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 #include <linux/kcore.h>
  36
  37 #include <asm/processor.h>
  38 #include <asm/bios_ebda.h>
  39 #include <asm/uaccess.h>
  40 #include <asm/pgtable.h>
  41 #include <asm/pgalloc.h>
  42 #include <asm/dma.h>
  43 #include <asm/fixmap.h>
  44 #include <asm/e820.h>
  45 #include <asm/apic.h>
  46 #include <asm/tlb.h>
  47 #include <asm/mmu_context.h>
  48 #include <asm/proto.h>
  49 #include <asm/smp.h>
  50 #include <asm/sections.h>
  51 #include <asm/kdebug.h>
  52 #include <asm/numa.h>
  53 #include <asm/cacheflush.h>
  54 #include <asm/init.h>
  55 #include <asm/uv/uv.h>
  56 #include <asm/setup.h>
  57
  58 #include "mm_internal.h"
  59
  60 static void ident_pmd_init(unsigned long pmd_flag, pmd_t *pmd_page,
  61                            unsigned long addr, unsigned long end)
  62 {
  63         addr &= PMD_MASK;
  64         for (; addr < end; addr += PMD_SIZE) {
  65                 pmd_t *pmd = pmd_page + pmd_index(addr);
  66
  67                 if (!pmd_present(*pmd))
  68                         set_pmd(pmd, __pmd(addr | pmd_flag));
  69         }
  70 }
  71 static int ident_pud_init(struct x86_mapping_info *info, pud_t *pud_page,
  72                           unsigned long addr, unsigned long end)
  73 {
  74         unsigned long next;
  75
  76         for (; addr < end; addr = next) {
  77                 pud_t *pud = pud_page + pud_index(addr);
  78                 pmd_t *pmd;
  79
  80                 next = (addr & PUD_MASK) + PUD_SIZE;
  81                 if (next > end)
  82                         next = end;
  83
  84                 if (pud_present(*pud)) {
  85                         pmd = pmd_offset(pud, 0);
  86                         ident_pmd_init(info->pmd_flag, pmd, addr, next);
  87                         continue;
  88                 }
  89                 pmd = (pmd_t *)info->alloc_pgt_page(info->context);
  90                 if (!pmd)
  91                         return -ENOMEM;
  92                 ident_pmd_init(info->pmd_flag, pmd, addr, next);
  93                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
  94         }
  95
  96         return 0;
  97 }
  98
  99 int kernel_ident_mapping_init(struct x86_mapping_info *info, pgd_t *pgd_page,
 100                               unsigned long addr, unsigned long end)
 101 {
 102         unsigned long next;
 103         int result;
 104         int off = info->kernel_mapping ? pgd_index(__PAGE_OFFSET) : 0;
 105
 106         for (; addr < end; addr = next) {
 107                 pgd_t *pgd = pgd_page + pgd_index(addr) + off;
 108                 pud_t *pud;
 109
 110                 next = (addr & PGDIR_MASK) + PGDIR_SIZE;
 111                 if (next > end)
 112                         next = end;
 113
 114                 if (pgd_present(*pgd)) {
 115                         pud = pud_offset(pgd, 0);
 116                         result = ident_pud_init(info, pud, addr, next);
 117                         if (result)
 118                                 return result;
 119                         continue;
 120                 }
 121
 122                 pud = (pud_t *)info->alloc_pgt_page(info->context);
 123                 if (!pud)
 124                         return -ENOMEM;
 125                 result = ident_pud_init(info, pud, addr, next);
 126                 if (result)
 127                         return result;
 128                 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE));
 129         }
 130
 131         return 0;
 132 }
 133
 134 static int __init parse_direct_gbpages_off(char *arg)
 135 {
 136         direct_gbpages = 0;
 137         return 0;
 138 }
 139 early_param("nogbpages", parse_direct_gbpages_off);
 140
 141 static int __init parse_direct_gbpages_on(char *arg)
 142 {
 143         direct_gbpages = 1;
 144         return 0;
 145 }
 146 early_param("gbpages", parse_direct_gbpages_on);
 147
 148 /*
 149  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
 150  * physical space so we can cache the place of the first one and move
 151  * around without checking the pgd every time.
 152  */
 153
 154 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
 155 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 156
 157 int force_personality32;
 158
 159 /*
 160  * noexec32=on|off
 161  * Control non executable heap for 32bit processes.
 162  * To control the stack too use noexec=off
 163  *
 164  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 165  * off  PROT_READ implies PROT_EXEC
 166  */
 167 static int __init nonx32_setup(char *str)
 168 {
 169         if (!strcmp(str, "on"))
 170                 force_personality32 &= ~READ_IMPLIES_EXEC;
 171         else if (!strcmp(str, "off"))
 172                 force_personality32 |= READ_IMPLIES_EXEC;
 173         return 1;
 174 }
 175 __setup("noexec32=", nonx32_setup);
 176
 177 /*
 178  * When memory was added/removed make sure all the processes MM have
 179  * suitable PGD entries in the local PGD level page.
