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
  82 EXPORT_SYMBOL(num_physpages);
  83 EXPORT_SYMBOL(high_memory);
  84
  85 int randomize_va_space __read_mostly = 1;
  86
  87 static int __init disable_randmaps(char *s)
  88 {
  89         randomize_va_space = 0;
  90         return 1;
  91 }
  92 __setup("norandmaps", disable_randmaps);
  93
  94
  95 /*
  96  * If a p?d_bad entry is found while walking page tables, report
  97  * the error, before resetting entry to p?d_none.  Usually (but
  98  * very seldom) called out from the p?d_none_or_clear_bad macros.
  99  */
 100
 101 void pgd_clear_bad(pgd_t *pgd)
 102 {
 103         pgd_ERROR(*pgd);
 104         pgd_clear(pgd);
 105 }
 106
 107 void pud_clear_bad(pud_t *pud)
 108 {
 109         pud_ERROR(*pud);
 110         pud_clear(pud);
 111 }
 112
 113 void pmd_clear_bad(pmd_t *pmd)
 114 {
 115         pmd_ERROR(*pmd);
 116         pmd_clear(pmd);
 117 }
 118
 119 /*
 120  * Note: this doesn't free the actual pages themselves. That
 121  * has been handled earlier when unmapping all the memory regions.
 122  */
 123 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
 124 {
 125         struct page *page = pmd_page(*pmd);
 126         pmd_clear(pmd);
 127         pte_lock_deinit(page);
 128         pte_free_tlb(tlb, page);
 129         dec_zone_page_state(page, NR_PAGETABLE);
 130         tlb->mm->nr_ptes--;
 131 }
 132
 133 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 134                                 unsigned long addr, unsigned long end,
 135                                 unsigned long floor, unsigned long ceiling)
 136 {
 137         pmd_t *pmd;
 138         unsigned long next;
 139         unsigned long start;
 140
 141         start = addr;
 142         pmd = pmd_offset(pud, addr);
 143         do {
 144                 next = pmd_addr_end(addr, end);
 145                 if (pmd_none_or_clear_bad(pmd))
 146                         continue;
 147                 free_pte_range(tlb, pmd);
 148         } while (pmd++, addr = next, addr != end);
 149
 150         start &= PUD_MASK;
 151         if (start < floor)
 152                 return;
 153         if (ceiling) {
 154                 ceiling &= PUD_MASK;
 155                 if (!ceiling)
 156                         return;
 157         }
 158         if (end - 1 > ceiling - 1)
 159                 return;
 160
 161         pmd = pmd_offset(pud, start);
 162         pud_clear(pud);
 163         pmd_free_tlb(tlb, pmd);
 164 }
 165
 166 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 167                                 unsigned long addr, unsigned long end,
 168                                 unsigned long floor, unsigned long ceiling)
 169 {
 170         pud_t *pud;
 171         unsigned long next;
 172         unsigned long start;
 173
 174         start = addr;
 175         pud = pud_offset(pgd, addr);
 176         do {
 177                 next = pud_addr_end(addr, end);
 178                 if (pud_none_or_clear_bad(pud))
 179                         continue;
 180                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 181         } while (pud++, addr = next, addr != end);
 182
 183         start &= PGDIR_MASK;
 184         if (start < floor)
 185                 return;
 186         if (ceiling) {
 187                 ceiling &= PGDIR_MASK;
 188                 if (!ceiling)
 189                         return;
 190         }
 191         if (end - 1 > ceiling - 1)
 192                 return;
 193
 194         pud = pud_offset(pgd, start);
 195         pgd_clear(pgd);
 196         pud_free_tlb(tlb, pud);
 197 }
 198
 199 /*
 200  * This function frees user-level page tables of a process.
 201  *
 202  * Must be called with pagetable lock held.
 203  */
 204 void free_pgd_range(struct mmu_gather **tlb,
 205                         unsigned long addr, unsigned long end,
 206                         unsigned long floor, unsigned long ceiling)
 207 {
 208         pgd_t *pgd;
 209         unsigned long next;
 210         unsigned long start;
 211
 212         /*
 213          * The next few lines have given us lots of grief...
 214          *
 215          * Why are we testing PMD* at this top level?  Because often
 216          * there will be no work to do at all, and we'd prefer not to
 217          * go all the way down to the bottom just to discover that.
 218          *
 219          * Why all these "- 1"s?  Because 0 represents both the bottom
 220          * of the address space and the top of it (using -1 for the
 221          * top wouldn't help much: the masks would do the wrong thing).
 222          * The rule is that addr 0 and floor 0 refer to the bottom of
 223          * the address space, but end 0 and ceiling 0 refer to the top
 224          * Comparisons need to use "end - 1" and "ceiling - 1" (though
 225          * that end 0 case should be mythical).
 226          *
 227          * Wherever addr is brought up or ceiling brought down, we must
 228          * be careful to reject "the opposite 0" before it confuses the
 229          * subsequent tests.  But what about where end is brought down
 230          * by PMD_SIZE below? no, end can't go down to 0 there.
 231          *
 232          * Whereas we round start (addr) and ceiling down, by different
 233          * masks at different levels, in order to test whether a table
 234          * now has no other vmas using it, so can be freed, we don't
 235          * bother to round floor or end up - the tests don't need that.
 236          */
 237
 238         addr &= PMD_MASK;
 239         if (addr < floor) {
 240                 addr += PMD_SIZE;
 241                 if (!addr)
 242                         return;
 243         }
 244         if (ceiling) {
 245                 ceiling &= PMD_MASK;
 246                 if (!ceiling)
 247                         return;
 248         }
 249         if (end - 1 > ceiling - 1)
 250                 end -= PMD_SIZE;
 251         if (addr > end - 1)
 252                 return;
 253
 254         start = addr;
 255         pgd = pgd_offset((*tlb)->mm, addr);
 256         do {
 257                 next = pgd_addr_end(addr, end);
 258                 if (pgd_none_or_clear_bad(pgd))
 259                         continue;
 260                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
 261         } while (pgd++, addr = next, addr != end);
 262 }
 263
 264 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
 265                 unsigned long floor, unsigned long ceiling)
 266 {
 267         while (vma) {
 268                 struct vm_area_struct *next = vma->vm_next;
 269                 unsigned long addr = vma->vm_start;
 270
 271                 /*
 272                  * Hide vma from rmap and vmtruncate before freeing pgtables
 273                  */
 274                 anon_vma_unlink(vma);
 275                 unlink_file_vma(vma);
 276
 277                 if (is_vm_hugetlb_page(vma)) {
 278                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 279                                 floor, next? next->vm_start: ceiling);
 280                 } else {
 281                         /*
 282                          * Optimization: gather nearby vmas into one call down
 283                          */
 284                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 285                                && !is_vm_hugetlb_page(next)) {
 286                                 vma = next;
 287                                 next = vma->vm_next;
 288                                 anon_vma_unlink(vma);
 289                                 unlink_file_vma(vma);
 290                         }
 291                         free_pgd_range(tlb, addr, vma->vm_end,
 292                                 floor, next? next->vm_start: ceiling);
 293                 }
 294                 vma = next;
 295         }
 296 }
 297
 298 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 299 {
 300         struct page *new = pte_alloc_one(mm, address);
 301         if (!new)
 302                 return -ENOMEM;
 303
 304         pte_lock_init(new);
 305         spin_lock(&mm->page_table_lock);
 306         if (pmd_present(*pmd)) {        /* Another has populated it */
 307                 pte_lock_deinit(new);
 308                 pte_free(new);
 309         } else {
 310                 mm->nr_ptes++;
 311                 inc_zone_page_state(new, NR_PAGETABLE);
 312                 pmd_populate(mm, pmd, new);
 313         }
 314         spin_unlock(&mm->page_table_lock);
 315         return 0;
 316 }
 317
 318 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 319 {
 320         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 321         if (!new)
 322                 return -ENOMEM;
 323
 324         spin_lock(&init_mm.page_table_lock);
 325         if (pmd_present(*pmd))          /* Another has populated it */
 326                 pte_free_kernel(new);
 327         else
 328                 pmd_populate_kernel(&init_mm, pmd, new);
 329         spin_unlock(&init_mm.page_table_lock);
 330         return 0;
 331 }
 332
 333 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
 334 {
 335         if (file_rss)
 336                 add_mm_counter(mm, file_rss, file_rss);
 337         if (anon_rss)
 338                 add_mm_counter(mm, anon_rss, anon_rss);
 339 }
 340
 341 /*
 342  * This function is called to print an error when a bad pte
 343  * is found. For example, we might have a PFN-mapped pte in
 344  * a region that doesn't allow it.
 345  *
 346  * The calling function must still handle the error.
 347  */
 348 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
 349 {
 350         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
 351                         "vm_flags = %lx, vaddr = %lx\n",
 352                 (long long)pte_val(pte),
 353                 (vma->vm_mm == current->mm ? current->comm : "???"),
 354                 vma->vm_flags, vaddr);
 355         dump_stack();
 356 }
 357
 358 static inline int is_cow_mapping(unsigned int flags)
 359 {
 360         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 361 }
 362
 363 /*
 364  * This function gets the "struct page" associated with a pte.
 365  *
 366  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
 367  * will have each page table entry just pointing to a raw page frame
 368  * number, and as far as the VM layer is concerned, those do not have
 369  * pages associated with them - even if the PFN might point to memory
 370  * that otherwise is perfectly fine and has a "struct page".
 371  *
 372  * The way we recognize those mappings is through the rules set up
 373  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
 374  * and the vm_pgoff will point to the first PFN mapped: thus every
 375  * page that is a raw mapping will always honor the rule
 376  *
 377  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 378  *
 379  * and if that isn't true, the page has been COW'ed (in which case it
 380  * _does_ have a "struct page" associated with it even if it is in a
 381  * VM_PFNMAP range).
 382  */
 383 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
 384 {
 385         unsigned long pfn = pte_pfn(pte);
 386
 387         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
 388                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
 389                 if (pfn == vma->vm_pgoff + off)
 390                         return NULL;
 391                 if (!is_cow_mapping(vma->vm_flags))
 392                         return NULL;
 393         }
 394
 395 #ifdef CONFIG_DEBUG_VM
 396         /*
 397          * Add some anal sanity checks for now. Eventually,
 398          * we should just do "return pfn_to_page(pfn)", but
 399          * in the meantime we check that we get a valid pfn,
 400          * and that the resulting page looks ok.
 401          */
 402         if (unlikely(!pfn_valid(pfn))) {
 403                 print_bad_pte(vma, pte, addr);
 404                 return NULL;
 405         }
 406 #endif
 407
 408         /*
 409          * NOTE! We still have PageReserved() pages in the page
 410          * tables.
 411          *
 412          * The PAGE_ZERO() pages and various VDSO mappings can
 413          * cause them to exist.
 414          */
 415         return pfn_to_page(pfn);
 416 }
 417
 418 /*
 419  * copy one vm_area from one task to the other. Assumes the page tables
 420  * already present in the new task to be cleared in the whole range
 421  * covered by this vma.
