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 #include <linux/memcontrol.h>
  54
  55 #include <asm/pgalloc.h>
  56 #include <asm/uaccess.h>
  57 #include <asm/tlb.h>
  58 #include <asm/tlbflush.h>
  59 #include <asm/pgtable.h>
  60
  61 #include <linux/swapops.h>
  62 #include <linux/elf.h>
  63
  64 #ifndef CONFIG_NEED_MULTIPLE_NODES
  65 /* use the per-pgdat data instead for discontigmem - mbligh */
  66 unsigned long max_mapnr;
  67 struct page *mem_map;
  68
  69 EXPORT_SYMBOL(max_mapnr);
  70 EXPORT_SYMBOL(mem_map);
  71 #endif
  72
  73 unsigned long num_physpages;
  74 /*
  75  * A number of key systems in x86 including ioremap() rely on the assumption
  76  * that high_memory defines the upper bound on direct map memory, then end
  77  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  78  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  79  * and ZONE_HIGHMEM.
  80  */
  81 void * high_memory;
  82
  83 EXPORT_SYMBOL(num_physpages);
  84 EXPORT_SYMBOL(high_memory);
  85
  86 /*
  87  * Randomize the address space (stacks, mmaps, brk, etc.).
  88  *
  89  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
  90  *   as ancient (libc5 based) binaries can segfault. )
  91  */
  92 int randomize_va_space __read_mostly =
  93 #ifdef CONFIG_COMPAT_BRK
  94                                         1;
  95 #else
  96                                         2;
  97 #endif
  98
  99 static int __init disable_randmaps(char *s)
 100 {
 101         randomize_va_space = 0;
 102         return 1;
 103 }
 104 __setup("norandmaps", disable_randmaps);
 105
 106
 107 /*
 108  * If a p?d_bad entry is found while walking page tables, report
 109  * the error, before resetting entry to p?d_none.  Usually (but
 110  * very seldom) called out from the p?d_none_or_clear_bad macros.
 111  */
 112
 113 void pgd_clear_bad(pgd_t *pgd)
 114 {
 115         pgd_ERROR(*pgd);
 116         pgd_clear(pgd);
 117 }
 118
 119 void pud_clear_bad(pud_t *pud)
 120 {
 121         pud_ERROR(*pud);
 122         pud_clear(pud);
 123 }
 124
 125 void pmd_clear_bad(pmd_t *pmd)
 126 {
 127         pmd_ERROR(*pmd);
 128         pmd_clear(pmd);
 129 }
 130
 131 /*
 132  * Note: this doesn't free the actual pages themselves. That
 133  * has been handled earlier when unmapping all the memory regions.
 134  */
 135 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
 136 {
 137         pgtable_t token = pmd_pgtable(*pmd);
 138         pmd_clear(pmd);
 139         pte_free_tlb(tlb, token);
 140         tlb->mm->nr_ptes--;
 141 }
 142
 143 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 144                                 unsigned long addr, unsigned long end,
 145                                 unsigned long floor, unsigned long ceiling)
 146 {
 147         pmd_t *pmd;
 148         unsigned long next;
 149         unsigned long start;
 150
 151         start = addr;
 152         pmd = pmd_offset(pud, addr);
 153         do {
 154                 next = pmd_addr_end(addr, end);
 155                 if (pmd_none_or_clear_bad(pmd))
 156                         continue;
 157                 free_pte_range(tlb, pmd);
 158         } while (pmd++, addr = next, addr != end);
 159
 160         start &= PUD_MASK;
 161         if (start < floor)
 162                 return;
 163         if (ceiling) {
 164                 ceiling &= PUD_MASK;
 165                 if (!ceiling)
 166                         return;
 167         }
 168         if (end - 1 > ceiling - 1)
 169                 return;
 170
 171         pmd = pmd_offset(pud, start);
 172         pud_clear(pud);
 173         pmd_free_tlb(tlb, pmd);
 174 }
 175
 176 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 177                                 unsigned long addr, unsigned long end,
 178                                 unsigned long floor, unsigned long ceiling)
 179 {
 180         pud_t *pud;
 181         unsigned long next;
 182         unsigned long start;
 183
 184         start = addr;
 185         pud = pud_offset(pgd, addr);
 186         do {
 187                 next = pud_addr_end(addr, end);
 188                 if (pud_none_or_clear_bad(pud))
 189                         continue;
 190                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 191         } while (pud++, addr = next, addr != end);
 192
 193         start &= PGDIR_MASK;
 194         if (start < floor)
 195                 return;
 196         if (ceiling) {
 197                 ceiling &= PGDIR_MASK;
 198                 if (!ceiling)
 199                         return;
 200         }
 201         if (end - 1 > ceiling - 1)
 202                 return;
 203
 204         pud = pud_offset(pgd, start);
 205         pgd_clear(pgd);
 206         pud_free_tlb(tlb, pud);
 207 }
 208
 209 /*
 210  * This function frees user-level page tables of a process.
 211  *
 212  * Must be called with pagetable lock held.
 213  */
 214 void free_pgd_range(struct mmu_gather **tlb,
 215                         unsigned long addr, unsigned long end,
 216                         unsigned long floor, unsigned long ceiling)
 217 {
 218         pgd_t *pgd;
 219         unsigned long next;
 220         unsigned long start;
 221
 222         /*
 223          * The next few lines have given us lots of grief...
 224          *
 225          * Why are we testing PMD* at this top level?  Because often
 226          * there will be no work to do at all, and we'd prefer not to
 227          * go all the way down to the bottom just to discover that.
 228          *
 229          * Why all these "- 1"s?  Because 0 represents both the bottom
 230          * of the address space and the top of it (using -1 for the
 231          * top wouldn't help much: the masks would do the wrong thing).
 232          * The rule is that addr 0 and floor 0 refer to the bottom of
 233          * the address space, but end 0 and ceiling 0 refer to the top
 234          * Comparisons need to use "end - 1" and "ceiling - 1" (though
 235          * that end 0 case should be mythical).
 236          *
 237          * Wherever addr is brought up or ceiling brought down, we must
 238          * be careful to reject "the opposite 0" before it confuses the
 239          * subsequent tests.  But what about where end is brought down
 240          * by PMD_SIZE below? no, end can't go down to 0 there.
 241          *
 242          * Whereas we round start (addr) and ceiling down, by different
 243          * masks at different levels, in order to test whether a table
 244          * now has no other vmas using it, so can be freed, we don't
 245          * bother to round floor or end up - the tests don't need that.
 246          */
 247
 248         addr &= PMD_MASK;
 249         if (addr < floor) {
 250                 addr += PMD_SIZE;
 251                 if (!addr)
 252                         return;
 253         }
 254         if (ceiling) {
 255                 ceiling &= PMD_MASK;
 256                 if (!ceiling)
 257                         return;
 258         }
 259         if (end - 1 > ceiling - 1)
 260                 end -= PMD_SIZE;
 261         if (addr > end - 1)
 262                 return;
 263
 264         start = addr;
 265         pgd = pgd_offset((*tlb)->mm, addr);
 266         do {
 267                 next = pgd_addr_end(addr, end);
 268                 if (pgd_none_or_clear_bad(pgd))
 269                         continue;
 270                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
 271         } while (pgd++, addr = next, addr != end);
 272 }
 273
 274 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
 275                 unsigned long floor, unsigned long ceiling)
 276 {
 277         while (vma) {
 278                 struct vm_area_struct *next = vma->vm_next;
 279                 unsigned long addr = vma->vm_start;
 280
 281                 /*
 282                  * Hide vma from rmap and vmtruncate before freeing pgtables
 283                  */
 284                 anon_vma_unlink(vma);
 285                 unlink_file_vma(vma);
 286
 287                 if (is_vm_hugetlb_page(vma)) {
 288                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 289                                 floor, next? next->vm_start: ceiling);
 290                 } else {
 291                         /*
 292                          * Optimization: gather nearby vmas into one call down
 293                          */
 294                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 295                                && !is_vm_hugetlb_page(next)) {
 296                                 vma = next;
 297                                 next = vma->vm_next;
 298                                 anon_vma_unlink(vma);
 299                                 unlink_file_vma(vma);
 300                         }
 301                         free_pgd_range(tlb, addr, vma->vm_end,
 302                                 floor, next? next->vm_start: ceiling);
 303                 }
 304                 vma = next;
 305         }
 306 }
 307
 308 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 309 {
 310         pgtable_t new = pte_alloc_one(mm, address);
 311         if (!new)
 312                 return -ENOMEM;
 313
 314         spin_lock(&mm->page_table_lock);
 315         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
 316                 mm->nr_ptes++;
 317                 pmd_populate(mm, pmd, new);
 318                 new = NULL;
 319         }
 320         spin_unlock(&mm->page_table_lock);
 321         if (new)
 322                 pte_free(mm, new);
 323         return 0;
 324 }
 325
 326 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 327 {
 328         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 329         if (!new)
 330                 return -ENOMEM;
 331
 332         spin_lock(&init_mm.page_table_lock);
 333         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
 334                 pmd_populate_kernel(&init_mm, pmd, new);
 335                 new = NULL;
 336         }
 337         spin_unlock(&init_mm.page_table_lock);
 338         if (new)
 339                 pte_free_kernel(&init_mm, new);
 340         return 0;
 341 }
 342
 343 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
 344 {
 345         if (file_rss)
 346                 add_mm_counter(mm, file_rss, file_rss);
 347         if (anon_rss)
 348                 add_mm_counter(mm, anon_rss, anon_rss);
 349 }
 350
 351 /*
 352  * This function is called to print an error when a bad pte
 353  * is found. For example, we might have a PFN-mapped pte in
 354  * a region that doesn't allow it.
 355  *
 356  * The calling function must still handle the error.
 357  */
 358 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
 359 {
 360         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
 361                         "vm_flags = %lx, vaddr = %lx\n",
 362                 (long long)pte_val(pte),
 363                 (vma->vm_mm == current->mm ? current->comm : "???"),
 364                 vma->vm_flags, vaddr);
 365         dump_stack();
 366 }
 367
 368 static inline int is_cow_mapping(unsigned int flags)
 369 {
 370         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 371 }
 372
 373 /*
 374  * This function gets the "struct page" associated with a pte.
 375  *
 376  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
 377  * will have each page table entry just pointing to a raw page frame
 378  * number, and as far as the VM layer is concerned, those do not have
 379  * pages associated with them - even if the PFN might point to memory
 380  * that otherwise is perfectly fine and has a "struct page".
 381  *
 382  * The way we recognize those mappings is through the rules set up
 383  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
 384  * and the vm_pgoff will point to the first PFN mapped: thus every
 385  * page that is a raw mapping will always honor the rule
 386  *
 387  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 388  *
 389  * and if that isn't true, the page has been COW'ed (in which case it
 390  * _does_ have a "struct page" associated with it even if it is in a
 391  * VM_PFNMAP range).