 180  */
 181 void sync_global_pgds(unsigned long start, unsigned long end)
 182 {
 183         unsigned long address;
 184
 185         for (address = start; address <= end; address += PGDIR_SIZE) {
 186                 const pgd_t *pgd_ref = pgd_offset_k(address);
 187                 struct page *page;
 188
 189                 if (pgd_none(*pgd_ref))
 190                         continue;
 191
 192                 spin_lock(&pgd_lock);
 193                 list_for_each_entry(page, &pgd_list, lru) {
 194                         pgd_t *pgd;
 195                         spinlock_t *pgt_lock;
 196
 197                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 198                         /* the pgt_lock only for Xen */
 199                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 200                         spin_lock(pgt_lock);
 201
 202                         if (pgd_none(*pgd))
 203                                 set_pgd(pgd, *pgd_ref);
 204                         else
 205                                 BUG_ON(pgd_page_vaddr(*pgd)
 206                                        != pgd_page_vaddr(*pgd_ref));
 207
 208                         spin_unlock(pgt_lock);
 209                 }
 210                 spin_unlock(&pgd_lock);
 211         }
 212 }
 213
 214 /*
 215  * NOTE: This function is marked __ref because it calls __init function
 216  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 217  */
 218 static __ref void *spp_getpage(void)
 219 {
 220         void *ptr;
 221
 222         if (after_bootmem)
 223                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 224         else
 225                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 226
 227         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 228                 panic("set_pte_phys: cannot allocate page data %s\n",
 229                         after_bootmem ? "after bootmem" : "");
 230         }
 231
 232         pr_debug("spp_getpage %p\n", ptr);
 233
 234         return ptr;
 235 }
 236
 237 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 238 {
 239         if (pgd_none(*pgd)) {
 240                 pud_t *pud = (pud_t *)spp_getpage();
 241                 pgd_populate(&init_mm, pgd, pud);
 242                 if (pud != pud_offset(pgd, 0))
 243                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 244                                pud, pud_offset(pgd, 0));
 245         }
 246         return pud_offset(pgd, vaddr);
 247 }
 248
 249 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 250 {
 251         if (pud_none(*pud)) {
 252                 pmd_t *pmd = (pmd_t *) spp_getpage();
 253                 pud_populate(&init_mm, pud, pmd);
 254                 if (pmd != pmd_offset(pud, 0))
 255                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 256                                pmd, pmd_offset(pud, 0));
 257         }
 258         return pmd_offset(pud, vaddr);
 259 }
 260
 261 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 262 {
 263         if (pmd_none(*pmd)) {
 264                 pte_t *pte = (pte_t *) spp_getpage();
 265                 pmd_populate_kernel(&init_mm, pmd, pte);
 266                 if (pte != pte_offset_kernel(pmd, 0))
 267                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 268         }
 269         return pte_offset_kernel(pmd, vaddr);
 270 }
 271
 272 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 273 {
 274         pud_t *pud;
 275         pmd_t *pmd;
 276         pte_t *pte;
 277
 278         pud = pud_page + pud_index(vaddr);
 279         pmd = fill_pmd(pud, vaddr);
 280         pte = fill_pte(pmd, vaddr);
 281
 282         set_pte(pte, new_pte);
 283
 284         /*
 285          * It's enough to flush this one mapping.
 286          * (PGE mappings get flushed as well)
 287          */
 288         __flush_tlb_one(vaddr);
 289 }
 290
 291 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 292 {
 293         pgd_t *pgd;
 294         pud_t *pud_page;
 295
 296         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 297
 298         pgd = pgd_offset_k(vaddr);
 299         if (pgd_none(*pgd)) {
 300                 printk(KERN_ERR
 301                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 302                 return;
 303         }
 304         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 305         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 306 }
 307
 308 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 309 {
 310         pgd_t *pgd;
 311         pud_t *pud;
 312
 313         pgd = pgd_offset_k(vaddr);
 314         pud = fill_pud(pgd, vaddr);
 315         return fill_pmd(pud, vaddr);
 316 }
 317
 318 pte_t * __init populate_extra_pte(unsigned long vaddr)
 319 {
 320         pmd_t *pmd;
 321
 322         pmd = populate_extra_pmd(vaddr);
 323         return fill_pte(pmd, vaddr);
 324 }
 325
 326 /*
 327  * Create large page table mappings for a range of physical addresses.
 328  */
 329 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 330                                                 pgprot_t prot)
 331 {
 332         pgd_t *pgd;
 333         pud_t *pud;
 334         pmd_t *pmd;
 335
 336         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 337         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 338                 pgd = pgd_offset_k((unsigned long)__va(phys));
 339                 if (pgd_none(*pgd)) {
 340                         pud = (pud_t *) spp_getpage();
 341                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 342                                                 _PAGE_USER));
 343                 }
 344                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 345                 if (pud_none(*pud)) {
 346                         pmd = (pmd_t *) spp_getpage();
 347                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 348                                                 _PAGE_USER));
 349                 }
 350                 pmd = pmd_offset(pud, phys);
 351                 BUG_ON(!pmd_none(*pmd));
 352                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 353         }
 354 }
 355
 356 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 357 {
 358         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 359 }
 360
 361 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 362 {
 363         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 364 }
 365
 366 /*
 367  * The head.S code sets up the kernel high mapping:
 368  *
 369  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 370  *
 371  * phys_addr holds the negative offset to the kernel, which is added
 372  * to the compile time generated pmds. This results in invalid pmds up
 373  * to the point where we hit the physaddr 0 mapping.
 374  *
 375  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 376  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 377  * well, as they are located before _text:
 378  */
 379 void __init cleanup_highmap(void)
 380 {
 381         unsigned long vaddr = __START_KERNEL_map;
 382         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 383         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 384         pmd_t *pmd = level2_kernel_pgt;
 385
 386         /*
 387          * Native path, max_pfn_mapped is not set yet.
