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/init.h>
  51
  52 #include <asm/pgalloc.h>
  53 #include <asm/uaccess.h>
  54 #include <asm/tlb.h>
  55 #include <asm/tlbflush.h>
  56 #include <asm/pgtable.h>
  57
  58 #include <linux/swapops.h>
  59 #include <linux/elf.h>
  60
  61 #ifndef CONFIG_NEED_MULTIPLE_NODES
  62 /* use the per-pgdat data instead for discontigmem - mbligh */
  63 unsigned long max_mapnr;
  64 struct page *mem_map;
  65
  66 EXPORT_SYMBOL(max_mapnr);
  67 EXPORT_SYMBOL(mem_map);
  68 #endif
  69
  70 unsigned long num_physpages;
  71 /*
  72  * A number of key systems in x86 including ioremap() rely on the assumption
  73  * that high_memory defines the upper bound on direct map memory, then end
  74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  76  * and ZONE_HIGHMEM.
  77  */
  78 void * high_memory;
  79 unsigned long vmalloc_earlyreserve;
  80
  81 EXPORT_SYMBOL(num_physpages);
  82 EXPORT_SYMBOL(high_memory);
  83 EXPORT_SYMBOL(vmalloc_earlyreserve);
  84
  85 /*
  86  * If a p?d_bad entry is found while walking page tables, report
  87  * the error, before resetting entry to p?d_none.  Usually (but
  88  * very seldom) called out from the p?d_none_or_clear_bad macros.
  89  */
  90
  91 void pgd_clear_bad(pgd_t *pgd)
  92 {
  93         pgd_ERROR(*pgd);
  94         pgd_clear(pgd);
  95 }
  96
  97 void pud_clear_bad(pud_t *pud)
  98 {
  99         pud_ERROR(*pud);
 100         pud_clear(pud);
 101 }
 102
 103 void pmd_clear_bad(pmd_t *pmd)
 104 {
 105         pmd_ERROR(*pmd);
 106         pmd_clear(pmd);
 107 }
 108
 109 /*
 110  * Note: this doesn't free the actual pages themselves. That
 111  * has been handled earlier when unmapping all the memory regions.
 112  */
 113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
 114 {
 115         struct page *page = pmd_page(*pmd);
 116         pmd_clear(pmd);
 117         pte_free_tlb(tlb, page);
 118         dec_page_state(nr_page_table_pages);
 119         tlb->mm->nr_ptes--;
 120 }
 121
 122 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 123                                 unsigned long addr, unsigned long end,
 124                                 unsigned long floor, unsigned long ceiling)
 125 {
 126         pmd_t *pmd;
 127         unsigned long next;
 128         unsigned long start;
 129
 130         start = addr;
 131         pmd = pmd_offset(pud, addr);
 132         do {
 133                 next = pmd_addr_end(addr, end);
 134                 if (pmd_none_or_clear_bad(pmd))
 135                         continue;
 136                 free_pte_range(tlb, pmd);
 137         } while (pmd++, addr = next, addr != end);
 138
 139         start &= PUD_MASK;
 140         if (start < floor)
 141                 return;
 142         if (ceiling) {
 143                 ceiling &= PUD_MASK;
 144                 if (!ceiling)
 145                         return;
 146         }
 147         if (end - 1 > ceiling - 1)
 148                 return;
 149
 150         pmd = pmd_offset(pud, start);
 151         pud_clear(pud);
 152         pmd_free_tlb(tlb, pmd);
 153 }
 154
 155 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 156                                 unsigned long addr, unsigned long end,
 157                                 unsigned long floor, unsigned long ceiling)
 158 {
 159         pud_t *pud;
 160         unsigned long next;
 161         unsigned long start;
 162
 163         start = addr;
 164         pud = pud_offset(pgd, addr);
 165         do {
 166                 next = pud_addr_end(addr, end);
 167                 if (pud_none_or_clear_bad(pud))
 168                         continue;
 169                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 170         } while (pud++, addr = next, addr != end);
 171
 172         start &= PGDIR_MASK;
 173         if (start < floor)
 174                 return;
 175         if (ceiling) {
 176                 ceiling &= PGDIR_MASK;
 177                 if (!ceiling)
 178                         return;
 179         }
 180         if (end - 1 > ceiling - 1)
 181                 return;
 182
 183         pud = pud_offset(pgd, start);
 184         pgd_clear(pgd);
 185         pud_free_tlb(tlb, pud);
 186 }
 187
 188 /*
 189  * This function frees user-level page tables of a process.
 190  *
 191  * Must be called with pagetable lock held.
 192  */
 193 void free_pgd_range(struct mmu_gather **tlb,
 194                         unsigned long addr, unsigned long end,
 195                         unsigned long floor, unsigned long ceiling)
 196 {
 197         pgd_t *pgd;
 198         unsigned long next;
 199         unsigned long start;
 200
 201         /*
 202          * The next few lines have given us lots of grief...
 203          *
 204          * Why are we testing PMD* at this top level?  Because often
 205          * there will be no work to do at all, and we'd prefer not to
 206          * go all the way down to the bottom just to discover that.
 207          *
 208          * Why all these "- 1"s?  Because 0 represents both the bottom
 209          * of the address space and the top of it (using -1 for the
 210          * top wouldn't help much: the masks would do the wrong thing).
