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 = ~0;
 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, int removed)
 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                 /*
 190                  * When it is called after memory hot remove, pgd_none()
 191                  * returns true. In this case (removed == 1), we must clear
 192                  * the PGD entries in the local PGD level page.
 193                  */
 194                 if (pgd_none(*pgd_ref) && !removed)
 195                         continue;
 196
 197                 spin_lock(&pgd_lock);
 198                 list_for_each_entry(page, &pgd_list, lru) {
 199                         pgd_t *pgd;
 200                         spinlock_t *pgt_lock;
 201
 202                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 203                         /* the pgt_lock only for Xen */
 204                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 205                         spin_lock(pgt_lock);
 206
 207                         if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
 208                                 BUG_ON(pgd_page_vaddr(*pgd)
 209                                        != pgd_page_vaddr(*pgd_ref));
 210
 211                         if (removed) {
 212                                 if (pgd_none(*pgd_ref) && !pgd_none(*pgd))
 213                                         pgd_clear(pgd);
 214                         } else {
 215                                 if (pgd_none(*pgd))
 216                                         set_pgd(pgd, *pgd_ref);
 217                         }
 218
 219                         spin_unlock(pgt_lock);
 220                 }
 221                 spin_unlock(&pgd_lock);
 222         }
 223 }
 224
 225 /*
 226  * NOTE: This function is marked __ref because it calls __init function
 227  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 228  */
 229 static __ref void *spp_getpage(void)
 230 {
 231         void *ptr;
 232
 233         if (after_bootmem)
 234                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 235         else
 236                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 237
 238         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 239                 panic("set_pte_phys: cannot allocate page data %s\n",
 240                         after_bootmem ? "after bootmem" : "");
 241         }
 242
 243         pr_debug("spp_getpage %p\n", ptr);
 244
 245         return ptr;
 246 }
 247
 248 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 249 {
 250         if (pgd_none(*pgd)) {
 251                 pud_t *pud = (pud_t *)spp_getpage();
 252                 pgd_populate(&init_mm, pgd, pud);
 253                 if (pud != pud_offset(pgd, 0))
 254                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 255                                pud, pud_offset(pgd, 0));
 256         }
 257         return pud_offset(pgd, vaddr);
 258 }
 259
 260 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 261 {
 262         if (pud_none(*pud)) {
 263                 pmd_t *pmd = (pmd_t *) spp_getpage();
 264                 pud_populate(&init_mm, pud, pmd);
 265                 if (pmd != pmd_offset(pud, 0))
 266                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 267                                pmd, pmd_offset(pud, 0));
 268         }
 269         return pmd_offset(pud, vaddr);
 270 }
 271
 272 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 273 {
 274         if (pmd_none(*pmd)) {
 275                 pte_t *pte = (pte_t *) spp_getpage();
 276                 pmd_populate_kernel(&init_mm, pmd, pte);
 277                 if (pte != pte_offset_kernel(pmd, 0))
 278                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 279         }
 280         return pte_offset_kernel(pmd, vaddr);
 281 }
 282
 283 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 284 {
 285         pud_t *pud;
 286         pmd_t *pmd;
 287         pte_t *pte;
 288
 289         pud = pud_page + pud_index(vaddr);
 290         pmd = fill_pmd(pud, vaddr);
 291         pte = fill_pte(pmd, vaddr);
 292
 293         set_pte(pte, new_pte);
 294
 295         /*
 296          * It's enough to flush this one mapping.
 297          * (PGE mappings get flushed as well)
 298          */
 299         __flush_tlb_one(vaddr);
 300 }
 301
 302 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 303 {
 304         pgd_t *pgd;
 305         pud_t *pud_page;
 306
 307         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 308
 309         pgd = pgd_offset_k(vaddr);
 310         if (pgd_none(*pgd)) {
 311                 printk(KERN_ERR
 312                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 313                 return;
 314         }
 315         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 316         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 317 }
 318
 319 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 320 {
 321         pgd_t *pgd;
 322         pud_t *pud;
 323
 324         pgd = pgd_offset_k(vaddr);
 325         pud = fill_pud(pgd, vaddr);
 326         return fill_pmd(pud, vaddr);
 327 }
 328
 329 pte_t * __init populate_extra_pte(unsigned long vaddr)
 330 {
 331         pmd_t *pmd;
 332
 333         pmd = populate_extra_pmd(vaddr);
 334         return fill_pte(pmd, vaddr);
 335 }
 336
 337 /*
 338  * Create large page table mappings for a range of physical addresses.