 422  */
 423
 424 static inline void
 425 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 426                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 427                 unsigned long addr, int *rss)
 428 {
 429         unsigned long vm_flags = vma->vm_flags;
 430         pte_t pte = *src_pte;
 431         struct page *page;
 432
 433         /* pte contains position in swap or file, so copy. */
 434         if (unlikely(!pte_present(pte))) {
 435                 if (!pte_file(pte)) {
 436                         swp_entry_t entry = pte_to_swp_entry(pte);
 437
 438                         swap_duplicate(entry);
 439                         /* make sure dst_mm is on swapoff's mmlist. */
 440                         if (unlikely(list_empty(&dst_mm->mmlist))) {
 441                                 spin_lock(&mmlist_lock);
 442                                 if (list_empty(&dst_mm->mmlist))
 443                                         list_add(&dst_mm->mmlist,
 444                                                  &src_mm->mmlist);
 445                                 spin_unlock(&mmlist_lock);
 446                         }
 447                         if (is_write_migration_entry(entry) &&
 448                                         is_cow_mapping(vm_flags)) {
 449                                 /*
 450                                  * COW mappings require pages in both parent
 451                                  * and child to be set to read.
 452                                  */
 453                                 make_migration_entry_read(&entry);
 454                                 pte = swp_entry_to_pte(entry);
 455                                 set_pte_at(src_mm, addr, src_pte, pte);
 456                         }
 457                 }
 458                 goto out_set_pte;
 459         }
 460
 461         /*
 462          * If it's a COW mapping, write protect it both
 463          * in the parent and the child
 464          */
 465         if (is_cow_mapping(vm_flags)) {
 466                 ptep_set_wrprotect(src_mm, addr, src_pte);
 467                 pte = pte_wrprotect(pte);
 468         }
 469
 470         /*
 471          * If it's a shared mapping, mark it clean in
 472          * the child
 473          */
 474         if (vm_flags & VM_SHARED)
 475                 pte = pte_mkclean(pte);
 476         pte = pte_mkold(pte);
 477
 478         page = vm_normal_page(vma, addr, pte);
 479         if (page) {
 480                 get_page(page);
 481                 page_dup_rmap(page, vma, addr);
 482                 rss[!!PageAnon(page)]++;
 483         }
 484
 485 out_set_pte:
 486         set_pte_at(dst_mm, addr, dst_pte, pte);
 487 }
 488
 489 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 490                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 491                 unsigned long addr, unsigned long end)
 492 {
 493         pte_t *src_pte, *dst_pte;
 494         spinlock_t *src_ptl, *dst_ptl;
 495         int progress = 0;
 496         int rss[2];
 497
 498 again:
 499         rss[1] = rss[0] = 0;
 500         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 501         if (!dst_pte)
 502                 return -ENOMEM;
 503         src_pte = pte_offset_map_nested(src_pmd, addr);
 504         src_ptl = pte_lockptr(src_mm, src_pmd);
 505         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 506         arch_enter_lazy_mmu_mode();
 507
 508         do {
 509                 /*
 510                  * We are holding two locks at this point - either of them
 511                  * could generate latencies in another task on another CPU.
 512                  */
 513                 if (progress >= 32) {
 514                         progress = 0;
 515                         if (need_resched() ||
 516                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 517                                 break;
 518                 }
 519                 if (pte_none(*src_pte)) {
 520                         progress++;
 521                         continue;
 522                 }
 523                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
 524                 progress += 8;
 525         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 526
 527         arch_leave_lazy_mmu_mode();
 528         spin_unlock(src_ptl);
 529         pte_unmap_nested(src_pte - 1);
 530         add_mm_rss(dst_mm, rss[0], rss[1]);
 531         pte_unmap_unlock(dst_pte - 1, dst_ptl);
 532         cond_resched();
 533         if (addr != end)
 534                 goto again;
 535         return 0;
 536 }
 537
 538 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 539                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 540                 unsigned long addr, unsigned long end)
 541 {
 542         pmd_t *src_pmd, *dst_pmd;
 543         unsigned long next;
 544
 545         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 546         if (!dst_pmd)
 547                 return -ENOMEM;
 548         src_pmd = pmd_offset(src_pud, addr);
 549         do {
 550                 next = pmd_addr_end(addr, end);
 551                 if (pmd_none_or_clear_bad(src_pmd))
 552                         continue;
 553                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
 554                                                 vma, addr, next))
 555                         return -ENOMEM;
 556         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
 557         return 0;
 558 }
 559
 560 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 561                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
 562                 unsigned long addr, unsigned long end)
 563 {
 564         pud_t *src_pud, *dst_pud;
 565         unsigned long next;
 566
 567         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
 568         if (!dst_pud)
 569                 return -ENOMEM;
 570         src_pud = pud_offset(src_pgd, addr);
 571         do {
 572                 next = pud_addr_end(addr, end);
 573                 if (pud_none_or_clear_bad(src_pud))
 574                         continue;
 575                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
 576                                                 vma, addr, next))
 577                         return -ENOMEM;
 578         } while (dst_pud++, src_pud++, addr = next, addr != end);
 579         return 0;
 580 }
 581
 582 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 583                 struct vm_area_struct *vma)
 584 {
 585         pgd_t *src_pgd, *dst_pgd;
 586         unsigned long next;
 587         unsigned long addr = vma->vm_start;
 588         unsigned long end = vma->vm_end;
 589
 590         /*
 591          * Don't copy ptes where a page fault will fill them correctly.
 592          * Fork becomes much lighter when there are big shared or private
 593          * readonly mappings. The tradeoff is that copy_page_range is more
 594          * efficient than faulting.
 595          */
 596         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
 597                 if (!vma->anon_vma)
 598                         return 0;
 599         }
 600
 601         if (is_vm_hugetlb_page(vma))
 602                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
 603
 604         dst_pgd = pgd_offset(dst_mm, addr);
 605         src_pgd = pgd_offset(src_mm, addr);
 606         do {
 607                 next = pgd_addr_end(addr, end);
 608                 if (pgd_none_or_clear_bad(src_pgd))
 609                         continue;
 610                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
 611                                                 vma, addr, next))
 612                         return -ENOMEM;
 613         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
 614         return 0;
 615 }
 616
 617 static unsigned long zap_pte_range(struct mmu_gather *tlb,
 618                                 struct vm_area_struct *vma, pmd_t *pmd,
 619                                 unsigned long addr, unsigned long end,
 620                                 long *zap_work, struct zap_details *details)
 621 {
 622         struct mm_struct *mm = tlb->mm;
 623         pte_t *pte;
 624         spinlock_t *ptl;
 625         int file_rss = 0;
 626         int anon_rss = 0;
 627
 628         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 629         arch_enter_lazy_mmu_mode();
 630         do {
 631                 pte_t ptent = *pte;
 632                 if (pte_none(ptent)) {
 633                         (*zap_work)--;
 634                         continue;
 635                 }
 636
 637                 (*zap_work) -= PAGE_SIZE;
 638
 639                 if (pte_present(ptent)) {
 640                         struct page *page;
 641
 642                         page = vm_normal_page(vma, addr, ptent);
 643                         if (unlikely(details) && page) {
 644                                 /*
 645                                  * unmap_shared_mapping_pages() wants to
 646                                  * invalidate cache without truncating:
 647                                  * unmap shared but keep private pages.
 648                                  */
 649                                 if (details->check_mapping &&
 650                                     details->check_mapping != page->mapping)
 651                                         continue;
 652                                 /*
 653                                  * Each page->index must be checked when
 654                                  * invalidating or truncating nonlinear.
 655                                  */
 656                                 if (details->nonlinear_vma &&
 657                                     (page->index < details->first_index ||
 658                                      page->index > details->last_index))
 659                                         continue;
 660                         }
 661                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 662                                                         tlb->fullmm);
 663                         tlb_remove_tlb_entry(tlb, pte, addr);
 664                         if (unlikely(!page))
 665                                 continue;
 666                         if (unlikely(details) && details->nonlinear_vma
 667                             && linear_page_index(details->nonlinear_vma,
 668                                                 addr) != page->index)
 669                                 set_pte_at(mm, addr, pte,
 670                                            pgoff_to_pte(page->index));
 671                         if (PageAnon(page))
 672                                 anon_rss--;
 673                         else {
 674                                 if (pte_dirty(ptent))
 675                                         set_page_dirty(page);
 676                                 if (pte_young(ptent))
 677                                         SetPageReferenced(page);
 678                                 file_rss--;
 679                         }
 680                         page_remove_rmap(page, vma);
 681                         tlb_remove_page(tlb, page);
 682                         continue;
 683                 }
 684                 /*
 685                  * If details->check_mapping, we leave swap entries;
 686                  * if details->nonlinear_vma, we leave file entries.
 687                  */
 688                 if (unlikely(details))
 689                         continue;
 690                 if (!pte_file(ptent))
 691                         free_swap_and_cache(pte_to_swp_entry(ptent));
 692                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 693         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
 694
 695         add_mm_rss(mm, file_rss, anon_rss);
 696         arch_leave_lazy_mmu_mode();
 697         pte_unmap_unlock(pte - 1, ptl);
 698
 699         return addr;
 700 }
 701
 702 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
 703                                 struct vm_area_struct *vma, pud_t *pud,
 704                                 unsigned long addr, unsigned long end,
 705                                 long *zap_work, struct zap_details *details)
 706 {
 707         pmd_t *pmd;
 708         unsigned long next;
 709
 710         pmd = pmd_offset(pud, addr);
 711         do {
 712                 next = pmd_addr_end(addr, end);
 713                 if (pmd_none_or_clear_bad(pmd)) {
 714                         (*zap_work)--;
 715                         continue;
 716                 }
 717                 next = zap_pte_range(tlb, vma, pmd, addr, next,
 718                                                 zap_work, details);
 719         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
 720
 721         return addr;
 722 }
 723
 724 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
 725                                 struct vm_area_struct *vma, pgd_t *pgd,
 726                                 unsigned long addr, unsigned long end,
 727                                 long *zap_work, struct zap_details *details)
 728 {
 729         pud_t *pud;
 730         unsigned long next;
 731
 732         pud = pud_offset(pgd, addr);
 733         do {
 734                 next = pud_addr_end(addr, end);
 735                 if (pud_none_or_clear_bad(pud)) {
 736                         (*zap_work)--;
 737                         continue;
 738                 }
 739                 next = zap_pmd_range(tlb, vma, pud, addr, next,
 740                                                 zap_work, details);
 741         } while (pud++, addr = next, (addr != end && *zap_work > 0));
 742
 743         return addr;
 744 }
 745
 746 static unsigned long unmap_page_range(struct mmu_gather *tlb,
 747                                 struct vm_area_struct *vma,
 748                                 unsigned long addr, unsigned long end,
 749                                 long *zap_work, struct zap_details *details)
 750 {
 751         pgd_t *pgd;
 752         unsigned long next;
 753
 754         if (details && !details->check_mapping && !details->nonlinear_vma)
 755                 details = NULL;
 756
 757         BUG_ON(addr >= end);
 758         tlb_start_vma(tlb, vma);
 759         pgd = pgd_offset(vma->vm_mm, addr);
 760         do {
 761                 next = pgd_addr_end(addr, end);
 762                 if (pgd_none_or_clear_bad(pgd)) {
 763                         (*zap_work)--;
 764                         continue;
 765                 }
 766                 next = zap_pud_range(tlb, vma, pgd, addr, next,
 767                                                 zap_work, details);
 768         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
 769         tlb_end_vma(tlb, vma);
 770
 771         return addr;
 772 }
 773
 774 #ifdef CONFIG_PREEMPT
 775 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
 776 #else
 777 /* No preempt: go for improved straight-line efficiency */
 778 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
 779 #endif
 780
 781 /**
 782  * unmap_vmas - unmap a range of memory covered by a list of vma's
 783  * @tlbp: address of the caller's struct mmu_gather
 784  * @vma: the starting vma
 785  * @start_addr: virtual address at which to start unmapping
 786  * @end_addr: virtual address at which to end unmapping
 787  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 788  * @details: details of nonlinear truncation or shared cache invalidation
 789  *
 790  * Returns the end address of the unmapping (restart addr if interrupted).