 392  */
 393 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
 394 {
 395         unsigned long pfn = pte_pfn(pte);
 396
 397         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
 398                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
 399                 if (pfn == vma->vm_pgoff + off)
 400                         return NULL;
 401                 if (!is_cow_mapping(vma->vm_flags))
 402                         return NULL;
 403         }
 404
 405 #ifdef CONFIG_DEBUG_VM
 406         /*
 407          * Add some anal sanity checks for now. Eventually,
 408          * we should just do "return pfn_to_page(pfn)", but
 409          * in the meantime we check that we get a valid pfn,
 410          * and that the resulting page looks ok.
 411          */
 412         if (unlikely(!pfn_valid(pfn))) {
 413                 print_bad_pte(vma, pte, addr);
 414                 return NULL;
 415         }
 416 #endif
 417
 418         /*
 419          * NOTE! We still have PageReserved() pages in the page
 420          * tables.
 421          *
 422          * The PAGE_ZERO() pages and various VDSO mappings can
 423          * cause them to exist.
 424          */
 425         return pfn_to_page(pfn);
 426 }
 427
 428 /*
 429  * copy one vm_area from one task to the other. Assumes the page tables
 430  * already present in the new task to be cleared in the whole range
 431  * covered by this vma.
 432  */
 433
 434 static inline void
 435 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 436                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 437                 unsigned long addr, int *rss)
 438 {
 439         unsigned long vm_flags = vma->vm_flags;
 440         pte_t pte = *src_pte;
 441         struct page *page;
 442
 443         /* pte contains position in swap or file, so copy. */
 444         if (unlikely(!pte_present(pte))) {
 445                 if (!pte_file(pte)) {
 446                         swp_entry_t entry = pte_to_swp_entry(pte);
 447
 448                         swap_duplicate(entry);
 449                         /* make sure dst_mm is on swapoff's mmlist. */
 450                         if (unlikely(list_empty(&dst_mm->mmlist))) {
 451                                 spin_lock(&mmlist_lock);
 452                                 if (list_empty(&dst_mm->mmlist))
 453                                         list_add(&dst_mm->mmlist,
 454                                                  &src_mm->mmlist);
 455                                 spin_unlock(&mmlist_lock);
 456                         }
 457                         if (is_write_migration_entry(entry) &&
 458                                         is_cow_mapping(vm_flags)) {
 459                                 /*
 460                                  * COW mappings require pages in both parent
 461                                  * and child to be set to read.
 462                                  */
 463                                 make_migration_entry_read(&entry);
 464                                 pte = swp_entry_to_pte(entry);
 465                                 set_pte_at(src_mm, addr, src_pte, pte);
 466                         }
 467                 }
 468                 goto out_set_pte;
 469         }
 470
 471         /*
 472          * If it's a COW mapping, write protect it both
 473          * in the parent and the child
 474          */
 475         if (is_cow_mapping(vm_flags)) {
 476                 ptep_set_wrprotect(src_mm, addr, src_pte);
 477                 pte = pte_wrprotect(pte);
 478         }
 479
 480         /*
 481          * If it's a shared mapping, mark it clean in
 482          * the child
 483          */
 484         if (vm_flags & VM_SHARED)
 485                 pte = pte_mkclean(pte);
 486         pte = pte_mkold(pte);
 487
 488         page = vm_normal_page(vma, addr, pte);
 489         if (page) {
 490                 get_page(page);
 491                 page_dup_rmap(page, vma, addr);
 492                 rss[!!PageAnon(page)]++;
 493         }
 494
 495 out_set_pte:
 496         set_pte_at(dst_mm, addr, dst_pte, pte);
 497 }
 498
 499 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 500                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 501                 unsigned long addr, unsigned long end)
 502 {
 503         pte_t *src_pte, *dst_pte;
 504         spinlock_t *src_ptl, *dst_ptl;
 505         int progress = 0;
 506         int rss[2];
 507
 508 again:
 509         rss[1] = rss[0] = 0;
 510         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 511         if (!dst_pte)
 512                 return -ENOMEM;
 513         src_pte = pte_offset_map_nested(src_pmd, addr);
 514         src_ptl = pte_lockptr(src_mm, src_pmd);
 515         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 516         arch_enter_lazy_mmu_mode();
 517
 518         do {
 519                 /*
 520                  * We are holding two locks at this point - either of them
 521                  * could generate latencies in another task on another CPU.
 522                  */
 523                 if (progress >= 32) {
 524                         progress = 0;
 525                         if (need_resched() ||
 526                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 527                                 break;
 528                 }
 529                 if (pte_none(*src_pte)) {
 530                         progress++;
 531                         continue;
 532                 }
 533                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
 534                 progress += 8;
 535         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 536
 537         arch_leave_lazy_mmu_mode();
 538         spin_unlock(src_ptl);
 539         pte_unmap_nested(src_pte - 1);
 540         add_mm_rss(dst_mm, rss[0], rss[1]);
 541         pte_unmap_unlock(dst_pte - 1, dst_ptl);
 542         cond_resched();
 543         if (addr != end)
 544                 goto again;
 545         return 0;
 546 }
 547
 548 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 549                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 550                 unsigned long addr, unsigned long end)
 551 {
 552         pmd_t *src_pmd, *dst_pmd;
 553         unsigned long next;
 554
 555         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 556         if (!dst_pmd)
 557                 return -ENOMEM;
 558         src_pmd = pmd_offset(src_pud, addr);
 559         do {
 560                 next = pmd_addr_end(addr, end);
 561                 if (pmd_none_or_clear_bad(src_pmd))
 562                         continue;
 563                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
 564                                                 vma, addr, next))
 565                         return -ENOMEM;
 566         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
 567         return 0;
 568 }
 569
 570 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 571                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
 572                 unsigned long addr, unsigned long end)
 573 {
 574         pud_t *src_pud, *dst_pud;
 575         unsigned long next;
 576
 577         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
 578         if (!dst_pud)
 579                 return -ENOMEM;
 580         src_pud = pud_offset(src_pgd, addr);
 581         do {
 582                 next = pud_addr_end(addr, end);
 583                 if (pud_none_or_clear_bad(src_pud))
 584                         continue;
 585                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
 586                                                 vma, addr, next))
 587                         return -ENOMEM;
 588         } while (dst_pud++, src_pud++, addr = next, addr != end);
 589         return 0;
 590 }
 591
 592 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 593                 struct vm_area_struct *vma)
 594 {
 595         pgd_t *src_pgd, *dst_pgd;
 596         unsigned long next;
 597         unsigned long addr = vma->vm_start;
 598         unsigned long end = vma->vm_end;
 599
 600         /*
 601          * Don't copy ptes where a page fault will fill them correctly.
 602          * Fork becomes much lighter when there are big shared or private
 603          * readonly mappings. The tradeoff is that copy_page_range is more
 604          * efficient than faulting.
 605          */
 606         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
 607                 if (!vma->anon_vma)
 608                         return 0;
 609         }
 610
 611         if (is_vm_hugetlb_page(vma))
 612                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
 613
 614         dst_pgd = pgd_offset(dst_mm, addr);
 615         src_pgd = pgd_offset(src_mm, addr);
 616         do {
 617                 next = pgd_addr_end(addr, end);
 618                 if (pgd_none_or_clear_bad(src_pgd))
 619                         continue;
 620                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
 621                                                 vma, addr, next))
 622                         return -ENOMEM;
 623         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
 624         return 0;
 625 }
 626
 627 static unsigned long zap_pte_range(struct mmu_gather *tlb,
 628                                 struct vm_area_struct *vma, pmd_t *pmd,
 629                                 unsigned long addr, unsigned long end,
 630                                 long *zap_work, struct zap_details *details)
 631 {
 632         struct mm_struct *mm = tlb->mm;
 633         pte_t *pte;
 634         spinlock_t *ptl;
 635         int file_rss = 0;
 636         int anon_rss = 0;
 637
 638         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 639         arch_enter_lazy_mmu_mode();
 640         do {
 641                 pte_t ptent = *pte;
 642                 if (pte_none(ptent)) {
 643                         (*zap_work)--;
 644                         continue;
 645                 }
 646
 647                 (*zap_work) -= PAGE_SIZE;
 648
 649                 if (pte_present(ptent)) {
 650                         struct page *page;
 651
 652                         page = vm_normal_page(vma, addr, ptent);
 653                         if (unlikely(details) && page) {
 654                                 /*
 655                                  * unmap_shared_mapping_pages() wants to
 656                                  * invalidate cache without truncating:
 657                                  * unmap shared but keep private pages.
 658                                  */
 659                                 if (details->check_mapping &&
 660                                     details->check_mapping != page->mapping)
 661                                         continue;
 662                                 /*
 663                                  * Each page->index must be checked when
 664                                  * invalidating or truncating nonlinear.
 665                                  */
 666                                 if (details->nonlinear_vma &&
 667                                     (page->index < details->first_index ||
 668                                      page->index > details->last_index))
 669                                         continue;
 670                         }
 671                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 672                                                         tlb->fullmm);
 673                         tlb_remove_tlb_entry(tlb, pte, addr);
 674                         if (unlikely(!page))
 675                                 continue;
 676                         if (unlikely(details) && details->nonlinear_vma
 677                             && linear_page_index(details->nonlinear_vma,
 678                                                 addr) != page->index)
 679                                 set_pte_at(mm, addr, pte,
 680                                            pgoff_to_pte(page->index));
 681                         if (PageAnon(page))
 682                                 anon_rss--;
 683                         else {
 684                                 if (pte_dirty(ptent))
 685                                         set_page_dirty(page);
 686                                 if (pte_young(ptent))
 687                                         SetPageReferenced(page);
 688                                 file_rss--;
 689                         }
 690                         page_remove_rmap(page, vma);
 691                         tlb_remove_page(tlb, page);
 692                         continue;
 693                 }
 694                 /*
 695                  * If details->check_mapping, we leave swap entries;
 696                  * if details->nonlinear_vma, we leave file entries.
