mm/memory.c

   1 /*
   2  *  linux/mm/memory.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  */
   6
   7 /*
   8  * demand-loading started 01.12.91 - seems it is high on the list of
   9  * things wanted, and it should be easy to implement. - Linus
  10  */
  11
  12 /*
  13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14  * pages started 02.12.91, seems to work. - Linus.
  15  *
  16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17  * would have taken more than the 6M I have free, but it worked well as
  18  * far as I could see.
  19  *
  20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21  */
  22
  23 /*
  24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25  * thought has to go into this. Oh, well..
  26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27  *              Found it. Everything seems to work now.
  28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29  */
  30
  31 /*
  32  * 05.04.94  -  Multi-page memory management added for v1.1.
  33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  34  *
  35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36  *              (Gerhard.Wichert@pdb.siemens.de)
  37  *
  38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  39  */
  40
  41 #include <linux/kernel_stat.h>
  42 #include <linux/mm.h>
  43 #include <linux/hugetlb.h>
  44 #include <linux/mman.h>
  45 #include <linux/swap.h>
  46 #include <linux/highmem.h>
  47 #include <linux/pagemap.h>
  48 #include <linux/rmap.h>
  49 #include <linux/module.h>
  50 #include <linux/delayacct.h>
  51 #include <linux/init.h>
  52 #include <linux/writeback.h>
  53
  54 #include <asm/pgalloc.h>
  55 #include <asm/uaccess.h>
  56 #include <asm/tlb.h>
  57 #include <asm/tlbflush.h>
  58 #include <asm/pgtable.h>
  59
  60 #include <linux/swapops.h>
  61 #include <linux/elf.h>
  62
  63 #ifndef CONFIG_NEED_MULTIPLE_NODES
  64 /* use the per-pgdat data instead for discontigmem - mbligh */
  65 unsigned long max_mapnr;
  66 struct page *mem_map;
  67
  68 EXPORT_SYMBOL(max_mapnr);
  69 EXPORT_SYMBOL(mem_map);
  70 #endif
  71
  72 unsigned long num_physpages;
  73 /*
  74  * A number of key systems in x86 including ioremap() rely on the assumption
  75  * that high_memory defines the upper bound on direct map memory, then end
  76  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  77  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  78  * and ZONE_HIGHMEM.
  79  */
  80 void * high_memory;
  81 unsigned long vmalloc_earlyreserve;
  82
  83 EXPORT_SYMBOL(num_physpages);
  84 EXPORT_SYMBOL(high_memory);
  85 EXPORT_SYMBOL(vmalloc_earlyreserve);
  86
  87 int randomize_va_space __read_mostly = 1;
  88
  89 static int __init disable_randmaps(char *s)
  90 {
  91         randomize_va_space = 0;
  92         return 1;
  93 }
  94 __setup("norandmaps", disable_randmaps);
  95
  96
  97 /*
  98  * If a p?d_bad entry is found while walking page tables, report
  99  * the error, before resetting entry to p?d_none.  Usually (but
 100  * very seldom) called out from the p?d_none_or_clear_bad macros.
 101  */
 102
 103 void pgd_clear_bad(pgd_t *pgd)
 104 {
 105         pgd_ERROR(*pgd);
 106         pgd_clear(pgd);
 107 }
 108
 109 void pud_clear_bad(pud_t *pud)
 110 {
 111         pud_ERROR(*pud);
 112         pud_clear(pud);
 113 }
 114
 115 void pmd_clear_bad(pmd_t *pmd)
 116 {
 117         pmd_ERROR(*pmd);
 118         pmd_clear(pmd);
 119 }
 120
 121 /*
 122  * Note: this doesn't free the actual pages themselves. That
 123  * has been handled earlier when unmapping all the memory regions.
 124  */
 125 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
 126 {
 127         struct page *page = pmd_page(*pmd);
 128         pmd_clear(pmd);
 129         pte_lock_deinit(page);
 130         pte_free_tlb(tlb, page);
 131         dec_zone_page_state(page, NR_PAGETABLE);
 132         tlb->mm->nr_ptes--;
 133 }
 134
 135 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 136                                 unsigned long addr, unsigned long end,
 137                                 unsigned long floor, unsigned long ceiling)
 138 {
 139         pmd_t *pmd;
 140         unsigned long next;
 141         unsigned long start;
 142
 143         start = addr;
 144         pmd = pmd_offset(pud, addr);
 145         do {
 146                 next = pmd_addr_end(addr, end);
 147                 if (pmd_none_or_clear_bad(pmd))
 148                         continue;
 149                 free_pte_range(tlb, pmd);
 150         } while (pmd++, addr = next, addr != end);
 151
 152         start &= PUD_MASK;
 153         if (start < floor)
 154                 return;
 155         if (ceiling) {
 156                 ceiling &= PUD_MASK;
 157                 if (!ceiling)
 158                         return;
 159         }
 160         if (end - 1 > ceiling - 1)
 161                 return;
 162
 163         pmd = pmd_offset(pud, start);
 164         pud_clear(pud);
 165         pmd_free_tlb(tlb, pmd);
 166 }
 167
 168 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 169                                 unsigned long addr, unsigned long end,
 170                                 unsigned long floor, unsigned long ceiling)
 171 {
 172         pud_t *pud;
 173         unsigned long next;
 174         unsigned long start;
 175
 176         start = addr;
 177         pud = pud_offset(pgd, addr);
 178         do {
 179                 next = pud_addr_end(addr, end);
 180                 if (pud_none_or_clear_bad(pud))
 181                         continue;
 182                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 183         } while (pud++, addr = next, addr != end);
 184
 185         start &= PGDIR_MASK;
 186         if (start < floor)
 187                 return;
 188         if (ceiling) {
 189                 ceiling &= PGDIR_MASK;
 190                 if (!ceiling)
 191                         return;
 192         }
 193         if (end - 1 > ceiling - 1)
 194                 return;
 195
 196         pud = pud_offset(pgd, start);
 197         pgd_clear(pgd);
 198         pud_free_tlb(tlb, pud);
 199 }
 200
 201 /*
 202  * This function frees user-level page tables of a process.
 203  *
 204  * Must be called with pagetable lock held.
 205  */
 206 void free_pgd_range(struct mmu_gather **tlb,
 207                         unsigned long addr, unsigned long end,
 208                         unsigned long floor, unsigned long ceiling)
 209 {
 210         pgd_t *pgd;
 211         unsigned long next;
 212         unsigned long start;
 213
 214         /*
 215          * The next few lines have given us lots of grief...
 216          *
 217          * Why are we testing PMD* at this top level?  Because often
 218          * there will be no work to do at all, and we'd prefer not to
 219          * go all the way down to the bottom just to discover that.
 220          *
 221          * Why all these "- 1"s?  Because 0 represents both the bottom
 222          * of the address space and the top of it (using -1 for the
 223          * top wouldn't help much: the masks would do the wrong thing).
 224          * The rule is that addr 0 and floor 0 refer to the bottom of
 225          * the address space, but end 0 and ceiling 0 refer to the top
 226          * Comparisons need to use "end - 1" and "ceiling - 1" (though
 227          * that end 0 case should be mythical).
 228          *
 229          * Wherever addr is brought up or ceiling brought down, we must
 230          * be careful to reject "the opposite 0" before it confuses the
 231          * subsequent tests.  But what about where end is brought down
 232          * by PMD_SIZE below? no, end can't go down to 0 there.
 233          *
 234          * Whereas we round start (addr) and ceiling down, by different
 235          * masks at different levels, in order to test whether a table
 236          * now has no other vmas using it, so can be freed, we don't
 237          * bother to round floor or end up - the tests don't need that.
 238          */
 239
 240         addr &= PMD_MASK;
 241         if (addr < floor) {
 242                 addr += PMD_SIZE;
 243                 if (!addr)
 244                         return;
 245         }
 246         if (ceiling) {
 247                 ceiling &= PMD_MASK;
 248                 if (!ceiling)
 249                         return;
 250         }
 251         if (end - 1 > ceiling - 1)
 252                 end -= PMD_SIZE;
 253         if (addr > end - 1)
 254                 return;
 255
 256         start = addr;
 257         pgd = pgd_offset((*tlb)->mm, addr);
 258         do {
 259                 next = pgd_addr_end(addr, end);
 260                 if (pgd_none_or_clear_bad(pgd))
 261                         continue;
 262                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
 263         } while (pgd++, addr = next, addr != end);
 264
 265         if (!(*tlb)->fullmm)
 266                 flush_tlb_pgtables((*tlb)->mm, start, end);
 267 }
 268
 269 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
 270                 unsigned long floor, unsigned long ceiling)
 271 {
 272         while (vma) {
 273                 struct vm_area_struct *next = vma->vm_next;
 274                 unsigned long addr = vma->vm_start;
 275
 276                 /*
 277                  * Hide vma from rmap and vmtruncate before freeing pgtables
 278                  */
 279                 anon_vma_unlink(vma);
 280                 unlink_file_vma(vma);
 281
 282                 if (is_vm_hugetlb_page(vma)) {
 283                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 284                                 floor, next? next->vm_start: ceiling);
 285                 } else {
 286                         /*
 287                          * Optimization: gather nearby vmas into one call down
 288                          */
 289                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 290                                && !is_vm_hugetlb_page(next)) {
 291                                 vma = next;
 292                                 next = vma->vm_next;
 293                                 anon_vma_unlink(vma);
 294                                 unlink_file_vma(vma);
 295                         }
 296                         free_pgd_range(tlb, addr, vma->vm_end,
 297                                 floor, next? next->vm_start: ceiling);
 298                 }
 299                 vma = next;
 300         }
 301 }
 302
 303 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 304 {
 305         struct page *new = pte_alloc_one(mm, address);
 306         if (!new)
 307                 return -ENOMEM;
 308
 309         pte_lock_init(new);
 310         spin_lock(&mm->page_table_lock);
 311         if (pmd_present(*pmd)) {        /* Another has populated it */
 312                 pte_lock_deinit(new);
 313                 pte_free(new);
 314         } else {
 315                 mm->nr_ptes++;
 316                 inc_zone_page_state(new, NR_PAGETABLE);
 317                 pmd_populate(mm, pmd, new);
 318         }
 319         spin_unlock(&mm->page_table_lock);
 320         return 0;
 321 }
 322
 323 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 324 {
 325         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 326         if (!new)
 327                 return -ENOMEM;
 328
 329         spin_lock(&init_mm.page_table_lock);
 330         if (pmd_present(*pmd))          /* Another has populated it */
 331                 pte_free_kernel(new);
 332         else
 333                 pmd_populate_kernel(&init_mm, pmd, new);
 334         spin_unlock(&init_mm.page_table_lock);
 335         return 0;
 336 }
 337
 338 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
 339 {
 340         if (file_rss)
 341                 add_mm_counter(mm, file_rss, file_rss);
 342         if (anon_rss)
 343                 add_mm_counter(mm, anon_rss, anon_rss);
 344 }
 345
 346 /*
 347  * This function is called to print an error when a bad pte
 348  * is found. For example, we might have a PFN-mapped pte in
 349  * a region that doesn't allow it.