 791  *
 792  * Unmap all pages in the vma list.
 793  *
 794  * We aim to not hold locks for too long (for scheduling latency reasons).
 795  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 796  * return the ending mmu_gather to the caller.
 797  *
 798  * Only addresses between `start' and `end' will be unmapped.
 799  *
 800  * The VMA list must be sorted in ascending virtual address order.
 801  *
 802  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 803  * range after unmap_vmas() returns.  So the only responsibility here is to
 804  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 805  * drops the lock and schedules.
 806  */
 807 unsigned long unmap_vmas(struct mmu_gather **tlbp,
 808                 struct vm_area_struct *vma, unsigned long start_addr,
 809                 unsigned long end_addr, unsigned long *nr_accounted,
 810                 struct zap_details *details)
 811 {
 812         long zap_work = ZAP_BLOCK_SIZE;
 813         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 814         int tlb_start_valid = 0;
 815         unsigned long start = start_addr;
 816         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
 817         int fullmm = (*tlbp)->fullmm;
 818
 819         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 820                 unsigned long end;
 821
 822                 start = max(vma->vm_start, start_addr);
 823                 if (start >= vma->vm_end)
 824                         continue;
 825                 end = min(vma->vm_end, end_addr);
 826                 if (end <= vma->vm_start)
 827                         continue;
 828
 829                 if (vma->vm_flags & VM_ACCOUNT)
 830                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 831
 832                 while (start != end) {
 833                         if (!tlb_start_valid) {
 834                                 tlb_start = start;
 835                                 tlb_start_valid = 1;
 836                         }
 837
 838                         if (unlikely(is_vm_hugetlb_page(vma))) {
 839                                 unmap_hugepage_range(vma, start, end);
 840                                 zap_work -= (end - start) /
 841                                                 (HPAGE_SIZE / PAGE_SIZE);
 842                                 start = end;
 843                         } else
 844                                 start = unmap_page_range(*tlbp, vma,
 845                                                 start, end, &zap_work, details);
 846
 847                         if (zap_work > 0) {
 848                                 BUG_ON(start != end);
 849                                 break;
 850                         }
 851
 852                         tlb_finish_mmu(*tlbp, tlb_start, start);
 853
 854                         if (need_resched() ||
 855                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
 856                                 if (i_mmap_lock) {
 857                                         *tlbp = NULL;
 858                                         goto out;
 859                                 }
 860                                 cond_resched();
 861                         }
 862
 863                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
 864                         tlb_start_valid = 0;
 865                         zap_work = ZAP_BLOCK_SIZE;
 866                 }
 867         }
 868 out:
 869         return start;   /* which is now the end (or restart) address */
 870 }
 871
 872 /**
 873  * zap_page_range - remove user pages in a given range
 874  * @vma: vm_area_struct holding the applicable pages
 875  * @address: starting address of pages to zap
 876  * @size: number of bytes to zap
 877  * @details: details of nonlinear truncation or shared cache invalidation
 878  */
 879 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
 880                 unsigned long size, struct zap_details *details)
 881 {
 882         struct mm_struct *mm = vma->vm_mm;
 883         struct mmu_gather *tlb;
 884         unsigned long end = address + size;
 885         unsigned long nr_accounted = 0;
 886
 887         lru_add_drain();
 888         tlb = tlb_gather_mmu(mm, 0);
 889         update_hiwater_rss(mm);
 890         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
 891         if (tlb)
 892                 tlb_finish_mmu(tlb, address, end);
 893         return end;
 894 }
 895
 896 /*
 897  * Do a quick page-table lookup for a single page.
 898  */
 899 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
 900                         unsigned int flags)
 901 {
 902         pgd_t *pgd;
 903         pud_t *pud;
 904         pmd_t *pmd;
 905         pte_t *ptep, pte;
 906         spinlock_t *ptl;
 907         struct page *page;
 908         struct mm_struct *mm = vma->vm_mm;
 909
 910         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 911         if (!IS_ERR(page)) {
 912                 BUG_ON(flags & FOLL_GET);
 913                 goto out;
 914         }
 915
 916         page = NULL;
 917         pgd = pgd_offset(mm, address);
 918         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 919                 goto no_page_table;
 920
 921         pud = pud_offset(pgd, address);
 922         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
 923                 goto no_page_table;
 924
 925         pmd = pmd_offset(pud, address);
 926         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 927                 goto no_page_table;
 928
 929         if (pmd_huge(*pmd)) {
 930                 BUG_ON(flags & FOLL_GET);
 931                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
 932                 goto out;
 933         }
 934
 935         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 936         if (!ptep)
 937                 goto out;
 938
 939         pte = *ptep;
 940         if (!pte_present(pte))
 941                 goto unlock;
 942         if ((flags & FOLL_WRITE) && !pte_write(pte))
 943                 goto unlock;
 944         page = vm_normal_page(vma, address, pte);
 945         if (unlikely(!page))
 946                 goto unlock;
 947
 948         if (flags & FOLL_GET)
 949                 get_page(page);
 950         if (flags & FOLL_TOUCH) {
 951                 if ((flags & FOLL_WRITE) &&
 952                     !pte_dirty(pte) && !PageDirty(page))
 953                         set_page_dirty(page);
 954                 mark_page_accessed(page);
 955         }
 956 unlock:
 957         pte_unmap_unlock(ptep, ptl);
 958 out:
 959         return page;
 960
 961 no_page_table:
 962         /*
 963          * When core dumping an enormous anonymous area that nobody
 964          * has touched so far, we don't want to allocate page tables.
 965          */
 966         if (flags & FOLL_ANON) {
 967                 page = ZERO_PAGE(0);
 968                 if (flags & FOLL_GET)
 969                         get_page(page);
 970                 BUG_ON(flags & FOLL_WRITE);
 971         }
 972         return page;
 973 }
 974
 975 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 976                 unsigned long start, int len, int write, int force,
 977                 struct page **pages, struct vm_area_struct **vmas)
 978 {
 979         int i;
 980         unsigned int vm_flags;
 981
 982         /*
 983          * Require read or write permissions.
 984          * If 'force' is set, we only require the "MAY" flags.
 985          */
 986         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 987         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 988         i = 0;
 989
 990         do {
 991                 struct vm_area_struct *vma;
 992                 unsigned int foll_flags;
 993
 994                 vma = find_extend_vma(mm, start);
 995                 if (!vma && in_gate_area(tsk, start)) {
 996                         unsigned long pg = start & PAGE_MASK;
 997                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
 998                         pgd_t *pgd;
 999                         pud_t *pud;
1000                         pmd_t *pmd;
1001                         pte_t *pte;
1002                         if (write) /* user gate pages are read-only */
1003                                 return i ? : -EFAULT;
1004                         if (pg > TASK_SIZE)
1005                                 pgd = pgd_offset_k(pg);
1006                         else
1007                                 pgd = pgd_offset_gate(mm, pg);
1008                         BUG_ON(pgd_none(*pgd));
1009                         pud = pud_offset(pgd, pg);
1010                         BUG_ON(pud_none(*pud));
1011                         pmd = pmd_offset(pud, pg);
1012                         if (pmd_none(*pmd))
1013                                 return i ? : -EFAULT;
1014                         pte = pte_offset_map(pmd, pg);
1015                         if (pte_none(*pte)) {
1016                                 pte_unmap(pte);
1017                                 return i ? : -EFAULT;
1018                         }
1019                         if (pages) {
1020                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1021                                 pages[i] = page;
1022                                 if (page)
1023                                         get_page(page);
1024                         }
1025                         pte_unmap(pte);
1026                         if (vmas)
1027                                 vmas[i] = gate_vma;
1028                         i++;
1029                         start += PAGE_SIZE;
1030                         len--;
1031                         continue;
1032                 }
1033
1034                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1035                                 || !(vm_flags & vma->vm_flags))
1036                         return i ? : -EFAULT;
1037
1038                 if (is_vm_hugetlb_page(vma)) {
1039                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1040                                                 &start, &len, i, write);
1041                         continue;
1042                 }
1043
1044                 foll_flags = FOLL_TOUCH;
1045                 if (pages)
1046                         foll_flags |= FOLL_GET;
1047                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1048                     (!vma->vm_ops || (!vma->vm_ops->nopage &&
1049                                         !vma->vm_ops->fault)))
1050                         foll_flags |= FOLL_ANON;
1051
1052                 do {
1053                         struct page *page;
1054
1055                         /*
1056                          * If tsk is ooming, cut off its access to large memory
1057                          * allocations. It has a pending SIGKILL, but it can't
1058                          * be processed until returning to user space.
1059                          */
1060                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1061                                 return -ENOMEM;
1062
1063                         if (write)
1064                                 foll_flags |= FOLL_WRITE;
1065
1066                         cond_resched();
1067                         while (!(page = follow_page(vma, start, foll_flags))) {
1068                                 int ret;
1069                                 ret = handle_mm_fault(mm, vma, start,
1070                                                 foll_flags & FOLL_WRITE);
1071                                 if (ret & VM_FAULT_ERROR) {
1072                                         if (ret & VM_FAULT_OOM)
1073                                                 return i ? i : -ENOMEM;
1074                                         else if (ret & VM_FAULT_SIGBUS)
1075                                                 return i ? i : -EFAULT;
1076                                         BUG();
1077                                 }
1078                                 if (ret & VM_FAULT_MAJOR)
1079                                         tsk->maj_flt++;
1080                                 else
1081                                         tsk->min_flt++;
1082
1083                                 /*
1084                                  * The VM_FAULT_WRITE bit tells us that
1085                                  * do_wp_page has broken COW when necessary,
1086                                  * even if maybe_mkwrite decided not to set
1087                                  * pte_write. We can thus safely do subsequent
1088                                  * page lookups as if they were reads.
1089                                  */
1090                                 if (ret & VM_FAULT_WRITE)
1091                                         foll_flags &= ~FOLL_WRITE;
1092
1093                                 cond_resched();
1094                         }
1095                         if (pages) {
1096                                 pages[i] = page;
1097
1098                                 flush_anon_page(vma, page, start);
1099                                 flush_dcache_page(page);
1100                         }
1101                         if (vmas)
1102                                 vmas[i] = vma;
1103                         i++;
1104                         start += PAGE_SIZE;
1105                         len--;
1106                 } while (len && start < vma->vm_end);
1107         } while (len);
1108         return i;
1109 }
1110 EXPORT_SYMBOL(get_user_pages);
1111
1112 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1113 {
1114         pgd_t * pgd = pgd_offset(mm, addr);
1115         pud_t * pud = pud_alloc(mm, pgd, addr);
1116         if (pud) {
1117                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1118                 if (pmd)
1119                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1120         }
1121         return NULL;
1122 }
1123
1124 /*
1125  * This is the old fallback for page remapping.