 339  */
 340 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 341                                                 pgprot_t prot)
 342 {
 343         pgd_t *pgd;
 344         pud_t *pud;
 345         pmd_t *pmd;
 346
 347         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 348         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 349                 pgd = pgd_offset_k((unsigned long)__va(phys));
 350                 if (pgd_none(*pgd)) {
 351                         pud = (pud_t *) spp_getpage();
 352                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 353                                                 _PAGE_USER));
 354                 }
 355                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 356                 if (pud_none(*pud)) {
 357                         pmd = (pmd_t *) spp_getpage();
 358                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 359                                                 _PAGE_USER));
 360                 }
 361                 pmd = pmd_offset(pud, phys);
 362                 BUG_ON(!pmd_none(*pmd));
 363                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 364         }
 365 }
 366
 367 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 368 {
 369         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 370 }
 371
 372 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 373 {
 374         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 375 }
 376
 377 /*
 378  * The head.S code sets up the kernel high mapping:
 379  *
 380  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 381  *
 382  * phys_base holds the negative offset to the kernel, which is added
 383  * to the compile time generated pmds. This results in invalid pmds up
 384  * to the point where we hit the physaddr 0 mapping.
 385  *
 386  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 387  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 388  * well, as they are located before _text:
 389  */
 390 void __init cleanup_highmap(void)
 391 {
 392         unsigned long vaddr = __START_KERNEL_map;
 393         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 394         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 395         pmd_t *pmd = level2_kernel_pgt;
 396
 397         /*
 398          * Native path, max_pfn_mapped is not set yet.
 399          * Xen has valid max_pfn_mapped set in
 400          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 401          */
 402         if (max_pfn_mapped)
 403                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 404
 405         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 406                 if (pmd_none(*pmd))
 407                         continue;
 408                 if (vaddr < (unsigned long) _text || vaddr > end)
 409                         set_pmd(pmd, __pmd(0));
 410         }
 411 }
 412
 413 static unsigned long __meminit
 414 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 415               pgprot_t prot)
 416 {
 417         unsigned long pages = 0, next;
 418         unsigned long last_map_addr = end;
 419         int i;
 420
 421         pte_t *pte = pte_page + pte_index(addr);
 422
 423         for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) {
 424                 next = (addr & PAGE_MASK) + PAGE_SIZE;
 425                 if (addr >= end) {
 426                         if (!after_bootmem &&
 427                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) &&
 428                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN))
 429                                 set_pte(pte, __pte(0));
 430                         continue;
 431                 }
 432
 433                 /*
 434                  * We will re-use the existing mapping.
 435                  * Xen for example has some special requirements, like mapping
 436                  * pagetable pages as RO. So assume someone who pre-setup
 437                  * these mappings are more intelligent.
 438                  */
 439                 if (pte_val(*pte)) {
 440                         if (!after_bootmem)
 441                                 pages++;
 442                         continue;
 443                 }
 444
 445                 if (0)
 446                         printk("   pte=%p addr=%lx pte=%016lx\n",
 447                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 448                 pages++;
 449                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 450                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 451         }
 452
 453         update_page_count(PG_LEVEL_4K, pages);
 454
 455         return last_map_addr;
 456 }
 457
 458 static unsigned long __meminit
 459 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 460               unsigned long page_size_mask, pgprot_t prot)
 461 {
 462         unsigned long pages = 0, next;
 463         unsigned long last_map_addr = end;
 464
 465         int i = pmd_index(address);
 466
 467         for (; i < PTRS_PER_PMD; i++, address = next) {
 468                 pmd_t *pmd = pmd_page + pmd_index(address);
 469                 pte_t *pte;
 470                 pgprot_t new_prot = prot;
 471
 472                 next = (address & PMD_MASK) + PMD_SIZE;
 473                 if (address >= end) {
 474                         if (!after_bootmem &&
 475                             !e820_any_mapped(address & PMD_MASK, next, E820_RAM) &&
 476                             !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN))
 477                                 set_pmd(pmd, __pmd(0));
 478                         continue;
 479                 }
 480
 481                 if (pmd_val(*pmd)) {
 482                         if (!pmd_large(*pmd)) {
 483                                 spin_lock(&init_mm.page_table_lock);
 484                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 485                                 last_map_addr = phys_pte_init(pte, address,
 486                                                                 end, prot);
 487                                 spin_unlock(&init_mm.page_table_lock);
 488                                 continue;
 489                         }
 490                         /*
 491                          * If we are ok with PG_LEVEL_2M mapping, then we will
 492                          * use the existing mapping,
 493                          *
 494                          * Otherwise, we will split the large page mapping but
 495                          * use the same existing protection bits except for
 496                          * large page, so that we don't violate Intel's TLB
 497                          * Application note (317080) which says, while changing
 498                          * the page sizes, new and old translations should
 499                          * not differ with respect to page frame and
 500                          * attributes.