 350  *
 351  * The calling function must still handle the error.
 352  */
 353 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
 354 {
 355         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
 356                         "vm_flags = %lx, vaddr = %lx\n",
 357                 (long long)pte_val(pte),
 358                 (vma->vm_mm == current->mm ? current->comm : "???"),
 359                 vma->vm_flags, vaddr);
 360         dump_stack();
 361 }
 362
 363 static inline int is_cow_mapping(unsigned int flags)
 364 {
 365         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 366 }
 367
 368 /*
 369  * This function gets the "struct page" associated with a pte.
 370  *
 371  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
 372  * will have each page table entry just pointing to a raw page frame
 373  * number, and as far as the VM layer is concerned, those do not have
 374  * pages associated with them - even if the PFN might point to memory
 375  * that otherwise is perfectly fine and has a "struct page".
 376  *
 377  * The way we recognize those mappings is through the rules set up
 378  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
 379  * and the vm_pgoff will point to the first PFN mapped: thus every
 380  * page that is a raw mapping will always honor the rule
 381  *
 382  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 383  *
 384  * and if that isn't true, the page has been COW'ed (in which case it
 385  * _does_ have a "struct page" associated with it even if it is in a
 386  * VM_PFNMAP range).
 387  */
 388 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
 389 {
 390         unsigned long pfn = pte_pfn(pte);
 391
 392         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
 393                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
 394                 if (pfn == vma->vm_pgoff + off)
 395                         return NULL;
 396                 if (!is_cow_mapping(vma->vm_flags))
 397                         return NULL;
 398         }
 399
 400         /*
 401          * Add some anal sanity checks for now. Eventually,
 402          * we should just do "return pfn_to_page(pfn)", but
 403          * in the meantime we check that we get a valid pfn,
 404          * and that the resulting page looks ok.
 405          */
 406         if (unlikely(!pfn_valid(pfn))) {
 407                 if (!(vma->vm_flags & VM_RESERVED))
 408                         print_bad_pte(vma, pte, addr);
 409                 return NULL;
 410         }
 411
 412         /*
 413          * NOTE! We still have PageReserved() pages in the page
 414          * tables.
 415          *
 416          * The PAGE_ZERO() pages and various VDSO mappings can
 417          * cause them to exist.
 418          */
 419         return pfn_to_page(pfn);
 420 }
 421
 422 /*
 423  * copy one vm_area from one task to the other. Assumes the page tables
 424  * already present in the new task to be cleared in the whole range
 425  * covered by this vma.
 426  */
 427
 428 static inline void
 429 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 430                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 431                 unsigned long addr, int *rss)
 432 {
 433         unsigned long vm_flags = vma->vm_flags;
 434         pte_t pte = *src_pte;
 435         struct page *page;
 436
 437         /* pte contains position in swap or file, so copy. */
 438         if (unlikely(!pte_present(pte))) {
 439                 if (!pte_file(pte)) {
 440                         swp_entry_t entry = pte_to_swp_entry(pte);
 441
 442                         swap_duplicate(entry);
 443                         /* make sure dst_mm is on swapoff's mmlist. */
 444                         if (unlikely(list_empty(&dst_mm->mmlist))) {
 445                                 spin_lock(&mmlist_lock);
 446                                 if (list_empty(&dst_mm->mmlist))
 447                                         list_add(&dst_mm->mmlist,
 448                                                  &src_mm->mmlist);
 449                                 spin_unlock(&mmlist_lock);
 450                         }
 451                         if (is_write_migration_entry(entry) &&
 452                                         is_cow_mapping(vm_flags)) {
 453                                 /*
 454                                  * COW mappings require pages in both parent
 455                                  * and child to be set to read.
 456                                  */
 457                                 make_migration_entry_read(&entry);
 458                                 pte = swp_entry_to_pte(entry);
 459                                 set_pte_at(src_mm, addr, src_pte, pte);
 460                         }
 461                 }
 462                 goto out_set_pte;
 463         }
 464
 465         /*
 466          * If it's a COW mapping, write protect it both
 467          * in the parent and the child
 468          */
 469         if (is_cow_mapping(vm_flags)) {
 470                 ptep_set_wrprotect(src_mm, addr, src_pte);
 471                 pte = *src_pte;
 472         }
 473
 474         /*
 475          * If it's a shared mapping, mark it clean in
 476          * the child
 477          */
 478         if (vm_flags & VM_SHARED)
 479                 pte = pte_mkclean(pte);
 480         pte = pte_mkold(pte);
 481
 482         page = vm_normal_page(vma, addr, pte);
 483         if (page) {
 484                 get_page(page);
 485                 page_dup_rmap(page);
 486                 rss[!!PageAnon(page)]++;
 487         }
 488
 489 out_set_pte:
 490         set_pte_at(dst_mm, addr, dst_pte, pte);
 491 }
 492
 493 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 494                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 495                 unsigned long addr, unsigned long end)
 496 {
 497         pte_t *src_pte, *dst_pte;
 498         spinlock_t *src_ptl, *dst_ptl;
 499         int progress = 0;
 500         int rss[2];
 501
 502 again:
 503         rss[1] = rss[0] = 0;
 504         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 505         if (!dst_pte)
 506                 return -ENOMEM;
 507         src_pte = pte_offset_map_nested(src_pmd, addr);
 508         src_ptl = pte_lockptr(src_mm, src_pmd);
 509         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 510
 511         do {
 512                 /*
 513                  * We are holding two locks at this point - either of them
 514                  * could generate latencies in another task on another CPU.
 515                  */
 516                 if (progress >= 32) {
 517                         progress = 0;
 518                         if (need_resched() ||
 519                             need_lockbreak(src_ptl) ||
 520                             need_lockbreak(dst_ptl))
 521                                 break;
 522                 }
 523                 if (pte_none(*src_pte)) {
 524                         progress++;
 525                         continue;
 526                 }
 527                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
 528                 progress += 8;
 529         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 530
 531         spin_unlock(src_ptl);
 532         pte_unmap_nested(src_pte - 1);
 533         add_mm_rss(dst_mm, rss[0], rss[1]);
 534         pte_unmap_unlock(dst_pte - 1, dst_ptl);
 535         cond_resched();
 536         if (addr != end)
 537                 goto again;
 538         return 0;
 539 }
 540
 541 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 542                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 543                 unsigned long addr, unsigned long end)
 544 {
 545         pmd_t *src_pmd, *dst_pmd;
 546         unsigned long next;
 547
 548         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 549         if (!dst_pmd)
 550                 return -ENOMEM;
 551         src_pmd = pmd_offset(src_pud, addr);
 552         do {
 553                 next = pmd_addr_end(addr, end);
 554                 if (pmd_none_or_clear_bad(src_pmd))
 555                         continue;
 556                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
 557                                                 vma, addr, next))
 558                         return -ENOMEM;
 559         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
 560         return 0;
 561 }
 562
 563 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 564                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
 565                 unsigned long addr, unsigned long end)
 566 {
 567         pud_t *src_pud, *dst_pud;
 568         unsigned long next;
 569
 570         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
 571         if (!dst_pud)
 572                 return -ENOMEM;
 573         src_pud = pud_offset(src_pgd, addr);
 574         do {
 575                 next = pud_addr_end(addr, end);
 576                 if (pud_none_or_clear_bad(src_pud))
 577                         continue;
 578                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
 579                                                 vma, addr, next))
 580                         return -ENOMEM;
 581         } while (dst_pud++, src_pud++, addr = next, addr != end);
 582         return 0;
 583 }
 584
 585 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 586                 struct vm_area_struct *vma)
 587 {
 588         pgd_t *src_pgd, *dst_pgd;
 589         unsigned long next;
 590         unsigned long addr = vma->vm_start;
 591         unsigned long end = vma->vm_end;
 592
 593         /*
 594          * Don't copy ptes where a page fault will fill them correctly.
 595          * Fork becomes much lighter when there are big shared or private
 596          * readonly mappings. The tradeoff is that copy_page_range is more
 597          * efficient than faulting.
 598          */
 599         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
 600                 if (!vma->anon_vma)
 601                         return 0;
 602         }
 603
 604         if (is_vm_hugetlb_page(vma))
 605                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
 606
 607         dst_pgd = pgd_offset(dst_mm, addr);
 608         src_pgd = pgd_offset(src_mm, addr);
 609         do {
 610                 next = pgd_addr_end(addr, end);
 611                 if (pgd_none_or_clear_bad(src_pgd))
 612                         continue;
 613                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
 614                                                 vma, addr, next))
 615                         return -ENOMEM;
 616         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
 617         return 0;
 618 }
 619
 620 static unsigned long zap_pte_range(struct mmu_gather *tlb,
 621                                 struct vm_area_struct *vma, pmd_t *pmd,
 622                                 unsigned long addr, unsigned long end,
 623                                 long *zap_work, struct zap_details *details)
 624 {
 625         struct mm_struct *mm = tlb->mm;
 626         pte_t *pte;
 627         spinlock_t *ptl;
 628         int file_rss = 0;
 629         int anon_rss = 0;
 630
 631         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 632         do {
 633                 pte_t ptent = *pte;
 634                 if (pte_none(ptent)) {
 635                         (*zap_work)--;
 636                         continue;
 637                 }
 638
 639                 (*zap_work) -= PAGE_SIZE;
 640
 641                 if (pte_present(ptent)) {
 642                         struct page *page;
 643
 644                         page = vm_normal_page(vma, addr, ptent);
 645                         if (unlikely(details) && page) {
 646                                 /*
 647                                  * unmap_shared_mapping_pages() wants to
 648                                  * invalidate cache without truncating:
 649                                  * unmap shared but keep private pages.
 650                                  */
 651                                 if (details->check_mapping &&
 652                                     details->check_mapping != page->mapping)
 653                                         continue;
 654                                 /*
 655                                  * Each page->index must be checked when
 656                                  * invalidating or truncating nonlinear.
 657                                  */
 658                                 if (details->nonlinear_vma &&
 659                                     (page->index < details->first_index ||
 660                                      page->index > details->last_index))
 661                                         continue;
 662                         }
 663                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 664                                                         tlb->fullmm);
 665                         tlb_remove_tlb_entry(tlb, pte, addr);
 666                         if (unlikely(!page))
 667                                 continue;
 668                         if (unlikely(details) && details->nonlinear_vma
 669                             && linear_page_index(details->nonlinear_vma,
 670                                                 addr) != page->index)
 671                                 set_pte_at(mm, addr, pte,
 672                                            pgoff_to_pte(page->index));
 673                         if (PageAnon(page))
 674                                 anon_rss--;
 675                         else {
 676                                 if (pte_dirty(ptent))
 677                                         set_page_dirty(page);
 678                                 if (pte_young(ptent))
 679                                         mark_page_accessed(page);
 680                                 file_rss--;
 681                         }
 682                         page_remove_rmap(page, vma);
 683                         tlb_remove_page(tlb, page);
 684                         continue;
 685                 }
 686                 /*
 687                  * If details->check_mapping, we leave swap entries;
 688                  * if details->nonlinear_vma, we leave file entries.
