mm/memory.c

   1 /*
   2  *  linux/mm/memory.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  */
   6
   7 /*
   8  * demand-loading started 01.12.91 - seems it is high on the list of
   9  * things wanted, and it should be easy to implement. - Linus
  10  */
  11
  12 /*
  13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14  * pages started 02.12.91, seems to work. - Linus.
  15  *
  16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17  * would have taken more than the 6M I have free, but it worked well as
  18  * far as I could see.
  19  *
  20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21  */
  22
  23 /*
  24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25  * thought has to go into this. Oh, well..
  26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27  *              Found it. Everything seems to work now.
  28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29  */
  30
  31 /*
  32  * 05.04.94  -  Multi-page memory management added for v1.1.
  33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  34  *
  35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36  *              (Gerhard.Wichert@pdb.siemens.de)
  37  *
  38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  39  */
  40
  41 #include <linux/kernel_stat.h>
  42 #include <linux/mm.h>
  43 #include <linux/hugetlb.h>
  44 #include <linux/mman.h>
  45 #include <linux/swap.h>
  46 #include <linux/highmem.h>
  47 #include <linux/pagemap.h>
  48 #include <linux/rmap.h>
  49 #include <linux/module.h>
  50 #include <linux/delayacct.h>
  51 #include <linux/init.h>
  52 #include <linux/writeback.h>
  53
  54 #include <asm/pgalloc.h>
  55 #include <asm/uaccess.h>
  56 #include <asm/tlb.h>
  57 #include <asm/tlbflush.h>
  58 #include <asm/pgtable.h>
  59
  60 #include <linux/swapops.h>
  61 #include <linux/elf.h>
  62
  63 #ifndef CONFIG_NEED_MULTIPLE_NODES
  64 /* use the per-pgdat data instead for discontigmem - mbligh */
  65 unsigned long max_mapnr;
  66 struct page *mem_map;
  67
  68 EXPORT_SYMBOL(max_mapnr);
  69 EXPORT_SYMBOL(mem_map);
  70 #endif
  71
  72 unsigned long num_physpages;
  73 /*
  74  * A number of key systems in x86 including ioremap() rely on the assumption
  75  * that high_memory defines the upper bound on direct map memory, then end
  76  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  77  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  78  * and ZONE_HIGHMEM.
  79  */
  80 void * high_memory;
  81
  82 EXPORT_SYMBOL(num_physpages);
  83 EXPORT_SYMBOL(high_memory);
  84
  85 /*
  86  * Randomize the address space (stacks, mmaps, brk, etc.).
  87  *
  88  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
  89  *   as ancient (libc5 based) binaries can segfault. )
  90  */
  91 int randomize_va_space __read_mostly =
  92 #ifdef CONFIG_COMPAT_BRK
  93                                         1;
  94 #else
  95                                         2;
  96 #endif
  97
  98 static int __init disable_randmaps(char *s)
  99 {
 100         randomize_va_space = 0;
 101         return 1;
 102 }
 103 __setup("norandmaps", disable_randmaps);
 104
 105
 106 /*
 107  * If a p?d_bad entry is found while walking page tables, report
 108  * the error, before resetting entry to p?d_none.  Usually (but
 109  * very seldom) called out from the p?d_none_or_clear_bad macros.
 110  */
 111
 112 void pgd_clear_bad(pgd_t *pgd)
 113 {
 114         pgd_ERROR(*pgd);
 115         pgd_clear(pgd);
 116 }
 117
 118 void pud_clear_bad(pud_t *pud)
 119 {
 120         pud_ERROR(*pud);
 121         pud_clear(pud);
 122 }
 123
 124 void pmd_clear_bad(pmd_t *pmd)
 125 {
 126         pmd_ERROR(*pmd);
 127         pmd_clear(pmd);
 128 }
 129
 130 /*
 131  * Note: this doesn't free the actual pages themselves. That
 132  * has been handled earlier when unmapping all the memory regions.
 133  */
 134 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
 135 {
 136         struct page *page = pmd_page(*pmd);
 137         pmd_clear(pmd);
 138         pte_lock_deinit(page);
 139         pte_free_tlb(tlb, page);
 140         dec_zone_page_state(page, NR_PAGETABLE);
 141         tlb->mm->nr_ptes--;
 142 }
 143
 144 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 145                                 unsigned long addr, unsigned long end,
 146                                 unsigned long floor, unsigned long ceiling)
 147 {
 148         pmd_t *pmd;
 149         unsigned long next;
 150         unsigned long start;
 151
 152         start = addr;
 153         pmd = pmd_offset(pud, addr);
 154         do {
 155                 next = pmd_addr_end(addr, end);
 156                 if (pmd_none_or_clear_bad(pmd))
 157                         continue;
 158                 free_pte_range(tlb, pmd);
 159         } while (pmd++, addr = next, addr != end);
 160
 161         start &= PUD_MASK;
 162         if (start < floor)
 163                 return;
 164         if (ceiling) {
 165                 ceiling &= PUD_MASK;
 166                 if (!ceiling)
 167                         return;
 168         }
 169         if (end - 1 > ceiling - 1)
 170                 return;
 171
 172         pmd = pmd_offset(pud, start);
 173         pud_clear(pud);
 174         pmd_free_tlb(tlb, pmd);
 175 }
 176
 177 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 178                                 unsigned long addr, unsigned long end,
 179                                 unsigned long floor, unsigned long ceiling)
 180 {
 181         pud_t *pud;
 182         unsigned long next;
 183         unsigned long start;
 184
 185         start = addr;
 186         pud = pud_offset(pgd, addr);
 187         do {
 188                 next = pud_addr_end(addr, end);
 189                 if (pud_none_or_clear_bad(pud))
 190                         continue;
 191                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 192         } while (pud++, addr = next, addr != end);
 193
 194         start &= PGDIR_MASK;
 195         if (start < floor)
 196                 return;
 197         if (ceiling) {
 198                 ceiling &= PGDIR_MASK;
 199                 if (!ceiling)
 200                         return;
 201         }
 202         if (end - 1 > ceiling - 1)
 203                 return;
 204
 205         pud = pud_offset(pgd, start);
 206         pgd_clear(pgd);
 207         pud_free_tlb(tlb, pud);
 208 }
 209
 210 /*
 211  * This function frees user-level page tables of a process.
 212  *
 213  * Must be called with pagetable lock held.
 214  */
 215 void free_pgd_range(struct mmu_gather **tlb,
 216                         unsigned long addr, unsigned long end,
 217                         unsigned long floor, unsigned long ceiling)
 218 {
 219         pgd_t *pgd;
 220         unsigned long next;
 221         unsigned long start;
 222
 223         /*
 224          * The next few lines have given us lots of grief...
 225          *
 226          * Why are we testing PMD* at this top level?  Because often
 227          * there will be no work to do at all, and we'd prefer not to
 228          * go all the way down to the bottom just to discover that.
 229          *
 230          * Why all these "- 1"s?  Because 0 represents both the bottom
 231          * of the address space and the top of it (using -1 for the
 232          * top wouldn't help much: the masks would do the wrong thing).
 233          * The rule is that addr 0 and floor 0 refer to the bottom of
 234          * the address space, but end 0 and ceiling 0 refer to the top
 235          * Comparisons need to use "end - 1" and "ceiling - 1" (though
 236          * that end 0 case should be mythical).
 237          *
 238          * Wherever addr is brought up or ceiling brought down, we must
 239          * be careful to reject "the opposite 0" before it confuses the
 240          * subsequent tests.  But what about where end is brought down
 241          * by PMD_SIZE below? no, end can't go down to 0 there.
 242          *
 243          * Whereas we round start (addr) and ceiling down, by different
 244          * masks at different levels, in order to test whether a table
 245          * now has no other vmas using it, so can be freed, we don't
 246          * bother to round floor or end up - the tests don't need that.
 247          */
 248
 249         addr &= PMD_MASK;
 250         if (addr < floor) {
 251                 addr += PMD_SIZE;
 252                 if (!addr)
 253                         return;
 254         }
 255         if (ceiling) {
 256                 ceiling &= PMD_MASK;
 257                 if (!ceiling)
 258                         return;
 259         }
 260         if (end - 1 > ceiling - 1)
 261                 end -= PMD_SIZE;
 262         if (addr > end - 1)
 263                 return;
 264
 265         start = addr;
 266         pgd = pgd_offset((*tlb)->mm, addr);
 267         do {
 268                 next = pgd_addr_end(addr, end);
 269                 if (pgd_none_or_clear_bad(pgd))
 270                         continue;
 271                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
 272         } while (pgd++, addr = next, addr != end);
 273 }
 274
 275 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
 276                 unsigned long floor, unsigned long ceiling)
 277 {
 278         while (vma) {
 279                 struct vm_area_struct *next = vma->vm_next;
 280                 unsigned long addr = vma->vm_start;
 281
 282                 /*
 283                  * Hide vma from rmap and vmtruncate before freeing pgtables
 284                  */
 285                 anon_vma_unlink(vma);
 286                 unlink_file_vma(vma);
 287
 288                 if (is_vm_hugetlb_page(vma)) {
 289                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 290                                 floor, next? next->vm_start: ceiling);
 291                 } else {
 292                         /*
 293                          * Optimization: gather nearby vmas into one call down
 294                          */
 295                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 296                                && !is_vm_hugetlb_page(next)) {
 297                                 vma = next;
 298                                 next = vma->vm_next;
 299                                 anon_vma_unlink(vma);
 300                                 unlink_file_vma(vma);
 301                         }
 302                         free_pgd_range(tlb, addr, vma->vm_end,
 303                                 floor, next? next->vm_start: ceiling);
 304                 }
 305                 vma = next;
 306         }
 307 }
 308
 309 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 310 {
 311         struct page *new = pte_alloc_one(mm, address);
 312         if (!new)
 313                 return -ENOMEM;
 314
 315         pte_lock_init(new);
 316         spin_lock(&mm->page_table_lock);
 317         if (pmd_present(*pmd)) {        /* Another has populated it */
 318                 pte_lock_deinit(new);
 319                 pte_free(mm, new);
 320         } else {
 321                 mm->nr_ptes++;
 322                 inc_zone_page_state(new, NR_PAGETABLE);
 323                 pmd_populate(mm, pmd, new);
 324         }
 325         spin_unlock(&mm->page_table_lock);
 326         return 0;
 327 }
 328
 329 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 330 {
 331         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 332         if (!new)
 333                 return -ENOMEM;
 334
 335         spin_lock(&init_mm.page_table_lock);
 336         if (pmd_present(*pmd))          /* Another has populated it */
 337                 pte_free_kernel(&init_mm, new);
 338         else
 339                 pmd_populate_kernel(&init_mm, pmd, new);
 340         spin_unlock(&init_mm.page_table_lock);
 341         return 0;
 342 }
 343
 344 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
 345 {
 346         if (file_rss)
 347                 add_mm_counter(mm, file_rss, file_rss);
 348         if (anon_rss)
 349                 add_mm_counter(mm, anon_rss, anon_rss);
 350 }
 351
 352 /*
 353  * This function is called to print an error when a bad pte
 354  * is found. For example, we might have a PFN-mapped pte in
 355  * a region that doesn't allow it.