 501                          */
 502                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 503                                 if (!after_bootmem)
 504                                         pages++;
 505                                 last_map_addr = next;
 506                                 continue;
 507                         }
 508                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 509                 }
 510
 511                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 512                         pages++;
 513                         spin_lock(&init_mm.page_table_lock);
 514                         set_pte((pte_t *)pmd,
 515                                 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT,
 516                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 517                         spin_unlock(&init_mm.page_table_lock);
 518                         last_map_addr = next;
 519                         continue;
 520                 }
 521
 522                 pte = alloc_low_page();
 523                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 524
 525                 spin_lock(&init_mm.page_table_lock);
 526                 pmd_populate_kernel(&init_mm, pmd, pte);
 527                 spin_unlock(&init_mm.page_table_lock);
 528         }
 529         update_page_count(PG_LEVEL_2M, pages);
 530         return last_map_addr;
 531 }
 532
 533 static unsigned long __meminit
 534 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 535                          unsigned long page_size_mask)
 536 {
 537         unsigned long pages = 0, next;
 538         unsigned long last_map_addr = end;
 539         int i = pud_index(addr);
 540
 541         for (; i < PTRS_PER_PUD; i++, addr = next) {
 542                 pud_t *pud = pud_page + pud_index(addr);
 543                 pmd_t *pmd;
 544                 pgprot_t prot = PAGE_KERNEL;
 545
 546                 next = (addr & PUD_MASK) + PUD_SIZE;
 547                 if (addr >= end) {
 548                         if (!after_bootmem &&
 549                             !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) &&
 550                             !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN))
 551                                 set_pud(pud, __pud(0));
 552                         continue;
 553                 }
 554
 555                 if (pud_val(*pud)) {
 556                         if (!pud_large(*pud)) {
 557                                 pmd = pmd_offset(pud, 0);
 558                                 last_map_addr = phys_pmd_init(pmd, addr, end,
 559                                                          page_size_mask, prot);
 560                                 __flush_tlb_all();
 561                                 continue;
 562                         }
 563                         /*
 564                          * If we are ok with PG_LEVEL_1G mapping, then we will
 565                          * use the existing mapping.
 566                          *
 567                          * Otherwise, we will split the gbpage mapping but use
 568                          * the same existing protection  bits except for large
 569                          * page, so that we don't violate Intel's TLB
 570                          * Application note (317080) which says, while changing
 571                          * the page sizes, new and old translations should
 572                          * not differ with respect to page frame and
 573                          * attributes.
 574                          */
 575                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 576                                 if (!after_bootmem)
 577                                         pages++;
 578                                 last_map_addr = next;
 579                                 continue;
 580                         }
 581                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 582                 }
 583
 584                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 585                         pages++;
 586                         spin_lock(&init_mm.page_table_lock);
 587                         set_pte((pte_t *)pud,
 588                                 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT,
 589                                         PAGE_KERNEL_LARGE));
 590                         spin_unlock(&init_mm.page_table_lock);
 591                         last_map_addr = next;
 592                         continue;
 593                 }
 594
 595                 pmd = alloc_low_page();
 596                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 597                                               prot);
 598
 599                 spin_lock(&init_mm.page_table_lock);
 600                 pud_populate(&init_mm, pud, pmd);
 601                 spin_unlock(&init_mm.page_table_lock);
 602         }
 603         __flush_tlb_all();
 604
 605         update_page_count(PG_LEVEL_1G, pages);
 606
 607         return last_map_addr;
 608 }
 609
 610 unsigned long __meminit
 611 kernel_physical_mapping_init(unsigned long start,
 612                              unsigned long end,
 613                              unsigned long page_size_mask)
 614 {
 615         bool pgd_changed = false;
 616         unsigned long next, last_map_addr = end;
 617         unsigned long addr;
 618
 619         start = (unsigned long)__va(start);
 620         end = (unsigned long)__va(end);
 621         addr = start;
 622
 623         for (; start < end; start = next) {
 624                 pgd_t *pgd = pgd_offset_k(start);
 625                 pud_t *pud;
 626
 627                 next = (start & PGDIR_MASK) + PGDIR_SIZE;
 628
 629                 if (pgd_val(*pgd)) {
 630                         pud = (pud_t *)pgd_page_vaddr(*pgd);
 631                         last_map_addr = phys_pud_init(pud, __pa(start),
 632                                                  __pa(end), page_size_mask);
 633                         continue;
 634                 }
 635
 636                 pud = alloc_low_page();
 637                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(end),
 638                                                  page_size_mask);
 639
 640                 spin_lock(&init_mm.page_table_lock);
 641                 pgd_populate(&init_mm, pgd, pud);
 642                 spin_unlock(&init_mm.page_table_lock);
 643                 pgd_changed = true;
 644         }
 645
 646         if (pgd_changed)
 647                 sync_global_pgds(addr, end - 1, 0);
 648
 649         __flush_tlb_all();
 650
 651         return last_map_addr;
 652 }
 653
 654 #ifndef CONFIG_NUMA
 655 void __init initmem_init(void)
 656 {
 657         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
 658 }
 659 #endif
 660
 661 void __init paging_init(void)
 662 {
 663         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 664         sparse_init();
 665
 666         /*
 667          * clear the default setting with node 0
 668          * note: don't use nodes_clear here, that is really clearing when
 669          *       numa support is not compiled in, and later node_set_state
 670          *       will not set it back.