 689                  */
 690                 if (unlikely(details))
 691                         continue;
 692                 if (!pte_file(ptent))
 693                         free_swap_and_cache(pte_to_swp_entry(ptent));
 694                 pte_clear_full(mm, addr, pte, tlb->fullmm);
 695         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
 696
 697         add_mm_rss(mm, file_rss, anon_rss);
 698         pte_unmap_unlock(pte - 1, ptl);
 699
 700         return addr;
 701 }
 702
 703 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
 704                                 struct vm_area_struct *vma, pud_t *pud,
 705                                 unsigned long addr, unsigned long end,
 706                                 long *zap_work, struct zap_details *details)
 707 {
 708         pmd_t *pmd;
 709         unsigned long next;
 710
 711         pmd = pmd_offset(pud, addr);
 712         do {
 713                 next = pmd_addr_end(addr, end);
 714                 if (pmd_none_or_clear_bad(pmd)) {
 715                         (*zap_work)--;
 716                         continue;
 717                 }
 718                 next = zap_pte_range(tlb, vma, pmd, addr, next,
 719                                                 zap_work, details);
 720         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
 721
 722         return addr;
 723 }
 724
 725 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
 726                                 struct vm_area_struct *vma, pgd_t *pgd,
 727                                 unsigned long addr, unsigned long end,
 728                                 long *zap_work, struct zap_details *details)
 729 {
 730         pud_t *pud;
 731         unsigned long next;
 732
 733         pud = pud_offset(pgd, addr);
 734         do {
 735                 next = pud_addr_end(addr, end);
 736                 if (pud_none_or_clear_bad(pud)) {
 737                         (*zap_work)--;
 738                         continue;
 739                 }
 740                 next = zap_pmd_range(tlb, vma, pud, addr, next,
 741                                                 zap_work, details);
 742         } while (pud++, addr = next, (addr != end && *zap_work > 0));
 743
 744         return addr;
 745 }
 746
 747 static unsigned long unmap_page_range(struct mmu_gather *tlb,
 748                                 struct vm_area_struct *vma,
 749                                 unsigned long addr, unsigned long end,
 750                                 long *zap_work, struct zap_details *details)
 751 {
 752         pgd_t *pgd;
 753         unsigned long next;
 754
 755         if (details && !details->check_mapping && !details->nonlinear_vma)
 756                 details = NULL;
 757
 758         BUG_ON(addr >= end);
 759         tlb_start_vma(tlb, vma);
 760         pgd = pgd_offset(vma->vm_mm, addr);
 761         do {
 762                 next = pgd_addr_end(addr, end);
 763                 if (pgd_none_or_clear_bad(pgd)) {
 764                         (*zap_work)--;
 765                         continue;
 766                 }
 767                 next = zap_pud_range(tlb, vma, pgd, addr, next,
 768                                                 zap_work, details);
 769         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
 770         tlb_end_vma(tlb, vma);
 771
 772         return addr;
 773 }
 774
 775 #ifdef CONFIG_PREEMPT
 776 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
 777 #else
 778 /* No preempt: go for improved straight-line efficiency */
 779 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
 780 #endif
 781
 782 /**
 783  * unmap_vmas - unmap a range of memory covered by a list of vma's
 784  * @tlbp: address of the caller's struct mmu_gather
 785  * @vma: the starting vma
 786  * @start_addr: virtual address at which to start unmapping
 787  * @end_addr: virtual address at which to end unmapping
 788  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 789  * @details: details of nonlinear truncation or shared cache invalidation
 790  *
 791  * Returns the end address of the unmapping (restart addr if interrupted).
 792  *
 793  * Unmap all pages in the vma list.
 794  *
 795  * We aim to not hold locks for too long (for scheduling latency reasons).
 796  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 797  * return the ending mmu_gather to the caller.
 798  *
 799  * Only addresses between `start' and `end' will be unmapped.
 800  *
 801  * The VMA list must be sorted in ascending virtual address order.
 802  *
 803  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 804  * range after unmap_vmas() returns.  So the only responsibility here is to
 805  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 806  * drops the lock and schedules.
 807  */
 808 unsigned long unmap_vmas(struct mmu_gather **tlbp,
 809                 struct vm_area_struct *vma, unsigned long start_addr,
 810                 unsigned long end_addr, unsigned long *nr_accounted,
 811                 struct zap_details *details)
 812 {
 813         long zap_work = ZAP_BLOCK_SIZE;
 814         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 815         int tlb_start_valid = 0;
 816         unsigned long start = start_addr;
 817         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
 818         int fullmm = (*tlbp)->fullmm;
 819
 820         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 821                 unsigned long end;
 822
 823                 start = max(vma->vm_start, start_addr);
 824                 if (start >= vma->vm_end)
 825                         continue;
 826                 end = min(vma->vm_end, end_addr);
 827                 if (end <= vma->vm_start)
 828                         continue;
 829
 830                 if (vma->vm_flags & VM_ACCOUNT)
 831                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 832
 833                 while (start != end) {
 834                         if (!tlb_start_valid) {
 835                                 tlb_start = start;
 836                                 tlb_start_valid = 1;
 837                         }
 838
 839                         if (unlikely(is_vm_hugetlb_page(vma))) {
 840                                 unmap_hugepage_range(vma, start, end);
 841                                 zap_work -= (end - start) /
 842                                                 (HPAGE_SIZE / PAGE_SIZE);
 843                                 start = end;
 844                         } else
 845                                 start = unmap_page_range(*tlbp, vma,
 846                                                 start, end, &zap_work, details);
 847
 848                         if (zap_work > 0) {
 849                                 BUG_ON(start != end);
 850                                 break;
 851                         }
 852
 853                         tlb_finish_mmu(*tlbp, tlb_start, start);
 854
 855                         if (need_resched() ||
 856                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
 857                                 if (i_mmap_lock) {
 858                                         *tlbp = NULL;
 859                                         goto out;
 860                                 }
 861                                 cond_resched();
 862                         }
 863
 864                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
 865                         tlb_start_valid = 0;
 866                         zap_work = ZAP_BLOCK_SIZE;
 867                 }
 868         }
 869 out:
 870         return start;   /* which is now the end (or restart) address */
 871 }
 872
 873 /**
 874  * zap_page_range - remove user pages in a given range
 875  * @vma: vm_area_struct holding the applicable pages
 876  * @address: starting address of pages to zap
 877  * @size: number of bytes to zap
 878  * @details: details of nonlinear truncation or shared cache invalidation
 879  */
 880 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
 881                 unsigned long size, struct zap_details *details)
 882 {
 883         struct mm_struct *mm = vma->vm_mm;
 884         struct mmu_gather *tlb;
 885         unsigned long end = address + size;
 886         unsigned long nr_accounted = 0;
 887
 888         lru_add_drain();
 889         tlb = tlb_gather_mmu(mm, 0);
 890         update_hiwater_rss(mm);
 891         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
 892         if (tlb)
 893                 tlb_finish_mmu(tlb, address, end);
 894         return end;
 895 }
 896 EXPORT_SYMBOL(zap_page_range);
 897
 898 /*
 899  * Do a quick page-table lookup for a single page.
 900  */
 901 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
 902                         unsigned int flags)
 903 {
 904         pgd_t *pgd;
 905         pud_t *pud;
 906         pmd_t *pmd;
 907         pte_t *ptep, pte;
 908         spinlock_t *ptl;
 909         struct page *page;
 910         struct mm_struct *mm = vma->vm_mm;
 911
 912         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 913         if (!IS_ERR(page)) {
 914                 BUG_ON(flags & FOLL_GET);
 915                 goto out;
 916         }
 917
 918         page = NULL;
 919         pgd = pgd_offset(mm, address);
 920         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 921                 goto no_page_table;
 922
 923         pud = pud_offset(pgd, address);
 924         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
 925                 goto no_page_table;
 926
 927         pmd = pmd_offset(pud, address);
 928         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 929                 goto no_page_table;
 930
 931         if (pmd_huge(*pmd)) {
 932                 BUG_ON(flags & FOLL_GET);
 933                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
 934                 goto out;
 935         }
 936
 937         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 938         if (!ptep)
 939                 goto out;
 940
 941         pte = *ptep;
 942         if (!pte_present(pte))
 943                 goto unlock;
 944         if ((flags & FOLL_WRITE) && !pte_write(pte))
 945                 goto unlock;
 946         page = vm_normal_page(vma, address, pte);
 947         if (unlikely(!page))
 948                 goto unlock;
 949
 950         if (flags & FOLL_GET)
 951                 get_page(page);
 952         if (flags & FOLL_TOUCH) {
 953                 if ((flags & FOLL_WRITE) &&
 954                     !pte_dirty(pte) && !PageDirty(page))
 955                         set_page_dirty(page);
 956                 mark_page_accessed(page);
 957         }
 958 unlock:
 959         pte_unmap_unlock(ptep, ptl);
 960 out:
 961         return page;
 962
 963 no_page_table:
 964         /*
 965          * When core dumping an enormous anonymous area that nobody
 966          * has touched so far, we don't want to allocate page tables.
 967          */
 968         if (flags & FOLL_ANON) {
 969                 page = ZERO_PAGE(address);
 970                 if (flags & FOLL_GET)
 971                         get_page(page);
 972                 BUG_ON(flags & FOLL_WRITE);
 973         }
 974         return page;
 975 }
 976
 977 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 978                 unsigned long start, int len, int write, int force,
 979                 struct page **pages, struct vm_area_struct **vmas)
 980 {
 981         int i;
 982         unsigned int vm_flags;
 983
 984         /*
 985          * Require read or write permissions.
 986          * If 'force' is set, we only require the "MAY" flags.