 388          * Xen has valid max_pfn_mapped set in
 389          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 390          */
 391         if (max_pfn_mapped)
 392                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 393
 394         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 395                 if (pmd_none(*pmd))
 396                         continue;
 397                 if (vaddr < (unsigned long) _text || vaddr > end)
 398                         set_pmd(pmd, __pmd(0));
 399         }
 400 }
 401
 402 static unsigned long __meminit
 403 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 404               pgprot_t prot)
 405 {
 406         unsigned long pages = 0, next;
 407         unsigned long last_map_addr = end;
 408         int i;
 409
 410         pte_t *pte = pte_page + pte_index(addr);
 411
 412         for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) {
 413                 next = (addr & PAGE_MASK) + PAGE_SIZE;
 414                 if (addr >= end) {
 415                         if (!after_bootmem &&
 416                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) &&
 417                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN))
 418                                 set_pte(pte, __pte(0));
 419                         continue;
 420                 }
 421
 422                 /*
 423                  * We will re-use the existing mapping.
 424                  * Xen for example has some special requirements, like mapping
 425                  * pagetable pages as RO. So assume someone who pre-setup
 426                  * these mappings are more intelligent.
 427                  */
 428                 if (pte_val(*pte)) {
 429                         if (!after_bootmem)
 430                                 pages++;
 431                         continue;
 432                 }
 433
 434                 if (0)
 435                         printk("   pte=%p addr=%lx pte=%016lx\n",
 436                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 437                 pages++;
 438                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 439                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 440         }
 441
 442         update_page_count(PG_LEVEL_4K, pages);
 443
 444         return last_map_addr;
 445 }
 446
 447 static unsigned long __meminit
 448 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 449               unsigned long page_size_mask, pgprot_t prot)
 450 {
 451         unsigned long pages = 0, next;
 452         unsigned long last_map_addr = end;
 453
 454         int i = pmd_index(address);
 455
 456         for (; i < PTRS_PER_PMD; i++, address = next) {
 457                 pmd_t *pmd = pmd_page + pmd_index(address);
 458                 pte_t *pte;
 459                 pgprot_t new_prot = prot;
 460
 461                 next = (address & PMD_MASK) + PMD_SIZE;
 462                 if (address >= end) {
 463                         if (!after_bootmem &&
 464                             !e820_any_mapped(address & PMD_MASK, next, E820_RAM) &&
 465                             !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN))
 466                                 set_pmd(pmd, __pmd(0));
 467                         continue;
 468                 }
 469
 470                 if (pmd_val(*pmd)) {
 471                         if (!pmd_large(*pmd)) {
 472                                 spin_lock(&init_mm.page_table_lock);
 473                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 474                                 last_map_addr = phys_pte_init(pte, address,
 475                                                                 end, prot);
 476                                 spin_unlock(&init_mm.page_table_lock);
 477                                 continue;
 478                         }
 479                         /*
 480                          * If we are ok with PG_LEVEL_2M mapping, then we will
 481                          * use the existing mapping,
 482                          *
 483                          * Otherwise, we will split the large page mapping but
 484                          * use the same existing protection bits except for
 485                          * large page, so that we don't violate Intel's TLB
 486                          * Application note (317080) which says, while changing
 487                          * the page sizes, new and old translations should
 488                          * not differ with respect to page frame and
 489                          * attributes.
 490                          */
 491                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 492                                 if (!after_bootmem)
 493                                         pages++;
 494                                 last_map_addr = next;
 495                                 continue;
 496                         }
 497                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 498                 }
 499
 500                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 501                         pages++;
 502                         spin_lock(&init_mm.page_table_lock);
 503                         set_pte((pte_t *)pmd,
 504                                 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT,
 505                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 506                         spin_unlock(&init_mm.page_table_lock);
 507                         last_map_addr = next;
 508                         continue;
 509                 }
 510
 511                 pte = alloc_low_page();
 512                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 513
 514                 spin_lock(&init_mm.page_table_lock);
 515                 pmd_populate_kernel(&init_mm, pmd, pte);
 516                 spin_unlock(&init_mm.page_table_lock);
 517         }
 518         update_page_count(PG_LEVEL_2M, pages);
 519         return last_map_addr;
 520 }
 521
 522 static unsigned long __meminit
 523 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 524                          unsigned long page_size_mask)
 525 {
 526         unsigned long pages = 0, next;
 527         unsigned long last_map_addr = end;
 528         int i = pud_index(addr);
 529
 530         for (; i < PTRS_PER_PUD; i++, addr = next) {
 531                 pud_t *pud = pud_page + pud_index(addr);
 532                 pmd_t *pmd;
 533                 pgprot_t prot = PAGE_KERNEL;
 534
 535                 next = (addr & PUD_MASK) + PUD_SIZE;
 536                 if (addr >= end) {
 537                         if (!after_bootmem &&
 538                             !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) &&
 539                             !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN))
 540                                 set_pud(pud, __pud(0));
 541                         continue;
 542                 }
 543
 544                 if (pud_val(*pud)) {
 545                         if (!pud_large(*pud)) {
 546                                 pmd = pmd_offset(pud, 0);
 547                                 last_map_addr = phys_pmd_init(pmd, addr, end,
 548                                                          page_size_mask, prot);
 549                                 __flush_tlb_all();
 550                                 continue;
 551                         }
 552                         /*
 553                          * If we are ok with PG_LEVEL_1G mapping, then we will
 554                          * use the existing mapping.
 555                          *
 556                          * Otherwise, we will split the gbpage mapping but use
 557                          * the same existing protection  bits except for large
 558                          * page, so that we don't violate Intel's TLB
 559                          * Application note (317080) which says, while changing
 560                          * the page sizes, new and old translations should
 561                          * not differ with respect to page frame and
 562                          * attributes.