 697                  */
 698                 if (unlikely(details))
 699                         continue;
 700                 if (!pte_file(ptent))
 701                         free_swap_and_cache(pte_to_swp_entry(ptent));
 702                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 703         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
 704
 705         add_mm_rss(mm, file_rss, anon_rss);
 706         arch_leave_lazy_mmu_mode();
 707         pte_unmap_unlock(pte - 1, ptl);
 708
 709         return addr;
 710 }
 711
 712 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
 713                                 struct vm_area_struct *vma, pud_t *pud,
 714                                 unsigned long addr, unsigned long end,
 715                                 long *zap_work, struct zap_details *details)
 716 {
 717         pmd_t *pmd;
 718         unsigned long next;
 719
 720         pmd = pmd_offset(pud, addr);
 721         do {
 722                 next = pmd_addr_end(addr, end);
 723                 if (pmd_none_or_clear_bad(pmd)) {
 724                         (*zap_work)--;
 725                         continue;
 726                 }
 727                 next = zap_pte_range(tlb, vma, pmd, addr, next,
 728                                                 zap_work, details);
 729         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
 730
 731         return addr;
 732 }
 733
 734 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
 735                                 struct vm_area_struct *vma, pgd_t *pgd,
 736                                 unsigned long addr, unsigned long end,
 737                                 long *zap_work, struct zap_details *details)
 738 {
 739         pud_t *pud;
 740         unsigned long next;
 741
 742         pud = pud_offset(pgd, addr);
 743         do {
 744                 next = pud_addr_end(addr, end);
 745                 if (pud_none_or_clear_bad(pud)) {
 746                         (*zap_work)--;
 747                         continue;
 748                 }
 749                 next = zap_pmd_range(tlb, vma, pud, addr, next,
 750                                                 zap_work, details);
 751         } while (pud++, addr = next, (addr != end && *zap_work > 0));
 752
 753         return addr;
 754 }
 755
 756 static unsigned long unmap_page_range(struct mmu_gather *tlb,
 757                                 struct vm_area_struct *vma,
 758                                 unsigned long addr, unsigned long end,
 759                                 long *zap_work, struct zap_details *details)
 760 {
 761         pgd_t *pgd;
 762         unsigned long next;
 763
 764         if (details && !details->check_mapping && !details->nonlinear_vma)
 765                 details = NULL;
 766
 767         BUG_ON(addr >= end);
 768         tlb_start_vma(tlb, vma);
 769         pgd = pgd_offset(vma->vm_mm, addr);
 770         do {
 771                 next = pgd_addr_end(addr, end);
 772                 if (pgd_none_or_clear_bad(pgd)) {
 773                         (*zap_work)--;
 774                         continue;
 775                 }
 776                 next = zap_pud_range(tlb, vma, pgd, addr, next,
 777                                                 zap_work, details);
 778         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
 779         tlb_end_vma(tlb, vma);
 780
 781         return addr;
 782 }
 783
 784 #ifdef CONFIG_PREEMPT
 785 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
 786 #else
 787 /* No preempt: go for improved straight-line efficiency */
 788 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
 789 #endif
 790
 791 /**
 792  * unmap_vmas - unmap a range of memory covered by a list of vma's
 793  * @tlbp: address of the caller's struct mmu_gather
 794  * @vma: the starting vma
 795  * @start_addr: virtual address at which to start unmapping
 796  * @end_addr: virtual address at which to end unmapping
 797  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 798  * @details: details of nonlinear truncation or shared cache invalidation
 799  *
 800  * Returns the end address of the unmapping (restart addr if interrupted).
 801  *
 802  * Unmap all pages in the vma list.
 803  *
 804  * We aim to not hold locks for too long (for scheduling latency reasons).
 805  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 806  * return the ending mmu_gather to the caller.
 807  *
 808  * Only addresses between `start' and `end' will be unmapped.
 809  *
 810  * The VMA list must be sorted in ascending virtual address order.
 811  *
 812  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 813  * range after unmap_vmas() returns.  So the only responsibility here is to
 814  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 815  * drops the lock and schedules.
 816  */
 817 unsigned long unmap_vmas(struct mmu_gather **tlbp,
 818                 struct vm_area_struct *vma, unsigned long start_addr,
 819                 unsigned long end_addr, unsigned long *nr_accounted,
 820                 struct zap_details *details)
 821 {
 822         long zap_work = ZAP_BLOCK_SIZE;
 823         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 824         int tlb_start_valid = 0;
 825         unsigned long start = start_addr;
 826         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
 827         int fullmm = (*tlbp)->fullmm;
 828
 829         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 830                 unsigned long end;
 831
 832                 start = max(vma->vm_start, start_addr);
 833                 if (start >= vma->vm_end)
 834                         continue;
 835                 end = min(vma->vm_end, end_addr);
 836                 if (end <= vma->vm_start)
 837                         continue;
 838
 839                 if (vma->vm_flags & VM_ACCOUNT)
 840                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 841
 842                 while (start != end) {
 843                         if (!tlb_start_valid) {
 844                                 tlb_start = start;
 845                                 tlb_start_valid = 1;
 846                         }
 847
 848                         if (unlikely(is_vm_hugetlb_page(vma))) {
 849                                 unmap_hugepage_range(vma, start, end);
 850                                 zap_work -= (end - start) /
 851                                                 (HPAGE_SIZE / PAGE_SIZE);
 852                                 start = end;
 853                         } else
 854                                 start = unmap_page_range(*tlbp, vma,
 855                                                 start, end, &zap_work, details);
 856
 857                         if (zap_work > 0) {
 858                                 BUG_ON(start != end);
 859                                 break;
 860                         }
 861
 862                         tlb_finish_mmu(*tlbp, tlb_start, start);
 863
 864                         if (need_resched() ||
 865                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
 866                                 if (i_mmap_lock) {
 867                                         *tlbp = NULL;
 868                                         goto out;
 869                                 }
 870                                 cond_resched();
 871                         }
 872
 873                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
 874                         tlb_start_valid = 0;
 875                         zap_work = ZAP_BLOCK_SIZE;
 876                 }
 877         }
 878 out:
 879         return start;   /* which is now the end (or restart) address */
 880 }
 881
 882 /**
 883  * zap_page_range - remove user pages in a given range
 884  * @vma: vm_area_struct holding the applicable pages
 885  * @address: starting address of pages to zap
 886  * @size: number of bytes to zap
 887  * @details: details of nonlinear truncation or shared cache invalidation
 888  */
 889 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
 890                 unsigned long size, struct zap_details *details)
 891 {
 892         struct mm_struct *mm = vma->vm_mm;
 893         struct mmu_gather *tlb;
 894         unsigned long end = address + size;
 895         unsigned long nr_accounted = 0;
 896
 897         lru_add_drain();
 898         tlb = tlb_gather_mmu(mm, 0);
 899         update_hiwater_rss(mm);
 900         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
 901         if (tlb)
 902                 tlb_finish_mmu(tlb, address, end);
 903         return end;
 904 }
 905
 906 /*
 907  * Do a quick page-table lookup for a single page.
 908  */
 909 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
 910                         unsigned int flags)
 911 {
 912         pgd_t *pgd;
 913         pud_t *pud;
 914         pmd_t *pmd;
 915         pte_t *ptep, pte;
 916         spinlock_t *ptl;
 917         struct page *page;
 918         struct mm_struct *mm = vma->vm_mm;
 919
 920         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 921         if (!IS_ERR(page)) {
 922                 BUG_ON(flags & FOLL_GET);
 923                 goto out;
 924         }
 925
 926         page = NULL;
 927         pgd = pgd_offset(mm, address);
 928         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 929                 goto no_page_table;
 930
 931         pud = pud_offset(pgd, address);
 932         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
 933                 goto no_page_table;
 934
 935         pmd = pmd_offset(pud, address);
 936         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 937                 goto no_page_table;
 938
 939         if (pmd_huge(*pmd)) {
 940                 BUG_ON(flags & FOLL_GET);
 941                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
 942                 goto out;
 943         }
 944
 945         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 946         if (!ptep)
 947                 goto out;
 948
 949         pte = *ptep;
 950         if (!pte_present(pte))
 951                 goto unlock;
 952         if ((flags & FOLL_WRITE) && !pte_write(pte))
 953                 goto unlock;
 954         page = vm_normal_page(vma, address, pte);
 955         if (unlikely(!page))
 956                 goto unlock;
 957
 958         if (flags & FOLL_GET)
 959                 get_page(page);
 960         if (flags & FOLL_TOUCH) {
 961                 if ((flags & FOLL_WRITE) &&
 962                     !pte_dirty(pte) && !PageDirty(page))
 963                         set_page_dirty(page);
 964                 mark_page_accessed(page);
 965         }
 966 unlock:
 967         pte_unmap_unlock(ptep, ptl);
 968 out:
 969         return page;
 970
 971 no_page_table:
 972         /*
 973          * When core dumping an enormous anonymous area that nobody
 974          * has touched so far, we don't want to allocate page tables.
 975          */
 976         if (flags & FOLL_ANON) {
 977                 page = ZERO_PAGE(0);
 978                 if (flags & FOLL_GET)
 979                         get_page(page);
 980                 BUG_ON(flags & FOLL_WRITE);
 981         }
 982         return page;
 983 }
 984
 985 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 986                 unsigned long start, int len, int write, int force,
 987                 struct page **pages, struct vm_area_struct **vmas)
 988 {
 989         int i;
 990         unsigned int vm_flags;
 991
 992         if (len <= 0)
 993                 return 0;
 994         /*
 995          * Require read or write permissions.
 996          * If 'force' is set, we only require the "MAY" flags.
 997          */
 998         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 999         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1000         i = 0;
1001
1002         do {
1003                 struct vm_area_struct *vma;
1004                 unsigned int foll_flags;
1005
1006                 vma = find_extend_vma(mm, start);
1007                 if (!vma && in_gate_area(tsk, start)) {
1008                         unsigned long pg = start & PAGE_MASK;
1009                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1010                         pgd_t *pgd;
1011                         pud_t *pud;
1012                         pmd_t *pmd;
1013                         pte_t *pte;
1014                         if (write) /* user gate pages are read-only */
1015                                 return i ? : -EFAULT;
1016                         if (pg > TASK_SIZE)
1017                                 pgd = pgd_offset_k(pg);
1018                         else
1019                                 pgd = pgd_offset_gate(mm, pg);
1020                         BUG_ON(pgd_none(*pgd));
1021                         pud = pud_offset(pgd, pg);
1022                         BUG_ON(pud_none(*pud));
1023                         pmd = pmd_offset(pud, pg);
1024                         if (pmd_none(*pmd))
1025                                 return i ? : -EFAULT;
1026                         pte = pte_offset_map(pmd, pg);
1027                         if (pte_none(*pte)) {
1028                                 pte_unmap(pte);
1029                                 return i ? : -EFAULT;
1030                         }
1031                         if (pages) {
1032                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1033                                 pages[i] = page;
1034                                 if (page)
1035                                         get_page(page);
1036                         }
1037                         pte_unmap(pte);
1038                         if (vmas)
1039                                 vmas[i] = gate_vma;
1040                         i++;
1041                         start += PAGE_SIZE;
1042                         len--;
1043                         continue;
1044                 }
1045
1046                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1047                                 || !(vm_flags & vma->vm_flags))
1048                         return i ? : -EFAULT;
1049
1050                 if (is_vm_hugetlb_page(vma)) {
1051                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1052                                                 &start, &len, i, write);
1053                         continue;
1054                 }
1055
1056                 foll_flags = FOLL_TOUCH;
1057                 if (pages)
1058                         foll_flags |= FOLL_GET;
1059                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1060                     (!vma->vm_ops || (!vma->vm_ops->nopage &&
1061                                         !vma->vm_ops->fault)))
1062                         foll_flags |= FOLL_ANON;
1063
1064                 do {
1065                         struct page *page;
1066
1067                         /*
1068                          * If tsk is ooming, cut off its access to large memory
1069                          * allocations. It has a pending SIGKILL, but it can't
1070                          * be processed until returning to user space.