 671          */
 672         node_clear_state(0, N_MEMORY);
 673         if (N_MEMORY != N_NORMAL_MEMORY)
 674                 node_clear_state(0, N_NORMAL_MEMORY);
 675
 676         zone_sizes_init();
 677 }
 678
 679 /*
 680  * Memory hotplug specific functions
 681  */
 682 #ifdef CONFIG_MEMORY_HOTPLUG
 683 /*
 684  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 685  * updating.
 686  */
 687 static void  update_end_of_memory_vars(u64 start, u64 size)
 688 {
 689         unsigned long end_pfn = PFN_UP(start + size);
 690
 691         if (end_pfn > max_pfn) {
 692                 max_pfn = end_pfn;
 693                 max_low_pfn = end_pfn;
 694                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 695         }
 696 }
 697
 698 /*
 699  * Memory is added always to NORMAL zone. This means you will never get
 700  * additional DMA/DMA32 memory.
 701  */
 702 int arch_add_memory(int nid, u64 start, u64 size)
 703 {
 704         struct pglist_data *pgdat = NODE_DATA(nid);
 705         struct zone *zone = pgdat->node_zones +
 706                 zone_for_memory(nid, start, size, ZONE_NORMAL);
 707         unsigned long start_pfn = start >> PAGE_SHIFT;
 708         unsigned long nr_pages = size >> PAGE_SHIFT;
 709         int ret;
 710
 711         init_memory_mapping(start, start + size);
 712
 713         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 714         WARN_ON_ONCE(ret);
 715
 716         /* update max_pfn, max_low_pfn and high_memory */
 717         update_end_of_memory_vars(start, size);
 718
 719         return ret;
 720 }
 721 EXPORT_SYMBOL_GPL(arch_add_memory);
 722
 723 #define PAGE_INUSE 0xFD
 724
 725 static void __meminit free_pagetable(struct page *page, int order)
 726 {
 727         unsigned long magic;
 728         unsigned int nr_pages = 1 << order;
 729
 730         /* bootmem page has reserved flag */
 731         if (PageReserved(page)) {
 732                 __ClearPageReserved(page);
 733
 734                 magic = (unsigned long)page->lru.next;
 735                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 736                         while (nr_pages--)
 737                                 put_page_bootmem(page++);
 738                 } else
 739                         while (nr_pages--)
 740                                 free_reserved_page(page++);
 741         } else
 742                 free_pages((unsigned long)page_address(page), order);
 743 }
 744
 745 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 746 {
 747         pte_t *pte;
 748         int i;
 749
 750         for (i = 0; i < PTRS_PER_PTE; i++) {
 751                 pte = pte_start + i;
 752                 if (pte_val(*pte))
 753                         return;
 754         }
 755
 756         /* free a pte talbe */
 757         free_pagetable(pmd_page(*pmd), 0);
 758         spin_lock(&init_mm.page_table_lock);
 759         pmd_clear(pmd);
 760         spin_unlock(&init_mm.page_table_lock);
 761 }
 762
 763 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 764 {
 765         pmd_t *pmd;
 766         int i;
 767
 768         for (i = 0; i < PTRS_PER_PMD; i++) {
 769                 pmd = pmd_start + i;
 770                 if (pmd_val(*pmd))
 771                         return;
 772         }
 773
 774         /* free a pmd talbe */
 775         free_pagetable(pud_page(*pud), 0);
 776         spin_lock(&init_mm.page_table_lock);
 777         pud_clear(pud);
 778         spin_unlock(&init_mm.page_table_lock);
 779 }
 780
 781 /* Return true if pgd is changed, otherwise return false. */
 782 static bool __meminit free_pud_table(pud_t *pud_start, pgd_t *pgd)
 783 {
 784         pud_t *pud;
 785         int i;
 786
 787         for (i = 0; i < PTRS_PER_PUD; i++) {
 788                 pud = pud_start + i;
 789                 if (pud_val(*pud))
 790                         return false;
 791         }
 792
 793         /* free a pud table */
 794         free_pagetable(pgd_page(*pgd), 0);
 795         spin_lock(&init_mm.page_table_lock);
 796         pgd_clear(pgd);
 797         spin_unlock(&init_mm.page_table_lock);
 798
 799         return true;
 800 }
 801
 802 static void __meminit
 803 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 804                  bool direct)
 805 {
 806         unsigned long next, pages = 0;
 807         pte_t *pte;
 808         void *page_addr;
 809         phys_addr_t phys_addr;
 810
 811         pte = pte_start + pte_index(addr);
 812         for (; addr < end; addr = next, pte++) {
 813                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 814                 if (next > end)
 815                         next = end;
 816
 817                 if (!pte_present(*pte))
 818                         continue;
 819
 820                 /*
 821                  * We mapped [0,1G) memory as identity mapping when
 822                  * initializing, in arch/x86/kernel/head_64.S. These
 823                  * pagetables cannot be removed.