1126  *
1127  * For historical reasons, it only allows reserved pages. Only
1128  * old drivers should use this, and they needed to mark their
1129  * pages reserved for the old functions anyway.
1130  */
1131 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1132 {
1133         int retval;
1134         pte_t *pte;
1135         spinlock_t *ptl;
1136
1137         retval = -EINVAL;
1138         if (PageAnon(page))
1139                 goto out;
1140         retval = -ENOMEM;
1141         flush_dcache_page(page);
1142         pte = get_locked_pte(mm, addr, &ptl);
1143         if (!pte)
1144                 goto out;
1145         retval = -EBUSY;
1146         if (!pte_none(*pte))
1147                 goto out_unlock;
1148
1149         /* Ok, finally just insert the thing.. */
1150         get_page(page);
1151         inc_mm_counter(mm, file_rss);
1152         page_add_file_rmap(page);
1153         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1154
1155         retval = 0;
1156 out_unlock:
1157         pte_unmap_unlock(pte, ptl);
1158 out:
1159         return retval;
1160 }
1161
1162 /**
1163  * vm_insert_page - insert single page into user vma
1164  * @vma: user vma to map to
1165  * @addr: target user address of this page
1166  * @page: source kernel page
1167  *
1168  * This allows drivers to insert individual pages they've allocated
1169  * into a user vma.
1170  *
1171  * The page has to be a nice clean _individual_ kernel allocation.
1172  * If you allocate a compound page, you need to have marked it as
1173  * such (__GFP_COMP), or manually just split the page up yourself
1174  * (see split_page()).
1175  *
1176  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1177  * took an arbitrary page protection parameter. This doesn't allow
1178  * that. Your vma protection will have to be set up correctly, which
1179  * means that if you want a shared writable mapping, you'd better
1180  * ask for a shared writable mapping!
1181  *
1182  * The page does not need to be reserved.
1183  */
1184 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1185 {
1186         if (addr < vma->vm_start || addr >= vma->vm_end)
1187                 return -EFAULT;
1188         if (!page_count(page))
1189                 return -EINVAL;
1190         vma->vm_flags |= VM_INSERTPAGE;
1191         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1192 }
1193 EXPORT_SYMBOL(vm_insert_page);
1194
1195 /**
1196  * vm_insert_pfn - insert single pfn into user vma
1197  * @vma: user vma to map to
1198  * @addr: target user address of this page
1199  * @pfn: source kernel pfn
1200  *
1201  * Similar to vm_inert_page, this allows drivers to insert individual pages
1202  * they've allocated into a user vma. Same comments apply.
1203  *
1204  * This function should only be called from a vm_ops->fault handler, and
1205  * in that case the handler should return NULL.
1206  */
1207 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1208                 unsigned long pfn)
1209 {
1210         struct mm_struct *mm = vma->vm_mm;
1211         int retval;
1212         pte_t *pte, entry;
1213         spinlock_t *ptl;
1214
1215         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
1216         BUG_ON(is_cow_mapping(vma->vm_flags));
1217
1218         retval = -ENOMEM;
1219         pte = get_locked_pte(mm, addr, &ptl);
1220         if (!pte)
1221                 goto out;
1222         retval = -EBUSY;
1223         if (!pte_none(*pte))
1224                 goto out_unlock;
1225
1226         /* Ok, finally just insert the thing.. */
1227         entry = pfn_pte(pfn, vma->vm_page_prot);
1228         set_pte_at(mm, addr, pte, entry);
1229         update_mmu_cache(vma, addr, entry);
1230
1231         retval = 0;
1232 out_unlock:
1233         pte_unmap_unlock(pte, ptl);
1234
1235 out:
1236         return retval;
1237 }
1238 EXPORT_SYMBOL(vm_insert_pfn);
1239
1240 /*
1241  * maps a range of physical memory into the requested pages. the old
1242  * mappings are removed. any references to nonexistent pages results
1243  * in null mappings (currently treated as "copy-on-access")
1244  */
1245 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1246                         unsigned long addr, unsigned long end,
1247                         unsigned long pfn, pgprot_t prot)
1248 {
1249         pte_t *pte;
1250         spinlock_t *ptl;
1251
1252         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1253         if (!pte)
1254                 return -ENOMEM;
1255         arch_enter_lazy_mmu_mode();
1256         do {
1257                 BUG_ON(!pte_none(*pte));
1258                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1259                 pfn++;
1260         } while (pte++, addr += PAGE_SIZE, addr != end);
1261         arch_leave_lazy_mmu_mode();
1262         pte_unmap_unlock(pte - 1, ptl);
1263         return 0;
1264 }
1265
1266 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1267                         unsigned long addr, unsigned long end,
1268                         unsigned long pfn, pgprot_t prot)
1269 {
1270         pmd_t *pmd;
1271         unsigned long next;
1272
1273         pfn -= addr >> PAGE_SHIFT;
1274         pmd = pmd_alloc(mm, pud, addr);
1275         if (!pmd)
1276                 return -ENOMEM;
1277         do {
1278                 next = pmd_addr_end(addr, end);
1279                 if (remap_pte_range(mm, pmd, addr, next,
1280                                 pfn + (addr >> PAGE_SHIFT), prot))
1281                         return -ENOMEM;
1282         } while (pmd++, addr = next, addr != end);
1283         return 0;
1284 }
1285
1286 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1287                         unsigned long addr, unsigned long end,
1288                         unsigned long pfn, pgprot_t prot)
1289 {
1290         pud_t *pud;
1291         unsigned long next;
1292
1293         pfn -= addr >> PAGE_SHIFT;
1294         pud = pud_alloc(mm, pgd, addr);
1295         if (!pud)
1296                 return -ENOMEM;
1297         do {
1298                 next = pud_addr_end(addr, end);
1299                 if (remap_pmd_range(mm, pud, addr, next,
1300                                 pfn + (addr >> PAGE_SHIFT), prot))
1301                         return -ENOMEM;
1302         } while (pud++, addr = next, addr != end);
1303         return 0;
1304 }
1305
1306 /**
1307  * remap_pfn_range - remap kernel memory to userspace
1308  * @vma: user vma to map to
1309  * @addr: target user address to start at
1310  * @pfn: physical address of kernel memory
1311  * @size: size of map area
1312  * @prot: page protection flags for this mapping
1313  *
1314  *  Note: this is only safe if the mm semaphore is held when called.
1315  */
1316 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1317                     unsigned long pfn, unsigned long size, pgprot_t prot)
1318 {
1319         pgd_t *pgd;
1320         unsigned long next;
1321         unsigned long end = addr + PAGE_ALIGN(size);
1322         struct mm_struct *mm = vma->vm_mm;
1323         int err;
1324
1325         /*
1326          * Physically remapped pages are special. Tell the
1327          * rest of the world about it:
1328          *   VM_IO tells people not to look at these pages
1329          *      (accesses can have side effects).
1330          *   VM_RESERVED is specified all over the place, because
1331          *      in 2.4 it kept swapout's vma scan off this vma; but
1332          *      in 2.6 the LRU scan won't even find its pages, so this
1333          *      flag means no more than count its pages in reserved_vm,
1334          *      and omit it from core dump, even when VM_IO turned off.
1335          *   VM_PFNMAP tells the core MM that the base pages are just
1336          *      raw PFN mappings, and do not have a "struct page" associated
1337          *      with them.
1338          *
1339          * There's a horrible special case to handle copy-on-write
1340          * behaviour that some programs depend on. We mark the "original"
1341          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1342          */
1343         if (is_cow_mapping(vma->vm_flags)) {
1344                 if (addr != vma->vm_start || end != vma->vm_end)
1345                         return -EINVAL;
1346                 vma->vm_pgoff = pfn;
1347         }
1348
1349         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1350
1351         BUG_ON(addr >= end);
1352         pfn -= addr >> PAGE_SHIFT;
1353         pgd = pgd_offset(mm, addr);
1354         flush_cache_range(vma, addr, end);
1355         do {
1356                 next = pgd_addr_end(addr, end);
1357                 err = remap_pud_range(mm, pgd, addr, next,
1358                                 pfn + (addr >> PAGE_SHIFT), prot);
1359                 if (err)
1360                         break;
1361         } while (pgd++, addr = next, addr != end);
1362         return err;
1363 }
1364 EXPORT_SYMBOL(remap_pfn_range);
1365
1366 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1367                                      unsigned long addr, unsigned long end,
1368                                      pte_fn_t fn, void *data)
1369 {
1370         pte_t *pte;
1371         int err;
1372         struct page *pmd_page;
1373         spinlock_t *uninitialized_var(ptl);
1374
1375         pte = (mm == &init_mm) ?
1376                 pte_alloc_kernel(pmd, addr) :
1377                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1378         if (!pte)
1379                 return -ENOMEM;
1380
1381         BUG_ON(pmd_huge(*pmd));
1382
1383         pmd_page = pmd_page(*pmd);
1384
1385         do {
1386                 err = fn(pte, pmd_page, addr, data);
1387                 if (err)
1388                         break;
1389         } while (pte++, addr += PAGE_SIZE, addr != end);
1390
1391         if (mm != &init_mm)
1392                 pte_unmap_unlock(pte-1, ptl);
1393         return err;
1394 }
1395
1396 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1397                                      unsigned long addr, unsigned long end,
1398                                      pte_fn_t fn, void *data)
1399 {
1400         pmd_t *pmd;
1401         unsigned long next;
1402         int err;
1403
1404         pmd = pmd_alloc(mm, pud, addr);
1405         if (!pmd)
1406                 return -ENOMEM;
1407         do {
1408                 next = pmd_addr_end(addr, end);
1409                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1410                 if (err)
1411                         break;
1412         } while (pmd++, addr = next, addr != end);
1413         return err;
1414 }
1415
1416 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1417                                      unsigned long addr, unsigned long end,
1418                                      pte_fn_t fn, void *data)
1419 {
1420         pud_t *pud;
1421         unsigned long next;
1422         int err;
1423
1424         pud = pud_alloc(mm, pgd, addr);
1425         if (!pud)
1426                 return -ENOMEM;
1427         do {
1428                 next = pud_addr_end(addr, end);
1429                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1430                 if (err)
1431                         break;
1432         } while (pud++, addr = next, addr != end);
1433         return err;
1434 }
1435
1436 /*
1437  * Scan a region of virtual memory, filling in page tables as necessary
1438  * and calling a provided function on each leaf page table.