 356  *
 357  * The calling function must still handle the error.
 358  */
 359 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
 360 {
 361         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
 362                         "vm_flags = %lx, vaddr = %lx\n",
 363                 (long long)pte_val(pte),
 364                 (vma->vm_mm == current->mm ? current->comm : "???"),
 365                 vma->vm_flags, vaddr);
 366         dump_stack();
 367 }
 368
 369 static inline int is_cow_mapping(unsigned int flags)
 370 {
 371         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 372 }
 373
 374 /*
 375  * This function gets the "struct page" associated with a pte.
 376  *
 377  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
 378  * will have each page table entry just pointing to a raw page frame
 379  * number, and as far as the VM layer is concerned, those do not have
 380  * pages associated with them - even if the PFN might point to memory
 381  * that otherwise is perfectly fine and has a "struct page".
 382  *
 383  * The way we recognize those mappings is through the rules set up
 384  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
 385  * and the vm_pgoff will point to the first PFN mapped: thus every
 386  * page that is a raw mapping will always honor the rule
 387  *
 388  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 389  *
 390  * and if that isn't true, the page has been COW'ed (in which case it
 391  * _does_ have a "struct page" associated with it even if it is in a
 392  * VM_PFNMAP range).
 393  */
 394 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
 395 {
 396         unsigned long pfn = pte_pfn(pte);
 397
 398         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
 399                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
 400                 if (pfn == vma->vm_pgoff + off)
 401                         return NULL;
 402                 if (!is_cow_mapping(vma->vm_flags))
 403                         return NULL;
 404         }
 405
 406 #ifdef CONFIG_DEBUG_VM
 407         /*
 408          * Add some anal sanity checks for now. Eventually,
 409          * we should just do "return pfn_to_page(pfn)", but
 410          * in the meantime we check that we get a valid pfn,
 411          * and that the resulting page looks ok.
 412          */
 413         if (unlikely(!pfn_valid(pfn))) {
 414                 print_bad_pte(vma, pte, addr);
 415                 return NULL;
 416         }
 417 #endif
 418
 419         /*
 420          * NOTE! We still have PageReserved() pages in the page
 421          * tables.
 422          *
 423          * The PAGE_ZERO() pages and various VDSO mappings can
 424          * cause them to exist.
 425          */
 426         return pfn_to_page(pfn);
 427 }
 428
 429 /*
 430  * copy one vm_area from one task to the other. Assumes the page tables
 431  * already present in the new task to be cleared in the whole range
 432  * covered by this vma.
 433  */
 434
 435 static inline void
 436 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 437                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 438                 unsigned long addr, int *rss)
 439 {
 440         unsigned long vm_flags = vma->vm_flags;
 441         pte_t pte = *src_pte;
 442         struct page *page;
 443
 444         /* pte contains position in swap or file, so copy. */
 445         if (unlikely(!pte_present(pte))) {
 446                 if (!pte_file(pte)) {
 447                         swp_entry_t entry = pte_to_swp_entry(pte);
 448
 449                         swap_duplicate(entry);
 450                         /* make sure dst_mm is on swapoff's mmlist. */
 451                         if (unlikely(list_empty(&dst_mm->mmlist))) {
 452                                 spin_lock(&mmlist_lock);
 453                                 if (list_empty(&dst_mm->mmlist))
 454                                         list_add(&dst_mm->mmlist,
 455                                                  &src_mm->mmlist);
 456                                 spin_unlock(&mmlist_lock);
 457                         }
 458                         if (is_write_migration_entry(entry) &&
 459                                         is_cow_mapping(vm_flags)) {
 460                                 /*
 461                                  * COW mappings require pages in both parent
 462                                  * and child to be set to read.
 463                                  */
 464                                 make_migration_entry_read(&entry);
 465                                 pte = swp_entry_to_pte(entry);
 466                                 set_pte_at(src_mm, addr, src_pte, pte);
 467                         }
 468                 }
 469                 goto out_set_pte;
 470         }
 471
 472         /*
 473          * If it's a COW mapping, write protect it both
 474          * in the parent and the child
 475          */
 476         if (is_cow_mapping(vm_flags)) {
 477                 ptep_set_wrprotect(src_mm, addr, src_pte);
 478                 pte = pte_wrprotect(pte);
 479         }
 480
 481         /*
 482          * If it's a shared mapping, mark it clean in
 483          * the child
 484          */
 485         if (vm_flags & VM_SHARED)
 486                 pte = pte_mkclean(pte);
 487         pte = pte_mkold(pte);
 488
 489         page = vm_normal_page(vma, addr, pte);
 490         if (page) {
 491                 get_page(page);
 492                 page_dup_rmap(page, vma, addr);
 493                 rss[!!PageAnon(page)]++;
 494         }
 495
 496 out_set_pte:
 497         set_pte_at(dst_mm, addr, dst_pte, pte);
 498 }
 499
 500 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 501                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 502                 unsigned long addr, unsigned long end)
 503 {
 504         pte_t *src_pte, *dst_pte;
 505         spinlock_t *src_ptl, *dst_ptl;
 506         int progress = 0;
 507         int rss[2];
 508
 509 again:
 510         rss[1] = rss[0] = 0;
 511         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 512         if (!dst_pte)
 513                 return -ENOMEM;
 514         src_pte = pte_offset_map_nested(src_pmd, addr);
 515         src_ptl = pte_lockptr(src_mm, src_pmd);
 516         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 517         arch_enter_lazy_mmu_mode();
 518
 519         do {
 520                 /*
 521                  * We are holding two locks at this point - either of them
 522                  * could generate latencies in another task on another CPU.
 523                  */
 524                 if (progress >= 32) {
 525                         progress = 0;
 526                         if (need_resched() ||
 527                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 528                                 break;
 529                 }
 530                 if (pte_none(*src_pte)) {
 531                         progress++;
 532                         continue;
 533                 }
 534                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
 535                 progress += 8;
 536         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 537
 538         arch_leave_lazy_mmu_mode();
 539         spin_unlock(src_ptl);
 540         pte_unmap_nested(src_pte - 1);
 541         add_mm_rss(dst_mm, rss[0], rss[1]);
 542         pte_unmap_unlock(dst_pte - 1, dst_ptl);
 543         cond_resched();
 544         if (addr != end)
 545                 goto again;
 546         return 0;
 547 }
 548
 549 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 550                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 551                 unsigned long addr, unsigned long end)
 552 {
 553         pmd_t *src_pmd, *dst_pmd;
 554         unsigned long next;
 555
 556         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 557         if (!dst_pmd)
 558                 return -ENOMEM;
 559         src_pmd = pmd_offset(src_pud, addr);
 560         do {
 561                 next = pmd_addr_end(addr, end);
 562                 if (pmd_none_or_clear_bad(src_pmd))
 563                         continue;
 564                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
 565                                                 vma, addr, next))
 566                         return -ENOMEM;
 567         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
 568         return 0;
 569 }
 570
 571 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 572                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
 573                 unsigned long addr, unsigned long end)
 574 {
 575         pud_t *src_pud, *dst_pud;
 576         unsigned long next;
 577
 578         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
 579         if (!dst_pud)
 580                 return -ENOMEM;
 581         src_pud = pud_offset(src_pgd, addr);
 582         do {
 583                 next = pud_addr_end(addr, end);
 584                 if (pud_none_or_clear_bad(src_pud))
 585                         continue;
 586                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
 587                                                 vma, addr, next))
 588                         return -ENOMEM;
 589         } while (dst_pud++, src_pud++, addr = next, addr != end);
 590         return 0;
 591 }
 592
 593 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 594                 struct vm_area_struct *vma)
 595 {
 596         pgd_t *src_pgd, *dst_pgd;
 597         unsigned long next;
 598         unsigned long addr = vma->vm_start;
 599         unsigned long end = vma->vm_end;
 600
 601         /*
 602          * Don't copy ptes where a page fault will fill them correctly.
 603          * Fork becomes much lighter when there are big shared or private
 604          * readonly mappings. The tradeoff is that copy_page_range is more
 605          * efficient than faulting.
 606          */
 607         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
 608                 if (!vma->anon_vma)
 609                         return 0;
 610         }
 611
 612         if (is_vm_hugetlb_page(vma))
 613                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
 614
 615         dst_pgd = pgd_offset(dst_mm, addr);
 616         src_pgd = pgd_offset(src_mm, addr);
 617         do {
 618                 next = pgd_addr_end(addr, end);
 619                 if (pgd_none_or_clear_bad(src_pgd))
 620                         continue;
 621                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
 622                                                 vma, addr, next))
 623                         return -ENOMEM;
 624         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
 625         return 0;
 626 }
 627
 628 static unsigned long zap_pte_range(struct mmu_gather *tlb,
 629                                 struct vm_area_struct *vma, pmd_t *pmd,
 630                                 unsigned long addr, unsigned long end,
 631                                 long *zap_work, struct zap_details *details)
 632 {
 633         struct mm_struct *mm = tlb->mm;
 634         pte_t *pte;
 635         spinlock_t *ptl;
 636         int file_rss = 0;
 637         int anon_rss = 0;
 638
 639         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 640         arch_enter_lazy_mmu_mode();
 641         do {
 642                 pte_t ptent = *pte;
 643                 if (pte_none(ptent)) {
 644                         (*zap_work)--;
 645                         continue;
 646                 }
 647
 648                 (*zap_work) -= PAGE_SIZE;
 649
 650                 if (pte_present(ptent)) {
 651                         struct page *page;
 652
 653                         page = vm_normal_page(vma, addr, ptent);
 654                         if (unlikely(details) && page) {
 655                                 /*
 656                                  * unmap_shared_mapping_pages() wants to
 657                                  * invalidate cache without truncating:
 658                                  * unmap shared but keep private pages.
 659                                  */
 660                                 if (details->check_mapping &&
 661                                     details->check_mapping != page->mapping)
 662                                         continue;
 663                                 /*
 664                                  * Each page->index must be checked when
 665                                  * invalidating or truncating nonlinear.
 666                                  */
 667                                 if (details->nonlinear_vma &&
 668                                     (page->index < details->first_index ||
 669                                      page->index > details->last_index))
 670                                         continue;
 671                         }
 672                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 673                                                         tlb->fullmm);
 674                         tlb_remove_tlb_entry(tlb, pte, addr);
 675                         if (unlikely(!page))
 676                                 continue;
 677                         if (unlikely(details) && details->nonlinear_vma
 678                             && linear_page_index(details->nonlinear_vma,
 679                                                 addr) != page->index)
 680                                 set_pte_at(mm, addr, pte,
 681                                            pgoff_to_pte(page->index));
 682                         if (PageAnon(page))
 683                                 anon_rss--;
 684                         else {
 685                                 if (pte_dirty(ptent))
 686                                         set_page_dirty(page);
 687                                 if (pte_young(ptent))
 688                                         SetPageReferenced(page);
 689                                 file_rss--;
 690                         }
 691                         page_remove_rmap(page, vma);
 692                         tlb_remove_page(tlb, page);
 693                         continue;
 694                 }
 695                 /*
 696                  * If details->check_mapping, we leave swap entries;
 697                  * if details->nonlinear_vma, we leave file entries.