 987          */
 988         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 989         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 990         i = 0;
 991
 992         do {
 993                 struct vm_area_struct *vma;
 994                 unsigned int foll_flags;
 995
 996                 vma = find_extend_vma(mm, start);
 997                 if (!vma && in_gate_area(tsk, start)) {
 998                         unsigned long pg = start & PAGE_MASK;
 999                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1000                         pgd_t *pgd;
1001                         pud_t *pud;
1002                         pmd_t *pmd;
1003                         pte_t *pte;
1004                         if (write) /* user gate pages are read-only */
1005                                 return i ? : -EFAULT;
1006                         if (pg > TASK_SIZE)
1007                                 pgd = pgd_offset_k(pg);
1008                         else
1009                                 pgd = pgd_offset_gate(mm, pg);
1010                         BUG_ON(pgd_none(*pgd));
1011                         pud = pud_offset(pgd, pg);
1012                         BUG_ON(pud_none(*pud));
1013                         pmd = pmd_offset(pud, pg);
1014                         if (pmd_none(*pmd))
1015                                 return i ? : -EFAULT;
1016                         pte = pte_offset_map(pmd, pg);
1017                         if (pte_none(*pte)) {
1018                                 pte_unmap(pte);
1019                                 return i ? : -EFAULT;
1020                         }
1021                         if (pages) {
1022                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1023                                 pages[i] = page;
1024                                 if (page)
1025                                         get_page(page);
1026                         }
1027                         pte_unmap(pte);
1028                         if (vmas)
1029                                 vmas[i] = gate_vma;
1030                         i++;
1031                         start += PAGE_SIZE;
1032                         len--;
1033                         continue;
1034                 }
1035
1036 #ifdef CONFIG_XEN
1037                 if (vma && (vma->vm_flags & VM_FOREIGN)) {
1038                         struct page **map = vma->vm_private_data;
1039                         int offset = (start - vma->vm_start) >> PAGE_SHIFT;
1040                         if (map[offset] != NULL) {
1041                                 if (pages) {
1042                                         struct page *page = map[offset];
1043
1044                                         pages[i] = page;
1045                                         get_page(page);
1046                                 }
1047                                 if (vmas)
1048                                         vmas[i] = vma;
1049                                 i++;
1050                                 start += PAGE_SIZE;
1051                                 len--;
1052                                 continue;
1053                         }
1054                 }
1055 #endif
1056                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1057                                 || !(vm_flags & vma->vm_flags))
1058                         return i ? : -EFAULT;
1059
1060                 if (is_vm_hugetlb_page(vma)) {
1061                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1062                                                 &start, &len, i);
1063                         continue;
1064                 }
1065
1066                 foll_flags = FOLL_TOUCH;
1067                 if (pages)
1068                         foll_flags |= FOLL_GET;
1069                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1070                     (!vma->vm_ops || !vma->vm_ops->nopage))
1071                         foll_flags |= FOLL_ANON;
1072
1073                 do {
1074                         struct page *page;
1075
1076                         if (write)
1077                                 foll_flags |= FOLL_WRITE;
1078
1079                         cond_resched();
1080                         while (!(page = follow_page(vma, start, foll_flags))) {
1081                                 int ret;
1082                                 ret = __handle_mm_fault(mm, vma, start,
1083                                                 foll_flags & FOLL_WRITE);
1084                                 /*
1085                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1086                                  * broken COW when necessary, even if maybe_mkwrite
1087                                  * decided not to set pte_write. We can thus safely do
1088                                  * subsequent page lookups as if they were reads.
1089                                  */
1090                                 if (ret & VM_FAULT_WRITE)
1091                                         foll_flags &= ~FOLL_WRITE;
1092
1093                                 switch (ret & ~VM_FAULT_WRITE) {
1094                                 case VM_FAULT_MINOR:
1095                                         tsk->min_flt++;
1096                                         break;
1097                                 case VM_FAULT_MAJOR:
1098                                         tsk->maj_flt++;
1099                                         break;
1100                                 case VM_FAULT_SIGBUS:
1101                                         return i ? i : -EFAULT;
1102                                 case VM_FAULT_OOM:
1103                                         return i ? i : -ENOMEM;
1104                                 default:
1105                                         BUG();
1106                                 }
1107                         }
1108                         if (pages) {
1109                                 pages[i] = page;
1110
1111                                 flush_anon_page(page, start);
1112                                 flush_dcache_page(page);
1113                         }
1114                         if (vmas)
1115                                 vmas[i] = vma;
1116                         i++;
1117                         start += PAGE_SIZE;
1118                         len--;
1119                 } while (len && start < vma->vm_end);
1120         } while (len);
1121         return i;
1122 }
1123 EXPORT_SYMBOL(get_user_pages);
1124
1125 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1126                         unsigned long addr, unsigned long end, pgprot_t prot)
1127 {
1128         pte_t *pte;
1129         spinlock_t *ptl;
1130
1131         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1132         if (!pte)
1133                 return -ENOMEM;
1134         do {
1135                 struct page *page = ZERO_PAGE(addr);
1136                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1137                 page_cache_get(page);
1138                 page_add_file_rmap(page);
1139                 inc_mm_counter(mm, file_rss);
1140                 BUG_ON(!pte_none(*pte));
1141                 set_pte_at(mm, addr, pte, zero_pte);
1142         } while (pte++, addr += PAGE_SIZE, addr != end);
1143         pte_unmap_unlock(pte - 1, ptl);
1144         return 0;
1145 }
1146
1147 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1148                         unsigned long addr, unsigned long end, pgprot_t prot)
1149 {
1150         pmd_t *pmd;
1151         unsigned long next;
1152
1153         pmd = pmd_alloc(mm, pud, addr);
1154         if (!pmd)
1155                 return -ENOMEM;
1156         do {
1157                 next = pmd_addr_end(addr, end);
1158                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1159                         return -ENOMEM;
1160         } while (pmd++, addr = next, addr != end);
1161         return 0;
1162 }
1163
1164 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1165                         unsigned long addr, unsigned long end, pgprot_t prot)
1166 {
1167         pud_t *pud;
1168         unsigned long next;
1169
1170         pud = pud_alloc(mm, pgd, addr);
1171         if (!pud)
1172                 return -ENOMEM;
1173         do {
1174                 next = pud_addr_end(addr, end);
1175                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1176                         return -ENOMEM;
1177         } while (pud++, addr = next, addr != end);
1178         return 0;
1179 }
1180
1181 int zeromap_page_range(struct vm_area_struct *vma,
1182                         unsigned long addr, unsigned long size, pgprot_t prot)
1183 {
1184         pgd_t *pgd;
1185         unsigned long next;
1186         unsigned long end = addr + size;
1187         struct mm_struct *mm = vma->vm_mm;
1188         int err;
1189
1190         BUG_ON(addr >= end);
1191         pgd = pgd_offset(mm, addr);
1192         flush_cache_range(vma, addr, end);
1193         do {
1194                 next = pgd_addr_end(addr, end);
1195                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1196                 if (err)
1197                         break;
1198         } while (pgd++, addr = next, addr != end);
1199         return err;
1200 }
1201
1202 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1203 {
1204         pgd_t * pgd = pgd_offset(mm, addr);
1205         pud_t * pud = pud_alloc(mm, pgd, addr);
1206         if (pud) {
1207                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1208                 if (pmd)
1209                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1210         }
1211         return NULL;
1212 }
1213
1214 /*
1215  * This is the old fallback for page remapping.
1216  *
1217  * For historical reasons, it only allows reserved pages. Only
1218  * old drivers should use this, and they needed to mark their
1219  * pages reserved for the old functions anyway.
1220  */
1221 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1222 {
1223         int retval;
1224         pte_t *pte;
1225         spinlock_t *ptl;
1226
1227         retval = -EINVAL;
1228         if (PageAnon(page))
1229                 goto out;
1230         retval = -ENOMEM;
1231         flush_dcache_page(page);
1232         pte = get_locked_pte(mm, addr, &ptl);
1233         if (!pte)
1234                 goto out;
1235         retval = -EBUSY;
1236         if (!pte_none(*pte))
1237                 goto out_unlock;
1238
1239         /* Ok, finally just insert the thing.. */
1240         get_page(page);
1241         inc_mm_counter(mm, file_rss);
1242         page_add_file_rmap(page);
1243         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1244
1245         retval = 0;
1246 out_unlock:
1247         pte_unmap_unlock(pte, ptl);
1248 out:
1249         return retval;
1250 }
1251
1252 /*
1253  * This allows drivers to insert individual pages they've allocated
1254  * into a user vma.
1255  *
1256  * The page has to be a nice clean _individual_ kernel allocation.
1257  * If you allocate a compound page, you need to have marked it as
1258  * such (__GFP_COMP), or manually just split the page up yourself
1259  * (see split_page()).
1260  *
1261  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1262  * took an arbitrary page protection parameter. This doesn't allow
1263  * that. Your vma protection will have to be set up correctly, which
1264  * means that if you want a shared writable mapping, you'd better
1265  * ask for a shared writable mapping!
1266  *
1267  * The page does not need to be reserved.
1268  */
1269 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1270 {
1271         if (addr < vma->vm_start || addr >= vma->vm_end)
1272                 return -EFAULT;
1273         if (!page_count(page))
1274                 return -EINVAL;
1275         vma->vm_flags |= VM_INSERTPAGE;
1276         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1277 }
1278 EXPORT_SYMBOL(vm_insert_page);
1279
1280 /*
1281  * maps a range of physical memory into the requested pages. the old
1282  * mappings are removed. any references to nonexistent pages results
1283  * in null mappings (currently treated as "copy-on-access")
1284  */
1285 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1286                         unsigned long addr, unsigned long end,
1287                         unsigned long pfn, pgprot_t prot)
1288 {
1289         pte_t *pte;
1290         spinlock_t *ptl;
1291
1292         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1293         if (!pte)
1294                 return -ENOMEM;
1295         do {
1296                 BUG_ON(!pte_none(*pte));
1297                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1298                 pfn++;
1299         } while (pte++, addr += PAGE_SIZE, addr != end);
1300         pte_unmap_unlock(pte - 1, ptl);
1301         return 0;
1302 }
1303
1304 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1305                         unsigned long addr, unsigned long end,
1306                         unsigned long pfn, pgprot_t prot)
1307 {
1308         pmd_t *pmd;
1309         unsigned long next;
1310
1311         pfn -= addr >> PAGE_SHIFT;
1312         pmd = pmd_alloc(mm, pud, addr);
1313         if (!pmd)
1314                 return -ENOMEM;
1315         do {
1316                 next = pmd_addr_end(addr, end);
1317                 if (remap_pte_range(mm, pmd, addr, next,
1318                                 pfn + (addr >> PAGE_SHIFT), prot))
1319                         return -ENOMEM;
1320         } while (pmd++, addr = next, addr != end);
1321         return 0;
1322 }
1323
1324 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1325                         unsigned long addr, unsigned long end,
1326                         unsigned long pfn, pgprot_t prot)
1327 {
1328         pud_t *pud;
1329         unsigned long next;
1330
1331         pfn -= addr >> PAGE_SHIFT;
1332         pud = pud_alloc(mm, pgd, addr);
1333         if (!pud)
1334                 return -ENOMEM;
1335         do {
1336                 next = pud_addr_end(addr, end);
1337                 if (remap_pmd_range(mm, pud, addr, next,
1338                                 pfn + (addr >> PAGE_SHIFT), prot))
1339                         return -ENOMEM;
1340         } while (pud++, addr = next, addr != end);
1341         return 0;
1342 }
1343
1344 /*  Note: this is only safe if the mm semaphore is held when called. */
1345 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1346                     unsigned long pfn, unsigned long size, pgprot_t prot)
1347 {
1348         pgd_t *pgd;
1349         unsigned long next;
1350         unsigned long end = addr + PAGE_ALIGN(size);
1351         struct mm_struct *mm = vma->vm_mm;
1352         int err;
1353
1354         /*
1355          * Physically remapped pages are special. Tell the
1356          * rest of the world about it:
1357          *   VM_IO tells people not to look at these pages
1358          *      (accesses can have side effects).