1071                          */
1072                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1073                                 return -ENOMEM;
1074
1075                         if (write)
1076                                 foll_flags |= FOLL_WRITE;
1077
1078                         cond_resched();
1079                         while (!(page = follow_page(vma, start, foll_flags))) {
1080                                 int ret;
1081                                 ret = handle_mm_fault(mm, vma, start,
1082                                                 foll_flags & FOLL_WRITE);
1083                                 if (ret & VM_FAULT_ERROR) {
1084                                         if (ret & VM_FAULT_OOM)
1085                                                 return i ? i : -ENOMEM;
1086                                         else if (ret & VM_FAULT_SIGBUS)
1087                                                 return i ? i : -EFAULT;
1088                                         BUG();
1089                                 }
1090                                 if (ret & VM_FAULT_MAJOR)
1091                                         tsk->maj_flt++;
1092                                 else
1093                                         tsk->min_flt++;
1094
1095                                 /*
1096                                  * The VM_FAULT_WRITE bit tells us that
1097                                  * do_wp_page has broken COW when necessary,
1098                                  * even if maybe_mkwrite decided not to set
1099                                  * pte_write. We can thus safely do subsequent
1100                                  * page lookups as if they were reads.
1101                                  */
1102                                 if (ret & VM_FAULT_WRITE)
1103                                         foll_flags &= ~FOLL_WRITE;
1104
1105                                 cond_resched();
1106                         }
1107                         if (pages) {
1108                                 pages[i] = page;
1109
1110                                 flush_anon_page(vma, page, start);
1111                                 flush_dcache_page(page);
1112                         }
1113                         if (vmas)
1114                                 vmas[i] = vma;
1115                         i++;
1116                         start += PAGE_SIZE;
1117                         len--;
1118                 } while (len && start < vma->vm_end);
1119         } while (len);
1120         return i;
1121 }
1122 EXPORT_SYMBOL(get_user_pages);
1123
1124 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1125                         spinlock_t **ptl)
1126 {
1127         pgd_t * pgd = pgd_offset(mm, addr);
1128         pud_t * pud = pud_alloc(mm, pgd, addr);
1129         if (pud) {
1130                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1131                 if (pmd)
1132                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1133         }
1134         return NULL;
1135 }
1136
1137 /*
1138  * This is the old fallback for page remapping.
1139  *
1140  * For historical reasons, it only allows reserved pages. Only
1141  * old drivers should use this, and they needed to mark their
1142  * pages reserved for the old functions anyway.
1143  */
1144 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1145 {
1146         int retval;
1147         pte_t *pte;
1148         spinlock_t *ptl;
1149
1150         retval = mem_cgroup_charge(page, mm, GFP_KERNEL);
1151         if (retval)
1152                 goto out;
1153
1154         retval = -EINVAL;
1155         if (PageAnon(page))
1156                 goto out_uncharge;
1157         retval = -ENOMEM;
1158         flush_dcache_page(page);
1159         pte = get_locked_pte(mm, addr, &ptl);
1160         if (!pte)
1161                 goto out_uncharge;
1162         retval = -EBUSY;
1163         if (!pte_none(*pte))
1164                 goto out_unlock;
1165
1166         /* Ok, finally just insert the thing.. */
1167         get_page(page);
1168         inc_mm_counter(mm, file_rss);
1169         page_add_file_rmap(page);
1170         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1171
1172         retval = 0;
1173         pte_unmap_unlock(pte, ptl);
1174         return retval;
1175 out_unlock:
1176         pte_unmap_unlock(pte, ptl);
1177 out_uncharge:
1178         mem_cgroup_uncharge_page(page);
1179 out:
1180         return retval;
1181 }
1182
1183 /**
1184  * vm_insert_page - insert single page into user vma
1185  * @vma: user vma to map to
1186  * @addr: target user address of this page
1187  * @page: source kernel page
1188  *
1189  * This allows drivers to insert individual pages they've allocated
1190  * into a user vma.
1191  *
1192  * The page has to be a nice clean _individual_ kernel allocation.
1193  * If you allocate a compound page, you need to have marked it as
1194  * such (__GFP_COMP), or manually just split the page up yourself
1195  * (see split_page()).
1196  *
1197  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1198  * took an arbitrary page protection parameter. This doesn't allow
1199  * that. Your vma protection will have to be set up correctly, which
1200  * means that if you want a shared writable mapping, you'd better
1201  * ask for a shared writable mapping!
1202  *
1203  * The page does not need to be reserved.
1204  */
1205 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1206 {
1207         if (addr < vma->vm_start || addr >= vma->vm_end)
1208                 return -EFAULT;
1209         if (!page_count(page))
1210                 return -EINVAL;
1211         vma->vm_flags |= VM_INSERTPAGE;
1212         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1213 }
1214 EXPORT_SYMBOL(vm_insert_page);
1215
1216 /**
1217  * vm_insert_pfn - insert single pfn into user vma
1218  * @vma: user vma to map to
1219  * @addr: target user address of this page
1220  * @pfn: source kernel pfn
1221  *
1222  * Similar to vm_inert_page, this allows drivers to insert individual pages
1223  * they've allocated into a user vma. Same comments apply.
1224  *
1225  * This function should only be called from a vm_ops->fault handler, and
1226  * in that case the handler should return NULL.
1227  */
1228 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1229                 unsigned long pfn)
1230 {
1231         struct mm_struct *mm = vma->vm_mm;
1232         int retval;
1233         pte_t *pte, entry;
1234         spinlock_t *ptl;
1235
1236         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
1237         BUG_ON(is_cow_mapping(vma->vm_flags));
1238
1239         retval = -ENOMEM;
1240         pte = get_locked_pte(mm, addr, &ptl);
1241         if (!pte)
1242                 goto out;
1243         retval = -EBUSY;
1244         if (!pte_none(*pte))
1245                 goto out_unlock;
1246
1247         /* Ok, finally just insert the thing.. */
1248         entry = pfn_pte(pfn, vma->vm_page_prot);
1249         set_pte_at(mm, addr, pte, entry);
1250         update_mmu_cache(vma, addr, entry);
1251
1252         retval = 0;
1253 out_unlock:
1254         pte_unmap_unlock(pte, ptl);
1255
1256 out:
1257         return retval;
1258 }
1259 EXPORT_SYMBOL(vm_insert_pfn);
1260
1261 /*
1262  * maps a range of physical memory into the requested pages. the old
1263  * mappings are removed. any references to nonexistent pages results
1264  * in null mappings (currently treated as "copy-on-access")
1265  */
1266 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1267                         unsigned long addr, unsigned long end,
1268                         unsigned long pfn, pgprot_t prot)
1269 {
1270         pte_t *pte;
1271         spinlock_t *ptl;
1272
1273         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1274         if (!pte)
1275                 return -ENOMEM;
1276         arch_enter_lazy_mmu_mode();
1277         do {
1278                 BUG_ON(!pte_none(*pte));
1279                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1280                 pfn++;
1281         } while (pte++, addr += PAGE_SIZE, addr != end);
1282         arch_leave_lazy_mmu_mode();
1283         pte_unmap_unlock(pte - 1, ptl);
1284         return 0;
1285 }
1286
1287 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1288                         unsigned long addr, unsigned long end,
1289                         unsigned long pfn, pgprot_t prot)
1290 {
1291         pmd_t *pmd;
1292         unsigned long next;
1293
1294         pfn -= addr >> PAGE_SHIFT;
1295         pmd = pmd_alloc(mm, pud, addr);
1296         if (!pmd)
1297                 return -ENOMEM;
1298         do {
1299                 next = pmd_addr_end(addr, end);
1300                 if (remap_pte_range(mm, pmd, addr, next,
1301                                 pfn + (addr >> PAGE_SHIFT), prot))
1302                         return -ENOMEM;
1303         } while (pmd++, addr = next, addr != end);
1304         return 0;
1305 }
1306
1307 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1308                         unsigned long addr, unsigned long end,
1309                         unsigned long pfn, pgprot_t prot)
1310 {
1311         pud_t *pud;
1312         unsigned long next;
1313
1314         pfn -= addr >> PAGE_SHIFT;
1315         pud = pud_alloc(mm, pgd, addr);
1316         if (!pud)
1317                 return -ENOMEM;
1318         do {
1319                 next = pud_addr_end(addr, end);
1320                 if (remap_pmd_range(mm, pud, addr, next,
1321                                 pfn + (addr >> PAGE_SHIFT), prot))
1322                         return -ENOMEM;
1323         } while (pud++, addr = next, addr != end);
1324         return 0;
1325 }
1326
1327 /**
1328  * remap_pfn_range - remap kernel memory to userspace
1329  * @vma: user vma to map to
1330  * @addr: target user address to start at
1331  * @pfn: physical address of kernel memory
1332  * @size: size of map area
1333  * @prot: page protection flags for this mapping
1334  *
1335  *  Note: this is only safe if the mm semaphore is held when called.
1336  */
1337 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1338                     unsigned long pfn, unsigned long size, pgprot_t prot)
1339 {
1340         pgd_t *pgd;
1341         unsigned long next;
1342         unsigned long end = addr + PAGE_ALIGN(size);
1343         struct mm_struct *mm = vma->vm_mm;
1344         int err;
1345
1346         /*
1347          * Physically remapped pages are special. Tell the
1348          * rest of the world about it:
1349          *   VM_IO tells people not to look at these pages
1350          *      (accesses can have side effects).
1351          *   VM_RESERVED is specified all over the place, because
1352          *      in 2.4 it kept swapout's vma scan off this vma; but
1353          *      in 2.6 the LRU scan won't even find its pages, so this
1354          *      flag means no more than count its pages in reserved_vm,
1355          *      and omit it from core dump, even when VM_IO turned off.
1356          *   VM_PFNMAP tells the core MM that the base pages are just
1357          *      raw PFN mappings, and do not have a "struct page" associated
1358          *      with them.
1359          *
1360          * There's a horrible special case to handle copy-on-write
1361          * behaviour that some programs depend on. We mark the "original"
1362          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1363          */
1364         if (is_cow_mapping(vma->vm_flags)) {
1365                 if (addr != vma->vm_start || end != vma->vm_end)
1366                         return -EINVAL;
1367                 vma->vm_pgoff = pfn;
1368         }
1369
1370         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1371
1372         BUG_ON(addr >= end);
1373         pfn -= addr >> PAGE_SHIFT;
1374         pgd = pgd_offset(mm, addr);
1375         flush_cache_range(vma, addr, end);
1376         do {
1377                 next = pgd_addr_end(addr, end);
1378                 err = remap_pud_range(mm, pgd, addr, next,
1379                                 pfn + (addr >> PAGE_SHIFT), prot);
1380                 if (err)
1381                         break;
1382         } while (pgd++, addr = next, addr != end);
1383         return err;
1384 }
1385 EXPORT_SYMBOL(remap_pfn_range);
1386
1387 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1388                                      unsigned long addr, unsigned long end,
1389                                      pte_fn_t fn, void *data)
1390 {
1391         pte_t *pte;
1392         int err;
1393         pgtable_t token;
1394         spinlock_t *uninitialized_var(ptl);
1395
1396         pte = (mm == &init_mm) ?