 698                  */
 699                 if (unlikely(details))
 700                         continue;
 701                 if (!pte_file(ptent))
 702                         free_swap_and_cache(pte_to_swp_entry(ptent));
 703                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 704         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
 705
 706         add_mm_rss(mm, file_rss, anon_rss);
 707         arch_leave_lazy_mmu_mode();
 708         pte_unmap_unlock(pte - 1, ptl);
 709
 710         return addr;
 711 }
 712
 713 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
 714                                 struct vm_area_struct *vma, pud_t *pud,
 715                                 unsigned long addr, unsigned long end,
 716                                 long *zap_work, struct zap_details *details)
 717 {
 718         pmd_t *pmd;
 719         unsigned long next;
 720
 721         pmd = pmd_offset(pud, addr);
 722         do {
 723                 next = pmd_addr_end(addr, end);
 724                 if (pmd_none_or_clear_bad(pmd)) {
 725                         (*zap_work)--;
 726                         continue;
 727                 }
 728                 next = zap_pte_range(tlb, vma, pmd, addr, next,
 729                                                 zap_work, details);
 730         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
 731
 732         return addr;
 733 }
 734
 735 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
 736                                 struct vm_area_struct *vma, pgd_t *pgd,
 737                                 unsigned long addr, unsigned long end,
 738                                 long *zap_work, struct zap_details *details)
 739 {
 740         pud_t *pud;
 741         unsigned long next;
 742
 743         pud = pud_offset(pgd, addr);
 744         do {
 745                 next = pud_addr_end(addr, end);
 746                 if (pud_none_or_clear_bad(pud)) {
 747                         (*zap_work)--;
 748                         continue;
 749                 }
 750                 next = zap_pmd_range(tlb, vma, pud, addr, next,
 751                                                 zap_work, details);
 752         } while (pud++, addr = next, (addr != end && *zap_work > 0));
 753
 754         return addr;
 755 }
 756
 757 static unsigned long unmap_page_range(struct mmu_gather *tlb,
 758                                 struct vm_area_struct *vma,
 759                                 unsigned long addr, unsigned long end,
 760                                 long *zap_work, struct zap_details *details)
 761 {
 762         pgd_t *pgd;
 763         unsigned long next;
 764
 765         if (details && !details->check_mapping && !details->nonlinear_vma)
 766                 details = NULL;
 767
 768         BUG_ON(addr >= end);
 769         tlb_start_vma(tlb, vma);
 770         pgd = pgd_offset(vma->vm_mm, addr);
 771         do {
 772                 next = pgd_addr_end(addr, end);
 773                 if (pgd_none_or_clear_bad(pgd)) {
 774                         (*zap_work)--;
 775                         continue;
 776                 }
 777                 next = zap_pud_range(tlb, vma, pgd, addr, next,
 778                                                 zap_work, details);
 779         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
 780         tlb_end_vma(tlb, vma);
 781
 782         return addr;
 783 }
 784
 785 #ifdef CONFIG_PREEMPT
 786 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
 787 #else
 788 /* No preempt: go for improved straight-line efficiency */
 789 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
 790 #endif
 791
 792 /**
 793  * unmap_vmas - unmap a range of memory covered by a list of vma's
 794  * @tlbp: address of the caller's struct mmu_gather
 795  * @vma: the starting vma
 796  * @start_addr: virtual address at which to start unmapping
 797  * @end_addr: virtual address at which to end unmapping
 798  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 799  * @details: details of nonlinear truncation or shared cache invalidation
 800  *
 801  * Returns the end address of the unmapping (restart addr if interrupted).
 802  *
 803  * Unmap all pages in the vma list.
 804  *
 805  * We aim to not hold locks for too long (for scheduling latency reasons).
 806  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 807  * return the ending mmu_gather to the caller.
 808  *
 809  * Only addresses between `start' and `end' will be unmapped.
 810  *
 811  * The VMA list must be sorted in ascending virtual address order.
 812  *
 813  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 814  * range after unmap_vmas() returns.  So the only responsibility here is to
 815  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 816  * drops the lock and schedules.
 817  */
 818 unsigned long unmap_vmas(struct mmu_gather **tlbp,
 819                 struct vm_area_struct *vma, unsigned long start_addr,
 820                 unsigned long end_addr, unsigned long *nr_accounted,
 821                 struct zap_details *details)
 822 {
 823         long zap_work = ZAP_BLOCK_SIZE;
 824         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 825         int tlb_start_valid = 0;
 826         unsigned long start = start_addr;
 827         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
 828         int fullmm = (*tlbp)->fullmm;
 829
 830         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 831                 unsigned long end;
 832
 833                 start = max(vma->vm_start, start_addr);
 834                 if (start >= vma->vm_end)
 835                         continue;
 836                 end = min(vma->vm_end, end_addr);
 837                 if (end <= vma->vm_start)
 838                         continue;
 839
 840                 if (vma->vm_flags & VM_ACCOUNT)
 841                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 842
 843                 while (start != end) {
 844                         if (!tlb_start_valid) {
 845                                 tlb_start = start;
 846                                 tlb_start_valid = 1;
 847                         }
 848
 849                         if (unlikely(is_vm_hugetlb_page(vma))) {
 850                                 unmap_hugepage_range(vma, start, end);
 851                                 zap_work -= (end - start) /
 852                                                 (HPAGE_SIZE / PAGE_SIZE);
 853                                 start = end;
 854                         } else
 855                                 start = unmap_page_range(*tlbp, vma,
 856                                                 start, end, &zap_work, details);
 857
 858                         if (zap_work > 0) {
 859                                 BUG_ON(start != end);
 860                                 break;
 861                         }
 862
 863                         tlb_finish_mmu(*tlbp, tlb_start, start);
 864
 865                         if (need_resched() ||
 866                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
 867                                 if (i_mmap_lock) {
 868                                         *tlbp = NULL;
 869                                         goto out;
 870                                 }
 871                                 cond_resched();
 872                         }
 873
 874                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
 875                         tlb_start_valid = 0;
 876                         zap_work = ZAP_BLOCK_SIZE;
 877                 }
 878         }
 879 out:
 880         return start;   /* which is now the end (or restart) address */
 881 }
 882
 883 /**
 884  * zap_page_range - remove user pages in a given range
 885  * @vma: vm_area_struct holding the applicable pages
 886  * @address: starting address of pages to zap
 887  * @size: number of bytes to zap
 888  * @details: details of nonlinear truncation or shared cache invalidation
 889  */
 890 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
 891                 unsigned long size, struct zap_details *details)
 892 {
 893         struct mm_struct *mm = vma->vm_mm;
 894         struct mmu_gather *tlb;
 895         unsigned long end = address + size;
 896         unsigned long nr_accounted = 0;
 897
 898         lru_add_drain();
 899         tlb = tlb_gather_mmu(mm, 0);
 900         update_hiwater_rss(mm);
 901         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
 902         if (tlb)
 903                 tlb_finish_mmu(tlb, address, end);
 904         return end;
 905 }
 906
 907 /*
 908  * Do a quick page-table lookup for a single page.
 909  */
 910 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
 911                         unsigned int flags)
 912 {
 913         pgd_t *pgd;
 914         pud_t *pud;
 915         pmd_t *pmd;
 916         pte_t *ptep, pte;
 917         spinlock_t *ptl;
 918         struct page *page;
 919         struct mm_struct *mm = vma->vm_mm;
 920
 921         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 922         if (!IS_ERR(page)) {
 923                 BUG_ON(flags & FOLL_GET);
 924                 goto out;
 925         }
 926
 927         page = NULL;
 928         pgd = pgd_offset(mm, address);
 929         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 930                 goto no_page_table;
 931
 932         pud = pud_offset(pgd, address);
 933         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
 934                 goto no_page_table;
 935
 936         pmd = pmd_offset(pud, address);
 937         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 938                 goto no_page_table;
 939
 940         if (pmd_huge(*pmd)) {
 941                 BUG_ON(flags & FOLL_GET);
 942                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
 943                 goto out;
 944         }
 945
 946         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 947         if (!ptep)
 948                 goto out;
 949
 950         pte = *ptep;
 951         if (!pte_present(pte))
 952                 goto unlock;
 953         if ((flags & FOLL_WRITE) && !pte_write(pte))
 954                 goto unlock;
 955         page = vm_normal_page(vma, address, pte);
 956         if (unlikely(!page))
 957                 goto unlock;
 958
 959         if (flags & FOLL_GET)
 960                 get_page(page);
 961         if (flags & FOLL_TOUCH) {
 962                 if ((flags & FOLL_WRITE) &&
 963                     !pte_dirty(pte) && !PageDirty(page))
 964                         set_page_dirty(page);
 965                 mark_page_accessed(page);
 966         }
 967 unlock:
 968         pte_unmap_unlock(ptep, ptl);
 969 out:
 970         return page;
 971
 972 no_page_table:
 973         /*
 974          * When core dumping an enormous anonymous area that nobody
 975          * has touched so far, we don't want to allocate page tables.
 976          */
 977         if (flags & FOLL_ANON) {
 978                 page = ZERO_PAGE(0);
 979                 if (flags & FOLL_GET)
 980                         get_page(page);
 981                 BUG_ON(flags & FOLL_WRITE);
 982         }
 983         return page;
 984 }
 985
 986 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 987                 unsigned long start, int len, int write, int force,
 988                 struct page **pages, struct vm_area_struct **vmas)
 989 {
 990         int i;
 991         unsigned int vm_flags;
 992
 993         /*
 994          * Require read or write permissions.
 995          * If 'force' is set, we only require the "MAY" flags.
 996          */
 997         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 998         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 999         i = 0;
1000
1001         do {
1002                 struct vm_area_struct *vma;
1003                 unsigned int foll_flags;
1004
1005                 vma = find_extend_vma(mm, start);
1006                 if (!vma && in_gate_area(tsk, start)) {
1007                         unsigned long pg = start & PAGE_MASK;
1008                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1009                         pgd_t *pgd;
1010                         pud_t *pud;
1011                         pmd_t *pmd;
1012                         pte_t *pte;
1013                         if (write) /* user gate pages are read-only */
1014                                 return i ? : -EFAULT;
1015                         if (pg > TASK_SIZE)
1016                                 pgd = pgd_offset_k(pg);
1017                         else
1018                                 pgd = pgd_offset_gate(mm, pg);
1019                         BUG_ON(pgd_none(*pgd));
1020                         pud = pud_offset(pgd, pg);
1021                         BUG_ON(pud_none(*pud));
1022                         pmd = pmd_offset(pud, pg);
1023                         if (pmd_none(*pmd))
1024                                 return i ? : -EFAULT;
1025                         pte = pte_offset_map(pmd, pg);
1026                         if (pte_none(*pte)) {
1027                                 pte_unmap(pte);
1028                                 return i ? : -EFAULT;
1029                         }
1030                         if (pages) {
1031                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1032                                 pages[i] = page;
1033                                 if (page)
1034                                         get_page(page);
1035                         }
1036                         pte_unmap(pte);
1037                         if (vmas)
1038                                 vmas[i] = gate_vma;
1039                         i++;
1040                         start += PAGE_SIZE;
1041                         len--;
1042                         continue;
1043                 }
1044
1045                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1046                                 || !(vm_flags & vma->vm_flags))
1047                         return i ? : -EFAULT;
1048
1049                 if (is_vm_hugetlb_page(vma)) {
1050                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1051                                                 &start, &len, i, write);
1052                         continue;
1053                 }
1054
1055                 foll_flags = FOLL_TOUCH;
1056                 if (pages)
1057                         foll_flags |= FOLL_GET;
1058                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1059                     (!vma->vm_ops || (!vma->vm_ops->nopage &&
1060                                         !vma->vm_ops->fault)))
1061                         foll_flags |= FOLL_ANON;
1062
1063                 do {
1064                         struct page *page;
1065
1066                         /*
1067                          * If tsk is ooming, cut off its access to large memory
1068                          * allocations. It has a pending SIGKILL, but it can't
1069                          * be processed until returning to user space.