1359          *   VM_RESERVED is specified all over the place, because
1360          *      in 2.4 it kept swapout's vma scan off this vma; but
1361          *      in 2.6 the LRU scan won't even find its pages, so this
1362          *      flag means no more than count its pages in reserved_vm,
1363          *      and omit it from core dump, even when VM_IO turned off.
1364          *   VM_PFNMAP tells the core MM that the base pages are just
1365          *      raw PFN mappings, and do not have a "struct page" associated
1366          *      with them.
1367          *
1368          * There's a horrible special case to handle copy-on-write
1369          * behaviour that some programs depend on. We mark the "original"
1370          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1371          */
1372         if (is_cow_mapping(vma->vm_flags)) {
1373                 if (addr != vma->vm_start || end != vma->vm_end)
1374                         return -EINVAL;
1375                 vma->vm_pgoff = pfn;
1376         }
1377
1378         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1379
1380         BUG_ON(addr >= end);
1381         pfn -= addr >> PAGE_SHIFT;
1382         pgd = pgd_offset(mm, addr);
1383         flush_cache_range(vma, addr, end);
1384         do {
1385                 next = pgd_addr_end(addr, end);
1386                 err = remap_pud_range(mm, pgd, addr, next,
1387                                 pfn + (addr >> PAGE_SHIFT), prot);
1388                 if (err)
1389                         break;
1390         } while (pgd++, addr = next, addr != end);
1391         return err;
1392 }
1393 EXPORT_SYMBOL(remap_pfn_range);
1394
1395 #ifdef CONFIG_XEN
1396 static inline int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1397                                      unsigned long addr, unsigned long end,
1398                                      pte_fn_t fn, void *data)
1399 {
1400         pte_t *pte;
1401         int err;
1402         struct page *pmd_page;
1403         spinlock_t *ptl;
1404
1405         pte = (mm == &init_mm) ?
1406                 pte_alloc_kernel(pmd, addr) :
1407                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1408         if (!pte)
1409                 return -ENOMEM;
1410
1411         BUG_ON(pmd_huge(*pmd));
1412
1413         pmd_page = pmd_page(*pmd);
1414
1415         do {
1416                 err = fn(pte, pmd_page, addr, data);
1417                 if (err)
1418                         break;
1419         } while (pte++, addr += PAGE_SIZE, addr != end);
1420
1421         if (mm != &init_mm)
1422                 pte_unmap_unlock(pte-1, ptl);
1423         return err;
1424 }
1425
1426 static inline int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1427                                      unsigned long addr, unsigned long end,
1428                                      pte_fn_t fn, void *data)
1429 {
1430         pmd_t *pmd;
1431         unsigned long next;
1432         int err;
1433
1434         pmd = pmd_alloc(mm, pud, addr);
1435         if (!pmd)
1436                 return -ENOMEM;
1437         do {
1438                 next = pmd_addr_end(addr, end);
1439                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1440                 if (err)
1441                         break;
1442         } while (pmd++, addr = next, addr != end);
1443         return err;
1444 }
1445
1446 static inline int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1447                                      unsigned long addr, unsigned long end,
1448                                      pte_fn_t fn, void *data)
1449 {
1450         pud_t *pud;
1451         unsigned long next;
1452         int err;
1453
1454         pud = pud_alloc(mm, pgd, addr);
1455         if (!pud)
1456                 return -ENOMEM;
1457         do {
1458                 next = pud_addr_end(addr, end);
1459                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1460                 if (err)
1461                         break;
1462         } while (pud++, addr = next, addr != end);
1463         return err;
1464 }
1465
1466 /*
1467  * Scan a region of virtual memory, filling in page tables as necessary
1468  * and calling a provided function on each leaf page table.
1469  */
1470 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1471                         unsigned long size, pte_fn_t fn, void *data)
1472 {
1473         pgd_t *pgd;
1474         unsigned long next;
1475         unsigned long end = addr + size;
1476         int err;
1477
1478         BUG_ON(addr >= end);
1479         pgd = pgd_offset(mm, addr);
1480         do {
1481                 next = pgd_addr_end(addr, end);
1482                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1483                 if (err)
1484                         break;
1485         } while (pgd++, addr = next, addr != end);
1486         return err;
1487 }
1488 EXPORT_SYMBOL_GPL(apply_to_page_range);
1489 #endif
1490
1491 /*
1492  * handle_pte_fault chooses page fault handler according to an entry
1493  * which was read non-atomically.  Before making any commitment, on
1494  * those architectures or configurations (e.g. i386 with PAE) which
1495  * might give a mix of unmatched parts, do_swap_page and do_file_page
1496  * must check under lock before unmapping the pte and proceeding
1497  * (but do_wp_page is only called after already making such a check;
1498  * and do_anonymous_page and do_no_page can safely check later on).
1499  */
1500 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1501                                 pte_t *page_table, pte_t orig_pte)
1502 {
1503         int same = 1;
1504 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1505         if (sizeof(pte_t) > sizeof(unsigned long)) {
1506                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1507                 spin_lock(ptl);
1508                 same = pte_same(*page_table, orig_pte);
1509                 spin_unlock(ptl);
1510         }
1511 #endif
1512         pte_unmap(page_table);
1513         return same;
1514 }
1515
1516 /*
1517  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1518  * servicing faults for write access.  In the normal case, do always want
1519  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1520  * that do not have writing enabled, when used by access_process_vm.
1521  */
1522 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1523 {
1524         if (likely(vma->vm_flags & VM_WRITE))
1525                 pte = pte_mkwrite(pte);
1526         return pte;
1527 }
1528
1529 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1530 {
1531         /*
1532          * If the source page was a PFN mapping, we don't have
1533          * a "struct page" for it. We do a best-effort copy by
1534          * just copying from the original user address. If that
1535          * fails, we just zero-fill it. Live with it.
1536          */
1537         if (unlikely(!src)) {
1538                 void *kaddr = kmap_atomic(dst, KM_USER0);
1539                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1540
1541                 /*
1542                  * This really shouldn't fail, because the page is there
1543                  * in the page tables. But it might just be unreadable,
1544                  * in which case we just give up and fill the result with
1545                  * zeroes.
1546                  */
1547                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1548                         memset(kaddr, 0, PAGE_SIZE);
1549                 kunmap_atomic(kaddr, KM_USER0);
1550                 return;
1551
1552         }
1553         copy_user_highpage(dst, src, va);
1554 }
1555
1556 /*
1557  * This routine handles present pages, when users try to write
1558  * to a shared page. It is done by copying the page to a new address
1559  * and decrementing the shared-page counter for the old page.
1560  *
1561  * Note that this routine assumes that the protection checks have been
1562  * done by the caller (the low-level page fault routine in most cases).
1563  * Thus we can safely just mark it writable once we've done any necessary
1564  * COW.
1565  *
1566  * We also mark the page dirty at this point even though the page will
1567  * change only once the write actually happens. This avoids a few races,
1568  * and potentially makes it more efficient.
1569  *
1570  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1571  * but allow concurrent faults), with pte both mapped and locked.
1572  * We return with mmap_sem still held, but pte unmapped and unlocked.
1573  */
1574 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1575                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1576                 spinlock_t *ptl, pte_t orig_pte)
1577 {
1578         struct page *old_page, *new_page;
1579         pte_t entry;
1580         int reuse = 0, ret = VM_FAULT_MINOR;
1581         struct page *dirty_page = NULL;
1582
1583         old_page = vm_normal_page(vma, address, orig_pte);
1584         if (!old_page)
1585                 goto gotten;
1586
1587         /*
1588          * Take out anonymous pages first, anonymous shared vmas are
1589          * not dirty accountable.
1590          */
1591         if (PageAnon(old_page)) {
1592                 if (!TestSetPageLocked(old_page)) {
1593                         reuse = can_share_swap_page(old_page);
1594                         unlock_page(old_page);
1595                 }
1596         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1597                                         (VM_WRITE|VM_SHARED))) {
1598                 /*
1599                  * Only catch write-faults on shared writable pages,
1600                  * read-only shared pages can get COWed by
1601                  * get_user_pages(.write=1, .force=1).
1602                  */
1603                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1604                         /*
1605                          * Notify the address space that the page is about to
1606                          * become writable so that it can prohibit this or wait
1607                          * for the page to get into an appropriate state.
1608                          *
1609                          * We do this without the lock held, so that it can
1610                          * sleep if it needs to.
1611                          */
1612                         page_cache_get(old_page);
1613                         pte_unmap_unlock(page_table, ptl);
1614
1615                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1616                                 goto unwritable_page;
1617
1618                         page_cache_release(old_page);
1619
1620                         /*
1621                          * Since we dropped the lock we need to revalidate
1622                          * the PTE as someone else may have changed it.  If
1623                          * they did, we just return, as we can count on the
1624                          * MMU to tell us if they didn't also make it writable.
1625                          */
1626                         page_table = pte_offset_map_lock(mm, pmd, address,
1627                                                          &ptl);
1628                         if (!pte_same(*page_table, orig_pte))
1629                                 goto unlock;
1630                 }
1631                 dirty_page = old_page;
1632                 get_page(dirty_page);
1633                 reuse = 1;
1634         }
1635
1636         if (reuse) {
1637                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1638                 entry = pte_mkyoung(orig_pte);
1639                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1640                 ptep_set_access_flags(vma, address, page_table, entry, 1);
1641                 update_mmu_cache(vma, address, entry);
1642                 lazy_mmu_prot_update(entry);
1643                 ret |= VM_FAULT_WRITE;
1644                 goto unlock;
1645         }
1646
1647         /*
1648          * Ok, we need to copy. Oh, well..
1649          */
1650         page_cache_get(old_page);
1651 gotten:
1652         pte_unmap_unlock(page_table, ptl);
1653
1654         if (unlikely(anon_vma_prepare(vma)))
1655                 goto oom;
1656         if (old_page == ZERO_PAGE(address)) {
1657                 new_page = alloc_zeroed_user_highpage(vma, address);
1658                 if (!new_page)
1659                         goto oom;
1660         } else {
1661                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1662                 if (!new_page)
1663                         goto oom;
1664                 cow_user_page(new_page, old_page, address);
1665         }
1666
1667         /*
1668          * Re-check the pte - we dropped the lock
1669          */
1670         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1671         if (likely(pte_same(*page_table, orig_pte))) {
1672                 if (old_page) {
1673                         page_remove_rmap(old_page, vma);
1674                         if (!PageAnon(old_page)) {
1675                                 dec_mm_counter(mm, file_rss);
1676                                 inc_mm_counter(mm, anon_rss);
1677                         }
1678                 } else
1679                         inc_mm_counter(mm, anon_rss);
1680                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1681                 entry = mk_pte(new_page, vma->vm_page_prot);
1682                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1683                 lazy_mmu_prot_update(entry);
1684                 /*
1685                  * Clear the pte entry and flush it first, before updating the
1686                  * pte with the new entry. This will avoid a race condition
1687                  * seen in the presence of one thread doing SMC and another
1688                  * thread doing COW.