 824                  */
 825                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 826                 if (phys_addr < (phys_addr_t)0x40000000)
 827                         return;
 828
 829                 if (IS_ALIGNED(addr, PAGE_SIZE) &&
 830                     IS_ALIGNED(next, PAGE_SIZE)) {
 831                         /*
 832                          * Do not free direct mapping pages since they were
 833                          * freed when offlining, or simplely not in use.
 834                          */
 835                         if (!direct)
 836                                 free_pagetable(pte_page(*pte), 0);
 837
 838                         spin_lock(&init_mm.page_table_lock);
 839                         pte_clear(&init_mm, addr, pte);
 840                         spin_unlock(&init_mm.page_table_lock);
 841
 842                         /* For non-direct mapping, pages means nothing. */
 843                         pages++;
 844                 } else {
 845                         /*
 846                          * If we are here, we are freeing vmemmap pages since
 847                          * direct mapped memory ranges to be freed are aligned.
 848                          *
 849                          * If we are not removing the whole page, it means
 850                          * other page structs in this page are being used and
 851                          * we canot remove them. So fill the unused page_structs
 852                          * with 0xFD, and remove the page when it is wholly
 853                          * filled with 0xFD.
 854                          */
 855                         memset((void *)addr, PAGE_INUSE, next - addr);
 856
 857                         page_addr = page_address(pte_page(*pte));
 858                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
 859                                 free_pagetable(pte_page(*pte), 0);
 860
 861                                 spin_lock(&init_mm.page_table_lock);
 862                                 pte_clear(&init_mm, addr, pte);
 863                                 spin_unlock(&init_mm.page_table_lock);
 864                         }
 865                 }
 866         }
 867
 868         /* Call free_pte_table() in remove_pmd_table(). */
 869         flush_tlb_all();
 870         if (direct)
 871                 update_page_count(PG_LEVEL_4K, -pages);
 872 }
 873
 874 static void __meminit
 875 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
 876                  bool direct)
 877 {
 878         unsigned long next, pages = 0;
 879         pte_t *pte_base;
 880         pmd_t *pmd;
 881         void *page_addr;
 882
 883         pmd = pmd_start + pmd_index(addr);
 884         for (; addr < end; addr = next, pmd++) {
 885                 next = pmd_addr_end(addr, end);
 886
 887                 if (!pmd_present(*pmd))
 888                         continue;
 889
 890                 if (pmd_large(*pmd)) {
 891                         if (IS_ALIGNED(addr, PMD_SIZE) &&
 892                             IS_ALIGNED(next, PMD_SIZE)) {
 893                                 if (!direct)
 894                                         free_pagetable(pmd_page(*pmd),
 895                                                        get_order(PMD_SIZE));
 896
 897                                 spin_lock(&init_mm.page_table_lock);
 898                                 pmd_clear(pmd);
 899                                 spin_unlock(&init_mm.page_table_lock);
 900                                 pages++;
 901                         } else {
 902                                 /* If here, we are freeing vmemmap pages. */
 903                                 memset((void *)addr, PAGE_INUSE, next - addr);
 904
 905                                 page_addr = page_address(pmd_page(*pmd));
 906                                 if (!memchr_inv(page_addr, PAGE_INUSE,
 907                                                 PMD_SIZE)) {
 908                                         free_pagetable(pmd_page(*pmd),
 909                                                        get_order(PMD_SIZE));
 910
 911                                         spin_lock(&init_mm.page_table_lock);
 912                                         pmd_clear(pmd);
 913                                         spin_unlock(&init_mm.page_table_lock);
 914                                 }
 915                         }
 916
 917                         continue;
 918                 }
 919
 920                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
 921                 remove_pte_table(pte_base, addr, next, direct);
 922                 free_pte_table(pte_base, pmd);
 923         }
 924
 925         /* Call free_pmd_table() in remove_pud_table(). */
 926         if (direct)
 927                 update_page_count(PG_LEVEL_2M, -pages);
 928 }
 929
 930 static void __meminit
 931 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
 932                  bool direct)
 933 {
 934         unsigned long next, pages = 0;
 935         pmd_t *pmd_base;
 936         pud_t *pud;
 937         void *page_addr;
 938
 939         pud = pud_start + pud_index(addr);
 940         for (; addr < end; addr = next, pud++) {
 941                 next = pud_addr_end(addr, end);
 942
 943                 if (!pud_present(*pud))
 944                         continue;
 945
 946                 if (pud_large(*pud)) {