1070                          */
1071                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1072                                 return -ENOMEM;
1073
1074                         if (write)
1075                                 foll_flags |= FOLL_WRITE;
1076
1077                         cond_resched();
1078                         while (!(page = follow_page(vma, start, foll_flags))) {
1079                                 int ret;
1080                                 ret = handle_mm_fault(mm, vma, start,
1081                                                 foll_flags & FOLL_WRITE);
1082                                 if (ret & VM_FAULT_ERROR) {
1083                                         if (ret & VM_FAULT_OOM)
1084                                                 return i ? i : -ENOMEM;
1085                                         else if (ret & VM_FAULT_SIGBUS)
1086                                                 return i ? i : -EFAULT;
1087                                         BUG();
1088                                 }
1089                                 if (ret & VM_FAULT_MAJOR)
1090                                         tsk->maj_flt++;
1091                                 else
1092                                         tsk->min_flt++;
1093
1094                                 /*
1095                                  * The VM_FAULT_WRITE bit tells us that
1096                                  * do_wp_page has broken COW when necessary,
1097                                  * even if maybe_mkwrite decided not to set
1098                                  * pte_write. We can thus safely do subsequent
1099                                  * page lookups as if they were reads.
1100                                  */
1101                                 if (ret & VM_FAULT_WRITE)
1102                                         foll_flags &= ~FOLL_WRITE;
1103
1104                                 cond_resched();
1105                         }
1106                         if (pages) {
1107                                 pages[i] = page;
1108
1109                                 flush_anon_page(vma, page, start);
1110                                 flush_dcache_page(page);
1111                         }
1112                         if (vmas)
1113                                 vmas[i] = vma;
1114                         i++;
1115                         start += PAGE_SIZE;
1116                         len--;
1117                 } while (len && start < vma->vm_end);
1118         } while (len);
1119         return i;
1120 }
1121 EXPORT_SYMBOL(get_user_pages);
1122
1123 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1124                         spinlock_t **ptl)
1125 {
1126         pgd_t * pgd = pgd_offset(mm, addr);
1127         pud_t * pud = pud_alloc(mm, pgd, addr);
1128         if (pud) {
1129                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1130                 if (pmd)
1131                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1132         }
1133         return NULL;
1134 }
1135
1136 /*
1137  * This is the old fallback for page remapping.
1138  *
1139  * For historical reasons, it only allows reserved pages. Only
1140  * old drivers should use this, and they needed to mark their
1141  * pages reserved for the old functions anyway.
1142  */
1143 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1144 {
1145         int retval;
1146         pte_t *pte;
1147         spinlock_t *ptl;
1148
1149         retval = -EINVAL;
1150         if (PageAnon(page))
1151                 goto out;
1152         retval = -ENOMEM;
1153         flush_dcache_page(page);
1154         pte = get_locked_pte(mm, addr, &ptl);
1155         if (!pte)
1156                 goto out;
1157         retval = -EBUSY;
1158         if (!pte_none(*pte))
1159                 goto out_unlock;
1160
1161         /* Ok, finally just insert the thing.. */
1162         get_page(page);
1163         inc_mm_counter(mm, file_rss);
1164         page_add_file_rmap(page);
1165         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1166
1167         retval = 0;
1168 out_unlock:
1169         pte_unmap_unlock(pte, ptl);
1170 out:
1171         return retval;
1172 }
1173
1174 /**
1175  * vm_insert_page - insert single page into user vma
1176  * @vma: user vma to map to
1177  * @addr: target user address of this page
1178  * @page: source kernel page
1179  *
1180  * This allows drivers to insert individual pages they've allocated
1181  * into a user vma.
1182  *
1183  * The page has to be a nice clean _individual_ kernel allocation.
1184  * If you allocate a compound page, you need to have marked it as
1185  * such (__GFP_COMP), or manually just split the page up yourself
1186  * (see split_page()).
1187  *
1188  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1189  * took an arbitrary page protection parameter. This doesn't allow
1190  * that. Your vma protection will have to be set up correctly, which
1191  * means that if you want a shared writable mapping, you'd better
1192  * ask for a shared writable mapping!
1193  *
1194  * The page does not need to be reserved.
1195  */
1196 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1197 {
1198         if (addr < vma->vm_start || addr >= vma->vm_end)
1199                 return -EFAULT;
1200         if (!page_count(page))
1201                 return -EINVAL;
1202         vma->vm_flags |= VM_INSERTPAGE;
1203         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1204 }
1205 EXPORT_SYMBOL(vm_insert_page);
1206
1207 /**
1208  * vm_insert_pfn - insert single pfn into user vma
1209  * @vma: user vma to map to
1210  * @addr: target user address of this page
1211  * @pfn: source kernel pfn
1212  *
1213  * Similar to vm_inert_page, this allows drivers to insert individual pages
1214  * they've allocated into a user vma. Same comments apply.
1215  *
1216  * This function should only be called from a vm_ops->fault handler, and
1217  * in that case the handler should return NULL.
1218  */
1219 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1220                 unsigned long pfn)
1221 {
1222         struct mm_struct *mm = vma->vm_mm;
1223         int retval;
1224         pte_t *pte, entry;
1225         spinlock_t *ptl;
1226
1227         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
1228         BUG_ON(is_cow_mapping(vma->vm_flags));
1229
1230         retval = -ENOMEM;
1231         pte = get_locked_pte(mm, addr, &ptl);
1232         if (!pte)
1233                 goto out;
1234         retval = -EBUSY;
1235         if (!pte_none(*pte))
1236                 goto out_unlock;
1237
1238         /* Ok, finally just insert the thing.. */
1239         entry = pfn_pte(pfn, vma->vm_page_prot);
1240         set_pte_at(mm, addr, pte, entry);
1241         update_mmu_cache(vma, addr, entry);
1242
1243         retval = 0;
1244 out_unlock:
1245         pte_unmap_unlock(pte, ptl);
1246
1247 out:
1248         return retval;
1249 }
1250 EXPORT_SYMBOL(vm_insert_pfn);
1251
1252 /*
1253  * maps a range of physical memory into the requested pages. the old
1254  * mappings are removed. any references to nonexistent pages results
1255  * in null mappings (currently treated as "copy-on-access")
1256  */
1257 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1258                         unsigned long addr, unsigned long end,
1259                         unsigned long pfn, pgprot_t prot)
1260 {
1261         pte_t *pte;
1262         spinlock_t *ptl;
1263
1264         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1265         if (!pte)
1266                 return -ENOMEM;
1267         arch_enter_lazy_mmu_mode();
1268         do {
1269                 BUG_ON(!pte_none(*pte));
1270                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1271                 pfn++;
1272         } while (pte++, addr += PAGE_SIZE, addr != end);
1273         arch_leave_lazy_mmu_mode();
1274         pte_unmap_unlock(pte - 1, ptl);
1275         return 0;
1276 }
1277
1278 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1279                         unsigned long addr, unsigned long end,
1280                         unsigned long pfn, pgprot_t prot)
1281 {
1282         pmd_t *pmd;
1283         unsigned long next;
1284
1285         pfn -= addr >> PAGE_SHIFT;
1286         pmd = pmd_alloc(mm, pud, addr);
1287         if (!pmd)
1288                 return -ENOMEM;
1289         do {
1290                 next = pmd_addr_end(addr, end);
1291                 if (remap_pte_range(mm, pmd, addr, next,
1292                                 pfn + (addr >> PAGE_SHIFT), prot))
1293                         return -ENOMEM;
1294         } while (pmd++, addr = next, addr != end);
1295         return 0;
1296 }
1297
1298 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1299                         unsigned long addr, unsigned long end,
1300                         unsigned long pfn, pgprot_t prot)
1301 {
1302         pud_t *pud;
1303         unsigned long next;
1304
1305         pfn -= addr >> PAGE_SHIFT;
1306         pud = pud_alloc(mm, pgd, addr);
1307         if (!pud)
1308                 return -ENOMEM;
1309         do {
1310                 next = pud_addr_end(addr, end);
1311                 if (remap_pmd_range(mm, pud, addr, next,
1312                                 pfn + (addr >> PAGE_SHIFT), prot))
1313                         return -ENOMEM;
1314         } while (pud++, addr = next, addr != end);
1315         return 0;
1316 }
1317
1318 /**
1319  * remap_pfn_range - remap kernel memory to userspace
1320  * @vma: user vma to map to
1321  * @addr: target user address to start at
1322  * @pfn: physical address of kernel memory
1323  * @size: size of map area
1324  * @prot: page protection flags for this mapping
1325  *
1326  *  Note: this is only safe if the mm semaphore is held when called.
1327  */
1328 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1329                     unsigned long pfn, unsigned long size, pgprot_t prot)
1330 {
1331         pgd_t *pgd;
1332         unsigned long next;
1333         unsigned long end = addr + PAGE_ALIGN(size);
1334         struct mm_struct *mm = vma->vm_mm;
1335         int err;
1336
1337         /*
1338          * Physically remapped pages are special. Tell the
1339          * rest of the world about it:
1340          *   VM_IO tells people not to look at these pages
1341          *      (accesses can have side effects).
1342          *   VM_RESERVED is specified all over the place, because
1343          *      in 2.4 it kept swapout's vma scan off this vma; but
1344          *      in 2.6 the LRU scan won't even find its pages, so this
1345          *      flag means no more than count its pages in reserved_vm,
1346          *      and omit it from core dump, even when VM_IO turned off.
1347          *   VM_PFNMAP tells the core MM that the base pages are just
1348          *      raw PFN mappings, and do not have a "struct page" associated
1349          *      with them.