1689                  */
1690                 ptep_clear_flush(vma, address, page_table);
1691                 set_pte_at(mm, address, page_table, entry);
1692                 update_mmu_cache(vma, address, entry);
1693                 lru_cache_add_active(new_page);
1694                 page_add_new_anon_rmap(new_page, vma, address);
1695
1696                 /* Free the old page.. */
1697                 new_page = old_page;
1698                 ret |= VM_FAULT_WRITE;
1699         }
1700         if (new_page)
1701                 page_cache_release(new_page);
1702         if (old_page)
1703                 page_cache_release(old_page);
1704 unlock:
1705         pte_unmap_unlock(page_table, ptl);
1706         if (dirty_page) {
1707                 set_page_dirty_balance(dirty_page);
1708                 put_page(dirty_page);
1709         }
1710         return ret;
1711 oom:
1712         if (old_page)
1713                 page_cache_release(old_page);
1714         return VM_FAULT_OOM;
1715
1716 unwritable_page:
1717         page_cache_release(old_page);
1718         return VM_FAULT_SIGBUS;
1719 }
1720
1721 /*
1722  * Helper functions for unmap_mapping_range().
1723  *
1724  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1725  *
1726  * We have to restart searching the prio_tree whenever we drop the lock,
1727  * since the iterator is only valid while the lock is held, and anyway
1728  * a later vma might be split and reinserted earlier while lock dropped.
1729  *
1730  * The list of nonlinear vmas could be handled more efficiently, using
1731  * a placeholder, but handle it in the same way until a need is shown.
1732  * It is important to search the prio_tree before nonlinear list: a vma
1733  * may become nonlinear and be shifted from prio_tree to nonlinear list
1734  * while the lock is dropped; but never shifted from list to prio_tree.
1735  *
1736  * In order to make forward progress despite restarting the search,
1737  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1738  * quickly skip it next time around.  Since the prio_tree search only
1739  * shows us those vmas affected by unmapping the range in question, we
1740  * can't efficiently keep all vmas in step with mapping->truncate_count:
1741  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1742  * mapping->truncate_count and vma->vm_truncate_count are protected by
1743  * i_mmap_lock.
1744  *
1745  * In order to make forward progress despite repeatedly restarting some
1746  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1747  * and restart from that address when we reach that vma again.  It might
1748  * have been split or merged, shrunk or extended, but never shifted: so
1749  * restart_addr remains valid so long as it remains in the vma's range.
1750  * unmap_mapping_range forces truncate_count to leap over page-aligned
1751  * values so we can save vma's restart_addr in its truncate_count field.
1752  */
1753 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1754
1755 static void reset_vma_truncate_counts(struct address_space *mapping)
1756 {
1757         struct vm_area_struct *vma;
1758         struct prio_tree_iter iter;
1759
1760         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1761                 vma->vm_truncate_count = 0;
1762         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1763                 vma->vm_truncate_count = 0;
1764 }
1765
1766 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1767                 unsigned long start_addr, unsigned long end_addr,
1768                 struct zap_details *details)
1769 {
1770         unsigned long restart_addr;
1771         int need_break;
1772
1773 again:
1774         restart_addr = vma->vm_truncate_count;
1775         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1776                 start_addr = restart_addr;
1777                 if (start_addr >= end_addr) {
1778                         /* Top of vma has been split off since last time */
1779                         vma->vm_truncate_count = details->truncate_count;
1780                         return 0;
1781                 }
1782         }
1783
1784         restart_addr = zap_page_range(vma, start_addr,
1785                                         end_addr - start_addr, details);
1786         need_break = need_resched() ||
1787                         need_lockbreak(details->i_mmap_lock);
1788
1789         if (restart_addr >= end_addr) {
1790                 /* We have now completed this vma: mark it so */
1791                 vma->vm_truncate_count = details->truncate_count;
1792                 if (!need_break)
1793                         return 0;
1794         } else {
1795                 /* Note restart_addr in vma's truncate_count field */
1796                 vma->vm_truncate_count = restart_addr;
1797                 if (!need_break)
1798                         goto again;
1799         }
1800
1801         spin_unlock(details->i_mmap_lock);
1802         cond_resched();
1803         spin_lock(details->i_mmap_lock);
1804         return -EINTR;
1805 }
1806
1807 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1808                                             struct zap_details *details)
1809 {
1810         struct vm_area_struct *vma;
1811         struct prio_tree_iter iter;
1812         pgoff_t vba, vea, zba, zea;
1813
1814 restart:
1815         vma_prio_tree_foreach(vma, &iter, root,
1816                         details->first_index, details->last_index) {
1817                 /* Skip quickly over those we have already dealt with */
1818                 if (vma->vm_truncate_count == details->truncate_count)
1819                         continue;
1820
1821                 vba = vma->vm_pgoff;
1822                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1823                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1824                 zba = details->first_index;
1825                 if (zba < vba)
1826                         zba = vba;
1827                 zea = details->last_index;
1828                 if (zea > vea)
1829                         zea = vea;
1830
1831                 if (unmap_mapping_range_vma(vma,
1832                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1833                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1834                                 details) < 0)
1835                         goto restart;
1836         }
1837 }
1838
1839 static inline void unmap_mapping_range_list(struct list_head *head,
1840                                             struct zap_details *details)
1841 {
1842         struct vm_area_struct *vma;
1843
1844         /*
1845          * In nonlinear VMAs there is no correspondence between virtual address
1846          * offset and file offset.  So we must perform an exhaustive search
1847          * across *all* the pages in each nonlinear VMA, not just the pages
1848          * whose virtual address lies outside the file truncation point.
1849          */
1850 restart:
1851         list_for_each_entry(vma, head, shared.vm_set.list) {
1852                 /* Skip quickly over those we have already dealt with */
1853                 if (vma->vm_truncate_count == details->truncate_count)
1854                         continue;
1855                 details->nonlinear_vma = vma;
1856                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1857                                         vma->vm_end, details) < 0)
1858                         goto restart;
1859         }
1860 }
1861
1862 /**
1863  * unmap_mapping_range - unmap the portion of all mmaps
1864  * in the specified address_space corresponding to the specified
1865  * page range in the underlying file.
1866  * @mapping: the address space containing mmaps to be unmapped.
1867  * @holebegin: byte in first page to unmap, relative to the start of
1868  * the underlying file.  This will be rounded down to a PAGE_SIZE
1869  * boundary.  Note that this is different from vmtruncate(), which
1870  * must keep the partial page.  In contrast, we must get rid of
1871  * partial pages.
1872  * @holelen: size of prospective hole in bytes.  This will be rounded
1873  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1874  * end of the file.
1875  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1876  * but 0 when invalidating pagecache, don't throw away private data.
1877  */
1878 void unmap_mapping_range(struct address_space *mapping,
1879                 loff_t const holebegin, loff_t const holelen, int even_cows)
1880 {
1881         struct zap_details details;
1882         pgoff_t hba = holebegin >> PAGE_SHIFT;
1883         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1884
1885         /* Check for overflow. */
1886         if (sizeof(holelen) > sizeof(hlen)) {
1887                 long long holeend =
1888                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1889                 if (holeend & ~(long long)ULONG_MAX)
1890                         hlen = ULONG_MAX - hba + 1;
1891         }
1892
1893         details.check_mapping = even_cows? NULL: mapping;
1894         details.nonlinear_vma = NULL;
1895         details.first_index = hba;
1896         details.last_index = hba + hlen - 1;
1897         if (details.last_index < details.first_index)
1898                 details.last_index = ULONG_MAX;
1899         details.i_mmap_lock = &mapping->i_mmap_lock;
1900
1901         spin_lock(&mapping->i_mmap_lock);
1902
1903         /* serialize i_size write against truncate_count write */
1904         smp_wmb();
1905         /* Protect against page faults, and endless unmapping loops */
1906         mapping->truncate_count++;
1907         /*
1908          * For archs where spin_lock has inclusive semantics like ia64
1909          * this smp_mb() will prevent to read pagetable contents
1910          * before the truncate_count increment is visible to
1911          * other cpus.
1912          */
1913         smp_mb();
1914         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1915                 if (mapping->truncate_count == 0)
1916                         reset_vma_truncate_counts(mapping);
1917                 mapping->truncate_count++;
1918         }
1919         details.truncate_count = mapping->truncate_count;
1920
1921         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1922                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1923         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1924                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1925         spin_unlock(&mapping->i_mmap_lock);
1926 }
1927 EXPORT_SYMBOL(unmap_mapping_range);
1928
1929 /*
1930  * Handle all mappings that got truncated by a "truncate()"
1931  * system call.
1932  *
1933  * NOTE! We have to be ready to update the memory sharing
1934  * between the file and the memory map for a potential last
1935  * incomplete page.  Ugly, but necessary.
1936  */
1937 int vmtruncate(struct inode * inode, loff_t offset)
1938 {
1939         struct address_space *mapping = inode->i_mapping;
1940         unsigned long limit;
1941
1942         if (inode->i_size < offset)
1943                 goto do_expand;
1944         /*
1945          * truncation of in-use swapfiles is disallowed - it would cause
1946          * subsequent swapout to scribble on the now-freed blocks.
1947          */
1948         if (IS_SWAPFILE(inode))
1949                 goto out_busy;
1950         i_size_write(inode, offset);
1951         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1952         truncate_inode_pages(mapping, offset);
1953         goto out_truncate;
1954
1955 do_expand:
1956         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1957         if (limit != RLIM_INFINITY && offset > limit)
1958                 goto out_sig;
1959         if (offset > inode->i_sb->s_maxbytes)
1960                 goto out_big;
1961         i_size_write(inode, offset);
1962
1963 out_truncate:
1964         if (inode->i_op && inode->i_op->truncate)
1965                 inode->i_op->truncate(inode);
1966         return 0;
1967 out_sig:
1968         send_sig(SIGXFSZ, current, 0);
1969 out_big:
1970         return -EFBIG;
1971 out_busy:
1972         return -ETXTBSY;
1973 }
1974 EXPORT_SYMBOL(vmtruncate);
1975
1976 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1977 {
1978         struct address_space *mapping = inode->i_mapping;
1979
1980         /*
1981          * If the underlying filesystem is not going to provide
1982          * a way to truncate a range of blocks (punch a hole) -
1983          * we should return failure right now.
1984          */
1985         if (!inode->i_op || !inode->i_op->truncate_range)
1986                 return -ENOSYS;
1987
1988         mutex_lock(&inode->i_mutex);
1989         down_write(&inode->i_alloc_sem);
1990         unmap_mapping_range(mapping, offset, (end - offset), 1);
1991         truncate_inode_pages_range(mapping, offset, end);
1992         inode->i_op->truncate_range(inode, offset, end);
1993         up_write(&inode->i_alloc_sem);
1994         mutex_unlock(&inode->i_mutex);
1995
1996         return 0;
1997 }
1998 EXPORT_UNUSED_SYMBOL(vmtruncate_range);  /*  June 2006  */
1999
2000 /*
2001  * Primitive swap readahead code. We simply read an aligned block of
2002  * (1 << page_cluster) entries in the swap area. This method is chosen
2003  * because it doesn't cost us any seek time.  We also make sure to queue
2004  * the 'original' request together with the readahead ones...