 563                          */
 564                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 565                                 if (!after_bootmem)
 566                                         pages++;
 567                                 last_map_addr = next;
 568                                 continue;
 569                         }
 570                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 571                 }
 572
 573                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 574                         pages++;
 575                         spin_lock(&init_mm.page_table_lock);
 576                         set_pte((pte_t *)pud,
 577                                 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT,
 578                                         PAGE_KERNEL_LARGE));
 579                         spin_unlock(&init_mm.page_table_lock);
 580                         last_map_addr = next;
 581                         continue;
 582                 }
 583
 584                 pmd = alloc_low_page();
 585                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 586                                               prot);
 587
 588                 spin_lock(&init_mm.page_table_lock);
 589                 pud_populate(&init_mm, pud, pmd);
 590                 spin_unlock(&init_mm.page_table_lock);
 591         }
 592         __flush_tlb_all();
 593
 594         update_page_count(PG_LEVEL_1G, pages);
 595
 596         return last_map_addr;
 597 }
 598
 599 unsigned long __meminit
 600 kernel_physical_mapping_init(unsigned long start,
 601                              unsigned long end,
 602                              unsigned long page_size_mask)
 603 {
 604         bool pgd_changed = false;
 605         unsigned long next, last_map_addr = end;
 606         unsigned long addr;
 607
 608         start = (unsigned long)__va(start);
 609         end = (unsigned long)__va(end);
 610         addr = start;
 611
 612         for (; start < end; start = next) {
 613                 pgd_t *pgd = pgd_offset_k(start);
 614                 pud_t *pud;
 615
 616                 next = (start & PGDIR_MASK) + PGDIR_SIZE;
 617
 618                 if (pgd_val(*pgd)) {
 619                         pud = (pud_t *)pgd_page_vaddr(*pgd);
 620                         last_map_addr = phys_pud_init(pud, __pa(start),
 621                                                  __pa(end), page_size_mask);
 622                         continue;
 623                 }
 624
 625                 pud = alloc_low_page();
 626                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(end),
 627                                                  page_size_mask);
 628
 629                 spin_lock(&init_mm.page_table_lock);
 630                 pgd_populate(&init_mm, pgd, pud);
 631                 spin_unlock(&init_mm.page_table_lock);
 632                 pgd_changed = true;
 633         }
 634
 635         if (pgd_changed)
 636                 sync_global_pgds(addr, end - 1);
 637
 638         __flush_tlb_all();
 639
 640         return last_map_addr;
 641 }
 642
 643 #ifndef CONFIG_NUMA
 644 void __init initmem_init(void)
 645 {
 646         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
 647 }
 648 #endif
 649
 650 void __init paging_init(void)
 651 {
 652         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 653         sparse_init();
 654
 655         /*
 656          * clear the default setting with node 0
 657          * note: don't use nodes_clear here, that is really clearing when
 658          *       numa support is not compiled in, and later node_set_state
 659          *       will not set it back.
 660          */
 661         node_clear_state(0, N_MEMORY);
 662         if (N_MEMORY != N_NORMAL_MEMORY)
 663                 node_clear_state(0, N_NORMAL_MEMORY);
 664
 665         zone_sizes_init();
 666 }
 667
 668 /*
 669  * Memory hotplug specific functions
 670  */
 671 #ifdef CONFIG_MEMORY_HOTPLUG
 672 /*
 673  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 674  * updating.
 675  */
 676 static void  update_end_of_memory_vars(u64 start, u64 size)
 677 {
 678         unsigned long end_pfn = PFN_UP(start + size);
 679
 680         if (end_pfn > max_pfn) {
 681                 max_pfn = end_pfn;
 682                 max_low_pfn = end_pfn;
 683                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 684         }
 685 }
 686
 687 /*
 688  * Memory is added always to NORMAL zone. This means you will never get
 689  * additional DMA/DMA32 memory.
 690  */
 691 int arch_add_memory(int nid, u64 start, u64 size)
 692 {
 693         struct pglist_data *pgdat = NODE_DATA(nid);
 694         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 695         unsigned long start_pfn = start >> PAGE_SHIFT;
 696         unsigned long nr_pages = size >> PAGE_SHIFT;
 697         int ret;
 698
 699         init_memory_mapping(start, start + size);
 700
 701         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 702         WARN_ON_ONCE(ret);
 703
 704         /* update max_pfn, max_low_pfn and high_memory */
 705         update_end_of_memory_vars(start, size);
 706
 707         return ret;
 708 }
 709 EXPORT_SYMBOL_GPL(arch_add_memory);
 710
 711 #define PAGE_INUSE 0xFD
 712
 713 static void __meminit free_pagetable(struct page *page, int order)
 714 {
 715         struct zone *zone;
 716         bool bootmem = false;
 717         unsigned long magic;
 718         unsigned int nr_pages = 1 << order;
 719
 720         /* bootmem page has reserved flag */
 721         if (PageReserved(page)) {
 722                 __ClearPageReserved(page);
 723                 bootmem = true;
 724
 725                 magic = (unsigned long)page->lru.next;
 726                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 727                         while (nr_pages--)
 728                                 put_page_bootmem(page++);
 729                 } else
 730                         __free_pages_bootmem(page, order);
 731         } else
 732                 free_pages((unsigned long)page_address(page), order);
 733
 734         /*
 735          * SECTION_INFO pages and MIX_SECTION_INFO pages
 736          * are all allocated by bootmem.
