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