1397                 pte_alloc_kernel(pmd, addr) :
1398                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1399         if (!pte)
1400                 return -ENOMEM;
1401
1402         BUG_ON(pmd_huge(*pmd));
1403
1404         token = pmd_pgtable(*pmd);
1405
1406         do {
1407                 err = fn(pte, token, addr, data);
1408                 if (err)
1409                         break;
1410         } while (pte++, addr += PAGE_SIZE, addr != end);
1411
1412         if (mm != &init_mm)
1413                 pte_unmap_unlock(pte-1, ptl);
1414         return err;
1415 }
1416
1417 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1418                                      unsigned long addr, unsigned long end,
1419                                      pte_fn_t fn, void *data)
1420 {
1421         pmd_t *pmd;
1422         unsigned long next;
1423         int err;
1424
1425         pmd = pmd_alloc(mm, pud, addr);
1426         if (!pmd)
1427                 return -ENOMEM;
1428         do {
1429                 next = pmd_addr_end(addr, end);
1430                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1431                 if (err)
1432                         break;
1433         } while (pmd++, addr = next, addr != end);
1434         return err;
1435 }
1436
1437 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1438                                      unsigned long addr, unsigned long end,
1439                                      pte_fn_t fn, void *data)
1440 {
1441         pud_t *pud;
1442         unsigned long next;
1443         int err;
1444
1445         pud = pud_alloc(mm, pgd, addr);
1446         if (!pud)
1447                 return -ENOMEM;
1448         do {
1449                 next = pud_addr_end(addr, end);
1450                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1451                 if (err)
1452                         break;
1453         } while (pud++, addr = next, addr != end);
1454         return err;
1455 }
1456
1457 /*
1458  * Scan a region of virtual memory, filling in page tables as necessary
1459  * and calling a provided function on each leaf page table.
1460  */
1461 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1462                         unsigned long size, pte_fn_t fn, void *data)
1463 {
1464         pgd_t *pgd;
1465         unsigned long next;
1466         unsigned long end = addr + size;
1467         int err;
1468
1469         BUG_ON(addr >= end);
1470         pgd = pgd_offset(mm, addr);
1471         do {
1472                 next = pgd_addr_end(addr, end);
1473                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1474                 if (err)
1475                         break;
1476         } while (pgd++, addr = next, addr != end);
1477         return err;
1478 }
1479 EXPORT_SYMBOL_GPL(apply_to_page_range);
1480
1481 /*
1482  * handle_pte_fault chooses page fault handler according to an entry
1483  * which was read non-atomically.  Before making any commitment, on
1484  * those architectures or configurations (e.g. i386 with PAE) which
1485  * might give a mix of unmatched parts, do_swap_page and do_file_page
1486  * must check under lock before unmapping the pte and proceeding
1487  * (but do_wp_page is only called after already making such a check;
1488  * and do_anonymous_page and do_no_page can safely check later on).
1489  */
1490 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1491                                 pte_t *page_table, pte_t orig_pte)
1492 {
1493         int same = 1;
1494 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1495         if (sizeof(pte_t) > sizeof(unsigned long)) {
1496                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1497                 spin_lock(ptl);
1498                 same = pte_same(*page_table, orig_pte);
1499                 spin_unlock(ptl);
1500         }
1501 #endif
1502         pte_unmap(page_table);
1503         return same;
1504 }
1505
1506 /*
1507  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1508  * servicing faults for write access.  In the normal case, do always want
1509  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1510  * that do not have writing enabled, when used by access_process_vm.
1511  */
1512 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1513 {
1514         if (likely(vma->vm_flags & VM_WRITE))
1515                 pte = pte_mkwrite(pte);
1516         return pte;
1517 }
1518
1519 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1520 {
1521         /*
1522          * If the source page was a PFN mapping, we don't have
1523          * a "struct page" for it. We do a best-effort copy by
1524          * just copying from the original user address. If that
1525          * fails, we just zero-fill it. Live with it.
1526          */
1527         if (unlikely(!src)) {
1528                 void *kaddr = kmap_atomic(dst, KM_USER0);
1529                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1530
1531                 /*
1532                  * This really shouldn't fail, because the page is there
1533                  * in the page tables. But it might just be unreadable,
1534                  * in which case we just give up and fill the result with
1535                  * zeroes.
1536                  */
1537                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1538                         memset(kaddr, 0, PAGE_SIZE);
1539                 kunmap_atomic(kaddr, KM_USER0);
1540                 flush_dcache_page(dst);
1541         } else
1542                 copy_user_highpage(dst, src, va, vma);
1543 }
1544
1545 /*
1546  * This routine handles present pages, when users try to write
1547  * to a shared page. It is done by copying the page to a new address
1548  * and decrementing the shared-page counter for the old page.
1549  *
1550  * Note that this routine assumes that the protection checks have been
1551  * done by the caller (the low-level page fault routine in most cases).
1552  * Thus we can safely just mark it writable once we've done any necessary
1553  * COW.
1554  *
1555  * We also mark the page dirty at this point even though the page will
1556  * change only once the write actually happens. This avoids a few races,
1557  * and potentially makes it more efficient.
1558  *
1559  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1560  * but allow concurrent faults), with pte both mapped and locked.
1561  * We return with mmap_sem still held, but pte unmapped and unlocked.
1562  */
1563 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1564                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1565                 spinlock_t *ptl, pte_t orig_pte)
1566 {
1567         struct page *old_page, *new_page;
1568         pte_t entry;
1569         int reuse = 0, ret = 0;
1570         int page_mkwrite = 0;
1571         struct page *dirty_page = NULL;
1572
1573         old_page = vm_normal_page(vma, address, orig_pte);
1574         if (!old_page)
1575                 goto gotten;
1576
1577         /*
1578          * Take out anonymous pages first, anonymous shared vmas are
1579          * not dirty accountable.
1580          */
1581         if (PageAnon(old_page)) {
1582                 if (!TestSetPageLocked(old_page)) {
1583                         reuse = can_share_swap_page(old_page);
1584                         unlock_page(old_page);
1585                 }
1586         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1587                                         (VM_WRITE|VM_SHARED))) {
1588                 /*
1589                  * Only catch write-faults on shared writable pages,
1590                  * read-only shared pages can get COWed by
1591                  * get_user_pages(.write=1, .force=1).
1592                  */
1593                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1594                         /*
1595                          * Notify the address space that the page is about to
1596                          * become writable so that it can prohibit this or wait
1597                          * for the page to get into an appropriate state.
1598                          *
1599                          * We do this without the lock held, so that it can
1600                          * sleep if it needs to.
1601                          */
1602                         page_cache_get(old_page);
1603                         pte_unmap_unlock(page_table, ptl);
1604
1605                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1606                                 goto unwritable_page;
1607
1608                         /*
1609                          * Since we dropped the lock we need to revalidate
1610                          * the PTE as someone else may have changed it.  If
1611                          * they did, we just return, as we can count on the
1612                          * MMU to tell us if they didn't also make it writable.
1613                          */
1614                         page_table = pte_offset_map_lock(mm, pmd, address,
1615                                                          &ptl);
1616                         page_cache_release(old_page);
1617                         if (!pte_same(*page_table, orig_pte))
1618                                 goto unlock;
1619
1620                         page_mkwrite = 1;
1621                 }
1622                 dirty_page = old_page;
1623                 get_page(dirty_page);
1624                 reuse = 1;
1625         }
1626
1627         if (reuse) {
1628                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1629                 entry = pte_mkyoung(orig_pte);
1630                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1631                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1632                         update_mmu_cache(vma, address, entry);
1633                 ret |= VM_FAULT_WRITE;
1634                 goto unlock;
1635         }
1636
1637         /*
1638          * Ok, we need to copy. Oh, well..
1639          */
1640         page_cache_get(old_page);
1641 gotten:
1642         pte_unmap_unlock(page_table, ptl);
1643
1644         if (unlikely(anon_vma_prepare(vma)))
1645                 goto oom;
1646         VM_BUG_ON(old_page == ZERO_PAGE(0));
1647         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1648         if (!new_page)
1649                 goto oom;
1650         cow_user_page(new_page, old_page, address, vma);
1651         __SetPageUptodate(new_page);
1652
1653         if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
1654                 goto oom_free_new;
1655
1656         /*
1657          * Re-check the pte - we dropped the lock
1658          */
1659         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1660         if (likely(pte_same(*page_table, orig_pte))) {
1661                 if (old_page) {
1662                         page_remove_rmap(old_page, vma);
1663                         if (!PageAnon(old_page)) {
1664                                 dec_mm_counter(mm, file_rss);
1665                                 inc_mm_counter(mm, anon_rss);
1666                         }
1667                 } else
1668                         inc_mm_counter(mm, anon_rss);
1669                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1670                 entry = mk_pte(new_page, vma->vm_page_prot);
1671                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1672                 /*
1673                  * Clear the pte entry and flush it first, before updating the
1674                  * pte with the new entry. This will avoid a race condition
1675                  * seen in the presence of one thread doing SMC and another
1676                  * thread doing COW.
1677                  */
1678                 ptep_clear_flush(vma, address, page_table);
1679                 set_pte_at(mm, address, page_table, entry);
1680                 update_mmu_cache(vma, address, entry);
1681                 lru_cache_add_active(new_page);
1682                 page_add_new_anon_rmap(new_page, vma, address);
1683
1684                 /* Free the old page.. */
1685                 new_page = old_page;
1686                 ret |= VM_FAULT_WRITE;
1687         } else
1688                 mem_cgroup_uncharge_page(new_page);
1689
1690         if (new_page)
1691                 page_cache_release(new_page);
1692         if (old_page)
1693                 page_cache_release(old_page);
1694 unlock:
1695         pte_unmap_unlock(page_table, ptl);
1696         if (dirty_page) {
1697                 if (vma->vm_file)
1698                         file_update_time(vma->vm_file);
1699
1700                 /*
1701                  * Yes, Virginia, this is actually required to prevent a race
1702                  * with clear_page_dirty_for_io() from clearing the page dirty
1703                  * bit after it clear all dirty ptes, but before a racing
1704                  * do_wp_page installs a dirty pte.
1705                  *
1706                  * do_no_page is protected similarly.
1707                  */
1708                 wait_on_page_locked(dirty_page);
1709                 set_page_dirty_balance(dirty_page, page_mkwrite);
1710                 put_page(dirty_page);
1711         }
1712         return ret;
1713 oom_free_new:
1714         page_cache_release(new_page);
1715 oom:
1716         if (old_page)
1717                 page_cache_release(old_page);
1718         return VM_FAULT_OOM;
1719
1720 unwritable_page:
1721         page_cache_release(old_page);
1722         return VM_FAULT_SIGBUS;
1723 }
1724
1725 /*
1726  * Helper functions for unmap_mapping_range().
1727  *
1728  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1729  *
1730  * We have to restart searching the prio_tree whenever we drop the lock,
1731  * since the iterator is only valid while the lock is held, and anyway
1732  * a later vma might be split and reinserted earlier while lock dropped.
1733  *
1734  * The list of nonlinear vmas could be handled more efficiently, using
1735  * a placeholder, but handle it in the same way until a need is shown.
1736  * It is important to search the prio_tree before nonlinear list: a vma
1737  * may become nonlinear and be shifted from prio_tree to nonlinear list
1738  * while the lock is dropped; but never shifted from list to prio_tree.