 737          */
 738         if (bootmem) {
 739                 zone = page_zone(page);
 740                 zone_span_writelock(zone);
 741                 zone->present_pages += nr_pages;
 742                 zone_span_writeunlock(zone);
 743                 totalram_pages += nr_pages;
 744         }
 745 }
 746
 747 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 748 {
 749         pte_t *pte;
 750         int i;
 751
 752         for (i = 0; i < PTRS_PER_PTE; i++) {
 753                 pte = pte_start + i;
 754                 if (pte_val(*pte))
 755                         return;
 756         }
 757
 758         /* free a pte talbe */
 759         free_pagetable(pmd_page(*pmd), 0);
 760         spin_lock(&init_mm.page_table_lock);
 761         pmd_clear(pmd);
 762         spin_unlock(&init_mm.page_table_lock);
 763 }
 764
 765 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 766 {
 767         pmd_t *pmd;
 768         int i;
 769
 770         for (i = 0; i < PTRS_PER_PMD; i++) {
 771                 pmd = pmd_start + i;
 772                 if (pmd_val(*pmd))
 773                         return;
 774         }
 775
 776         /* free a pmd talbe */
 777         free_pagetable(pud_page(*pud), 0);
 778         spin_lock(&init_mm.page_table_lock);
 779         pud_clear(pud);
 780         spin_unlock(&init_mm.page_table_lock);
 781 }
 782
 783 /* Return true if pgd is changed, otherwise return false. */
 784 static bool __meminit free_pud_table(pud_t *pud_start, pgd_t *pgd)
 785 {
 786         pud_t *pud;
 787         int i;
 788
 789         for (i = 0; i < PTRS_PER_PUD; i++) {
 790                 pud = pud_start + i;
 791                 if (pud_val(*pud))
 792                         return false;
 793         }
 794
 795         /* free a pud table */
 796         free_pagetable(pgd_page(*pgd), 0);
 797         spin_lock(&init_mm.page_table_lock);
 798         pgd_clear(pgd);
 799         spin_unlock(&init_mm.page_table_lock);
 800
 801         return true;
 802 }
 803
 804 static void __meminit
 805 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 806                  bool direct)
 807 {
 808         unsigned long next, pages = 0;
 809         pte_t *pte;
 810         void *page_addr;
 811         phys_addr_t phys_addr;
 812
 813         pte = pte_start + pte_index(addr);
 814         for (; addr < end; addr = next, pte++) {
 815                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 816                 if (next > end)
 817                         next = end;
 818
 819                 if (!pte_present(*pte))
 820                         continue;
 821
 822                 /*
 823                  * We mapped [0,1G) memory as identity mapping when
 824                  * initializing, in arch/x86/kernel/head_64.S. These
 825                  * pagetables cannot be removed.
 826                  */
 827                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 828                 if (phys_addr < (phys_addr_t)0x40000000)
 829                         return;
 830
 831                 if (IS_ALIGNED(addr, PAGE_SIZE) &&
 832                     IS_ALIGNED(next, PAGE_SIZE)) {
 833                         /*
 834                          * Do not free direct mapping pages since they were
 835                          * freed when offlining, or simplely not in use.
 836                          */
 837                         if (!direct)
 838                                 free_pagetable(pte_page(*pte), 0);
 839
 840                         spin_lock(&init_mm.page_table_lock);
 841                         pte_clear(&init_mm, addr, pte);
 842                         spin_unlock(&init_mm.page_table_lock);
 843
 844                         /* For non-direct mapping, pages means nothing. */
 845                         pages++;
 846                 } else {
 847                         /*
 848                          * If we are here, we are freeing vmemmap pages since
 849                          * direct mapped memory ranges to be freed are aligned.
 850                          *
 851                          * If we are not removing the whole page, it means
 852                          * other page structs in this page are being used and
 853                          * we canot remove them. So fill the unused page_structs
 854                          * with 0xFD, and remove the page when it is wholly
 855                          * filled with 0xFD.
 856                          */
 857                         memset((void *)addr, PAGE_INUSE, next - addr);
 858
 859                         page_addr = page_address(pte_page(*pte));
 860                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
 861                                 free_pagetable(pte_page(*pte), 0);
 862
 863                                 spin_lock(&init_mm.page_table_lock);
 864                                 pte_clear(&init_mm, addr, pte);
 865                                 spin_unlock(&init_mm.page_table_lock);
 866                         }
 867                 }
 868         }
 869
 870         /* Call free_pte_table() in remove_pmd_table(). */
 871         flush_tlb_all();
 872         if (direct)
 873                 update_page_count(PG_LEVEL_4K, -pages);
 874 }
 875
 876 static void __meminit
 877 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
 878                  bool direct)
 879 {
 880         unsigned long next, pages = 0;
 881         pte_t *pte_base;
 882         pmd_t *pmd;
 883         void *page_addr;
 884
 885         pmd = pmd_start + pmd_index(addr);
 886         for (; addr < end; addr = next, pmd++) {
 887                 next = pmd_addr_end(addr, end);
 888
 889                 if (!pmd_present(*pmd))
 890                         continue;
 891
 892                 if (pmd_large(*pmd)) {
 893                         if (IS_ALIGNED(addr, PMD_SIZE) &&
 894                             IS_ALIGNED(next, PMD_SIZE)) {
 895                                 if (!direct)
 896                                         free_pagetable(pmd_page(*pmd),
 897                                                        get_order(PMD_SIZE));
 898
 899                                 spin_lock(&init_mm.page_table_lock);
 900                                 pmd_clear(pmd);
 901                                 spin_unlock(&init_mm.page_table_lock);