1350          *
1351          * There's a horrible special case to handle copy-on-write
1352          * behaviour that some programs depend on. We mark the "original"
1353          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1354          */
1355         if (is_cow_mapping(vma->vm_flags)) {
1356                 if (addr != vma->vm_start || end != vma->vm_end)
1357                         return -EINVAL;
1358                 vma->vm_pgoff = pfn;
1359         }
1360
1361         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1362
1363         BUG_ON(addr >= end);
1364         pfn -= addr >> PAGE_SHIFT;
1365         pgd = pgd_offset(mm, addr);
1366         flush_cache_range(vma, addr, end);
1367         do {
1368                 next = pgd_addr_end(addr, end);
1369                 err = remap_pud_range(mm, pgd, addr, next,
1370                                 pfn + (addr >> PAGE_SHIFT), prot);
1371                 if (err)
1372                         break;
1373         } while (pgd++, addr = next, addr != end);
1374         return err;
1375 }
1376 EXPORT_SYMBOL(remap_pfn_range);
1377
1378 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1379                                      unsigned long addr, unsigned long end,
1380                                      pte_fn_t fn, void *data)
1381 {
1382         pte_t *pte;
1383         int err;
1384         struct page *pmd_page;
1385         spinlock_t *uninitialized_var(ptl);
1386
1387         pte = (mm == &init_mm) ?
1388                 pte_alloc_kernel(pmd, addr) :
1389                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1390         if (!pte)
1391                 return -ENOMEM;
1392
1393         BUG_ON(pmd_huge(*pmd));
1394
1395         pmd_page = pmd_page(*pmd);
1396
1397         do {
1398                 err = fn(pte, pmd_page, addr, data);
1399                 if (err)
1400                         break;
1401         } while (pte++, addr += PAGE_SIZE, addr != end);
1402
1403         if (mm != &init_mm)
1404                 pte_unmap_unlock(pte-1, ptl);
1405         return err;
1406 }
1407
1408 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1409                                      unsigned long addr, unsigned long end,
1410                                      pte_fn_t fn, void *data)
1411 {
1412         pmd_t *pmd;
1413         unsigned long next;
1414         int err;
1415
1416         pmd = pmd_alloc(mm, pud, addr);
1417         if (!pmd)
1418                 return -ENOMEM;
1419         do {
1420                 next = pmd_addr_end(addr, end);
1421                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1422                 if (err)
1423                         break;
1424         } while (pmd++, addr = next, addr != end);
1425         return err;
1426 }
1427
1428 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1429                                      unsigned long addr, unsigned long end,
1430                                      pte_fn_t fn, void *data)
1431 {
1432         pud_t *pud;
1433         unsigned long next;
1434         int err;
1435
1436         pud = pud_alloc(mm, pgd, addr);
1437         if (!pud)
1438                 return -ENOMEM;
1439         do {
1440                 next = pud_addr_end(addr, end);
1441                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1442                 if (err)
1443                         break;
1444         } while (pud++, addr = next, addr != end);
1445         return err;
1446 }
1447
1448 /*
1449  * Scan a region of virtual memory, filling in page tables as necessary
1450  * and calling a provided function on each leaf page table.
1451  */
1452 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1453                         unsigned long size, pte_fn_t fn, void *data)
1454 {
1455         pgd_t *pgd;
1456         unsigned long next;
1457         unsigned long end = addr + size;
1458         int err;
1459
1460         BUG_ON(addr >= end);
1461         pgd = pgd_offset(mm, addr);
1462         do {
1463                 next = pgd_addr_end(addr, end);
1464                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1465                 if (err)
1466                         break;
1467         } while (pgd++, addr = next, addr != end);
1468         return err;
1469 }
1470 EXPORT_SYMBOL_GPL(apply_to_page_range);
1471
1472 /*
1473  * handle_pte_fault chooses page fault handler according to an entry
1474  * which was read non-atomically.  Before making any commitment, on
1475  * those architectures or configurations (e.g. i386 with PAE) which
1476  * might give a mix of unmatched parts, do_swap_page and do_file_page
1477  * must check under lock before unmapping the pte and proceeding
1478  * (but do_wp_page is only called after already making such a check;
1479  * and do_anonymous_page and do_no_page can safely check later on).
1480  */
1481 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1482                                 pte_t *page_table, pte_t orig_pte)
1483 {
1484         int same = 1;
1485 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1486         if (sizeof(pte_t) > sizeof(unsigned long)) {
1487                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1488                 spin_lock(ptl);
1489                 same = pte_same(*page_table, orig_pte);
1490                 spin_unlock(ptl);
1491         }
1492 #endif
1493         pte_unmap(page_table);
1494         return same;
1495 }
1496
1497 /*
1498  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1499  * servicing faults for write access.  In the normal case, do always want
1500  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1501  * that do not have writing enabled, when used by access_process_vm.
1502  */
1503 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1504 {
1505         if (likely(vma->vm_flags & VM_WRITE))
1506                 pte = pte_mkwrite(pte);
1507         return pte;
1508 }
1509
1510 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1511 {
1512         /*
1513          * If the source page was a PFN mapping, we don't have
1514          * a "struct page" for it. We do a best-effort copy by
1515          * just copying from the original user address. If that
1516          * fails, we just zero-fill it. Live with it.
1517          */
1518         if (unlikely(!src)) {
1519                 void *kaddr = kmap_atomic(dst, KM_USER0);
1520                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1521
1522                 /*
1523                  * This really shouldn't fail, because the page is there
1524                  * in the page tables. But it might just be unreadable,
1525                  * in which case we just give up and fill the result with
1526                  * zeroes.
1527                  */
1528                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1529                         memset(kaddr, 0, PAGE_SIZE);
1530                 kunmap_atomic(kaddr, KM_USER0);
1531                 flush_dcache_page(dst);
1532         } else
1533                 copy_user_highpage(dst, src, va, vma);
1534 }
1535
1536 /*
1537  * This routine handles present pages, when users try to write
1538  * to a shared page. It is done by copying the page to a new address
1539  * and decrementing the shared-page counter for the old page.
1540  *
1541  * Note that this routine assumes that the protection checks have been
1542  * done by the caller (the low-level page fault routine in most cases).
1543  * Thus we can safely just mark it writable once we've done any necessary
1544  * COW.
1545  *
1546  * We also mark the page dirty at this point even though the page will
1547  * change only once the write actually happens. This avoids a few races,
1548  * and potentially makes it more efficient.
1549  *
1550  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1551  * but allow concurrent faults), with pte both mapped and locked.
1552  * We return with mmap_sem still held, but pte unmapped and unlocked.
1553  */
1554 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1555                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1556                 spinlock_t *ptl, pte_t orig_pte)
1557 {
1558         struct page *old_page, *new_page;
1559         pte_t entry;
1560         int reuse = 0, ret = 0;
1561         int page_mkwrite = 0;
1562         struct page *dirty_page = NULL;
1563
1564         old_page = vm_normal_page(vma, address, orig_pte);
1565         if (!old_page)
1566                 goto gotten;
1567
1568         /*
1569          * Take out anonymous pages first, anonymous shared vmas are
1570          * not dirty accountable.
1571          */
1572         if (PageAnon(old_page)) {
1573                 if (!TestSetPageLocked(old_page)) {
1574                         reuse = can_share_swap_page(old_page);
1575                         unlock_page(old_page);
1576                 }
1577         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1578                                         (VM_WRITE|VM_SHARED))) {
1579                 /*
1580                  * Only catch write-faults on shared writable pages,
1581                  * read-only shared pages can get COWed by
1582                  * get_user_pages(.write=1, .force=1).
1583                  */
1584                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1585                         /*
1586                          * Notify the address space that the page is about to
1587                          * become writable so that it can prohibit this or wait
1588                          * for the page to get into an appropriate state.
1589                          *
1590                          * We do this without the lock held, so that it can
1591                          * sleep if it needs to.
1592                          */
1593                         page_cache_get(old_page);
1594                         pte_unmap_unlock(page_table, ptl);
1595
1596                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1597                                 goto unwritable_page;
1598
1599                         /*
1600                          * Since we dropped the lock we need to revalidate
1601                          * the PTE as someone else may have changed it.  If
1602                          * they did, we just return, as we can count on the
1603                          * MMU to tell us if they didn't also make it writable.
1604                          */
1605                         page_table = pte_offset_map_lock(mm, pmd, address,
1606                                                          &ptl);
1607                         page_cache_release(old_page);
1608                         if (!pte_same(*page_table, orig_pte))
1609                                 goto unlock;
1610
1611                         page_mkwrite = 1;
1612                 }
1613                 dirty_page = old_page;
1614                 get_page(dirty_page);
1615                 reuse = 1;
1616         }
1617
1618         if (reuse) {
1619                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1620                 entry = pte_mkyoung(orig_pte);
1621                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1622                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1623                         update_mmu_cache(vma, address, entry);
1624                 ret |= VM_FAULT_WRITE;
1625                 goto unlock;
1626         }
1627
1628         /*
1629          * Ok, we need to copy. Oh, well..
1630          */
1631         page_cache_get(old_page);
1632 gotten:
1633         pte_unmap_unlock(page_table, ptl);
1634
1635         if (unlikely(anon_vma_prepare(vma)))
1636                 goto oom;
1637         VM_BUG_ON(old_page == ZERO_PAGE(0));
1638         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1639         if (!new_page)
1640                 goto oom;
1641         cow_user_page(new_page, old_page, address, vma);
1642         __SetPageUptodate(new_page);
1643
1644         /*
1645          * Re-check the pte - we dropped the lock
1646          */
1647         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1648         if (likely(pte_same(*page_table, orig_pte))) {
1649                 if (old_page) {
1650                         page_remove_rmap(old_page, vma);
1651                         if (!PageAnon(old_page)) {
1652                                 dec_mm_counter(mm, file_rss);
1653                                 inc_mm_counter(mm, anon_rss);
1654                         }
1655                 } else
1656                         inc_mm_counter(mm, anon_rss);
1657                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1658                 entry = mk_pte(new_page, vma->vm_page_prot);
1659                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1660                 /*
1661                  * Clear the pte entry and flush it first, before updating the
1662                  * pte with the new entry. This will avoid a race condition
1663                  * seen in the presence of one thread doing SMC and another
1664                  * thread doing COW.
1665                  */
1666                 ptep_clear_flush(vma, address, page_table);
1667                 set_pte_at(mm, address, page_table, entry);
1668                 update_mmu_cache(vma, address, entry);
1669                 lru_cache_add_active(new_page);
1670                 page_add_new_anon_rmap(new_page, vma, address);
1671
1672                 /* Free the old page.. */
1673                 new_page = old_page;
1674                 ret |= VM_FAULT_WRITE;
1675         }
1676         if (new_page)
1677                 page_cache_release(new_page);
1678         if (old_page)
1679                 page_cache_release(old_page);
1680 unlock:
1681         pte_unmap_unlock(page_table, ptl);
1682         if (dirty_page) {
1683                 if (vma->vm_file)
1684                         file_update_time(vma->vm_file);
1685
1686                 /*
1687                  * Yes, Virginia, this is actually required to prevent a race
1688                  * with clear_page_dirty_for_io() from clearing the page dirty
1689                  * bit after it clear all dirty ptes, but before a racing
1690                  * do_wp_page installs a dirty pte.
1691                  *
1692                  * do_no_page is protected similarly.
1693                  */
1694                 wait_on_page_locked(dirty_page);
1695                 set_page_dirty_balance(dirty_page, page_mkwrite);
1696                 put_page(dirty_page);
1697         }
1698         return ret;
1699 oom:
1700         if (old_page)
1701                 page_cache_release(old_page);
1702         return VM_FAULT_OOM;
1703
1704 unwritable_page:
1705         page_cache_release(old_page);
1706         return VM_FAULT_SIGBUS;
1707 }
1708
1709 /*
1710  * Helper functions for unmap_mapping_range().