1739  *
1740  * In order to make forward progress despite restarting the search,
1741  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1742  * quickly skip it next time around.  Since the prio_tree search only
1743  * shows us those vmas affected by unmapping the range in question, we
1744  * can't efficiently keep all vmas in step with mapping->truncate_count:
1745  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1746  * mapping->truncate_count and vma->vm_truncate_count are protected by
1747  * i_mmap_lock.
1748  *
1749  * In order to make forward progress despite repeatedly restarting some
1750  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1751  * and restart from that address when we reach that vma again.  It might
1752  * have been split or merged, shrunk or extended, but never shifted: so
1753  * restart_addr remains valid so long as it remains in the vma's range.
1754  * unmap_mapping_range forces truncate_count to leap over page-aligned
1755  * values so we can save vma's restart_addr in its truncate_count field.
1756  */
1757 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1758
1759 static void reset_vma_truncate_counts(struct address_space *mapping)
1760 {
1761         struct vm_area_struct *vma;
1762         struct prio_tree_iter iter;
1763
1764         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1765                 vma->vm_truncate_count = 0;
1766         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1767                 vma->vm_truncate_count = 0;
1768 }
1769
1770 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1771                 unsigned long start_addr, unsigned long end_addr,
1772                 struct zap_details *details)
1773 {
1774         unsigned long restart_addr;
1775         int need_break;
1776
1777         /*
1778          * files that support invalidating or truncating portions of the
1779          * file from under mmaped areas must have their ->fault function
1780          * return a locked page (and set VM_FAULT_LOCKED in the return).
1781          * This provides synchronisation against concurrent unmapping here.
1782          */
1783
1784 again:
1785         restart_addr = vma->vm_truncate_count;
1786         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1787                 start_addr = restart_addr;
1788                 if (start_addr >= end_addr) {
1789                         /* Top of vma has been split off since last time */
1790                         vma->vm_truncate_count = details->truncate_count;
1791                         return 0;
1792                 }
1793         }
1794
1795         restart_addr = zap_page_range(vma, start_addr,
1796                                         end_addr - start_addr, details);
1797         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
1798
1799         if (restart_addr >= end_addr) {
1800                 /* We have now completed this vma: mark it so */
1801                 vma->vm_truncate_count = details->truncate_count;
1802                 if (!need_break)
1803                         return 0;
1804         } else {
1805                 /* Note restart_addr in vma's truncate_count field */
1806                 vma->vm_truncate_count = restart_addr;
1807                 if (!need_break)
1808                         goto again;
1809         }
1810
1811         spin_unlock(details->i_mmap_lock);
1812         cond_resched();
1813         spin_lock(details->i_mmap_lock);
1814         return -EINTR;
1815 }
1816
1817 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1818                                             struct zap_details *details)
1819 {
1820         struct vm_area_struct *vma;
1821         struct prio_tree_iter iter;
1822         pgoff_t vba, vea, zba, zea;
1823
1824 restart:
1825         vma_prio_tree_foreach(vma, &iter, root,
1826                         details->first_index, details->last_index) {
1827                 /* Skip quickly over those we have already dealt with */
1828                 if (vma->vm_truncate_count == details->truncate_count)
1829                         continue;
1830
1831                 vba = vma->vm_pgoff;
1832                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1833                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1834                 zba = details->first_index;
1835                 if (zba < vba)
1836                         zba = vba;
1837                 zea = details->last_index;
1838                 if (zea > vea)
1839                         zea = vea;
1840
1841                 if (unmap_mapping_range_vma(vma,
1842                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1843                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1844                                 details) < 0)
1845                         goto restart;
1846         }
1847 }
1848
1849 static inline void unmap_mapping_range_list(struct list_head *head,
1850                                             struct zap_details *details)
1851 {
1852         struct vm_area_struct *vma;
1853
1854         /*
1855          * In nonlinear VMAs there is no correspondence between virtual address
1856          * offset and file offset.  So we must perform an exhaustive search
1857          * across *all* the pages in each nonlinear VMA, not just the pages
1858          * whose virtual address lies outside the file truncation point.
1859          */
1860 restart:
1861         list_for_each_entry(vma, head, shared.vm_set.list) {
1862                 /* Skip quickly over those we have already dealt with */
1863                 if (vma->vm_truncate_count == details->truncate_count)
1864                         continue;
1865                 details->nonlinear_vma = vma;
1866                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1867                                         vma->vm_end, details) < 0)
1868                         goto restart;
1869         }
1870 }
1871
1872 /**
1873  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1874  * @mapping: the address space containing mmaps to be unmapped.
1875  * @holebegin: byte in first page to unmap, relative to the start of
1876  * the underlying file.  This will be rounded down to a PAGE_SIZE
1877  * boundary.  Note that this is different from vmtruncate(), which
1878  * must keep the partial page.  In contrast, we must get rid of
1879  * partial pages.
1880  * @holelen: size of prospective hole in bytes.  This will be rounded
1881  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1882  * end of the file.
1883  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1884  * but 0 when invalidating pagecache, don't throw away private data.
1885  */
1886 void unmap_mapping_range(struct address_space *mapping,
1887                 loff_t const holebegin, loff_t const holelen, int even_cows)
1888 {
1889         struct zap_details details;
1890         pgoff_t hba = holebegin >> PAGE_SHIFT;
1891         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1892
1893         /* Check for overflow. */
1894         if (sizeof(holelen) > sizeof(hlen)) {
1895                 long long holeend =
1896                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1897                 if (holeend & ~(long long)ULONG_MAX)
1898                         hlen = ULONG_MAX - hba + 1;
1899         }
1900
1901         details.check_mapping = even_cows? NULL: mapping;
1902         details.nonlinear_vma = NULL;
1903         details.first_index = hba;
1904         details.last_index = hba + hlen - 1;
1905         if (details.last_index < details.first_index)
1906                 details.last_index = ULONG_MAX;
1907         details.i_mmap_lock = &mapping->i_mmap_lock;
1908
1909         spin_lock(&mapping->i_mmap_lock);
1910
1911         /* Protect against endless unmapping loops */
1912         mapping->truncate_count++;
1913         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1914                 if (mapping->truncate_count == 0)
1915                         reset_vma_truncate_counts(mapping);
1916                 mapping->truncate_count++;
1917         }
1918         details.truncate_count = mapping->truncate_count;
1919
1920         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1921                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1922         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1923                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1924         spin_unlock(&mapping->i_mmap_lock);
1925 }
1926 EXPORT_SYMBOL(unmap_mapping_range);
1927
1928 /**
1929  * vmtruncate - unmap mappings "freed" by truncate() syscall
1930  * @inode: inode of the file used
1931  * @offset: file offset to start truncating
1932  *
1933  * NOTE! We have to be ready to update the memory sharing
1934  * between the file and the memory map for a potential last
1935  * incomplete page.  Ugly, but necessary.
1936  */
1937 int vmtruncate(struct inode * inode, loff_t offset)
1938 {
1939         if (inode->i_size < offset) {
1940                 unsigned long limit;
1941
1942                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1943                 if (limit != RLIM_INFINITY && offset > limit)
1944                         goto out_sig;
1945                 if (offset > inode->i_sb->s_maxbytes)
1946                         goto out_big;
1947                 i_size_write(inode, offset);
1948         } else {
1949                 struct address_space *mapping = inode->i_mapping;
1950
1951                 /*
1952                  * truncation of in-use swapfiles is disallowed - it would
1953                  * cause subsequent swapout to scribble on the now-freed
1954                  * blocks.
1955                  */
1956                 if (IS_SWAPFILE(inode))
1957                         return -ETXTBSY;
1958                 i_size_write(inode, offset);
1959
1960                 /*
1961                  * unmap_mapping_range is called twice, first simply for
1962                  * efficiency so that truncate_inode_pages does fewer
1963                  * single-page unmaps.  However after this first call, and
1964                  * before truncate_inode_pages finishes, it is possible for
1965                  * private pages to be COWed, which remain after
1966                  * truncate_inode_pages finishes, hence the second
1967                  * unmap_mapping_range call must be made for correctness.
1968                  */
1969                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1970                 truncate_inode_pages(mapping, offset);
1971                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1972         }
1973
1974         if (inode->i_op && inode->i_op->truncate)
1975                 inode->i_op->truncate(inode);
1976         return 0;
1977
1978 out_sig:
1979         send_sig(SIGXFSZ, current, 0);
1980 out_big:
1981         return -EFBIG;
1982 }
1983 EXPORT_SYMBOL(vmtruncate);
1984
1985 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1986 {
1987         struct address_space *mapping = inode->i_mapping;
1988
1989         /*
1990          * If the underlying filesystem is not going to provide
1991          * a way to truncate a range of blocks (punch a hole) -
1992          * we should return failure right now.
1993          */
1994         if (!inode->i_op || !inode->i_op->truncate_range)
1995                 return -ENOSYS;
1996
1997         mutex_lock(&inode->i_mutex);
1998         down_write(&inode->i_alloc_sem);
1999         unmap_mapping_range(mapping, offset, (end - offset), 1);
2000         truncate_inode_pages_range(mapping, offset, end);
2001         unmap_mapping_range(mapping, offset, (end - offset), 1);
2002         inode->i_op->truncate_range(inode, offset, end);
2003         up_write(&inode->i_alloc_sem);
2004         mutex_unlock(&inode->i_mutex);
2005
2006         return 0;
2007 }
2008
2009 /*
2010  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2011  * but allow concurrent faults), and pte mapped but not yet locked.
2012  * We return with mmap_sem still held, but pte unmapped and unlocked.
2013  */
2014 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2015                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2016                 int write_access, pte_t orig_pte)
2017 {
2018         spinlock_t *ptl;
2019         struct page *page;
2020         swp_entry_t entry;
2021         pte_t pte;
2022         int ret = 0;
2023
2024         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2025                 goto out;
2026
2027         entry = pte_to_swp_entry(orig_pte);
2028         if (is_migration_entry(entry)) {
2029                 migration_entry_wait(mm, pmd, address);
2030                 goto out;
2031         }
2032         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2033         page = lookup_swap_cache(entry);
2034         if (!page) {
2035                 grab_swap_token(); /* Contend for token _before_ read-in */
2036                 page = swapin_readahead(entry,
2037                                         GFP_HIGHUSER_MOVABLE, vma, address);
2038                 if (!page) {
2039                         /*
2040                          * Back out if somebody else faulted in this pte
2041                          * while we released the pte lock.
2042                          */
2043                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2044                         if (likely(pte_same(*page_table, orig_pte)))
2045                                 ret = VM_FAULT_OOM;
2046                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2047                         goto unlock;
2048                 }
2049
2050                 /* Had to read the page from swap area: Major fault */
2051                 ret = VM_FAULT_MAJOR;
2052                 count_vm_event(PGMAJFAULT);
2053         }
2054
2055         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2056                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2057                 ret = VM_FAULT_OOM;
2058                 goto out;
2059         }
2060
2061         mark_page_accessed(page);
2062         lock_page(page);
2063         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2064
2065         /*
2066          * Back out if somebody else already faulted in this pte.