 947                         if (IS_ALIGNED(addr, PUD_SIZE) &&
 948                             IS_ALIGNED(next, PUD_SIZE)) {
 949                                 if (!direct)
 950                                         free_pagetable(pud_page(*pud),
 951                                                        get_order(PUD_SIZE));
 952
 953                                 spin_lock(&init_mm.page_table_lock);
 954                                 pud_clear(pud);
 955                                 spin_unlock(&init_mm.page_table_lock);
 956                                 pages++;
 957                         } else {
 958                                 /* If here, we are freeing vmemmap pages. */
 959                                 memset((void *)addr, PAGE_INUSE, next - addr);
 960
 961                                 page_addr = page_address(pud_page(*pud));
 962                                 if (!memchr_inv(page_addr, PAGE_INUSE,
 963                                                 PUD_SIZE)) {
 964                                         free_pagetable(pud_page(*pud),
 965                                                        get_order(PUD_SIZE));
 966
 967                                         spin_lock(&init_mm.page_table_lock);
 968                                         pud_clear(pud);
 969                                         spin_unlock(&init_mm.page_table_lock);
 970                                 }
 971                         }
 972
 973                         continue;
 974                 }
 975
 976                 pmd_base = (pmd_t *)pud_page_vaddr(*pud);
 977                 remove_pmd_table(pmd_base, addr, next, direct);
 978                 free_pmd_table(pmd_base, pud);
 979         }
 980
 981         if (direct)
 982                 update_page_count(PG_LEVEL_1G, -pages);
 983 }
 984
 985 /* start and end are both virtual address. */
 986 static void __meminit
 987 remove_pagetable(unsigned long start, unsigned long end, bool direct)
 988 {
 989         unsigned long next;
 990         unsigned long addr;
 991         pgd_t *pgd;
 992         pud_t *pud;
 993         bool pgd_changed = false;
 994
 995         for (addr = start; addr < end; addr = next) {
 996                 next = pgd_addr_end(addr, end);
 997
 998                 pgd = pgd_offset_k(addr);
 999                 if (!pgd_present(*pgd))
1000                         continue;
1001
1002                 pud = (pud_t *)pgd_page_vaddr(*pgd);
1003                 remove_pud_table(pud, addr, 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, 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 #ifdef CONFIG_MEMORY_HOTREMOVE
1020 static void __meminit
1021 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1022 {
1023         start = (unsigned long)__va(start);
1024         end = (unsigned long)__va(end);
1025
1026         remove_pagetable(start, end, true);
1027 }
1028
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         pci_iommu_alloc();
1061
1062         /* clear_bss() already clear the empty_zero_page */
1063
1064         register_page_bootmem_info();
1065
1066         /* this will put all memory onto the freelists */
1067         free_all_bootmem();
1068         after_bootmem = 1;
1069
1070         /* Register memory areas for /proc/kcore */
1071         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR,
1072                          PAGE_SIZE, KCORE_OTHER);
1073
1074         mem_init_print_info(NULL);
1075 }
1076
1077 #ifdef CONFIG_DEBUG_RODATA
1078 const int rodata_test_data = 0xC3;
1079 EXPORT_SYMBOL_GPL(rodata_test_data);
1080
1081 int kernel_set_to_readonly;
1082
1083 void set_kernel_text_rw(void)
1084 {
1085         unsigned long start = PFN_ALIGN(_text);
1086         unsigned long end = PFN_ALIGN(__stop___ex_table);
1087
1088         if (!kernel_set_to_readonly)
1089                 return;
1090
1091         pr_debug("Set kernel text: %lx - %lx for read write\n",
1092                  start, end);
1093
1094         /*
1095          * Make the kernel identity mapping for text RW. Kernel text
1096          * mapping will always be RO. Refer to the comment in
1097          * static_protections() in pageattr.c
1098          */
1099         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1100 }
1101
1102 void set_kernel_text_ro(void)
1103 {
1104         unsigned long start = PFN_ALIGN(_text);
1105         unsigned long end = PFN_ALIGN(__stop___ex_table);
1106
1107         if (!kernel_set_to_readonly)
1108                 return;
1109
1110         pr_debug("Set kernel text: %lx - %lx for read only\n",
1111                  start, end);
1112
1113         /*
1114          * Set the kernel identity mapping for text RO.
1115          */
1116         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1117 }
1118
1119 void mark_rodata_ro(void)
1120 {
1121         unsigned long start = PFN_ALIGN(_text);
1122         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1123         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1124         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1125         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1126         unsigned long all_end;
1127
1128         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1129                (end - start) >> 10);
1130         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1131
1132         kernel_set_to_readonly = 1;
1133
1134         /*
1135          * The rodata/data/bss/brk section (but not the kernel text!)