 902                                 pages++;
 903                         } else {
 904                                 /* If here, we are freeing vmemmap pages. */
 905                                 memset((void *)addr, PAGE_INUSE, next - addr);
 906
 907                                 page_addr = page_address(pmd_page(*pmd));
 908                                 if (!memchr_inv(page_addr, PAGE_INUSE,
 909                                                 PMD_SIZE)) {
 910                                         free_pagetable(pmd_page(*pmd),
 911                                                        get_order(PMD_SIZE));
 912
 913                                         spin_lock(&init_mm.page_table_lock);
 914                                         pmd_clear(pmd);
 915                                         spin_unlock(&init_mm.page_table_lock);
 916                                 }
 917                         }
 918
 919                         continue;
 920                 }
 921
 922                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
 923                 remove_pte_table(pte_base, addr, next, direct);
 924                 free_pte_table(pte_base, pmd);
 925         }
 926
 927         /* Call free_pmd_table() in remove_pud_table(). */
 928         if (direct)
 929                 update_page_count(PG_LEVEL_2M, -pages);
 930 }
 931
 932 static void __meminit
 933 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
 934                  bool direct)
 935 {
 936         unsigned long next, pages = 0;
 937         pmd_t *pmd_base;
 938         pud_t *pud;
 939         void *page_addr;
 940
 941         pud = pud_start + pud_index(addr);
 942         for (; addr < end; addr = next, pud++) {
 943                 next = pud_addr_end(addr, end);
 944
 945                 if (!pud_present(*pud))
 946                         continue;
 947
 948                 if (pud_large(*pud)) {
 949                         if (IS_ALIGNED(addr, PUD_SIZE) &&
 950                             IS_ALIGNED(next, PUD_SIZE)) {
 951                                 if (!direct)
 952                                         free_pagetable(pud_page(*pud),
 953                                                        get_order(PUD_SIZE));
 954
 955                                 spin_lock(&init_mm.page_table_lock);
 956                                 pud_clear(pud);
 957                                 spin_unlock(&init_mm.page_table_lock);
 958                                 pages++;
 959                         } else {
 960                                 /* If here, we are freeing vmemmap pages. */
 961                                 memset((void *)addr, PAGE_INUSE, next - addr);
 962
 963                                 page_addr = page_address(pud_page(*pud));
 964                                 if (!memchr_inv(page_addr, PAGE_INUSE,
 965                                                 PUD_SIZE)) {
 966                                         free_pagetable(pud_page(*pud),
 967                                                        get_order(PUD_SIZE));
 968
 969                                         spin_lock(&init_mm.page_table_lock);
 970                                         pud_clear(pud);
 971                                         spin_unlock(&init_mm.page_table_lock);
 972                                 }
 973                         }
 974
 975                         continue;
 976                 }
 977
 978                 pmd_base = (pmd_t *)pud_page_vaddr(*pud);
 979                 remove_pmd_table(pmd_base, addr, next, direct);
 980                 free_pmd_table(pmd_base, pud);
 981         }
 982
 983         if (direct)
 984                 update_page_count(PG_LEVEL_1G, -pages);
 985 }
 986
 987 /* start and end are both virtual address. */
 988 static void __meminit
 989 remove_pagetable(unsigned long start, unsigned long end, bool direct)
 990 {
 991         unsigned long next;
 992         pgd_t *pgd;
 993         pud_t *pud;
 994         bool pgd_changed = false;
 995
 996         for (; start < end; start = next) {
 997                 next = pgd_addr_end(start, end);
 998
 999                 pgd = pgd_offset_k(start);
1000                 if (!pgd_present(*pgd))
1001                         continue;
1002
1003                 pud = (pud_t *)pgd_page_vaddr(*pgd);
1004                 remove_pud_table(pud, start, next, direct);
1005                 if (free_pud_table(pud, pgd))
1006                         pgd_changed = true;
1007         }
1008
1009         if (pgd_changed)
1010                 sync_global_pgds(start, end - 1);
1011
1012         flush_tlb_all();
1013 }
1014
1015 void __ref vmemmap_free(struct page *memmap, unsigned long nr_pages)
1016 {
1017         unsigned long start = (unsigned long)memmap;
1018         unsigned long end = (unsigned long)(memmap + nr_pages);
1019
1020         remove_pagetable(start, end, false);
1021 }
1022
1023 static void __meminit
1024 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1025 {
1026         start = (unsigned long)__va(start);
1027         end = (unsigned long)__va(end);
1028
1029         remove_pagetable(start, end, true);
1030 }
1031
1032 #ifdef CONFIG_MEMORY_HOTREMOVE
1033 int __ref arch_remove_memory(u64 start, u64 size)
1034 {
1035         unsigned long start_pfn = start >> PAGE_SHIFT;
1036         unsigned long nr_pages = size >> PAGE_SHIFT;
1037         struct zone *zone;
1038         int ret;
1039
1040         zone = page_zone(pfn_to_page(start_pfn));
1041         kernel_physical_mapping_remove(start, start + size);
1042         ret = __remove_pages(zone, start_pfn, nr_pages);
1043         WARN_ON_ONCE(ret);
1044
1045         return ret;
1046 }
1047 #endif
1048 #endif /* CONFIG_MEMORY_HOTPLUG */
1049
1050 static struct kcore_list kcore_vsyscall;
1051
1052 static void __init register_page_bootmem_info(void)
1053 {
1054 #ifdef CONFIG_NUMA
1055         int i;
1056
1057         for_each_online_node(i)
1058                 register_page_bootmem_info_node(NODE_DATA(i));
1059 #endif
1060 }
1061
1062 void __init mem_init(void)
1063 {
1064         long codesize, reservedpages, datasize, initsize;
1065         unsigned long absent_pages;
1066
1067         pci_iommu_alloc();
1068
1069         /* clear_bss() already clear the empty_zero_page */
1070