1439  */
1440 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1441                         unsigned long size, pte_fn_t fn, void *data)
1442 {
1443         pgd_t *pgd;
1444         unsigned long next;
1445         unsigned long end = addr + size;
1446         int err;
1447
1448         BUG_ON(addr >= end);
1449         pgd = pgd_offset(mm, addr);
1450         do {
1451                 next = pgd_addr_end(addr, end);
1452                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1453                 if (err)
1454                         break;
1455         } while (pgd++, addr = next, addr != end);
1456         return err;
1457 }
1458 EXPORT_SYMBOL_GPL(apply_to_page_range);
1459
1460 /*
1461  * handle_pte_fault chooses page fault handler according to an entry
1462  * which was read non-atomically.  Before making any commitment, on
1463  * those architectures or configurations (e.g. i386 with PAE) which
1464  * might give a mix of unmatched parts, do_swap_page and do_file_page
1465  * must check under lock before unmapping the pte and proceeding
1466  * (but do_wp_page is only called after already making such a check;
1467  * and do_anonymous_page and do_no_page can safely check later on).
1468  */
1469 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1470                                 pte_t *page_table, pte_t orig_pte)
1471 {
1472         int same = 1;
1473 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1474         if (sizeof(pte_t) > sizeof(unsigned long)) {
1475                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1476                 spin_lock(ptl);
1477                 same = pte_same(*page_table, orig_pte);
1478                 spin_unlock(ptl);
1479         }
1480 #endif
1481         pte_unmap(page_table);
1482         return same;
1483 }
1484
1485 /*
1486  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1487  * servicing faults for write access.  In the normal case, do always want
1488  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1489  * that do not have writing enabled, when used by access_process_vm.
1490  */
1491 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1492 {
1493         if (likely(vma->vm_flags & VM_WRITE))
1494                 pte = pte_mkwrite(pte);
1495         return pte;
1496 }
1497
1498 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1499 {
1500         /*
1501          * If the source page was a PFN mapping, we don't have
1502          * a "struct page" for it. We do a best-effort copy by
1503          * just copying from the original user address. If that
1504          * fails, we just zero-fill it. Live with it.
1505          */
1506         if (unlikely(!src)) {
1507                 void *kaddr = kmap_atomic(dst, KM_USER0);
1508                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1509
1510                 /*
1511                  * This really shouldn't fail, because the page is there
1512                  * in the page tables. But it might just be unreadable,
1513                  * in which case we just give up and fill the result with
1514                  * zeroes.
1515                  */
1516                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1517                         memset(kaddr, 0, PAGE_SIZE);
1518                 kunmap_atomic(kaddr, KM_USER0);
1519                 flush_dcache_page(dst);
1520                 return;
1521
1522         }
1523         copy_user_highpage(dst, src, va, vma);
1524 }
1525
1526 /*
1527  * This routine handles present pages, when users try to write
1528  * to a shared page. It is done by copying the page to a new address
1529  * and decrementing the shared-page counter for the old page.
1530  *
1531  * Note that this routine assumes that the protection checks have been
1532  * done by the caller (the low-level page fault routine in most cases).
1533  * Thus we can safely just mark it writable once we've done any necessary
1534  * COW.
1535  *
1536  * We also mark the page dirty at this point even though the page will
1537  * change only once the write actually happens. This avoids a few races,
1538  * and potentially makes it more efficient.
1539  *
1540  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1541  * but allow concurrent faults), with pte both mapped and locked.
1542  * We return with mmap_sem still held, but pte unmapped and unlocked.
1543  */
1544 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1545                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1546                 spinlock_t *ptl, pte_t orig_pte)
1547 {
1548         struct page *old_page, *new_page;
1549         pte_t entry;
1550         int reuse = 0, ret = 0;
1551         int page_mkwrite = 0;
1552         struct page *dirty_page = NULL;
1553
1554         old_page = vm_normal_page(vma, address, orig_pte);
1555         if (!old_page)
1556                 goto gotten;
1557
1558         /*
1559          * Take out anonymous pages first, anonymous shared vmas are
1560          * not dirty accountable.
1561          */
1562         if (PageAnon(old_page)) {
1563                 if (!TestSetPageLocked(old_page)) {
1564                         reuse = can_share_swap_page(old_page);
1565                         unlock_page(old_page);
1566                 }
1567         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1568                                         (VM_WRITE|VM_SHARED))) {
1569                 /*
1570                  * Only catch write-faults on shared writable pages,
1571                  * read-only shared pages can get COWed by
1572                  * get_user_pages(.write=1, .force=1).
1573                  */
1574                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1575                         /*
1576                          * Notify the address space that the page is about to
1577                          * become writable so that it can prohibit this or wait
1578                          * for the page to get into an appropriate state.
1579                          *
1580                          * We do this without the lock held, so that it can
1581                          * sleep if it needs to.
1582                          */
1583                         page_cache_get(old_page);
1584                         pte_unmap_unlock(page_table, ptl);
1585
1586                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1587                                 goto unwritable_page;
1588
1589                         /*
1590                          * Since we dropped the lock we need to revalidate
1591                          * the PTE as someone else may have changed it.  If
1592                          * they did, we just return, as we can count on the
1593                          * MMU to tell us if they didn't also make it writable.
1594                          */
1595                         page_table = pte_offset_map_lock(mm, pmd, address,
1596                                                          &ptl);
1597                         page_cache_release(old_page);
1598                         if (!pte_same(*page_table, orig_pte))
1599                                 goto unlock;
1600
1601                         page_mkwrite = 1;
1602                 }
1603                 dirty_page = old_page;
1604                 get_page(dirty_page);
1605                 reuse = 1;
1606         }
1607
1608         if (reuse) {
1609                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1610                 entry = pte_mkyoung(orig_pte);
1611                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1612                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1613                         update_mmu_cache(vma, address, entry);
1614                 ret |= VM_FAULT_WRITE;
1615                 goto unlock;
1616         }
1617
1618         /*
1619          * Ok, we need to copy. Oh, well..
1620          */
1621         page_cache_get(old_page);
1622 gotten:
1623         pte_unmap_unlock(page_table, ptl);
1624
1625         if (unlikely(anon_vma_prepare(vma)))
1626                 goto oom;
1627         VM_BUG_ON(old_page == ZERO_PAGE(0));
1628         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1629         if (!new_page)
1630                 goto oom;
1631         cow_user_page(new_page, old_page, address, vma);
1632
1633         /*
1634          * Re-check the pte - we dropped the lock
1635          */
1636         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1637         if (likely(pte_same(*page_table, orig_pte))) {
1638                 if (old_page) {
1639                         page_remove_rmap(old_page, vma);
1640                         if (!PageAnon(old_page)) {
1641                                 dec_mm_counter(mm, file_rss);
1642                                 inc_mm_counter(mm, anon_rss);
1643                         }
1644                 } else
1645                         inc_mm_counter(mm, anon_rss);
1646                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1647                 entry = mk_pte(new_page, vma->vm_page_prot);
1648                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1649                 /*
1650                  * Clear the pte entry and flush it first, before updating the
1651                  * pte with the new entry. This will avoid a race condition
1652                  * seen in the presence of one thread doing SMC and another
1653                  * thread doing COW.
1654                  */
1655                 ptep_clear_flush(vma, address, page_table);
1656                 set_pte_at(mm, address, page_table, entry);
1657                 update_mmu_cache(vma, address, entry);
1658                 lru_cache_add_active(new_page);
1659                 page_add_new_anon_rmap(new_page, vma, address);
1660
1661                 /* Free the old page.. */
1662                 new_page = old_page;
1663                 ret |= VM_FAULT_WRITE;
1664         }
1665         if (new_page)
1666                 page_cache_release(new_page);
1667         if (old_page)
1668                 page_cache_release(old_page);
1669 unlock:
1670         pte_unmap_unlock(page_table, ptl);
1671         if (dirty_page) {
1672                 if (vma->vm_file)
1673                         file_update_time(vma->vm_file);
1674
1675                 /*
1676                  * Yes, Virginia, this is actually required to prevent a race
1677                  * with clear_page_dirty_for_io() from clearing the page dirty
1678                  * bit after it clear all dirty ptes, but before a racing
1679                  * do_wp_page installs a dirty pte.
1680                  *
1681                  * do_no_page is protected similarly.
1682                  */
1683                 wait_on_page_locked(dirty_page);
1684                 set_page_dirty_balance(dirty_page, page_mkwrite);
1685                 put_page(dirty_page);
1686         }
1687         return ret;
1688 oom:
1689         if (old_page)
1690                 page_cache_release(old_page);
1691         return VM_FAULT_OOM;
1692
1693 unwritable_page:
1694         page_cache_release(old_page);
1695         return VM_FAULT_SIGBUS;
1696 }
1697
1698 /*
1699  * Helper functions for unmap_mapping_range().
1700  *
1701  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1702  *
1703  * We have to restart searching the prio_tree whenever we drop the lock,
1704  * since the iterator is only valid while the lock is held, and anyway
1705  * a later vma might be split and reinserted earlier while lock dropped.
1706  *
1707  * The list of nonlinear vmas could be handled more efficiently, using
1708  * a placeholder, but handle it in the same way until a need is shown.
1709  * It is important to search the prio_tree before nonlinear list: a vma
1710  * may become nonlinear and be shifted from prio_tree to nonlinear list
1711  * while the lock is dropped; but never shifted from list to prio_tree.
1712  *
1713  * In order to make forward progress despite restarting the search,
1714  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1715  * quickly skip it next time around.  Since the prio_tree search only
1716  * shows us those vmas affected by unmapping the range in question, we
1717  * can't efficiently keep all vmas in step with mapping->truncate_count:
1718  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1719  * mapping->truncate_count and vma->vm_truncate_count are protected by
1720  * i_mmap_lock.
1721  *
1722  * In order to make forward progress despite repeatedly restarting some
1723  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1724  * and restart from that address when we reach that vma again.  It might
1725  * have been split or merged, shrunk or extended, but never shifted: so
1726  * restart_addr remains valid so long as it remains in the vma's range.
1727  * unmap_mapping_range forces truncate_count to leap over page-aligned
1728  * values so we can save vma's restart_addr in its truncate_count field.
1729  */
1730 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1731
1732 static void reset_vma_truncate_counts(struct address_space *mapping)
1733 {
1734         struct vm_area_struct *vma;
1735         struct prio_tree_iter iter;
1736
1737         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1738                 vma->vm_truncate_count = 0;
1739         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1740                 vma->vm_truncate_count = 0;
1741 }
1742
1743 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1744                 unsigned long start_addr, unsigned long end_addr,
1745                 struct zap_details *details)
1746 {
1747         unsigned long restart_addr;
1748         int need_break;
1749
1750         /*
1751          * files that support invalidating or truncating portions of the
1752          * file from under mmaped areas must have their ->fault function
1753          * return a locked page (and set VM_FAULT_LOCKED in the return).
1754          * This provides synchronisation against concurrent unmapping here.