1136          * should also be not-executable.
1137          *
1138          * We align all_end to PMD_SIZE because the existing mapping
1139          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1140          * split the PMD and the reminder between _brk_end and the end
1141          * of the PMD will remain mapped executable.
1142          *
1143          * Any PMD which was setup after the one which covers _brk_end
1144          * has been zapped already via cleanup_highmem().
1145          */
1146         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1147         set_memory_nx(rodata_start, (all_end - rodata_start) >> PAGE_SHIFT);
1148
1149         rodata_test();
1150
1151 #ifdef CONFIG_CPA_DEBUG
1152         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1153         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1154
1155         printk(KERN_INFO "Testing CPA: again\n");
1156         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1157 #endif
1158
1159         free_init_pages("unused kernel",
1160                         (unsigned long) __va(__pa_symbol(text_end)),
1161                         (unsigned long) __va(__pa_symbol(rodata_start)));
1162         free_init_pages("unused kernel",
1163                         (unsigned long) __va(__pa_symbol(rodata_end)),
1164                         (unsigned long) __va(__pa_symbol(_sdata)));
1165 }
1166
1167 #endif
1168
1169 int kern_addr_valid(unsigned long addr)
1170 {
1171         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1172         pgd_t *pgd;
1173         pud_t *pud;
1174         pmd_t *pmd;
1175         pte_t *pte;
1176
1177         if (above != 0 && above != -1UL)
1178                 return 0;
1179
1180         pgd = pgd_offset_k(addr);
1181         if (pgd_none(*pgd))
1182                 return 0;
1183
1184         pud = pud_offset(pgd, addr);
1185         if (pud_none(*pud))
1186                 return 0;
1187
1188         if (pud_large(*pud))
1189                 return pfn_valid(pud_pfn(*pud));
1190
1191         pmd = pmd_offset(pud, addr);
1192         if (pmd_none(*pmd))
1193                 return 0;
1194
1195         if (pmd_large(*pmd))
1196                 return pfn_valid(pmd_pfn(*pmd));
1197
1198         pte = pte_offset_kernel(pmd, addr);
1199         if (pte_none(*pte))
1200                 return 0;
1201
1202         return pfn_valid(pte_pfn(*pte));
1203 }
1204
1205 /*
1206  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1207  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1208  * not need special handling anymore:
1209  */
1210 static const char *gate_vma_name(struct vm_area_struct *vma)
1211 {
1212         return "[vsyscall]";
1213 }
1214 static struct vm_operations_struct gate_vma_ops = {
1215         .name = gate_vma_name,
1216 };
1217 static struct vm_area_struct gate_vma = {
1218         .vm_start       = VSYSCALL_ADDR,
1219         .vm_end         = VSYSCALL_ADDR + PAGE_SIZE,
1220         .vm_page_prot   = PAGE_READONLY_EXEC,
1221         .vm_flags       = VM_READ | VM_EXEC,
1222         .vm_ops         = &gate_vma_ops,
1223 };
1224
1225 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
1226 {
1227 #ifdef CONFIG_IA32_EMULATION
1228         if (!mm || mm->context.ia32_compat)
1229                 return NULL;
1230 #endif
1231         return &gate_vma;
1232 }
1233
1234 int in_gate_area(struct mm_struct *mm, unsigned long addr)
1235 {
1236         struct vm_area_struct *vma = get_gate_vma(mm);
1237
1238         if (!vma)
1239                 return 0;
1240
1241         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1242 }
1243
1244 /*
1245  * Use this when you have no reliable mm, typically from interrupt
1246  * context. It is less reliable than using a task's mm and may give
1247  * false positives.