1711  *
1712  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1713  *
1714  * We have to restart searching the prio_tree whenever we drop the lock,
1715  * since the iterator is only valid while the lock is held, and anyway
1716  * a later vma might be split and reinserted earlier while lock dropped.
1717  *
1718  * The list of nonlinear vmas could be handled more efficiently, using
1719  * a placeholder, but handle it in the same way until a need is shown.
1720  * It is important to search the prio_tree before nonlinear list: a vma
1721  * may become nonlinear and be shifted from prio_tree to nonlinear list
1722  * while the lock is dropped; but never shifted from list to prio_tree.
1723  *
1724  * In order to make forward progress despite restarting the search,
1725  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1726  * quickly skip it next time around.  Since the prio_tree search only
1727  * shows us those vmas affected by unmapping the range in question, we
1728  * can't efficiently keep all vmas in step with mapping->truncate_count:
1729  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1730  * mapping->truncate_count and vma->vm_truncate_count are protected by
1731  * i_mmap_lock.
1732  *
1733  * In order to make forward progress despite repeatedly restarting some
1734  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1735  * and restart from that address when we reach that vma again.  It might
1736  * have been split or merged, shrunk or extended, but never shifted: so
1737  * restart_addr remains valid so long as it remains in the vma's range.
1738  * unmap_mapping_range forces truncate_count to leap over page-aligned
1739  * values so we can save vma's restart_addr in its truncate_count field.
1740  */
1741 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1742
1743 static void reset_vma_truncate_counts(struct address_space *mapping)
1744 {
1745         struct vm_area_struct *vma;
1746         struct prio_tree_iter iter;
1747
1748         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1749                 vma->vm_truncate_count = 0;
1750         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1751                 vma->vm_truncate_count = 0;
1752 }
1753
1754 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1755                 unsigned long start_addr, unsigned long end_addr,
1756                 struct zap_details *details)
1757 {
1758         unsigned long restart_addr;
1759         int need_break;
1760
1761         /*
1762          * files that support invalidating or truncating portions of the
1763          * file from under mmaped areas must have their ->fault function
1764          * return a locked page (and set VM_FAULT_LOCKED in the return).
1765          * This provides synchronisation against concurrent unmapping here.
1766          */
1767
1768 again:
1769         restart_addr = vma->vm_truncate_count;
1770         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1771                 start_addr = restart_addr;
1772                 if (start_addr >= end_addr) {
1773                         /* Top of vma has been split off since last time */
1774                         vma->vm_truncate_count = details->truncate_count;
1775                         return 0;
1776                 }
1777         }
1778
1779         restart_addr = zap_page_range(vma, start_addr,
1780                                         end_addr - start_addr, details);
1781         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
1782
1783         if (restart_addr >= end_addr) {
1784                 /* We have now completed this vma: mark it so */
1785                 vma->vm_truncate_count = details->truncate_count;
1786                 if (!need_break)
1787                         return 0;
1788         } else {
1789                 /* Note restart_addr in vma's truncate_count field */
1790                 vma->vm_truncate_count = restart_addr;
1791                 if (!need_break)
1792                         goto again;
1793         }
1794
1795         spin_unlock(details->i_mmap_lock);
1796         cond_resched();
1797         spin_lock(details->i_mmap_lock);
1798         return -EINTR;
1799 }
1800
1801 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1802                                             struct zap_details *details)
1803 {
1804         struct vm_area_struct *vma;
1805         struct prio_tree_iter iter;
1806         pgoff_t vba, vea, zba, zea;
1807
1808 restart:
1809         vma_prio_tree_foreach(vma, &iter, root,
1810                         details->first_index, details->last_index) {
1811                 /* Skip quickly over those we have already dealt with */
1812                 if (vma->vm_truncate_count == details->truncate_count)
1813                         continue;
1814
1815                 vba = vma->vm_pgoff;
1816                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1817                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1818                 zba = details->first_index;
1819                 if (zba < vba)
1820                         zba = vba;
1821                 zea = details->last_index;
1822                 if (zea > vea)
1823                         zea = vea;
1824
1825                 if (unmap_mapping_range_vma(vma,
1826                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1827                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1828                                 details) < 0)
1829                         goto restart;
1830         }
1831 }
1832
1833 static inline void unmap_mapping_range_list(struct list_head *head,
1834                                             struct zap_details *details)
1835 {
1836         struct vm_area_struct *vma;
1837
1838         /*
1839          * In nonlinear VMAs there is no correspondence between virtual address
1840          * offset and file offset.  So we must perform an exhaustive search
1841          * across *all* the pages in each nonlinear VMA, not just the pages
1842          * whose virtual address lies outside the file truncation point.
1843          */
1844 restart:
1845         list_for_each_entry(vma, head, shared.vm_set.list) {
1846                 /* Skip quickly over those we have already dealt with */
1847                 if (vma->vm_truncate_count == details->truncate_count)
1848                         continue;
1849                 details->nonlinear_vma = vma;
1850                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1851                                         vma->vm_end, details) < 0)
1852                         goto restart;
1853         }
1854 }
1855
1856 /**
1857  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1858  * @mapping: the address space containing mmaps to be unmapped.
1859  * @holebegin: byte in first page to unmap, relative to the start of
1860  * the underlying file.  This will be rounded down to a PAGE_SIZE
1861  * boundary.  Note that this is different from vmtruncate(), which
1862  * must keep the partial page.  In contrast, we must get rid of
1863  * partial pages.
1864  * @holelen: size of prospective hole in bytes.  This will be rounded
1865  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1866  * end of the file.
1867  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1868  * but 0 when invalidating pagecache, don't throw away private data.
1869  */
1870 void unmap_mapping_range(struct address_space *mapping,
1871                 loff_t const holebegin, loff_t const holelen, int even_cows)
1872 {
1873         struct zap_details details;
1874         pgoff_t hba = holebegin >> PAGE_SHIFT;
1875         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1876
1877         /* Check for overflow. */
1878         if (sizeof(holelen) > sizeof(hlen)) {
1879                 long long holeend =
1880                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1881                 if (holeend & ~(long long)ULONG_MAX)
1882                         hlen = ULONG_MAX - hba + 1;
1883         }
1884
1885         details.check_mapping = even_cows? NULL: mapping;
1886         details.nonlinear_vma = NULL;
1887         details.first_index = hba;
1888         details.last_index = hba + hlen - 1;
1889         if (details.last_index < details.first_index)
1890                 details.last_index = ULONG_MAX;
1891         details.i_mmap_lock = &mapping->i_mmap_lock;
1892
1893         spin_lock(&mapping->i_mmap_lock);
1894
1895         /* Protect against endless unmapping loops */
1896         mapping->truncate_count++;
1897         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1898                 if (mapping->truncate_count == 0)
1899                         reset_vma_truncate_counts(mapping);
1900                 mapping->truncate_count++;
1901         }
1902         details.truncate_count = mapping->truncate_count;
1903
1904         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1905                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1906         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1907                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1908         spin_unlock(&mapping->i_mmap_lock);
1909 }
1910 EXPORT_SYMBOL(unmap_mapping_range);
1911
1912 /**
1913  * vmtruncate - unmap mappings "freed" by truncate() syscall
1914  * @inode: inode of the file used
1915  * @offset: file offset to start truncating
1916  *
1917  * NOTE! We have to be ready to update the memory sharing
1918  * between the file and the memory map for a potential last
1919  * incomplete page.  Ugly, but necessary.
1920  */
1921 int vmtruncate(struct inode * inode, loff_t offset)
1922 {
1923         if (inode->i_size < offset) {
1924                 unsigned long limit;
1925
1926                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1927                 if (limit != RLIM_INFINITY && offset > limit)
1928                         goto out_sig;
1929                 if (offset > inode->i_sb->s_maxbytes)
1930                         goto out_big;
1931                 i_size_write(inode, offset);
1932         } else {
1933                 struct address_space *mapping = inode->i_mapping;
1934
1935                 /*
1936                  * truncation of in-use swapfiles is disallowed - it would
1937                  * cause subsequent swapout to scribble on the now-freed
1938                  * blocks.
1939                  */
1940                 if (IS_SWAPFILE(inode))
1941                         return -ETXTBSY;
1942                 i_size_write(inode, offset);
1943
1944                 /*
1945                  * unmap_mapping_range is called twice, first simply for
1946                  * efficiency so that truncate_inode_pages does fewer
1947                  * single-page unmaps.  However after this first call, and
1948                  * before truncate_inode_pages finishes, it is possible for
1949                  * private pages to be COWed, which remain after
1950                  * truncate_inode_pages finishes, hence the second
1951                  * unmap_mapping_range call must be made for correctness.
1952                  */
1953                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1954                 truncate_inode_pages(mapping, offset);
1955                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1956         }
1957
1958         if (inode->i_op && inode->i_op->truncate)
1959                 inode->i_op->truncate(inode);
1960         return 0;
1961
1962 out_sig:
1963         send_sig(SIGXFSZ, current, 0);
1964 out_big:
1965         return -EFBIG;
1966 }
1967 EXPORT_SYMBOL(vmtruncate);
1968
1969 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1970 {
1971         struct address_space *mapping = inode->i_mapping;
1972
1973         /*
1974          * If the underlying filesystem is not going to provide
1975          * a way to truncate a range of blocks (punch a hole) -
1976          * we should return failure right now.
1977          */
1978         if (!inode->i_op || !inode->i_op->truncate_range)
1979                 return -ENOSYS;
1980
1981         mutex_lock(&inode->i_mutex);
1982         down_write(&inode->i_alloc_sem);
1983         unmap_mapping_range(mapping, offset, (end - offset), 1);
1984         truncate_inode_pages_range(mapping, offset, end);
1985         unmap_mapping_range(mapping, offset, (end - offset), 1);
1986         inode->i_op->truncate_range(inode, offset, end);
1987         up_write(&inode->i_alloc_sem);
1988         mutex_unlock(&inode->i_mutex);
1989
1990         return 0;
1991 }
1992
1993 /*
1994  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1995  * but allow concurrent faults), and pte mapped but not yet locked.
1996  * We return with mmap_sem still held, but pte unmapped and unlocked.
1997  */
1998 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1999                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2000                 int write_access, pte_t orig_pte)
2001 {
2002         spinlock_t *ptl;
2003         struct page *page;
2004         swp_entry_t entry;
2005         pte_t pte;
2006         int ret = 0;
2007
2008         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2009                 goto out;
2010
2011         entry = pte_to_swp_entry(orig_pte);
2012         if (is_migration_entry(entry)) {
2013                 migration_entry_wait(mm, pmd, address);
2014                 goto out;
2015         }
2016         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2017         page = lookup_swap_cache(entry);
2018         if (!page) {
2019                 grab_swap_token(); /* Contend for token _before_ read-in */
2020                 page = swapin_readahead(entry,
2021                                         GFP_HIGHUSER_MOVABLE, vma, address);
2022                 if (!page) {
2023                         /*
2024                          * Back out if somebody else faulted in this pte
2025                          * while we released the pte lock.