1755          */
1756
1757 again:
1758         restart_addr = vma->vm_truncate_count;
1759         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1760                 start_addr = restart_addr;
1761                 if (start_addr >= end_addr) {
1762                         /* Top of vma has been split off since last time */
1763                         vma->vm_truncate_count = details->truncate_count;
1764                         return 0;
1765                 }
1766         }
1767
1768         restart_addr = zap_page_range(vma, start_addr,
1769                                         end_addr - start_addr, details);
1770         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
1771
1772         if (restart_addr >= end_addr) {
1773                 /* We have now completed this vma: mark it so */
1774                 vma->vm_truncate_count = details->truncate_count;
1775                 if (!need_break)
1776                         return 0;
1777         } else {
1778                 /* Note restart_addr in vma's truncate_count field */
1779                 vma->vm_truncate_count = restart_addr;
1780                 if (!need_break)
1781                         goto again;
1782         }
1783
1784         spin_unlock(details->i_mmap_lock);
1785         cond_resched();
1786         spin_lock(details->i_mmap_lock);
1787         return -EINTR;
1788 }
1789
1790 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1791                                             struct zap_details *details)
1792 {
1793         struct vm_area_struct *vma;
1794         struct prio_tree_iter iter;
1795         pgoff_t vba, vea, zba, zea;
1796
1797 restart:
1798         vma_prio_tree_foreach(vma, &iter, root,
1799                         details->first_index, details->last_index) {
1800                 /* Skip quickly over those we have already dealt with */
1801                 if (vma->vm_truncate_count == details->truncate_count)
1802                         continue;
1803
1804                 vba = vma->vm_pgoff;
1805                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1806                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1807                 zba = details->first_index;
1808                 if (zba < vba)
1809                         zba = vba;
1810                 zea = details->last_index;
1811                 if (zea > vea)
1812                         zea = vea;
1813
1814                 if (unmap_mapping_range_vma(vma,
1815                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1816                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1817                                 details) < 0)
1818                         goto restart;
1819         }
1820 }
1821
1822 static inline void unmap_mapping_range_list(struct list_head *head,
1823                                             struct zap_details *details)
1824 {
1825         struct vm_area_struct *vma;
1826
1827         /*
1828          * In nonlinear VMAs there is no correspondence between virtual address
1829          * offset and file offset.  So we must perform an exhaustive search
1830          * across *all* the pages in each nonlinear VMA, not just the pages
1831          * whose virtual address lies outside the file truncation point.
1832          */
1833 restart:
1834         list_for_each_entry(vma, head, shared.vm_set.list) {
1835                 /* Skip quickly over those we have already dealt with */
1836                 if (vma->vm_truncate_count == details->truncate_count)
1837                         continue;
1838                 details->nonlinear_vma = vma;
1839                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1840                                         vma->vm_end, details) < 0)
1841                         goto restart;
1842         }
1843 }
1844
1845 /**
1846  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1847  * @mapping: the address space containing mmaps to be unmapped.
1848  * @holebegin: byte in first page to unmap, relative to the start of
1849  * the underlying file.  This will be rounded down to a PAGE_SIZE
1850  * boundary.  Note that this is different from vmtruncate(), which
1851  * must keep the partial page.  In contrast, we must get rid of
1852  * partial pages.
1853  * @holelen: size of prospective hole in bytes.  This will be rounded
1854  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1855  * end of the file.
1856  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1857  * but 0 when invalidating pagecache, don't throw away private data.
1858  */
1859 void unmap_mapping_range(struct address_space *mapping,
1860                 loff_t const holebegin, loff_t const holelen, int even_cows)
1861 {
1862         struct zap_details details;
1863         pgoff_t hba = holebegin >> PAGE_SHIFT;
1864         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1865
1866         /* Check for overflow. */
1867         if (sizeof(holelen) > sizeof(hlen)) {
1868                 long long holeend =
1869                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1870                 if (holeend & ~(long long)ULONG_MAX)
1871                         hlen = ULONG_MAX - hba + 1;
1872         }
1873
1874         details.check_mapping = even_cows? NULL: mapping;
1875         details.nonlinear_vma = NULL;
1876         details.first_index = hba;
1877         details.last_index = hba + hlen - 1;
1878         if (details.last_index < details.first_index)
1879                 details.last_index = ULONG_MAX;
1880         details.i_mmap_lock = &mapping->i_mmap_lock;
1881
1882         spin_lock(&mapping->i_mmap_lock);
1883
1884         /* Protect against endless unmapping loops */
1885         mapping->truncate_count++;
1886         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1887                 if (mapping->truncate_count == 0)
1888                         reset_vma_truncate_counts(mapping);
1889                 mapping->truncate_count++;
1890         }
1891         details.truncate_count = mapping->truncate_count;
1892
1893         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1894                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1895         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1896                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1897         spin_unlock(&mapping->i_mmap_lock);
1898 }
1899 EXPORT_SYMBOL(unmap_mapping_range);
1900
1901 /**
1902  * vmtruncate - unmap mappings "freed" by truncate() syscall
1903  * @inode: inode of the file used
1904  * @offset: file offset to start truncating
1905  *
1906  * NOTE! We have to be ready to update the memory sharing
1907  * between the file and the memory map for a potential last
1908  * incomplete page.  Ugly, but necessary.
1909  */
1910 int vmtruncate(struct inode * inode, loff_t offset)
1911 {
1912         struct address_space *mapping = inode->i_mapping;
1913         unsigned long limit;
1914
1915         if (inode->i_size < offset)
1916                 goto do_expand;
1917         /*
1918          * truncation of in-use swapfiles is disallowed - it would cause
1919          * subsequent swapout to scribble on the now-freed blocks.
1920          */
1921         if (IS_SWAPFILE(inode))
1922                 goto out_busy;
1923         i_size_write(inode, offset);
1924
1925         /*
1926          * unmap_mapping_range is called twice, first simply for efficiency
1927          * so that truncate_inode_pages does fewer single-page unmaps. However
1928          * after this first call, and before truncate_inode_pages finishes,
1929          * it is possible for private pages to be COWed, which remain after
1930          * truncate_inode_pages finishes, hence the second unmap_mapping_range
1931          * call must be made for correctness.
1932          */
1933         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1934         truncate_inode_pages(mapping, offset);
1935         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1936         goto out_truncate;
1937
1938 do_expand:
1939         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1940         if (limit != RLIM_INFINITY && offset > limit)
1941                 goto out_sig;
1942         if (offset > inode->i_sb->s_maxbytes)
1943                 goto out_big;
1944         i_size_write(inode, offset);
1945
1946 out_truncate:
1947         if (inode->i_op && inode->i_op->truncate)
1948                 inode->i_op->truncate(inode);
1949         return 0;
1950 out_sig:
1951         send_sig(SIGXFSZ, current, 0);
1952 out_big:
1953         return -EFBIG;
1954 out_busy:
1955         return -ETXTBSY;
1956 }
1957 EXPORT_SYMBOL(vmtruncate);
1958
1959 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1960 {
1961         struct address_space *mapping = inode->i_mapping;
1962
1963         /*
1964          * If the underlying filesystem is not going to provide
1965          * a way to truncate a range of blocks (punch a hole) -
1966          * we should return failure right now.
1967          */
1968         if (!inode->i_op || !inode->i_op->truncate_range)
1969                 return -ENOSYS;
1970
1971         mutex_lock(&inode->i_mutex);
1972         down_write(&inode->i_alloc_sem);
1973         unmap_mapping_range(mapping, offset, (end - offset), 1);
1974         truncate_inode_pages_range(mapping, offset, end);
1975         unmap_mapping_range(mapping, offset, (end - offset), 1);
1976         inode->i_op->truncate_range(inode, offset, end);
1977         up_write(&inode->i_alloc_sem);
1978         mutex_unlock(&inode->i_mutex);
1979
1980         return 0;
1981 }
1982
1983 /**
1984  * swapin_readahead - swap in pages in hope we need them soon
1985  * @entry: swap entry of this memory
1986  * @addr: address to start
1987  * @vma: user vma this addresses belong to
1988  *
1989  * Primitive swap readahead code. We simply read an aligned block of
1990  * (1 << page_cluster) entries in the swap area. This method is chosen
1991  * because it doesn't cost us any seek time.  We also make sure to queue
1992  * the 'original' request together with the readahead ones...
1993  *
1994  * This has been extended to use the NUMA policies from the mm triggering
1995  * the readahead.
1996  *
1997  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1998  */
1999 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
2000 {
2001 #ifdef CONFIG_NUMA
2002         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
2003 #endif
2004         int i, num;
2005         struct page *new_page;
2006         unsigned long offset;
2007
2008         /*
2009          * Get the number of handles we should do readahead io to.
2010          */
2011         num = valid_swaphandles(entry, &offset);
2012         for (i = 0; i < num; offset++, i++) {
2013                 /* Ok, do the async read-ahead now */
2014                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
2015                                                            offset), vma, addr);
2016                 if (!new_page)
2017                         break;
2018                 page_cache_release(new_page);
2019 #ifdef CONFIG_NUMA
2020                 /*
2021                  * Find the next applicable VMA for the NUMA policy.
2022                  */
2023                 addr += PAGE_SIZE;
2024                 if (addr == 0)
2025                         vma = NULL;
2026                 if (vma) {
2027                         if (addr >= vma->vm_end) {
2028                                 vma = next_vma;
2029                                 next_vma = vma ? vma->vm_next : NULL;
2030                         }
2031                         if (vma && addr < vma->vm_start)
2032                                 vma = NULL;
2033                 } else {
2034                         if (next_vma && addr >= next_vma->vm_start) {
2035                                 vma = next_vma;
2036                                 next_vma = vma->vm_next;
2037                         }
2038                 }
2039 #endif
2040         }
2041         lru_add_drain();        /* Push any new pages onto the LRU now */
2042 }
2043
2044 /*
2045  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2046  * but allow concurrent faults), and pte mapped but not yet locked.
2047  * We return with mmap_sem still held, but pte unmapped and unlocked.
2048  */
2049 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2050                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2051                 int write_access, pte_t orig_pte)
2052 {
2053         spinlock_t *ptl;
2054         struct page *page;
2055         swp_entry_t entry;
2056         pte_t pte;
2057         int ret = 0;
2058
2059         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2060                 goto out;
2061
2062         entry = pte_to_swp_entry(orig_pte);
2063         if (is_migration_entry(entry)) {
2064                 migration_entry_wait(mm, pmd, address);
2065                 goto out;
2066         }
2067         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2068         page = lookup_swap_cache(entry);
2069         if (!page) {
2070                 grab_swap_token(); /* Contend for token _before_ read-in */
2071                 swapin_readahead(entry, address, vma);
2072                 page = read_swap_cache_async(entry, vma, address);
2073                 if (!page) {
2074                         /*
2075                          * Back out if somebody else faulted in this pte
2076                          * while we released the pte lock.
2077                          */
2078                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2079                         if (likely(pte_same(*page_table, orig_pte)))
2080                                 ret = VM_FAULT_OOM;
2081                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2082                         goto unlock;
2083                 }
2084
2085                 /* Had to read the page from swap area: Major fault */
2086                 ret = VM_FAULT_MAJOR;
2087                 count_vm_event(PGMAJFAULT);
2088         }
2089
2090         mark_page_accessed(page);
2091         lock_page(page);
2092         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2093
2094         /*
2095          * Back out if somebody else already faulted in this pte.