2067          */
2068         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2069         if (unlikely(!pte_same(*page_table, orig_pte)))
2070                 goto out_nomap;
2071
2072         if (unlikely(!PageUptodate(page))) {
2073                 ret = VM_FAULT_SIGBUS;
2074                 goto out_nomap;
2075         }
2076
2077         /* The page isn't present yet, go ahead with the fault. */
2078
2079         inc_mm_counter(mm, anon_rss);
2080         pte = mk_pte(page, vma->vm_page_prot);
2081         if (write_access && can_share_swap_page(page)) {
2082                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2083                 write_access = 0;
2084         }
2085
2086         flush_icache_page(vma, page);
2087         set_pte_at(mm, address, page_table, pte);
2088         page_add_anon_rmap(page, vma, address);
2089
2090         swap_free(entry);
2091         if (vm_swap_full())
2092                 remove_exclusive_swap_page(page);
2093         unlock_page(page);
2094
2095         if (write_access) {
2096                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2097                 if (ret & VM_FAULT_ERROR)
2098                         ret &= VM_FAULT_ERROR;
2099                 goto out;
2100         }
2101
2102         /* No need to invalidate - it was non-present before */
2103         update_mmu_cache(vma, address, pte);
2104 unlock:
2105         pte_unmap_unlock(page_table, ptl);
2106 out:
2107         return ret;
2108 out_nomap:
2109         mem_cgroup_uncharge_page(page);
2110         pte_unmap_unlock(page_table, ptl);
2111         unlock_page(page);
2112         page_cache_release(page);
2113         return ret;
2114 }
2115
2116 /*
2117  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2118  * but allow concurrent faults), and pte mapped but not yet locked.
2119  * We return with mmap_sem still held, but pte unmapped and unlocked.
2120  */
2121 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2122                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2123                 int write_access)
2124 {
2125         struct page *page;
2126         spinlock_t *ptl;
2127         pte_t entry;
2128
2129         /* Allocate our own private page. */
2130         pte_unmap(page_table);
2131
2132         if (unlikely(anon_vma_prepare(vma)))
2133                 goto oom;
2134         page = alloc_zeroed_user_highpage_movable(vma, address);
2135         if (!page)
2136                 goto oom;
2137         __SetPageUptodate(page);
2138
2139         if (mem_cgroup_charge(page, mm, GFP_KERNEL))
2140                 goto oom_free_page;
2141
2142         entry = mk_pte(page, vma->vm_page_prot);
2143         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2144
2145         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2146         if (!pte_none(*page_table))
2147                 goto release;
2148         inc_mm_counter(mm, anon_rss);
2149         lru_cache_add_active(page);
2150         page_add_new_anon_rmap(page, vma, address);
2151         set_pte_at(mm, address, page_table, entry);
2152
2153         /* No need to invalidate - it was non-present before */
2154         update_mmu_cache(vma, address, entry);
2155 unlock:
2156         pte_unmap_unlock(page_table, ptl);
2157         return 0;
2158 release:
2159         mem_cgroup_uncharge_page(page);
2160         page_cache_release(page);
2161         goto unlock;
2162 oom_free_page:
2163         page_cache_release(page);
2164 oom:
2165         return VM_FAULT_OOM;
2166 }
2167
2168 /*
2169  * __do_fault() tries to create a new page mapping. It aggressively
2170  * tries to share with existing pages, but makes a separate copy if
2171  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2172  * the next page fault.
2173  *
2174  * As this is called only for pages that do not currently exist, we
2175  * do not need to flush old virtual caches or the TLB.
2176  *
2177  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2178  * but allow concurrent faults), and pte neither mapped nor locked.
2179  * We return with mmap_sem still held, but pte unmapped and unlocked.
2180  */
2181 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2182                 unsigned long address, pmd_t *pmd,
2183                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2184 {
2185         pte_t *page_table;
2186         spinlock_t *ptl;
2187         struct page *page;
2188         pte_t entry;
2189         int anon = 0;
2190         struct page *dirty_page = NULL;
2191         struct vm_fault vmf;
2192         int ret;
2193         int page_mkwrite = 0;
2194
2195         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2196         vmf.pgoff = pgoff;
2197         vmf.flags = flags;
2198         vmf.page = NULL;
2199
2200         BUG_ON(vma->vm_flags & VM_PFNMAP);
2201
2202         if (likely(vma->vm_ops->fault)) {
2203                 ret = vma->vm_ops->fault(vma, &vmf);
2204                 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2205                         return ret;
2206         } else {
2207                 /* Legacy ->nopage path */
2208                 ret = 0;
2209                 vmf.page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2210                 /* no page was available -- either SIGBUS or OOM */
2211                 if (unlikely(vmf.page == NOPAGE_SIGBUS))
2212                         return VM_FAULT_SIGBUS;
2213                 else if (unlikely(vmf.page == NOPAGE_OOM))
2214                         return VM_FAULT_OOM;
2215         }
2216
2217         /*
2218          * For consistency in subsequent calls, make the faulted page always
2219          * locked.
2220          */
2221         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2222                 lock_page(vmf.page);
2223         else
2224                 VM_BUG_ON(!PageLocked(vmf.page));
2225
2226         /*
2227          * Should we do an early C-O-W break?
2228          */
2229         page = vmf.page;
2230         if (flags & FAULT_FLAG_WRITE) {
2231                 if (!(vma->vm_flags & VM_SHARED)) {
2232                         anon = 1;
2233                         if (unlikely(anon_vma_prepare(vma))) {
2234                                 ret = VM_FAULT_OOM;
2235                                 goto out;
2236                         }
2237                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2238                                                 vma, address);
2239                         if (!page) {
2240                                 ret = VM_FAULT_OOM;
2241                                 goto out;
2242                         }
2243                         copy_user_highpage(page, vmf.page, address, vma);
2244                         __SetPageUptodate(page);
2245                 } else {
2246                         /*
2247                          * If the page will be shareable, see if the backing
2248                          * address space wants to know that the page is about
2249                          * to become writable
2250                          */
2251                         if (vma->vm_ops->page_mkwrite) {
2252                                 unlock_page(page);
2253                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2254                                         ret = VM_FAULT_SIGBUS;
2255                                         anon = 1; /* no anon but release vmf.page */
2256                                         goto out_unlocked;
2257                                 }
2258                                 lock_page(page);
2259                                 /*
2260                                  * XXX: this is not quite right (racy vs
2261                                  * invalidate) to unlock and relock the page
2262                                  * like this, however a better fix requires
2263                                  * reworking page_mkwrite locking API, which
2264                                  * is better done later.
2265                                  */
2266                                 if (!page->mapping) {
2267                                         ret = 0;
2268                                         anon = 1; /* no anon but release vmf.page */
2269                                         goto out;
2270                                 }
2271                                 page_mkwrite = 1;
2272                         }
2273                 }
2274
2275         }
2276
2277         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2278                 ret = VM_FAULT_OOM;
2279                 goto out;
2280         }
2281
2282         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2283
2284         /*
2285          * This silly early PAGE_DIRTY setting removes a race
2286          * due to the bad i386 page protection. But it's valid
2287          * for other architectures too.
2288          *
2289          * Note that if write_access is true, we either now have
2290          * an exclusive copy of the page, or this is a shared mapping,
2291          * so we can make it writable and dirty to avoid having to
2292          * handle that later.
2293          */
2294         /* Only go through if we didn't race with anybody else... */
2295         if (likely(pte_same(*page_table, orig_pte))) {
2296                 flush_icache_page(vma, page);
2297                 entry = mk_pte(page, vma->vm_page_prot);
2298                 if (flags & FAULT_FLAG_WRITE)
2299                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2300                 set_pte_at(mm, address, page_table, entry);
2301                 if (anon) {
2302                         inc_mm_counter(mm, anon_rss);
2303                         lru_cache_add_active(page);
2304                         page_add_new_anon_rmap(page, vma, address);
2305                 } else {
2306                         inc_mm_counter(mm, file_rss);
2307                         page_add_file_rmap(page);
2308                         if (flags & FAULT_FLAG_WRITE) {
2309                                 dirty_page = page;
2310                                 get_page(dirty_page);
2311                         }
2312                 }
2313
2314                 /* no need to invalidate: a not-present page won't be cached */
2315                 update_mmu_cache(vma, address, entry);
2316         } else {
2317                 mem_cgroup_uncharge_page(page);
2318                 if (anon)
2319                         page_cache_release(page);
2320                 else
2321                         anon = 1; /* no anon but release faulted_page */
2322         }
2323
2324         pte_unmap_unlock(page_table, ptl);
2325
2326 out:
2327         unlock_page(vmf.page);
2328 out_unlocked:
2329         if (anon)
2330                 page_cache_release(vmf.page);
2331         else if (dirty_page) {
2332                 if (vma->vm_file)
2333                         file_update_time(vma->vm_file);
2334
2335                 set_page_dirty_balance(dirty_page, page_mkwrite);
2336                 put_page(dirty_page);
2337         }
2338
2339         return ret;
2340 }
2341
2342 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2343                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2344                 int write_access, pte_t orig_pte)
2345 {
2346         pgoff_t pgoff = (((address & PAGE_MASK)
2347                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2348         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2349
2350         pte_unmap(page_table);
2351         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2352 }
2353
2354
2355 /*
2356  * do_no_pfn() tries to create a new page mapping for a page without
2357  * a struct_page backing it
2358  *
2359  * As this is called only for pages that do not currently exist, we
2360  * do not need to flush old virtual caches or the TLB.
2361  *
2362  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2363  * but allow concurrent faults), and pte mapped but not yet locked.
2364  * We return with mmap_sem still held, but pte unmapped and unlocked.
2365  *
2366  * It is expected that the ->nopfn handler always returns the same pfn
2367  * for a given virtual mapping.
2368  *
2369  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2370  */
2371 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2372                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2373                      int write_access)
2374 {
2375         spinlock_t *ptl;
2376         pte_t entry;
2377         unsigned long pfn;
2378
2379         pte_unmap(page_table);
2380         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
2381         BUG_ON(is_cow_mapping(vma->vm_flags));
2382
2383         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2384         if (unlikely(pfn == NOPFN_OOM))
2385                 return VM_FAULT_OOM;
2386         else if (unlikely(pfn == NOPFN_SIGBUS))
2387                 return VM_FAULT_SIGBUS;
2388         else if (unlikely(pfn == NOPFN_REFAULT))
2389                 return 0;
2390
2391         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2392
2393         /* Only go through if we didn't race with anybody else... */
2394         if (pte_none(*page_table)) {
2395                 entry = pfn_pte(pfn, vma->vm_page_prot);
2396                 if (write_access)
2397                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2398                 set_pte_at(mm, address, page_table, entry);
2399         }
2400         pte_unmap_unlock(page_table, ptl);
2401         return 0;
2402 }
2403
2404 /*
2405  * Fault of a previously existing named mapping. Repopulate the pte
2406  * from the encoded file_pte if possible. This enables swappable
2407  * nonlinear vmas.
2408  *
2409  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2410  * but allow concurrent faults), and pte mapped but not yet locked.