2026                          */
2027                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2028                         if (likely(pte_same(*page_table, orig_pte)))
2029                                 ret = VM_FAULT_OOM;
2030                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2031                         goto unlock;
2032                 }
2033
2034                 /* Had to read the page from swap area: Major fault */
2035                 ret = VM_FAULT_MAJOR;
2036                 count_vm_event(PGMAJFAULT);
2037         }
2038
2039         mark_page_accessed(page);
2040         lock_page(page);
2041         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2042
2043         /*
2044          * Back out if somebody else already faulted in this pte.
2045          */
2046         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2047         if (unlikely(!pte_same(*page_table, orig_pte)))
2048                 goto out_nomap;
2049
2050         if (unlikely(!PageUptodate(page))) {
2051                 ret = VM_FAULT_SIGBUS;
2052                 goto out_nomap;
2053         }
2054
2055         /* The page isn't present yet, go ahead with the fault. */
2056
2057         inc_mm_counter(mm, anon_rss);
2058         pte = mk_pte(page, vma->vm_page_prot);
2059         if (write_access && can_share_swap_page(page)) {
2060                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2061                 write_access = 0;
2062         }
2063
2064         flush_icache_page(vma, page);
2065         set_pte_at(mm, address, page_table, pte);
2066         page_add_anon_rmap(page, vma, address);
2067
2068         swap_free(entry);
2069         if (vm_swap_full())
2070                 remove_exclusive_swap_page(page);
2071         unlock_page(page);
2072
2073         if (write_access) {
2074                 /* XXX: We could OR the do_wp_page code with this one? */
2075                 if (do_wp_page(mm, vma, address,
2076                                 page_table, pmd, ptl, pte) & VM_FAULT_OOM)
2077                         ret = VM_FAULT_OOM;
2078                 goto out;
2079         }
2080
2081         /* No need to invalidate - it was non-present before */
2082         update_mmu_cache(vma, address, pte);
2083 unlock:
2084         pte_unmap_unlock(page_table, ptl);
2085 out:
2086         return ret;
2087 out_nomap:
2088         pte_unmap_unlock(page_table, ptl);
2089         unlock_page(page);
2090         page_cache_release(page);
2091         return ret;
2092 }
2093
2094 /*
2095  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2096  * but allow concurrent faults), and pte mapped but not yet locked.
2097  * We return with mmap_sem still held, but pte unmapped and unlocked.
2098  */
2099 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2100                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2101                 int write_access)
2102 {
2103         struct page *page;
2104         spinlock_t *ptl;
2105         pte_t entry;
2106
2107         /* Allocate our own private page. */
2108         pte_unmap(page_table);
2109
2110         if (unlikely(anon_vma_prepare(vma)))
2111                 goto oom;
2112         page = alloc_zeroed_user_highpage_movable(vma, address);
2113         if (!page)
2114                 goto oom;
2115         __SetPageUptodate(page);
2116
2117         entry = mk_pte(page, vma->vm_page_prot);
2118         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2119
2120         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2121         if (!pte_none(*page_table))
2122                 goto release;
2123         inc_mm_counter(mm, anon_rss);
2124         lru_cache_add_active(page);
2125         page_add_new_anon_rmap(page, vma, address);
2126         set_pte_at(mm, address, page_table, entry);
2127
2128         /* No need to invalidate - it was non-present before */
2129         update_mmu_cache(vma, address, entry);
2130 unlock:
2131         pte_unmap_unlock(page_table, ptl);
2132         return 0;
2133 release:
2134         page_cache_release(page);
2135         goto unlock;
2136 oom:
2137         return VM_FAULT_OOM;
2138 }
2139
2140 /*
2141  * __do_fault() tries to create a new page mapping. It aggressively
2142  * tries to share with existing pages, but makes a separate copy if
2143  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2144  * the next page fault.
2145  *
2146  * As this is called only for pages that do not currently exist, we
2147  * do not need to flush old virtual caches or the TLB.
2148  *
2149  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2150  * but allow concurrent faults), and pte neither mapped nor locked.
2151  * We return with mmap_sem still held, but pte unmapped and unlocked.
2152  */
2153 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2154                 unsigned long address, pmd_t *pmd,
2155                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2156 {
2157         pte_t *page_table;
2158         spinlock_t *ptl;
2159         struct page *page;
2160         pte_t entry;
2161         int anon = 0;
2162         struct page *dirty_page = NULL;
2163         struct vm_fault vmf;
2164         int ret;
2165         int page_mkwrite = 0;
2166
2167         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2168         vmf.pgoff = pgoff;
2169         vmf.flags = flags;
2170         vmf.page = NULL;
2171
2172         BUG_ON(vma->vm_flags & VM_PFNMAP);
2173
2174         if (likely(vma->vm_ops->fault)) {
2175                 ret = vma->vm_ops->fault(vma, &vmf);
2176                 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2177                         return ret;
2178         } else {
2179                 /* Legacy ->nopage path */
2180                 ret = 0;
2181                 vmf.page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2182                 /* no page was available -- either SIGBUS or OOM */
2183                 if (unlikely(vmf.page == NOPAGE_SIGBUS))
2184                         return VM_FAULT_SIGBUS;
2185                 else if (unlikely(vmf.page == NOPAGE_OOM))
2186                         return VM_FAULT_OOM;
2187         }
2188
2189         /*
2190          * For consistency in subsequent calls, make the faulted page always
2191          * locked.
2192          */
2193         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2194                 lock_page(vmf.page);
2195         else
2196                 VM_BUG_ON(!PageLocked(vmf.page));
2197
2198         /*
2199          * Should we do an early C-O-W break?
2200          */
2201         page = vmf.page;
2202         if (flags & FAULT_FLAG_WRITE) {
2203                 if (!(vma->vm_flags & VM_SHARED)) {
2204                         anon = 1;
2205                         if (unlikely(anon_vma_prepare(vma))) {
2206                                 ret = VM_FAULT_OOM;
2207                                 goto out;
2208                         }
2209                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2210                                                 vma, address);
2211                         if (!page) {
2212                                 ret = VM_FAULT_OOM;
2213                                 goto out;
2214                         }
2215                         copy_user_highpage(page, vmf.page, address, vma);
2216                         __SetPageUptodate(page);
2217                 } else {
2218                         /*
2219                          * If the page will be shareable, see if the backing
2220                          * address space wants to know that the page is about
2221                          * to become writable
2222                          */
2223                         if (vma->vm_ops->page_mkwrite) {
2224                                 unlock_page(page);
2225                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2226                                         ret = VM_FAULT_SIGBUS;
2227                                         anon = 1; /* no anon but release vmf.page */
2228                                         goto out_unlocked;
2229                                 }
2230                                 lock_page(page);
2231                                 /*
2232                                  * XXX: this is not quite right (racy vs
2233                                  * invalidate) to unlock and relock the page
2234                                  * like this, however a better fix requires
2235                                  * reworking page_mkwrite locking API, which
2236                                  * is better done later.
2237                                  */
2238                                 if (!page->mapping) {
2239                                         ret = 0;
2240                                         anon = 1; /* no anon but release vmf.page */
2241                                         goto out;
2242                                 }
2243                                 page_mkwrite = 1;
2244                         }
2245                 }
2246
2247         }
2248
2249         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2250
2251         /*
2252          * This silly early PAGE_DIRTY setting removes a race
2253          * due to the bad i386 page protection. But it's valid
2254          * for other architectures too.
2255          *
2256          * Note that if write_access is true, we either now have
2257          * an exclusive copy of the page, or this is a shared mapping,
2258          * so we can make it writable and dirty to avoid having to
2259          * handle that later.
2260          */
2261         /* Only go through if we didn't race with anybody else... */
2262         if (likely(pte_same(*page_table, orig_pte))) {
2263                 flush_icache_page(vma, page);
2264                 entry = mk_pte(page, vma->vm_page_prot);
2265                 if (flags & FAULT_FLAG_WRITE)
2266                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2267                 set_pte_at(mm, address, page_table, entry);
2268                 if (anon) {
2269                         inc_mm_counter(mm, anon_rss);
2270                         lru_cache_add_active(page);
2271                         page_add_new_anon_rmap(page, vma, address);
2272                 } else {
2273                         inc_mm_counter(mm, file_rss);
2274                         page_add_file_rmap(page);
2275                         if (flags & FAULT_FLAG_WRITE) {
2276                                 dirty_page = page;
2277                                 get_page(dirty_page);
2278                         }
2279                 }
2280
2281                 /* no need to invalidate: a not-present page won't be cached */
2282                 update_mmu_cache(vma, address, entry);
2283         } else {
2284                 if (anon)
2285                         page_cache_release(page);
2286                 else
2287                         anon = 1; /* no anon but release faulted_page */
2288         }
2289
2290         pte_unmap_unlock(page_table, ptl);
2291
2292 out:
2293         unlock_page(vmf.page);
2294 out_unlocked:
2295         if (anon)
2296                 page_cache_release(vmf.page);
2297         else if (dirty_page) {
2298                 if (vma->vm_file)
2299                         file_update_time(vma->vm_file);
2300
2301                 set_page_dirty_balance(dirty_page, page_mkwrite);
2302                 put_page(dirty_page);
2303         }
2304
2305         return ret;
2306 }
2307
2308 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2309                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2310                 int write_access, pte_t orig_pte)
2311 {
2312         pgoff_t pgoff = (((address & PAGE_MASK)
2313                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2314         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2315
2316         pte_unmap(page_table);
2317         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2318 }
2319
2320
2321 /*
2322  * do_no_pfn() tries to create a new page mapping for a page without
2323  * a struct_page backing it
2324  *
2325  * As this is called only for pages that do not currently exist, we
2326  * do not need to flush old virtual caches or the TLB.
2327  *
2328  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2329  * but allow concurrent faults), and pte mapped but not yet locked.
2330  * We return with mmap_sem still held, but pte unmapped and unlocked.
2331  *
2332  * It is expected that the ->nopfn handler always returns the same pfn
2333  * for a given virtual mapping.
2334  *
2335  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2336  */
2337 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2338                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2339                      int write_access)
2340 {
2341         spinlock_t *ptl;
2342         pte_t entry;
2343         unsigned long pfn;
2344
2345         pte_unmap(page_table);
2346         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
2347         BUG_ON(is_cow_mapping(vma->vm_flags));
2348
2349         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2350         if (unlikely(pfn == NOPFN_OOM))
2351                 return VM_FAULT_OOM;
2352         else if (unlikely(pfn == NOPFN_SIGBUS))
2353                 return VM_FAULT_SIGBUS;
2354         else if (unlikely(pfn == NOPFN_REFAULT))
2355                 return 0;
2356
2357         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2358
2359         /* Only go through if we didn't race with anybody else... */
2360         if (pte_none(*page_table)) {
2361                 entry = pfn_pte(pfn, vma->vm_page_prot);
2362                 if (write_access)
2363                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2364                 set_pte_at(mm, address, page_table, entry);
2365         }
2366         pte_unmap_unlock(page_table, ptl);
2367         return 0;
2368 }
2369
2370 /*
2371  * Fault of a previously existing named mapping. Repopulate the pte
2372  * from the encoded file_pte if possible. This enables swappable
2373  * nonlinear vmas.