1071         reservedpages = 0;
1072
1073         /* this will put all low memory onto the freelists */
1074         register_page_bootmem_info();
1075         totalram_pages = free_all_bootmem();
1076
1077         absent_pages = absent_pages_in_range(0, max_pfn);
1078         reservedpages = max_pfn - totalram_pages - absent_pages;
1079         after_bootmem = 1;
1080
1081         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
1082         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
1083         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
1084
1085         /* Register memory areas for /proc/kcore */
1086         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
1087                          VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
1088
1089         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
1090                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
1091                 nr_free_pages() << (PAGE_SHIFT-10),
1092                 max_pfn << (PAGE_SHIFT-10),
1093                 codesize >> 10,
1094                 absent_pages << (PAGE_SHIFT-10),
1095                 reservedpages << (PAGE_SHIFT-10),
1096                 datasize >> 10,
1097                 initsize >> 10);
1098 }
1099
1100 #ifdef CONFIG_DEBUG_RODATA
1101 const int rodata_test_data = 0xC3;
1102 EXPORT_SYMBOL_GPL(rodata_test_data);
1103
1104 int kernel_set_to_readonly;
1105
1106 void set_kernel_text_rw(void)
1107 {
1108         unsigned long start = PFN_ALIGN(_text);
1109         unsigned long end = PFN_ALIGN(__stop___ex_table);
1110
1111         if (!kernel_set_to_readonly)
1112                 return;
1113
1114         pr_debug("Set kernel text: %lx - %lx for read write\n",
1115                  start, end);
1116
1117         /*
1118          * Make the kernel identity mapping for text RW. Kernel text
1119          * mapping will always be RO. Refer to the comment in
1120          * static_protections() in pageattr.c
1121          */
1122         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1123 }
1124
1125 void set_kernel_text_ro(void)
1126 {
1127         unsigned long start = PFN_ALIGN(_text);
1128         unsigned long end = PFN_ALIGN(__stop___ex_table);
1129
1130         if (!kernel_set_to_readonly)
1131                 return;
1132
1133         pr_debug("Set kernel text: %lx - %lx for read only\n",
1134                  start, end);
1135
1136         /*
1137          * Set the kernel identity mapping for text RO.
1138          */
1139         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1140 }
1141
1142 void mark_rodata_ro(void)
1143 {
1144         unsigned long start = PFN_ALIGN(_text);
1145         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1146         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1147         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1148         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1149         unsigned long all_end = PFN_ALIGN(&_end);
1150
1151         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1152                (end - start) >> 10);
1153         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1154
1155         kernel_set_to_readonly = 1;
1156
1157         /*
1158          * The rodata/data/bss/brk section (but not the kernel text!)
1159          * should also be not-executable.
1160          */
1161         set_memory_nx(rodata_start, (all_end - rodata_start) >> PAGE_SHIFT);
1162
1163         rodata_test();
1164
1165 #ifdef CONFIG_CPA_DEBUG
1166         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1167         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1168
1169         printk(KERN_INFO "Testing CPA: again\n");
1170         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1171 #endif
1172
1173         free_init_pages("unused kernel memory",
1174                         (unsigned long) __va(__pa_symbol(text_end)),
1175                         (unsigned long) __va(__pa_symbol(rodata_start)));
1176
1177         free_init_pages("unused kernel memory",
1178                         (unsigned long) __va(__pa_symbol(rodata_end)),
1179                         (unsigned long) __va(__pa_symbol(_sdata)));
1180 }
1181
1182 #endif
1183
1184 int kern_addr_valid(unsigned long addr)
1185 {
1186         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1187         pgd_t *pgd;
1188         pud_t *pud;
1189         pmd_t *pmd;
1190         pte_t *pte;
1191
1192         if (above != 0 && above != -1UL)
1193                 return 0;
1194
1195         pgd = pgd_offset_k(addr);
1196         if (pgd_none(*pgd))
1197                 return 0;
1198
1199         pud = pud_offset(pgd, addr);
1200         if (pud_none(*pud))
1201                 return 0;
1202
1203         if (pud_large(*pud))
1204                 return pfn_valid(pud_pfn(*pud));
1205
1206         pmd = pmd_offset(pud, addr);
1207         if (pmd_none(*pmd))
1208                 return 0;
1209
1210         if (pmd_large(*pmd))
1211                 return pfn_valid(pmd_pfn(*pmd));
1212
1213         pte = pte_offset_kernel(pmd, addr);
1214         if (pte_none(*pte))
1215                 return 0;
1216
1217         return pfn_valid(pte_pfn(*pte));
1218 }
1219
1220 /*
1221  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1222  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1223  * not need special handling anymore:
1224  */
1225 static struct vm_area_struct gate_vma = {
1226         .vm_start       = VSYSCALL_START,
1227         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1228         .vm_page_prot   = PAGE_READONLY_EXEC,
1229         .vm_flags       = VM_READ | VM_EXEC
1230 };
1231
1232 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
1233 {
1234 #ifdef CONFIG_IA32_EMULATION
1235         if (!mm || mm->context.ia32_compat)
1236                 return NULL;
1237 #endif
1238         return &gate_vma;
1239 }
1240
1241 int in_gate_area(struct mm_struct *mm, unsigned long addr)
1242 {
1243         struct vm_area_struct *vma = get_gate_vma(mm);
1244
1245         if (!vma)
1246                 return 0;
1247
1248         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1249 }
1250
1251 /*
1252  * Use this when you have no reliable mm, typically from interrupt
1253  * context. It is less reliable than using a task's mm and may give
1254  * false positives.