1248  */
1249 int in_gate_area_no_mm(unsigned long addr)
1250 {
1251         return (addr & PAGE_MASK) == VSYSCALL_ADDR;
1252 }
1253
1254 static unsigned long probe_memory_block_size(void)
1255 {
1256         /* start from 2g */
1257         unsigned long bz = 1UL<<31;
1258
1259 #ifdef CONFIG_X86_UV
1260         if (is_uv_system()) {
1261                 printk(KERN_INFO "UV: memory block size 2GB\n");
1262                 return 2UL * 1024 * 1024 * 1024;
1263         }
1264 #endif
1265
1266         /* less than 64g installed */
1267         if ((max_pfn << PAGE_SHIFT) < (16UL << 32))
1268                 return MIN_MEMORY_BLOCK_SIZE;
1269
1270         /* get the tail size */
1271         while (bz > MIN_MEMORY_BLOCK_SIZE) {
1272                 if (!((max_pfn << PAGE_SHIFT) & (bz - 1)))
1273                         break;
1274                 bz >>= 1;
1275         }
1276
1277         printk(KERN_DEBUG "memory block size : %ldMB\n", bz >> 20);
1278
1279         return bz;
1280 }
1281
1282 static unsigned long memory_block_size_probed;
1283 unsigned long memory_block_size_bytes(void)
1284 {
1285         if (!memory_block_size_probed)
1286                 memory_block_size_probed = probe_memory_block_size();
1287
1288         return memory_block_size_probed;
1289 }
1290
1291 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1292 /*
1293  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1294  */
1295 static long __meminitdata addr_start, addr_end;
1296 static void __meminitdata *p_start, *p_end;
1297 static int __meminitdata node_start;
1298
1299 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1300                                                 unsigned long end, int node)
1301 {
1302         unsigned long addr;
1303         unsigned long next;
1304         pgd_t *pgd;
1305         pud_t *pud;
1306         pmd_t *pmd;
1307
1308         for (addr = start; addr < end; addr = next) {
1309                 next = pmd_addr_end(addr, end);
1310
1311                 pgd = vmemmap_pgd_populate(addr, node);
1312                 if (!pgd)
1313                         return -ENOMEM;
1314
1315                 pud = vmemmap_pud_populate(pgd, addr, node);
1316                 if (!pud)
1317                         return -ENOMEM;
1318
1319                 pmd = pmd_offset(pud, addr);
1320                 if (pmd_none(*pmd)) {
1321                         void *p;
1322
1323                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1324                         if (p) {
1325                                 pte_t entry;
1326
1327                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1328                                                 PAGE_KERNEL_LARGE);
1329                                 set_pmd(pmd, __pmd(pte_val(entry)));
1330
1331                                 /* check to see if we have contiguous blocks */
1332                                 if (p_end != p || node_start != node) {
1333                                         if (p_start)
1334                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1335                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1336                                         addr_start = addr;
1337                                         node_start = node;
1338                                         p_start = p;
1339                                 }
1340
1341                                 addr_end = addr + PMD_SIZE;
1342                                 p_end = p + PMD_SIZE;
1343                                 continue;
1344                         }
1345                 } else if (pmd_large(*pmd)) {
1346                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1347                         continue;
1348                 }
1349                 pr_warn_once("vmemmap: falling back to regular page backing\n");
1350                 if (vmemmap_populate_basepages(addr, next, node))
1351                         return -ENOMEM;
1352         }
1353         return 0;
1354 }
1355
1356 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
1357 {
1358         int err;
1359
1360         if (cpu_has_pse)
1361                 err = vmemmap_populate_hugepages(start, end, node);
1362         else
1363                 err = vmemmap_populate_basepages(start, end, node);
1364         if (!err)
1365                 sync_global_pgds(start, end - 1, 0);
1366         return err;
1367 }
1368
1369 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1370 void register_page_bootmem_memmap(unsigned long section_nr,
1371                                   struct page *start_page, unsigned long size)
1372 {
1373         unsigned long addr = (unsigned long)start_page;
1374         unsigned long end = (unsigned long)(start_page + size);
1375         unsigned long next;
1376         pgd_t *pgd;
1377         pud_t *pud;
1378         pmd_t *pmd;
1379         unsigned int nr_pages;
1380         struct page *page;
1381
1382         for (; addr < end; addr = next) {
1383                 pte_t *pte = NULL;
1384
1385                 pgd = pgd_offset_k(addr);
1386                 if (pgd_none(*pgd)) {
1387                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1388                         continue;
1389                 }
1390                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1391
1392                 pud = pud_offset(pgd, addr);
1393                 if (pud_none(*pud)) {
1394                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1395                         continue;
1396                 }
1397                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1398
1399                 if (!cpu_has_pse) {
1400                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1401                         pmd = pmd_offset(pud, addr);
1402                         if (pmd_none(*pmd))
1403                                 continue;
1404                         get_page_bootmem(section_nr, pmd_page(*pmd),
1405                                          MIX_SECTION_INFO);
1406
1407                         pte = pte_offset_kernel(pmd, addr);
1408                         if (pte_none(*pte))
1409                                 continue;
1410                         get_page_bootmem(section_nr, pte_page(*pte),
1411                                          SECTION_INFO);
1412                 } else {
1413                         next = pmd_addr_end(addr, end);
1414
1415                         pmd = pmd_offset(pud, addr);
1416                         if (pmd_none(*pmd))
1417                                 continue;
1418
1419                         nr_pages = 1 << (get_order(PMD_SIZE));
1420                         page = pmd_page(*pmd);
1421                         while (nr_pages--)
1422                                 get_page_bootmem(section_nr, page++,
1423                                                  SECTION_INFO);
1424                 }
1425         }
1426 }
1427 #endif
1428
1429 void __meminit vmemmap_populate_print_last(void)
1430 {
1431         if (p_start) {
1432                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1433                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1434                 p_start = NULL;
1435                 p_end = NULL;
1436                 node_start = 0;
1437         }
1438 }
1439 #endif