2005  *
2006  * This has been extended to use the NUMA policies from the mm triggering
2007  * the readahead.
2008  *
2009  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
2010  */
2011 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
2012 {
2013 #ifdef CONFIG_NUMA
2014         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
2015 #endif
2016         int i, num;
2017         struct page *new_page;
2018         unsigned long offset;
2019
2020         /*
2021          * Get the number of handles we should do readahead io to.
2022          */
2023         num = valid_swaphandles(entry, &offset);
2024         for (i = 0; i < num; offset++, i++) {
2025                 /* Ok, do the async read-ahead now */
2026                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
2027                                                            offset), vma, addr);
2028                 if (!new_page)
2029                         break;
2030                 page_cache_release(new_page);
2031 #ifdef CONFIG_NUMA
2032                 /*
2033                  * Find the next applicable VMA for the NUMA policy.
2034                  */
2035                 addr += PAGE_SIZE;
2036                 if (addr == 0)
2037                         vma = NULL;
2038                 if (vma) {
2039                         if (addr >= vma->vm_end) {
2040                                 vma = next_vma;
2041                                 next_vma = vma ? vma->vm_next : NULL;
2042                         }
2043                         if (vma && addr < vma->vm_start)
2044                                 vma = NULL;
2045                 } else {
2046                         if (next_vma && addr >= next_vma->vm_start) {
2047                                 vma = next_vma;
2048                                 next_vma = vma->vm_next;
2049                         }
2050                 }
2051 #endif
2052         }
2053         lru_add_drain();        /* Push any new pages onto the LRU now */
2054 }
2055
2056 /*
2057  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2058  * but allow concurrent faults), and pte mapped but not yet locked.
2059  * We return with mmap_sem still held, but pte unmapped and unlocked.
2060  */
2061 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2062                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2063                 int write_access, pte_t orig_pte)
2064 {
2065         spinlock_t *ptl;
2066         struct page *page;
2067         swp_entry_t entry;
2068         pte_t pte;
2069         int ret = VM_FAULT_MINOR;
2070
2071         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2072                 goto out;
2073
2074         entry = pte_to_swp_entry(orig_pte);
2075         if (is_migration_entry(entry)) {
2076                 migration_entry_wait(mm, pmd, address);
2077                 goto out;
2078         }
2079         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2080         page = lookup_swap_cache(entry);
2081         if (!page) {
2082                 swapin_readahead(entry, address, vma);
2083                 page = read_swap_cache_async(entry, vma, address);
2084                 if (!page) {
2085                         /*
2086                          * Back out if somebody else faulted in this pte
2087                          * while we released the pte lock.
2088                          */
2089                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2090                         if (likely(pte_same(*page_table, orig_pte)))
2091                                 ret = VM_FAULT_OOM;
2092                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2093                         goto unlock;
2094                 }
2095
2096                 /* Had to read the page from swap area: Major fault */
2097                 ret = VM_FAULT_MAJOR;
2098                 count_vm_event(PGMAJFAULT);
2099                 grab_swap_token();
2100         }
2101
2102         if (!vx_rsspages_avail(mm, 1)) {
2103                 ret = VM_FAULT_OOM;
2104                 goto out;
2105         }
2106
2107         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2108         mark_page_accessed(page);
2109         lock_page(page);
2110
2111         /*
2112          * Back out if somebody else already faulted in this pte.
2113          */
2114         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2115         if (unlikely(!pte_same(*page_table, orig_pte)))
2116                 goto out_nomap;
2117
2118         if (unlikely(!PageUptodate(page))) {
2119                 ret = VM_FAULT_SIGBUS;
2120                 goto out_nomap;
2121         }
2122
2123         /* The page isn't present yet, go ahead with the fault. */
2124
2125         inc_mm_counter(mm, anon_rss);
2126         pte = mk_pte(page, vma->vm_page_prot);
2127         if (write_access && can_share_swap_page(page)) {
2128                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2129                 write_access = 0;
2130         }
2131
2132         flush_icache_page(vma, page);
2133         set_pte_at(mm, address, page_table, pte);
2134         page_add_anon_rmap(page, vma, address);
2135
2136         swap_free(entry);
2137         if (vm_swap_full())
2138                 remove_exclusive_swap_page(page);
2139         unlock_page(page);
2140
2141         if (write_access) {
2142                 if (do_wp_page(mm, vma, address,
2143                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
2144                         ret = VM_FAULT_OOM;
2145                 goto out;
2146         }
2147
2148         /* No need to invalidate - it was non-present before */
2149         update_mmu_cache(vma, address, pte);
2150         lazy_mmu_prot_update(pte);
2151 unlock:
2152         pte_unmap_unlock(page_table, ptl);
2153 out:
2154         return ret;
2155 out_nomap:
2156         pte_unmap_unlock(page_table, ptl);
2157         unlock_page(page);
2158         page_cache_release(page);
2159         return ret;
2160 }
2161
2162 /*
2163  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2164  * but allow concurrent faults), and pte mapped but not yet locked.
2165  * We return with mmap_sem still held, but pte unmapped and unlocked.
2166  */
2167 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2168                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2169                 int write_access)
2170 {
2171         struct page *page;
2172         spinlock_t *ptl;
2173         pte_t entry;
2174
2175         if (write_access) {
2176                 /* Allocate our own private page. */
2177                 pte_unmap(page_table);
2178
2179                 if (!vx_rsspages_avail(mm, 1))
2180                         goto oom;
2181                 if (unlikely(anon_vma_prepare(vma)))
2182                         goto oom;
2183                 page = alloc_zeroed_user_highpage(vma, address);
2184                 if (!page)
2185                         goto oom;
2186
2187                 entry = mk_pte(page, vma->vm_page_prot);
2188                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2189
2190                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2191                 if (!pte_none(*page_table))
2192                         goto release;
2193                 inc_mm_counter(mm, anon_rss);
2194                 lru_cache_add_active(page);
2195                 page_add_new_anon_rmap(page, vma, address);
2196         } else {
2197                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2198                 page = ZERO_PAGE(address);
2199                 page_cache_get(page);
2200                 entry = mk_pte(page, vma->vm_page_prot);
2201
2202                 ptl = pte_lockptr(mm, pmd);
2203                 spin_lock(ptl);
2204                 if (!pte_none(*page_table))
2205                         goto release;
2206                 inc_mm_counter(mm, file_rss);
2207                 page_add_file_rmap(page);
2208         }
2209
2210         set_pte_at(mm, address, page_table, entry);
2211
2212         /* No need to invalidate - it was non-present before */
2213         update_mmu_cache(vma, address, entry);
2214         lazy_mmu_prot_update(entry);
2215 unlock:
2216         pte_unmap_unlock(page_table, ptl);
2217         return VM_FAULT_MINOR;
2218 release:
2219         page_cache_release(page);
2220         goto unlock;
2221 oom:
2222         return VM_FAULT_OOM;
2223 }
2224
2225 /*
2226  * do_no_page() tries to create a new page mapping. It aggressively
2227  * tries to share with existing pages, but makes a separate copy if
2228  * the "write_access" parameter is true in order to avoid the next
2229  * page fault.
2230  *
2231  * As this is called only for pages that do not currently exist, we
2232  * do not need to flush old virtual caches or the TLB.
2233  *
2234  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2235  * but allow concurrent faults), and pte mapped but not yet locked.
2236  * We return with mmap_sem still held, but pte unmapped and unlocked.
2237  */
2238 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2239                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2240                 int write_access)
2241 {
2242         spinlock_t *ptl;
2243         struct page *new_page;
2244         struct address_space *mapping = NULL;
2245         pte_t entry;
2246         unsigned int sequence = 0;
2247         int ret = VM_FAULT_MINOR;
2248         int anon = 0;
2249         struct page *dirty_page = NULL;
2250
2251         pte_unmap(page_table);
2252         BUG_ON(vma->vm_flags & VM_PFNMAP);
2253
2254         if (vma->vm_file) {
2255                 mapping = vma->vm_file->f_mapping;
2256                 sequence = mapping->truncate_count;
2257                 smp_rmb(); /* serializes i_size against truncate_count */
2258         }
2259 retry:
2260         /* FIXME: is that check useful here? */
2261         if (!vx_rsspages_avail(mm, 1))
2262                 return VM_FAULT_OOM;
2263         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2264         /*
2265          * No smp_rmb is needed here as long as there's a full
2266          * spin_lock/unlock sequence inside the ->nopage callback
2267          * (for the pagecache lookup) that acts as an implicit
2268          * smp_mb() and prevents the i_size read to happen
2269          * after the next truncate_count read.
2270          */
2271
2272         /* no page was available -- either SIGBUS or OOM */
2273         if (new_page == NOPAGE_SIGBUS)
2274                 return VM_FAULT_SIGBUS;
2275         if (new_page == NOPAGE_OOM)
2276                 return VM_FAULT_OOM;
2277
2278         /*
2279          * Should we do an early C-O-W break?
2280          */
2281         if (write_access) {
2282                 if (!(vma->vm_flags & VM_SHARED)) {
2283                         struct page *page;
2284
2285                         if (unlikely(anon_vma_prepare(vma)))
2286                                 goto oom;
2287                         page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2288                         if (!page)
2289                                 goto oom;
2290                         copy_user_highpage(page, new_page, address);
2291                         page_cache_release(new_page);
2292                         new_page = page;
2293                         anon = 1;
2294
2295                 } else {
2296                         /* if the page will be shareable, see if the backing
2297                          * address space wants to know that the page is about
2298                          * to become writable */
2299                         if (vma->vm_ops->page_mkwrite &&
2300                             vma->vm_ops->page_mkwrite(vma, new_page) < 0
2301                             ) {
2302                                 page_cache_release(new_page);
2303                                 return VM_FAULT_SIGBUS;
2304                         }
2305                 }
2306         }
2307
2308         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2309         /*
2310          * For a file-backed vma, someone could have truncated or otherwise
2311          * invalidated this page.  If unmap_mapping_range got called,
2312          * retry getting the page.
2313          */
2314         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2315                 pte_unmap_unlock(page_table, ptl);
2316                 page_cache_release(new_page);
2317                 cond_resched();
2318                 sequence = mapping->truncate_count;
2319                 smp_rmb();
2320                 goto retry;
2321         }
2322
2323         /*
2324          * This silly early PAGE_DIRTY setting removes a race
2325          * due to the bad i386 page protection. But it's valid
2326          * for other architectures too.
2327          *
2328          * Note that if write_access is true, we either now have
2329          * an exclusive copy of the page, or this is a shared mapping,
2330          * so we can make it writable and dirty to avoid having to
2331          * handle that later.