1255  */
1256 int in_gate_area_no_mm(unsigned long addr)
1257 {
1258         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1259 }
1260
1261 const char *arch_vma_name(struct vm_area_struct *vma)
1262 {
1263         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1264                 return "[vdso]";
1265         if (vma == &gate_vma)
1266                 return "[vsyscall]";
1267         return NULL;
1268 }
1269
1270 #ifdef CONFIG_X86_UV
1271 unsigned long memory_block_size_bytes(void)
1272 {
1273         if (is_uv_system()) {
1274                 printk(KERN_INFO "UV: memory block size 2GB\n");
1275                 return 2UL * 1024 * 1024 * 1024;
1276         }
1277         return MIN_MEMORY_BLOCK_SIZE;
1278 }
1279 #endif
1280
1281 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1282 /*
1283  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1284  */
1285 static long __meminitdata addr_start, addr_end;
1286 static void __meminitdata *p_start, *p_end;
1287 static int __meminitdata node_start;
1288
1289 int __meminit
1290 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1291 {
1292         unsigned long addr = (unsigned long)start_page;
1293         unsigned long end = (unsigned long)(start_page + size);
1294         unsigned long next;
1295         pgd_t *pgd;
1296         pud_t *pud;
1297         pmd_t *pmd;
1298
1299         for (; addr < end; addr = next) {
1300                 void *p = NULL;
1301
1302                 pgd = vmemmap_pgd_populate(addr, node);
1303                 if (!pgd)
1304                         return -ENOMEM;
1305
1306                 pud = vmemmap_pud_populate(pgd, addr, node);
1307                 if (!pud)
1308                         return -ENOMEM;
1309
1310                 if (!cpu_has_pse) {
1311                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1312                         pmd = vmemmap_pmd_populate(pud, addr, node);
1313
1314                         if (!pmd)
1315                                 return -ENOMEM;
1316
1317                         p = vmemmap_pte_populate(pmd, addr, node);
1318
1319                         if (!p)
1320                                 return -ENOMEM;
1321
1322                         addr_end = addr + PAGE_SIZE;
1323                         p_end = p + PAGE_SIZE;
1324                 } else {
1325                         next = pmd_addr_end(addr, end);
1326
1327                         pmd = pmd_offset(pud, addr);
1328                         if (pmd_none(*pmd)) {
1329                                 pte_t entry;
1330
1331                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1332                                 if (!p)
1333                                         return -ENOMEM;
1334
1335                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1336                                                 PAGE_KERNEL_LARGE);
1337                                 set_pmd(pmd, __pmd(pte_val(entry)));
1338
1339                                 /* check to see if we have contiguous blocks */
1340                                 if (p_end != p || node_start != node) {
1341                                         if (p_start)
1342                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1343                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1344                                         addr_start = addr;
1345                                         node_start = node;
1346                                         p_start = p;
1347                                 }
1348
1349                                 addr_end = addr + PMD_SIZE;
1350                                 p_end = p + PMD_SIZE;
1351                         } else
1352                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1353                 }
1354
1355         }
1356         sync_global_pgds((unsigned long)start_page, end - 1);
1357         return 0;
1358 }
1359
1360 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1361 void register_page_bootmem_memmap(unsigned long section_nr,
1362                                   struct page *start_page, unsigned long size)
1363 {
1364         unsigned long addr = (unsigned long)start_page;
1365         unsigned long end = (unsigned long)(start_page + size);
1366         unsigned long next;
1367         pgd_t *pgd;
1368         pud_t *pud;
1369         pmd_t *pmd;
1370         unsigned int nr_pages;
1371         struct page *page;
1372
1373         for (; addr < end; addr = next) {
1374                 pte_t *pte = NULL;
1375
1376                 pgd = pgd_offset_k(addr);
1377                 if (pgd_none(*pgd)) {
1378                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1379                         continue;
1380                 }
1381                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1382
1383                 pud = pud_offset(pgd, addr);
1384                 if (pud_none(*pud)) {
1385                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1386                         continue;
1387                 }
1388                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1389
1390                 if (!cpu_has_pse) {
1391                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1392                         pmd = pmd_offset(pud, addr);
1393                         if (pmd_none(*pmd))
1394                                 continue;
1395                         get_page_bootmem(section_nr, pmd_page(*pmd),
1396                                          MIX_SECTION_INFO);
1397
1398                         pte = pte_offset_kernel(pmd, addr);
1399                         if (pte_none(*pte))
1400                                 continue;
1401                         get_page_bootmem(section_nr, pte_page(*pte),
1402                                          SECTION_INFO);
1403                 } else {
1404                         next = pmd_addr_end(addr, end);
1405
1406                         pmd = pmd_offset(pud, addr);
1407                         if (pmd_none(*pmd))
1408                                 continue;
1409
1410                         nr_pages = 1 << (get_order(PMD_SIZE));
1411                         page = pmd_page(*pmd);
1412                         while (nr_pages--)
1413                                 get_page_bootmem(section_nr, page++,
1414                                                  SECTION_INFO);
1415                 }
1416         }
1417 }
1418 #endif
1419
1420 void __meminit vmemmap_populate_print_last(void)
1421 {
1422         if (p_start) {
1423                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1424                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1425                 p_start = NULL;
1426                 p_end = NULL;
1427                 node_start = 0;
1428         }
1429 }
1430 #endif