2096          */
2097         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2098         if (unlikely(!pte_same(*page_table, orig_pte)))
2099                 goto out_nomap;
2100
2101         if (unlikely(!PageUptodate(page))) {
2102                 ret = VM_FAULT_SIGBUS;
2103                 goto out_nomap;
2104         }
2105
2106         /* The page isn't present yet, go ahead with the fault. */
2107
2108         inc_mm_counter(mm, anon_rss);
2109         pte = mk_pte(page, vma->vm_page_prot);
2110         if (write_access && can_share_swap_page(page)) {
2111                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2112                 write_access = 0;
2113         }
2114
2115         flush_icache_page(vma, page);
2116         set_pte_at(mm, address, page_table, pte);
2117         page_add_anon_rmap(page, vma, address);
2118
2119         swap_free(entry);
2120         if (vm_swap_full())
2121                 remove_exclusive_swap_page(page);
2122         unlock_page(page);
2123
2124         if (write_access) {
2125                 /* XXX: We could OR the do_wp_page code with this one? */
2126                 if (do_wp_page(mm, vma, address,
2127                                 page_table, pmd, ptl, pte) & VM_FAULT_OOM)
2128                         ret = VM_FAULT_OOM;
2129                 goto out;
2130         }
2131
2132         /* No need to invalidate - it was non-present before */
2133         update_mmu_cache(vma, address, pte);
2134 unlock:
2135         pte_unmap_unlock(page_table, ptl);
2136 out:
2137         return ret;
2138 out_nomap:
2139         pte_unmap_unlock(page_table, ptl);
2140         unlock_page(page);
2141         page_cache_release(page);
2142         return ret;
2143 }
2144
2145 /*
2146  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2147  * but allow concurrent faults), and pte mapped but not yet locked.
2148  * We return with mmap_sem still held, but pte unmapped and unlocked.
2149  */
2150 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2151                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2152                 int write_access)
2153 {
2154         struct page *page;
2155         spinlock_t *ptl;
2156         pte_t entry;
2157
2158         /* Allocate our own private page. */
2159         pte_unmap(page_table);
2160
2161         if (unlikely(anon_vma_prepare(vma)))
2162                 goto oom;
2163         page = alloc_zeroed_user_highpage_movable(vma, address);
2164         if (!page)
2165                 goto oom;
2166
2167         entry = mk_pte(page, vma->vm_page_prot);
2168         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2169
2170         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2171         if (!pte_none(*page_table))
2172                 goto release;
2173         inc_mm_counter(mm, anon_rss);
2174         lru_cache_add_active(page);
2175         page_add_new_anon_rmap(page, vma, address);
2176         set_pte_at(mm, address, page_table, entry);
2177
2178         /* No need to invalidate - it was non-present before */
2179         update_mmu_cache(vma, address, entry);
2180 unlock:
2181         pte_unmap_unlock(page_table, ptl);
2182         return 0;
2183 release:
2184         page_cache_release(page);
2185         goto unlock;
2186 oom:
2187         return VM_FAULT_OOM;
2188 }
2189
2190 /*
2191  * __do_fault() tries to create a new page mapping. It aggressively
2192  * tries to share with existing pages, but makes a separate copy if
2193  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2194  * the next page fault.
2195  *
2196  * As this is called only for pages that do not currently exist, we
2197  * do not need to flush old virtual caches or the TLB.
2198  *
2199  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2200  * but allow concurrent faults), and pte neither mapped nor locked.
2201  * We return with mmap_sem still held, but pte unmapped and unlocked.
2202  */
2203 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2204                 unsigned long address, pmd_t *pmd,
2205                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2206 {
2207         pte_t *page_table;
2208         spinlock_t *ptl;
2209         struct page *page;
2210         pte_t entry;
2211         int anon = 0;
2212         struct page *dirty_page = NULL;
2213         struct vm_fault vmf;
2214         int ret;
2215         int page_mkwrite = 0;
2216
2217         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2218         vmf.pgoff = pgoff;
2219         vmf.flags = flags;
2220         vmf.page = NULL;
2221
2222         BUG_ON(vma->vm_flags & VM_PFNMAP);
2223
2224         if (likely(vma->vm_ops->fault)) {
2225                 ret = vma->vm_ops->fault(vma, &vmf);
2226                 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2227                         return ret;
2228         } else {
2229                 /* Legacy ->nopage path */
2230                 ret = 0;
2231                 vmf.page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2232                 /* no page was available -- either SIGBUS or OOM */
2233                 if (unlikely(vmf.page == NOPAGE_SIGBUS))
2234                         return VM_FAULT_SIGBUS;
2235                 else if (unlikely(vmf.page == NOPAGE_OOM))
2236                         return VM_FAULT_OOM;
2237         }
2238
2239         /*
2240          * For consistency in subsequent calls, make the faulted page always
2241          * locked.
2242          */
2243         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2244                 lock_page(vmf.page);
2245         else
2246                 VM_BUG_ON(!PageLocked(vmf.page));
2247
2248         /*
2249          * Should we do an early C-O-W break?
2250          */
2251         page = vmf.page;
2252         if (flags & FAULT_FLAG_WRITE) {
2253                 if (!(vma->vm_flags & VM_SHARED)) {
2254                         anon = 1;
2255                         if (unlikely(anon_vma_prepare(vma))) {
2256                                 ret = VM_FAULT_OOM;
2257                                 goto out;
2258                         }
2259                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2260                                                 vma, address);
2261                         if (!page) {
2262                                 ret = VM_FAULT_OOM;
2263                                 goto out;
2264                         }
2265                         copy_user_highpage(page, vmf.page, address, vma);
2266                 } else {
2267                         /*
2268                          * If the page will be shareable, see if the backing
2269                          * address space wants to know that the page is about
2270                          * to become writable
2271                          */
2272                         if (vma->vm_ops->page_mkwrite) {
2273                                 unlock_page(page);
2274                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2275                                         ret = VM_FAULT_SIGBUS;
2276                                         anon = 1; /* no anon but release vmf.page */
2277                                         goto out_unlocked;
2278                                 }
2279                                 lock_page(page);
2280                                 /*
2281                                  * XXX: this is not quite right (racy vs
2282                                  * invalidate) to unlock and relock the page
2283                                  * like this, however a better fix requires
2284                                  * reworking page_mkwrite locking API, which
2285                                  * is better done later.
2286                                  */
2287                                 if (!page->mapping) {
2288                                         ret = 0;
2289                                         anon = 1; /* no anon but release vmf.page */
2290                                         goto out;
2291                                 }
2292                                 page_mkwrite = 1;
2293                         }
2294                 }
2295
2296         }
2297
2298         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2299
2300         /*
2301          * This silly early PAGE_DIRTY setting removes a race
2302          * due to the bad i386 page protection. But it's valid
2303          * for other architectures too.
2304          *
2305          * Note that if write_access is true, we either now have
2306          * an exclusive copy of the page, or this is a shared mapping,
2307          * so we can make it writable and dirty to avoid having to
2308          * handle that later.
2309          */
2310         /* Only go through if we didn't race with anybody else... */
2311         if (likely(pte_same(*page_table, orig_pte))) {
2312                 flush_icache_page(vma, page);
2313                 entry = mk_pte(page, vma->vm_page_prot);
2314                 if (flags & FAULT_FLAG_WRITE)
2315                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2316                 set_pte_at(mm, address, page_table, entry);
2317                 if (anon) {
2318                         inc_mm_counter(mm, anon_rss);
2319                         lru_cache_add_active(page);
2320                         page_add_new_anon_rmap(page, vma, address);
2321                 } else {
2322                         inc_mm_counter(mm, file_rss);
2323                         page_add_file_rmap(page);
2324                         if (flags & FAULT_FLAG_WRITE) {
2325                                 dirty_page = page;
2326                                 get_page(dirty_page);
2327                         }
2328                 }
2329
2330                 /* no need to invalidate: a not-present page won't be cached */
2331                 update_mmu_cache(vma, address, entry);
2332         } else {
2333                 if (anon)
2334                         page_cache_release(page);
2335                 else
2336                         anon = 1; /* no anon but release faulted_page */
2337         }
2338
2339         pte_unmap_unlock(page_table, ptl);
2340
2341 out:
2342         unlock_page(vmf.page);
2343 out_unlocked:
2344         if (anon)
2345                 page_cache_release(vmf.page);
2346         else if (dirty_page) {
2347                 if (vma->vm_file)
2348                         file_update_time(vma->vm_file);
2349
2350                 set_page_dirty_balance(dirty_page, page_mkwrite);
2351                 put_page(dirty_page);
2352         }
2353
2354         return ret;
2355 }
2356
2357 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2358                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2359                 int write_access, pte_t orig_pte)
2360 {
2361         pgoff_t pgoff = (((address & PAGE_MASK)
2362                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2363         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2364
2365         pte_unmap(page_table);
2366         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2367 }
2368
2369
2370 /*
2371  * do_no_pfn() tries to create a new page mapping for a page without
2372  * a struct_page backing it
2373  *
2374  * As this is called only for pages that do not currently exist, we
2375  * do not need to flush old virtual caches or the TLB.
2376  *
2377  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2378  * but allow concurrent faults), and pte mapped but not yet locked.
2379  * We return with mmap_sem still held, but pte unmapped and unlocked.
2380  *
2381  * It is expected that the ->nopfn handler always returns the same pfn
2382  * for a given virtual mapping.
2383  *
2384  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2385  */
2386 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2387                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2388                      int write_access)
2389 {
2390         spinlock_t *ptl;
2391         pte_t entry;
2392         unsigned long pfn;
2393
2394         pte_unmap(page_table);
2395         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
2396         BUG_ON(is_cow_mapping(vma->vm_flags));
2397
2398         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2399         if (unlikely(pfn == NOPFN_OOM))
2400                 return VM_FAULT_OOM;
2401         else if (unlikely(pfn == NOPFN_SIGBUS))
2402                 return VM_FAULT_SIGBUS;
2403         else if (unlikely(pfn == NOPFN_REFAULT))
2404                 return 0;
2405
2406         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2407
2408         /* Only go through if we didn't race with anybody else... */
2409         if (pte_none(*page_table)) {
2410                 entry = pfn_pte(pfn, vma->vm_page_prot);
2411                 if (write_access)
2412                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2413                 set_pte_at(mm, address, page_table, entry);
2414         }
2415         pte_unmap_unlock(page_table, ptl);
2416         return 0;
2417 }
2418
2419 /*
2420  * Fault of a previously existing named mapping. Repopulate the pte
2421  * from the encoded file_pte if possible. This enables swappable
2422  * nonlinear vmas.
2423  *
2424  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2425  * but allow concurrent faults), and pte mapped but not yet locked.
2426  * We return with mmap_sem still held, but pte unmapped and unlocked.
2427  */
2428 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2429                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2430                 int write_access, pte_t orig_pte)
2431 {
2432         unsigned int flags = FAULT_FLAG_NONLINEAR |
2433                                 (write_access ? FAULT_FLAG_WRITE : 0);
2434         pgoff_t pgoff;
2435
2436         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2437                 return 0;
2438
2439         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2440                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2441                 /*
2442                  * Page table corrupted: show pte and kill process.
2443                  */
2444                 print_bad_pte(vma, orig_pte, address);
2445                 return VM_FAULT_OOM;
2446         }
2447
2448         pgoff = pte_to_pgoff(orig_pte);
2449         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2450 }
2451
2452 /*
2453  * These routines also need to handle stuff like marking pages dirty
2454  * and/or accessed for architectures that don't do it in hardware (most
2455  * RISC architectures).  The early dirtying is also good on the i386.
2456  *
2457  * There is also a hook called "update_mmu_cache()" that architectures
2458  * with external mmu caches can use to update those (ie the Sparc or
2459  * PowerPC hashed page tables that act as extended TLBs).
