr);
1209         if (pte_none(*pte))
1210                 return 0;
1211
1212         return pfn_valid(pte_pfn(*pte));
1213 }
1214
1215 /*
1216  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1217  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1218  * not need special handling anymore:
1219  */
1220 static struct vm_area_struct gate_vma = {
1221         .vm_start       = VSYSCALL_START,
1222         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1223         .vm_page_prot   = PAGE_READONLY_EXEC,
1224         .vm_flags       = VM_READ | VM_EXEC
1225 };
1226
1227 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
1228 {
1229 #ifdef CONFIG_IA32_EMULATION
1230         if (!mm || mm->context.ia32_compat)
1231                 return NULL;
1232 #endif
1233         return &gate_vma;
1234 }
1235
1236 int in_gate_area(struct mm_struct *mm, unsigned long addr)
1237 {
1238         struct vm_area_struct *vma = get_gate_vma(mm);
1239
1240         if (!vma)
1241                 return 0;
1242
1243         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1244 }
1245
1246 /*
1247  * Use this when you have no reliable mm, typically from interrupt
1248  * context. It is less reliable than using a task's mm and may give
1249  * false positives.
1250  */
1251 int in_gate_area_no_mm(unsigned long addr)
1252 {
1253         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1254 }
1255
1256 const char *arch_vma_name(struct vm_area_struct *vma)
1257 {
1258         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1259                 return "[vdso]";
1260         if (vma == &gate_vma)
1261                 return "[vsyscall]";
1262         return NULL;
1263 }
1264
1265 #ifdef CONFIG_X86_UV
1266 unsigned long memory_block_size_bytes(void)
1267 {
1268         if (is_uv_system()) {
1269                 printk(KERN_INFO "UV: memory block size 2GB\n");
1270                 return 2UL * 1024 * 1024 * 1024;
1271         }
1272         return MIN_MEMORY_BLOCK_SIZE;
1273 }
1274 #endif
1275
1276 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1277 /*
1278  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1279  */
1280 static long __meminitdata addr_start, addr_end;
1281 static void __meminitdata *p_start, *p_end;
1282 static int __meminitdata node_start;
1283
1284 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
1285 {
1286         unsigned long addr;
1287         unsigned long next;
1288         pgd_t *pgd;
1289         pud_t *pud;
1290         pmd_t *pmd;
1291
1292         for (addr = start; addr < end; addr = next) {
1293                 void *p = NULL;
1294
1295                 pgd = vmemmap_pgd_populate(addr, node);
1296                 if (!pgd)
1297                         return -ENOMEM;
1298
1299                 pud = vmemmap_pud_populate(pgd, addr, node);
1300                 if (!pud)
1301                         return -ENOMEM;
1302
1303                 if (!cpu_has_pse) {
1304                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1305                         pmd = vmemmap_pmd_populate(pud, addr, node);
1306
1307                         if (!pmd)
1308                                 return -ENOMEM;
1309
1310                         p = vmemmap_pte_populate(pmd, addr, node);
1311
1312                         if (!p)
1313                                 return -ENOMEM;
1314                 } else {
1315                         next = pmd_addr_end(addr, end);
1316
1317                         pmd = pmd_offset(pud, addr);
1318                         if (pmd_none(*pmd)) {
1319                                 pte_t entry;
1320
1321                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1322                                 if (!p)
1323                                         return -ENOMEM;
1324
1325                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1326                                                 PAGE_KERNEL_LARGE);
1327                                 set_pmd(pmd, __pmd(pte_val(entry)));
1328
1329                                 /* check to see if we have contiguous blocks */
1330                                 if (p_end != p || node_start != node) {
1331                                         if (p_start)
1332                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1333                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1334                                         addr_start = addr;
1335                                         node_start = node;
1336                                         p_start = p;
1337                                 }
1338
1339                                 addr_end = addr + PMD_SIZE;
1340                                 p_end = p + PMD_SIZE;
1341                         } else
1342                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1343                 }
1344
1345         }
1346         sync_global_pgds(start, end - 1);
1347         return 0;
1348 }
1349
1350 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1351 void register_page_bootmem_memmap(unsigned long section_nr,
1352                                   struct page *start_page, unsigned long size)
1353 {
1354         unsigned long addr = (unsigned long)start_page;
1355         unsigned long end = (unsigned long)(start_page + size);
1356         unsigned long next;
1357         pgd_t *pgd;
1358         pud_t *pud;
1359         pmd_t *pmd;
1360         unsigned int nr_pages;
1361         struct page *page;
1362
1363         for (; addr < end; addr = next) {
1364                 pte_t *pte = NULL;
1365
1366                 pgd = pgd_offset_k(addr);
1367                 if (pgd_none(*pgd)) {
1368                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1369                         continue;
1370                 }
1371                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1372
1373                 pud = pud_offset(pgd, addr);
1374                 if (pud_none(*pud)) {
1375                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1376                         continue;
1377                 }
1378                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1379
1380                 if (!cpu_has_pse) {
1381                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1382                         pmd = pmd_offset(pud, addr);
1383                         if (pmd_none(*pmd))
1384                                 continue;
1385                         get_page_bootmem(section_nr, pmd_page(*pmd),
1386                                          MIX_SECTION_INFO);
1387
1388                         pte = pte_offset_kernel(pmd, addr);
1389                         if (pte_none(*pte))
1390                                 continue;
1391                         get_page_bootmem(section_nr, pte_page(*pte),
1392                                          SECTION_INFO);
1393                 } else {
1394                         next = pmd_addr_end(addr, end);
1395
1396                         pmd = pmd_offset(pud, addr);
1397                         if (pmd_none(*pmd))
1398                                 continue;
1399
1400                         nr_pages = 1 << (get_order(PMD_SIZE));
1401                         page = pmd_page(*pmd);
1402                         while (nr_pages--)
1403                                 get_page_bootmem(section_nr, page++,
1404                                                  SECTION_INFO);
1405                 }
1406         }
1407 }
1408 #endif
1409
1410 void __meminit vmemmap_populate_print_last(void)
1411 {
1412         if (p_start) {
1413                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1414                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1415                 p_start = NULL;
1416                 p_end = NULL;
1417                 node_start = 0;
1418         }
1419 }
1420 #endif