2374  *
2375  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2376  * but allow concurrent faults), and pte mapped but not yet locked.
2377  * We return with mmap_sem still held, but pte unmapped and unlocked.
2378  */
2379 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2380                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2381                 int write_access, pte_t orig_pte)
2382 {
2383         unsigned int flags = FAULT_FLAG_NONLINEAR |
2384                                 (write_access ? FAULT_FLAG_WRITE : 0);
2385         pgoff_t pgoff;
2386
2387         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2388                 return 0;
2389
2390         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2391                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2392                 /*
2393                  * Page table corrupted: show pte and kill process.
2394                  */
2395                 print_bad_pte(vma, orig_pte, address);
2396                 return VM_FAULT_OOM;
2397         }
2398
2399         pgoff = pte_to_pgoff(orig_pte);
2400         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2401 }
2402
2403 /*
2404  * These routines also need to handle stuff like marking pages dirty
2405  * and/or accessed for architectures that don't do it in hardware (most
2406  * RISC architectures).  The early dirtying is also good on the i386.
2407  *
2408  * There is also a hook called "update_mmu_cache()" that architectures
2409  * with external mmu caches can use to update those (ie the Sparc or
2410  * PowerPC hashed page tables that act as extended TLBs).
2411  *
2412  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2413  * but allow concurrent faults), and pte mapped but not yet locked.
2414  * We return with mmap_sem still held, but pte unmapped and unlocked.
2415  */
2416 static inline int handle_pte_fault(struct mm_struct *mm,
2417                 struct vm_area_struct *vma, unsigned long address,
2418                 pte_t *pte, pmd_t *pmd, int write_access)
2419 {
2420         pte_t entry;
2421         spinlock_t *ptl;
2422
2423         entry = *pte;
2424         if (!pte_present(entry)) {
2425                 if (pte_none(entry)) {
2426                         if (vma->vm_ops) {
2427                                 if (vma->vm_ops->fault || vma->vm_ops->nopage)
2428                                         return do_linear_fault(mm, vma, address,
2429                                                 pte, pmd, write_access, entry);
2430                                 if (unlikely(vma->vm_ops->nopfn))
2431                                         return do_no_pfn(mm, vma, address, pte,
2432                                                          pmd, write_access);
2433                         }
2434                         return do_anonymous_page(mm, vma, address,
2435                                                  pte, pmd, write_access);
2436                 }
2437                 if (pte_file(entry))
2438                         return do_nonlinear_fault(mm, vma, address,
2439                                         pte, pmd, write_access, entry);
2440                 return do_swap_page(mm, vma, address,
2441                                         pte, pmd, write_access, entry);
2442         }
2443
2444         ptl = pte_lockptr(mm, pmd);
2445         spin_lock(ptl);
2446         if (unlikely(!pte_same(*pte, entry)))
2447                 goto unlock;
2448         if (write_access) {
2449                 if (!pte_write(entry))
2450                         return do_wp_page(mm, vma, address,
2451                                         pte, pmd, ptl, entry);
2452                 entry = pte_mkdirty(entry);
2453         }
2454         entry = pte_mkyoung(entry);
2455         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2456                 update_mmu_cache(vma, address, entry);
2457         } else {
2458                 /*
2459                  * This is needed only for protection faults but the arch code
2460                  * is not yet telling us if this is a protection fault or not.
2461                  * This still avoids useless tlb flushes for .text page faults
2462                  * with threads.
2463                  */
2464                 if (write_access)
2465                         flush_tlb_page(vma, address);
2466         }
2467 unlock:
2468         pte_unmap_unlock(pte, ptl);
2469         return 0;
2470 }
2471
2472 /*
2473  * By the time we get here, we already hold the mm semaphore
2474  */
2475 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2476                 unsigned long address, int write_access)
2477 {
2478         pgd_t *pgd;
2479         pud_t *pud;
2480         pmd_t *pmd;
2481         pte_t *pte;
2482
2483         __set_current_state(TASK_RUNNING);
2484
2485         count_vm_event(PGFAULT);
2486
2487         if (unlikely(is_vm_hugetlb_page(vma)))
2488                 return hugetlb_fault(mm, vma, address, write_access);
2489
2490         pgd = pgd_offset(mm, address);
2491         pud = pud_alloc(mm, pgd, address);
2492         if (!pud)
2493                 return VM_FAULT_OOM;
2494         pmd = pmd_alloc(mm, pud, address);
2495         if (!pmd)
2496                 return VM_FAULT_OOM;
2497         pte = pte_alloc_map(mm, pmd, address);
2498         if (!pte)
2499                 return VM_FAULT_OOM;
2500
2501         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2502 }
2503
2504 #ifndef __PAGETABLE_PUD_FOLDED
2505 /*
2506  * Allocate page upper directory.
2507  * We've already handled the fast-path in-line.
2508  */
2509 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2510 {
2511         pud_t *new = pud_alloc_one(mm, address);
2512         if (!new)
2513                 return -ENOMEM;
2514
2515         spin_lock(&mm->page_table_lock);
2516         if (pgd_present(*pgd))          /* Another has populated it */
2517                 pud_free(mm, new);
2518         else
2519                 pgd_populate(mm, pgd, new);
2520         spin_unlock(&mm->page_table_lock);
2521         return 0;
2522 }
2523 #endif /* __PAGETABLE_PUD_FOLDED */
2524
2525 #ifndef __PAGETABLE_PMD_FOLDED
2526 /*
2527  * Allocate page middle directory.
2528  * We've already handled the fast-path in-line.
2529  */
2530 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2531 {
2532         pmd_t *new = pmd_alloc_one(mm, address);
2533         if (!new)
2534                 return -ENOMEM;
2535
2536         spin_lock(&mm->page_table_lock);
2537 #ifndef __ARCH_HAS_4LEVEL_HACK
2538         if (pud_present(*pud))          /* Another has populated it */
2539                 pmd_free(mm, new);
2540         else
2541                 pud_populate(mm, pud, new);
2542 #else
2543         if (pgd_present(*pud))          /* Another has populated it */
2544                 pmd_free(mm, new);
2545         else
2546                 pgd_populate(mm, pud, new);
2547 #endif /* __ARCH_HAS_4LEVEL_HACK */
2548         spin_unlock(&mm->page_table_lock);
2549         return 0;
2550 }
2551 #endif /* __PAGETABLE_PMD_FOLDED */
2552
2553 int make_pages_present(unsigned long addr, unsigned long end)
2554 {
2555         int ret, len, write;
2556         struct vm_area_struct * vma;
2557
2558         vma = find_vma(current->mm, addr);
2559         if (!vma)
2560                 return -1;
2561         write = (vma->vm_flags & VM_WRITE) != 0;
2562         BUG_ON(addr >= end);
2563         BUG_ON(end > vma->vm_end);
2564         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2565         ret = get_user_pages(current, current->mm, addr,
2566                         len, write, 0, NULL, NULL);
2567         if (ret < 0)
2568                 return ret;
2569         return ret == len ? 0 : -1;
2570 }
2571
2572 #if !defined(__HAVE_ARCH_GATE_AREA)
2573
2574 #if defined(AT_SYSINFO_EHDR)
2575 static struct vm_area_struct gate_vma;
2576
2577 static int __init gate_vma_init(void)
2578 {
2579         gate_vma.vm_mm = NULL;
2580         gate_vma.vm_start = FIXADDR_USER_START;
2581         gate_vma.vm_end = FIXADDR_USER_END;
2582         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2583         gate_vma.vm_page_prot = __P101;
2584         /*
2585          * Make sure the vDSO gets into every core dump.
2586          * Dumping its contents makes post-mortem fully interpretable later
2587          * without matching up the same kernel and hardware config to see
2588          * what PC values meant.
2589          */
2590         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2591         return 0;
2592 }
2593 __initcall(gate_vma_init);
2594 #endif
2595
2596 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2597 {
2598 #ifdef AT_SYSINFO_EHDR
2599         return &gate_vma;
2600 #else
2601         return NULL;
2602 #endif
2603 }
2604
2605 int in_gate_area_no_task(unsigned long addr)
2606 {
2607 #ifdef AT_SYSINFO_EHDR
2608         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2609                 return 1;
2610 #endif
2611         return 0;
2612 }
2613
2614 #endif  /* __HAVE_ARCH_GATE_AREA */
2615
2616 /*
2617  * Access another process' address space.
2618  * Source/target buffer must be kernel space,
2619  * Do not walk the page table directly, use get_user_pages
2620  */
2621 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2622 {
2623         struct mm_struct *mm;
2624         struct vm_area_struct *vma;
2625         struct page *page;
2626         void *old_buf = buf;
2627
2628         mm = get_task_mm(tsk);
2629         if (!mm)
2630                 return 0;
2631
2632         down_read(&mm->mmap_sem);
2633         /* ignore errors, just check how much was successfully transferred */
2634         while (len) {
2635                 int bytes, ret, offset;
2636                 void *maddr;
2637
2638                 ret = get_user_pages(tsk, mm, addr, 1,
2639                                 write, 1, &page, &vma);
2640                 if (ret <= 0)
2641                         break;
2642
2643                 bytes = len;
2644                 offset = addr & (PAGE_SIZE-1);
2645                 if (bytes > PAGE_SIZE-offset)
2646                         bytes = PAGE_SIZE-offset;
2647
2648                 maddr = kmap(page);
2649                 if (write) {
2650                         copy_to_user_page(vma, page, addr,
2651                                           maddr + offset, buf, bytes);
2652                         set_page_dirty_lock(page);
2653                 } else {
2654                         copy_from_user_page(vma, page, addr,
2655                                             buf, maddr + offset, bytes);
2656                 }
2657                 kunmap(page);
2658                 page_cache_release(page);
2659                 len -= bytes;
2660                 buf += bytes;
2661                 addr += bytes;
2662         }
2663         up_read(&mm->mmap_sem);
2664         mmput(mm);
2665
2666         return buf - old_buf;
2667 }
2668
2669 /*
2670  * Print the name of a VMA.
2671  */
2672 void print_vma_addr(char *prefix, unsigned long ip)
2673 {
2674         struct mm_struct *mm = current->mm;
2675         struct vm_area_struct *vma;
2676
2677         down_read(&mm->mmap_sem);
2678         vma = find_vma(mm, ip);
2679         if (vma && vma->vm_file) {
2680                 struct file *f = vma->vm_file;
2681                 char *buf = (char *)__get_free_page(GFP_KERNEL);
2682                 if (buf) {
2683                         char *p, *s;
2684
2685                         p = d_path(f->f_dentry, f->f_vfsmnt, buf, PAGE_SIZE);
2686                         if (IS_ERR(p))
2687                                 p = "?";
2688                         s = strrchr(p, '/');
2689                         if (s)
2690                                 p = s+1;
2691                         printk("%s%s[%lx+%lx]", prefix, p,
2692                                         vma->vm_start,
2693                                         vma->vm_end - vma->vm_start);
2694                         free_page((unsigned long)buf);
2695                 }
2696         }
2697         up_read(&current->mm->mmap_sem);
2698 }