2460  *
2461  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2462  * but allow concurrent faults), and pte mapped but not yet locked.
2463  * We return with mmap_sem still held, but pte unmapped and unlocked.
2464  */
2465 static inline int handle_pte_fault(struct mm_struct *mm,
2466                 struct vm_area_struct *vma, unsigned long address,
2467                 pte_t *pte, pmd_t *pmd, int write_access)
2468 {
2469         pte_t entry;
2470         spinlock_t *ptl;
2471
2472         entry = *pte;
2473         if (!pte_present(entry)) {
2474                 if (pte_none(entry)) {
2475                         if (vma->vm_ops) {
2476                                 if (vma->vm_ops->fault || vma->vm_ops->nopage)
2477                                         return do_linear_fault(mm, vma, address,
2478                                                 pte, pmd, write_access, entry);
2479                                 if (unlikely(vma->vm_ops->nopfn))
2480                                         return do_no_pfn(mm, vma, address, pte,
2481                                                          pmd, write_access);
2482                         }
2483                         return do_anonymous_page(mm, vma, address,
2484                                                  pte, pmd, write_access);
2485                 }
2486                 if (pte_file(entry))
2487                         return do_nonlinear_fault(mm, vma, address,
2488                                         pte, pmd, write_access, entry);
2489                 return do_swap_page(mm, vma, address,
2490                                         pte, pmd, write_access, entry);
2491         }
2492
2493         ptl = pte_lockptr(mm, pmd);
2494         spin_lock(ptl);
2495         if (unlikely(!pte_same(*pte, entry)))
2496                 goto unlock;
2497         if (write_access) {
2498                 if (!pte_write(entry))
2499                         return do_wp_page(mm, vma, address,
2500                                         pte, pmd, ptl, entry);
2501                 entry = pte_mkdirty(entry);
2502         }
2503         entry = pte_mkyoung(entry);
2504         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2505                 update_mmu_cache(vma, address, entry);
2506         } else {
2507                 /*
2508                  * This is needed only for protection faults but the arch code
2509                  * is not yet telling us if this is a protection fault or not.
2510                  * This still avoids useless tlb flushes for .text page faults
2511                  * with threads.
2512                  */
2513                 if (write_access)
2514                         flush_tlb_page(vma, address);
2515         }
2516 unlock:
2517         pte_unmap_unlock(pte, ptl);
2518         return 0;
2519 }
2520
2521 /*
2522  * By the time we get here, we already hold the mm semaphore
2523  */
2524 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2525                 unsigned long address, int write_access)
2526 {
2527         pgd_t *pgd;
2528         pud_t *pud;
2529         pmd_t *pmd;
2530         pte_t *pte;
2531
2532         __set_current_state(TASK_RUNNING);
2533
2534         count_vm_event(PGFAULT);
2535
2536         if (unlikely(is_vm_hugetlb_page(vma)))
2537                 return hugetlb_fault(mm, vma, address, write_access);
2538
2539         pgd = pgd_offset(mm, address);
2540         pud = pud_alloc(mm, pgd, address);
2541         if (!pud)
2542                 return VM_FAULT_OOM;
2543         pmd = pmd_alloc(mm, pud, address);
2544         if (!pmd)
2545                 return VM_FAULT_OOM;
2546         pte = pte_alloc_map(mm, pmd, address);
2547         if (!pte)
2548                 return VM_FAULT_OOM;
2549
2550         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2551 }
2552
2553 #ifndef __PAGETABLE_PUD_FOLDED
2554 /*
2555  * Allocate page upper directory.
2556  * We've already handled the fast-path in-line.
2557  */
2558 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2559 {
2560         pud_t *new = pud_alloc_one(mm, address);
2561         if (!new)
2562                 return -ENOMEM;
2563
2564         spin_lock(&mm->page_table_lock);
2565         if (pgd_present(*pgd))          /* Another has populated it */
2566                 pud_free(new);
2567         else
2568                 pgd_populate(mm, pgd, new);
2569         spin_unlock(&mm->page_table_lock);
2570         return 0;
2571 }
2572 #endif /* __PAGETABLE_PUD_FOLDED */
2573
2574 #ifndef __PAGETABLE_PMD_FOLDED
2575 /*
2576  * Allocate page middle directory.
2577  * We've already handled the fast-path in-line.
2578  */
2579 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2580 {
2581         pmd_t *new = pmd_alloc_one(mm, address);
2582         if (!new)
2583                 return -ENOMEM;
2584
2585         spin_lock(&mm->page_table_lock);
2586 #ifndef __ARCH_HAS_4LEVEL_HACK
2587         if (pud_present(*pud))          /* Another has populated it */
2588                 pmd_free(new);
2589         else
2590                 pud_populate(mm, pud, new);
2591 #else
2592         if (pgd_present(*pud))          /* Another has populated it */
2593                 pmd_free(new);
2594         else
2595                 pgd_populate(mm, pud, new);
2596 #endif /* __ARCH_HAS_4LEVEL_HACK */
2597         spin_unlock(&mm->page_table_lock);
2598         return 0;
2599 }
2600 #endif /* __PAGETABLE_PMD_FOLDED */
2601
2602 int make_pages_present(unsigned long addr, unsigned long end)
2603 {
2604         int ret, len, write;
2605         struct vm_area_struct * vma;
2606
2607         vma = find_vma(current->mm, addr);
2608         if (!vma)
2609                 return -1;
2610         write = (vma->vm_flags & VM_WRITE) != 0;
2611         BUG_ON(addr >= end);
2612         BUG_ON(end > vma->vm_end);
2613         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2614         ret = get_user_pages(current, current->mm, addr,
2615                         len, write, 0, NULL, NULL);
2616         if (ret < 0)
2617                 return ret;
2618         return ret == len ? 0 : -1;
2619 }
2620
2621 /*
2622  * Map a vmalloc()-space virtual address to the physical page.
2623  */
2624 struct page * vmalloc_to_page(void * vmalloc_addr)
2625 {
2626         unsigned long addr = (unsigned long) vmalloc_addr;
2627         struct page *page = NULL;
2628         pgd_t *pgd = pgd_offset_k(addr);
2629         pud_t *pud;
2630         pmd_t *pmd;
2631         pte_t *ptep, pte;
2632
2633         if (!pgd_none(*pgd)) {
2634                 pud = pud_offset(pgd, addr);
2635                 if (!pud_none(*pud)) {
2636                         pmd = pmd_offset(pud, addr);
2637                         if (!pmd_none(*pmd)) {
2638                                 ptep = pte_offset_map(pmd, addr);
2639                                 pte = *ptep;
2640                                 if (pte_present(pte))
2641                                         page = pte_page(pte);
2642                                 pte_unmap(ptep);
2643                         }
2644                 }
2645         }
2646         return page;
2647 }
2648
2649 EXPORT_SYMBOL(vmalloc_to_page);
2650
2651 /*
2652  * Map a vmalloc()-space virtual address to the physical page frame number.
2653  */
2654 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2655 {
2656         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2657 }
2658
2659 EXPORT_SYMBOL(vmalloc_to_pfn);
2660
2661 #if !defined(__HAVE_ARCH_GATE_AREA)
2662
2663 #if defined(AT_SYSINFO_EHDR)
2664 static struct vm_area_struct gate_vma;
2665
2666 static int __init gate_vma_init(void)
2667 {
2668         gate_vma.vm_mm = NULL;
2669         gate_vma.vm_start = FIXADDR_USER_START;
2670         gate_vma.vm_end = FIXADDR_USER_END;
2671         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2672         gate_vma.vm_page_prot = __P101;
2673         /*
2674          * Make sure the vDSO gets into every core dump.
2675          * Dumping its contents makes post-mortem fully interpretable later
2676          * without matching up the same kernel and hardware config to see
2677          * what PC values meant.
2678          */
2679         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2680         return 0;
2681 }
2682 __initcall(gate_vma_init);
2683 #endif
2684
2685 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2686 {
2687 #ifdef AT_SYSINFO_EHDR
2688         return &gate_vma;
2689 #else
2690         return NULL;
2691 #endif
2692 }
2693
2694 int in_gate_area_no_task(unsigned long addr)
2695 {
2696 #ifdef AT_SYSINFO_EHDR
2697         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2698                 return 1;
2699 #endif
2700         return 0;
2701 }
2702
2703 #endif  /* __HAVE_ARCH_GATE_AREA */
2704
2705 /*
2706  * Access another process' address space.
2707  * Source/target buffer must be kernel space,
2708  * Do not walk the page table directly, use get_user_pages
2709  */
2710 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2711 {
2712         struct mm_struct *mm;
2713         struct vm_area_struct *vma;
2714         struct page *page;
2715         void *old_buf = buf;
2716
2717         mm = get_task_mm(tsk);
2718         if (!mm)
2719                 return 0;
2720
2721         down_read(&mm->mmap_sem);
2722         /* ignore errors, just check how much was successfully transferred */
2723         while (len) {
2724                 int bytes, ret, offset;
2725                 void *maddr;
2726
2727                 ret = get_user_pages(tsk, mm, addr, 1,
2728                                 write, 1, &page, &vma);
2729                 if (ret <= 0)
2730                         break;
2731
2732                 bytes = len;
2733                 offset = addr & (PAGE_SIZE-1);
2734                 if (bytes > PAGE_SIZE-offset)
2735                         bytes = PAGE_SIZE-offset;
2736
2737                 maddr = kmap(page);
2738                 if (write) {
2739                         copy_to_user_page(vma, page, addr,
2740                                           maddr + offset, buf, bytes);
2741                         set_page_dirty_lock(page);
2742                 } else {
2743                         copy_from_user_page(vma, page, addr,
2744                                             buf, maddr + offset, bytes);
2745                 }
2746                 kunmap(page);
2747                 page_cache_release(page);
2748                 len -= bytes;
2749                 buf += bytes;
2750                 addr += bytes;
2751         }
2752         up_read(&mm->mmap_sem);
2753         mmput(mm);
2754
2755         return buf - old_buf;
2756 }
2757
2758 /*
2759  * Print the name of a VMA.
2760  */
2761 void print_vma_addr(char *prefix, unsigned long ip)
2762 {
2763         struct mm_struct *mm = current->mm;
2764         struct vm_area_struct *vma;
2765
2766         down_read(&mm->mmap_sem);
2767         vma = find_vma(mm, ip);
2768         if (vma && vma->vm_file) {
2769                 struct file *f = vma->vm_file;
2770                 char *buf = (char *)__get_free_page(GFP_KERNEL);
2771                 if (buf) {
2772                         char *p, *s;
2773
2774                         p = d_path(f->f_dentry, f->f_vfsmnt, buf, PAGE_SIZE);
2775                         if (IS_ERR(p))
2776                                 p = "?";
2777                         s = strrchr(p, '/');
2778                         if (s)
2779                                 p = s+1;
2780                         printk("%s%s[%lx+%lx]", prefix, p,
2781                                         vma->vm_start,
2782                                         vma->vm_end - vma->vm_start);
2783                         free_page((unsigned long)buf);
2784                 }
2785         }
2786         up_read(&current->mm->mmap_sem);
2787 }