2332          */
2333         /* Only go through if we didn't race with anybody else... */
2334         if (pte_none(*page_table)) {
2335                 flush_icache_page(vma, new_page);
2336                 entry = mk_pte(new_page, vma->vm_page_prot);
2337                 if (write_access)
2338                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2339                 set_pte_at(mm, address, page_table, entry);
2340                 if (anon) {
2341                         inc_mm_counter(mm, anon_rss);
2342                         lru_cache_add_active(new_page);
2343                         page_add_new_anon_rmap(new_page, vma, address);
2344                 } else {
2345                         inc_mm_counter(mm, file_rss);
2346                         page_add_file_rmap(new_page);
2347                         if (write_access) {
2348                                 dirty_page = new_page;
2349                                 get_page(dirty_page);
2350                         }
2351                 }
2352         } else {
2353                 /* One of our sibling threads was faster, back out. */
2354                 page_cache_release(new_page);
2355                 goto unlock;
2356         }
2357
2358         /* no need to invalidate: a not-present page shouldn't be cached */
2359         update_mmu_cache(vma, address, entry);
2360         lazy_mmu_prot_update(entry);
2361 unlock:
2362         pte_unmap_unlock(page_table, ptl);
2363         if (dirty_page) {
2364                 set_page_dirty_balance(dirty_page);
2365                 put_page(dirty_page);
2366         }
2367         return ret;
2368 oom:
2369         page_cache_release(new_page);
2370         return VM_FAULT_OOM;
2371 }
2372
2373 /*
2374  * Fault of a previously existing named mapping. Repopulate the pte
2375  * from the encoded file_pte if possible. This enables swappable
2376  * nonlinear vmas.
2377  *
2378  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2379  * but allow concurrent faults), and pte mapped but not yet locked.
2380  * We return with mmap_sem still held, but pte unmapped and unlocked.
2381  */
2382 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2383                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2384                 int write_access, pte_t orig_pte)
2385 {
2386         pgoff_t pgoff;
2387         int err;
2388
2389         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2390                 return VM_FAULT_MINOR;
2391
2392         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2393                 /*
2394                  * Page table corrupted: show pte and kill process.
2395                  */
2396                 print_bad_pte(vma, orig_pte, address);
2397                 return VM_FAULT_OOM;
2398         }
2399         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2400
2401         pgoff = pte_to_pgoff(orig_pte);
2402         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2403                                         vma->vm_page_prot, pgoff, 0);
2404         if (err == -ENOMEM)
2405                 return VM_FAULT_OOM;
2406         if (err)
2407                 return VM_FAULT_SIGBUS;
2408         return VM_FAULT_MAJOR;
2409 }
2410
2411 /*
2412  * These routines also need to handle stuff like marking pages dirty
2413  * and/or accessed for architectures that don't do it in hardware (most
2414  * RISC architectures).  The early dirtying is also good on the i386.
2415  *
2416  * There is also a hook called "update_mmu_cache()" that architectures
2417  * with external mmu caches can use to update those (ie the Sparc or
2418  * PowerPC hashed page tables that act as extended TLBs).
2419  *
2420  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2421  * but allow concurrent faults), and pte mapped but not yet locked.
2422  * We return with mmap_sem still held, but pte unmapped and unlocked.
2423  */
2424 static inline int handle_pte_fault(struct mm_struct *mm,
2425                 struct vm_area_struct *vma, unsigned long address,
2426                 pte_t *pte, pmd_t *pmd, int write_access)
2427 {
2428         pte_t entry;
2429         pte_t old_entry;
2430         spinlock_t *ptl;
2431
2432         old_entry = entry = *pte;
2433         if (!pte_present(entry)) {
2434                 if (pte_none(entry)) {
2435                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2436                                 return do_anonymous_page(mm, vma, address,
2437                                         pte, pmd, write_access);
2438                         return do_no_page(mm, vma, address,
2439                                         pte, pmd, write_access);
2440                 }
2441                 if (pte_file(entry))
2442                         return do_file_page(mm, vma, address,
2443                                         pte, pmd, write_access, entry);
2444                 return do_swap_page(mm, vma, address,
2445                                         pte, pmd, write_access, entry);
2446         }
2447
2448         ptl = pte_lockptr(mm, pmd);
2449         spin_lock(ptl);
2450         if (unlikely(!pte_same(*pte, entry)))
2451                 goto unlock;
2452         if (write_access) {
2453                 if (!pte_write(entry))
2454                         return do_wp_page(mm, vma, address,
2455                                         pte, pmd, ptl, entry);
2456                 entry = pte_mkdirty(entry);
2457         }
2458         entry = pte_mkyoung(entry);
2459         if (!pte_same(old_entry, entry)) {
2460                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2461                 update_mmu_cache(vma, address, entry);
2462                 lazy_mmu_prot_update(entry);
2463         } else {
2464                 /*
2465                  * This is needed only for protection faults but the arch code
2466                  * is not yet telling us if this is a protection fault or not.
2467                  * This still avoids useless tlb flushes for .text page faults
2468                  * with threads.
2469                  */
2470                 if (write_access)
2471                         flush_tlb_page(vma, address);
2472         }
2473 unlock:
2474         pte_unmap_unlock(pte, ptl);
2475         return VM_FAULT_MINOR;
2476 }
2477
2478 /*
2479  * By the time we get here, we already hold the mm semaphore
2480  */
2481 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2482                 unsigned long address, int write_access)
2483 {
2484         pgd_t *pgd;
2485         pud_t *pud;
2486         pmd_t *pmd;
2487         pte_t *pte;
2488
2489         __set_current_state(TASK_RUNNING);
2490
2491         count_vm_event(PGFAULT);
2492
2493         if (unlikely(is_vm_hugetlb_page(vma)))
2494                 return hugetlb_fault(mm, vma, address, write_access);
2495
2496         pgd = pgd_offset(mm, address);
2497         pud = pud_alloc(mm, pgd, address);
2498         if (!pud)
2499                 return VM_FAULT_OOM;
2500         pmd = pmd_alloc(mm, pud, address);
2501         if (!pmd)
2502                 return VM_FAULT_OOM;
2503         pte = pte_alloc_map(mm, pmd, address);
2504         if (!pte)
2505                 return VM_FAULT_OOM;
2506
2507         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2508 }
2509
2510 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2511
2512 #ifndef __PAGETABLE_PUD_FOLDED
2513 /*
2514  * Allocate page upper directory.
2515  * We've already handled the fast-path in-line.
2516  */
2517 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2518 {
2519         pud_t *new = pud_alloc_one(mm, address);
2520         if (!new)
2521                 return -ENOMEM;
2522
2523         spin_lock(&mm->page_table_lock);
2524         if (pgd_present(*pgd))          /* Another has populated it */
2525                 pud_free(new);
2526         else
2527                 pgd_populate(mm, pgd, new);
2528         spin_unlock(&mm->page_table_lock);
2529         return 0;
2530 }
2531 #else
2532 /* Workaround for gcc 2.96 */
2533 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2534 {
2535         return 0;
2536 }
2537 #endif /* __PAGETABLE_PUD_FOLDED */
2538
2539 #ifndef __PAGETABLE_PMD_FOLDED
2540 /*
2541  * Allocate page middle directory.
2542  * We've already handled the fast-path in-line.
2543  */
2544 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2545 {
2546         pmd_t *new = pmd_alloc_one(mm, address);
2547         if (!new)
2548                 return -ENOMEM;
2549
2550         spin_lock(&mm->page_table_lock);
2551 #ifndef __ARCH_HAS_4LEVEL_HACK
2552         if (pud_present(*pud))          /* Another has populated it */
2553                 pmd_free(new);
2554         else
2555                 pud_populate(mm, pud, new);
2556 #else
2557         if (pgd_present(*pud))          /* Another has populated it */
2558                 pmd_free(new);
2559         else
2560                 pgd_populate(mm, pud, new);
2561 #endif /* __ARCH_HAS_4LEVEL_HACK */
2562         spin_unlock(&mm->page_table_lock);
2563         return 0;
2564 }
2565 #else
2566 /* Workaround for gcc 2.96 */
2567 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2568 {
2569         return 0;
2570 }
2571 #endif /* __PAGETABLE_PMD_FOLDED */
2572
2573 int make_pages_present(unsigned long addr, unsigned long end)
2574 {
2575         int ret, len, write;
2576         struct vm_area_struct * vma;
2577
2578         vma = find_vma(current->mm, addr);
2579         if (!vma)
2580                 return -1;
2581         write = (vma->vm_flags & VM_WRITE) != 0;
2582         BUG_ON(addr >= end);
2583         BUG_ON(end > vma->vm_end);
2584         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2585         ret = get_user_pages(current, current->mm, addr,
2586                         len, write, 0, NULL, NULL);
2587         if (ret < 0)
2588                 return ret;
2589         return ret == len ? 0 : -1;
2590 }
2591
2592 /*
2593  * Map a vmalloc()-space virtual address to the physical page.
2594  */
2595 struct page * vmalloc_to_page(void * vmalloc_addr)
2596 {
2597         unsigned long addr = (unsigned long) vmalloc_addr;
2598         struct page *page = NULL;
2599         pgd_t *pgd = pgd_offset_k(addr);
2600         pud_t *pud;
2601         pmd_t *pmd;
2602         pte_t *ptep, pte;
2603
2604         if (!pgd_none(*pgd)) {
2605                 pud = pud_offset(pgd, addr);
2606                 if (!pud_none(*pud)) {
2607                         pmd = pmd_offset(pud, addr);
2608                         if (!pmd_none(*pmd)) {
2609                                 ptep = pte_offset_map(pmd, addr);
2610                                 pte = *ptep;
2611                                 if (pte_present(pte))
2612                                         page = pte_page(pte);
2613                                 pte_unmap(ptep);
2614                         }
2615                 }
2616         }
2617         return page;
2618 }
2619
2620 EXPORT_SYMBOL(vmalloc_to_page);
2621
2622 /*
2623  * Map a vmalloc()-space virtual address to the physical page frame number.
2624  */
2625 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2626 {
2627         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2628 }
2629
2630 EXPORT_SYMBOL(vmalloc_to_pfn);
2631
2632 #if !defined(__HAVE_ARCH_GATE_AREA)
2633
2634 #if defined(AT_SYSINFO_EHDR)
2635 static struct vm_area_struct gate_vma;
2636
2637 static int __init gate_vma_init(void)
2638 {
2639         gate_vma.vm_mm = NULL;
2640         gate_vma.vm_start = FIXADDR_USER_START;
2641         gate_vma.vm_end = FIXADDR_USER_END;
2642         gate_vma.vm_page_prot = PAGE_READONLY;
2643         gate_vma.vm_flags = 0;
2644         return 0;
2645 }
2646 __initcall(gate_vma_init);
2647 #endif
2648
2649 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2650 {
2651 #ifdef AT_SYSINFO_EHDR
2652         return &gate_vma;
2653 #else
2654         return NULL;
2655 #endif
2656 }
2657
2658 int in_gate_area_no_task(unsigned long addr)
2659 {
2660 #ifdef AT_SYSINFO_EHDR
2661         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2662                 return 1;
2663 #endif
2664         return 0;
2665 }
2666
2667 #endif  /* __HAVE_ARCH_GATE_AREA */