2411  * We return with mmap_sem still held, but pte unmapped and unlocked.
2412  */
2413 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2414                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2415                 int write_access, pte_t orig_pte)
2416 {
2417         unsigned int flags = FAULT_FLAG_NONLINEAR |
2418                                 (write_access ? FAULT_FLAG_WRITE : 0);
2419         pgoff_t pgoff;
2420
2421         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2422                 return 0;
2423
2424         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2425                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2426                 /*
2427                  * Page table corrupted: show pte and kill process.
2428                  */
2429                 print_bad_pte(vma, orig_pte, address);
2430                 return VM_FAULT_OOM;
2431         }
2432
2433         pgoff = pte_to_pgoff(orig_pte);
2434         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2435 }
2436
2437 /*
2438  * These routines also need to handle stuff like marking pages dirty
2439  * and/or accessed for architectures that don't do it in hardware (most
2440  * RISC architectures).  The early dirtying is also good on the i386.
2441  *
2442  * There is also a hook called "update_mmu_cache()" that architectures
2443  * with external mmu caches can use to update those (ie the Sparc or
2444  * PowerPC hashed page tables that act as extended TLBs).
2445  *
2446  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2447  * but allow concurrent faults), and pte mapped but not yet locked.
2448  * We return with mmap_sem still held, but pte unmapped and unlocked.
2449  */
2450 static inline int handle_pte_fault(struct mm_struct *mm,
2451                 struct vm_area_struct *vma, unsigned long address,
2452                 pte_t *pte, pmd_t *pmd, int write_access)
2453 {
2454         pte_t entry;
2455         spinlock_t *ptl;
2456
2457         entry = *pte;
2458         if (!pte_present(entry)) {
2459                 if (pte_none(entry)) {
2460                         if (vma->vm_ops) {
2461                                 if (vma->vm_ops->fault || vma->vm_ops->nopage)
2462                                         return do_linear_fault(mm, vma, address,
2463                                                 pte, pmd, write_access, entry);
2464                                 if (unlikely(vma->vm_ops->nopfn))
2465                                         return do_no_pfn(mm, vma, address, pte,
2466                                                          pmd, write_access);
2467                         }
2468                         return do_anonymous_page(mm, vma, address,
2469                                                  pte, pmd, write_access);
2470                 }
2471                 if (pte_file(entry))
2472                         return do_nonlinear_fault(mm, vma, address,
2473                                         pte, pmd, write_access, entry);
2474                 return do_swap_page(mm, vma, address,
2475                                         pte, pmd, write_access, entry);
2476         }
2477
2478         ptl = pte_lockptr(mm, pmd);
2479         spin_lock(ptl);
2480         if (unlikely(!pte_same(*pte, entry)))
2481                 goto unlock;
2482         if (write_access) {
2483                 if (!pte_write(entry))
2484                         return do_wp_page(mm, vma, address,
2485                                         pte, pmd, ptl, entry);
2486                 entry = pte_mkdirty(entry);
2487         }
2488         entry = pte_mkyoung(entry);
2489         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2490                 update_mmu_cache(vma, address, entry);
2491         } else {
2492                 /*
2493                  * This is needed only for protection faults but the arch code
2494                  * is not yet telling us if this is a protection fault or not.
2495                  * This still avoids useless tlb flushes for .text page faults
2496                  * with threads.
2497                  */
2498                 if (write_access)
2499                         flush_tlb_page(vma, address);
2500         }
2501 unlock:
2502         pte_unmap_unlock(pte, ptl);
2503         return 0;
2504 }
2505
2506 /*
2507  * By the time we get here, we already hold the mm semaphore
2508  */
2509 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2510                 unsigned long address, int write_access)
2511 {
2512         pgd_t *pgd;
2513         pud_t *pud;
2514         pmd_t *pmd;
2515         pte_t *pte;
2516
2517         __set_current_state(TASK_RUNNING);
2518
2519         count_vm_event(PGFAULT);
2520
2521         if (unlikely(is_vm_hugetlb_page(vma)))
2522                 return hugetlb_fault(mm, vma, address, write_access);
2523
2524         pgd = pgd_offset(mm, address);
2525         pud = pud_alloc(mm, pgd, address);
2526         if (!pud)
2527                 return VM_FAULT_OOM;
2528         pmd = pmd_alloc(mm, pud, address);
2529         if (!pmd)
2530                 return VM_FAULT_OOM;
2531         pte = pte_alloc_map(mm, pmd, address);
2532         if (!pte)
2533                 return VM_FAULT_OOM;
2534
2535         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2536 }
2537
2538 #ifndef __PAGETABLE_PUD_FOLDED
2539 /*
2540  * Allocate page upper directory.
2541  * We've already handled the fast-path in-line.
2542  */
2543 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2544 {
2545         pud_t *new = pud_alloc_one(mm, address);
2546         if (!new)
2547                 return -ENOMEM;
2548
2549         spin_lock(&mm->page_table_lock);
2550         if (pgd_present(*pgd))          /* Another has populated it */
2551                 pud_free(mm, new);
2552         else
2553                 pgd_populate(mm, pgd, new);
2554         spin_unlock(&mm->page_table_lock);
2555         return 0;
2556 }
2557 #endif /* __PAGETABLE_PUD_FOLDED */
2558
2559 #ifndef __PAGETABLE_PMD_FOLDED
2560 /*
2561  * Allocate page middle directory.
2562  * We've already handled the fast-path in-line.
2563  */
2564 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2565 {
2566         pmd_t *new = pmd_alloc_one(mm, address);
2567         if (!new)
2568                 return -ENOMEM;
2569
2570         spin_lock(&mm->page_table_lock);
2571 #ifndef __ARCH_HAS_4LEVEL_HACK
2572         if (pud_present(*pud))          /* Another has populated it */
2573                 pmd_free(mm, new);
2574         else
2575                 pud_populate(mm, pud, new);
2576 #else
2577         if (pgd_present(*pud))          /* Another has populated it */
2578                 pmd_free(mm, new);
2579         else
2580                 pgd_populate(mm, pud, new);
2581 #endif /* __ARCH_HAS_4LEVEL_HACK */
2582         spin_unlock(&mm->page_table_lock);
2583         return 0;
2584 }
2585 #endif /* __PAGETABLE_PMD_FOLDED */
2586
2587 int make_pages_present(unsigned long addr, unsigned long end)
2588 {
2589         int ret, len, write;
2590         struct vm_area_struct * vma;
2591
2592         vma = find_vma(current->mm, addr);
2593         if (!vma)
2594                 return -1;
2595         write = (vma->vm_flags & VM_WRITE) != 0;
2596         BUG_ON(addr >= end);
2597         BUG_ON(end > vma->vm_end);
2598         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2599         ret = get_user_pages(current, current->mm, addr,
2600                         len, write, 0, NULL, NULL);
2601         if (ret < 0)
2602                 return ret;
2603         return ret == len ? 0 : -1;
2604 }
2605
2606 #if !defined(__HAVE_ARCH_GATE_AREA)
2607
2608 #if defined(AT_SYSINFO_EHDR)
2609 static struct vm_area_struct gate_vma;
2610
2611 static int __init gate_vma_init(void)
2612 {
2613         gate_vma.vm_mm = NULL;
2614         gate_vma.vm_start = FIXADDR_USER_START;
2615         gate_vma.vm_end = FIXADDR_USER_END;
2616         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2617         gate_vma.vm_page_prot = __P101;
2618         /*
2619          * Make sure the vDSO gets into every core dump.
2620          * Dumping its contents makes post-mortem fully interpretable later
2621          * without matching up the same kernel and hardware config to see
2622          * what PC values meant.
2623          */
2624         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2625         return 0;
2626 }
2627 __initcall(gate_vma_init);
2628 #endif
2629
2630 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2631 {
2632 #ifdef AT_SYSINFO_EHDR
2633         return &gate_vma;
2634 #else
2635         return NULL;
2636 #endif
2637 }
2638
2639 int in_gate_area_no_task(unsigned long addr)
2640 {
2641 #ifdef AT_SYSINFO_EHDR
2642         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2643                 return 1;
2644 #endif
2645         return 0;
2646 }
2647
2648 #endif  /* __HAVE_ARCH_GATE_AREA */
2649
2650 /*
2651  * Access another process' address space.
2652  * Source/target buffer must be kernel space,
2653  * Do not walk the page table directly, use get_user_pages
2654  */
2655 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2656 {
2657         struct mm_struct *mm;
2658         struct vm_area_struct *vma;
2659         struct page *page;
2660         void *old_buf = buf;
2661
2662         mm = get_task_mm(tsk);
2663         if (!mm)
2664                 return 0;
2665
2666         down_read(&mm->mmap_sem);
2667         /* ignore errors, just check how much was successfully transferred */
2668         while (len) {
2669                 int bytes, ret, offset;
2670                 void *maddr;
2671
2672                 ret = get_user_pages(tsk, mm, addr, 1,
2673                                 write, 1, &page, &vma);
2674                 if (ret <= 0)
2675                         break;
2676
2677                 bytes = len;
2678                 offset = addr & (PAGE_SIZE-1);
2679                 if (bytes > PAGE_SIZE-offset)
2680                         bytes = PAGE_SIZE-offset;
2681
2682                 maddr = kmap(page);
2683                 if (write) {
2684                         copy_to_user_page(vma, page, addr,
2685                                           maddr + offset, buf, bytes);
2686                         set_page_dirty_lock(page);
2687                 } else {
2688                         copy_from_user_page(vma, page, addr,
2689                                             buf, maddr + offset, bytes);
2690                 }
2691                 kunmap(page);
2692                 page_cache_release(page);
2693                 len -= bytes;
2694                 buf += bytes;
2695                 addr += bytes;
2696         }
2697         up_read(&mm->mmap_sem);
2698         mmput(mm);
2699
2700         return buf - old_buf;
2701 }
2702
2703 /*
2704  * Print the name of a VMA.
2705  */
2706 void print_vma_addr(char *prefix, unsigned long ip)
2707 {
2708         struct mm_struct *mm = current->mm;
2709         struct vm_area_struct *vma;
2710
2711         /*
2712          * Do not print if we are in atomic
2713          * contexts (in exception stacks, etc.):
2714          */
2715         if (preempt_count())
2716                 return;
2717
2718         down_read(&mm->mmap_sem);
2719         vma = find_vma(mm, ip);
2720         if (vma && vma->vm_file) {
2721                 struct file *f = vma->vm_file;
2722                 char *buf = (char *)__get_free_page(GFP_KERNEL);
2723                 if (buf) {
2724                         char *p, *s;
2725
2726                         p = d_path(&f->f_path, buf, PAGE_SIZE);
2727                         if (IS_ERR(p))
2728                                 p = "?";
2729                         s = strrchr(p, '/');
2730                         if (s)
2731                                 p = s+1;
2732                         printk("%s%s[%lx+%lx]", prefix, p,
2733                                         vma->vm_start,
2734                                         vma->vm_end - vma->vm_start);
2735                         free_page((unsigned long)buf);
2736                 }
2737         }
2738         up_read(&current->mm->mmap_sem);
2739 }