 211          * The rule is that addr 0 and floor 0 refer to the bottom of
 212          * the address space, but end 0 and ceiling 0 refer to the top
 213          * Comparisons need to use "end - 1" and "ceiling - 1" (though
 214          * that end 0 case should be mythical).
 215          *
 216          * Wherever addr is brought up or ceiling brought down, we must
 217          * be careful to reject "the opposite 0" before it confuses the
 218          * subsequent tests.  But what about where end is brought down
 219          * by PMD_SIZE below? no, end can't go down to 0 there.
 220          *
 221          * Whereas we round start (addr) and ceiling down, by different
 222          * masks at different levels, in order to test whether a table
 223          * now has no other vmas using it, so can be freed, we don't
 224          * bother to round floor or end up - the tests don't need that.
 225          */
 226
 227         addr &= PMD_MASK;
 228         if (addr < floor) {
 229                 addr += PMD_SIZE;
 230                 if (!addr)
 231                         return;
 232         }
 233         if (ceiling) {
 234                 ceiling &= PMD_MASK;
 235                 if (!ceiling)
 236                         return;
 237         }
 238         if (end - 1 > ceiling - 1)
 239                 end -= PMD_SIZE;
 240         if (addr > end - 1)
 241                 return;
 242
 243         start = addr;
 244         pgd = pgd_offset((*tlb)->mm, addr);
 245         do {
 246                 next = pgd_addr_end(addr, end);
 247                 if (pgd_none_or_clear_bad(pgd))
 248                         continue;
 249                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
 250         } while (pgd++, addr = next, addr != end);
 251
 252         if (!(*tlb)->fullmm)
 253                 flush_tlb_pgtables((*tlb)->mm, start, end);
 254 }
 255
 256 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
 257                 unsigned long floor, unsigned long ceiling)
 258 {
 259         while (vma) {
 260                 struct vm_area_struct *next = vma->vm_next;
 261                 unsigned long addr = vma->vm_start;
 262
 263                 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
 264                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 265                                 floor, next? next->vm_start: ceiling);
 266                 } else {
 267                         /*
 268                          * Optimization: gather nearby vmas into one call down
 269                          */
 270                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 271                           && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
 272                                                         HPAGE_SIZE)) {
 273                                 vma = next;
 274                                 next = vma->vm_next;
 275                         }
 276                         free_pgd_range(tlb, addr, vma->vm_end,
 277                                 floor, next? next->vm_start: ceiling);
 278                 }
 279                 vma = next;
 280         }
 281 }
 282
 283 pte_t fastcall *pte_alloc_map(struct mm_struct *mm, pmd_t *pmd,
 284                                 unsigned long address)
 285 {
 286         if (!pmd_present(*pmd)) {
 287                 struct page *new;
 288
 289                 spin_unlock(&mm->page_table_lock);
 290                 new = pte_alloc_one(mm, address);
 291                 spin_lock(&mm->page_table_lock);
 292                 if (!new)
 293                         return NULL;
 294                 /*
 295                  * Because we dropped the lock, we should re-check the
 296                  * entry, as somebody else could have populated it..
 297                  */
 298                 if (pmd_present(*pmd)) {
 299                         pte_free(new);
 300                         goto out;
 301                 }
 302                 mm->nr_ptes++;
 303                 inc_page_state(nr_page_table_pages);
 304                 pmd_populate(mm, pmd, new);
 305         }
 306 out:
 307         return pte_offset_map(pmd, address);
 308 }
 309
 310 pte_t fastcall * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 311 {
 312         if (!pmd_present(*pmd)) {
 313                 pte_t *new;
 314
 315                 spin_unlock(&mm->page_table_lock);
 316                 new = pte_alloc_one_kernel(mm, address);
 317                 spin_lock(&mm->page_table_lock);
 318                 if (!new)
 319                         return NULL;
 320
 321                 /*
 322                  * Because we dropped the lock, we should re-check the
 323                  * entry, as somebody else could have populated it..
 324                  */
 325                 if (pmd_present(*pmd)) {
 326                         pte_free_kernel(new);
 327                         goto out;
 328                 }
 329                 pmd_populate_kernel(mm, pmd, new);
 330         }
 331 out:
 332         return pte_offset_kernel(pmd, address);
 333 }
 334
 335 /*
 336  * copy one vm_area from one task to the other. Assumes the page tables
 337  * already present in the new task to be cleared in the whole range
 338  * covered by this vma.
 339  *
 340  * dst->page_table_lock is held on entry and exit,
 341  * but may be dropped within p[mg]d_alloc() and pte_alloc_map().
 342  */
 343
 344 static inline void
 345 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 346                 pte_t *dst_pte, pte_t *src_pte, unsigned long vm_flags,
 347                 unsigned long addr)
 348 {
 349         pte_t pte = *src_pte;
 350         struct page *page;
 351         unsigned long pfn;
 352
 353         /* pte contains position in swap or file, so copy. */
 354         if (unlikely(!pte_present(pte))) {
 355                 if (!pte_file(pte)) {
 356                         swap_duplicate(pte_to_swp_entry(pte));
 357                         /* make sure dst_mm is on swapoff's mmlist. */
 358                         if (unlikely(list_empty(&dst_mm->mmlist))) {
 359                                 spin_lock(&mmlist_lock);
 360                                 list_add(&dst_mm->mmlist, &src_mm->mmlist);
 361                                 spin_unlock(&mmlist_lock);
 362                         }
 363                 }
 364                 set_pte_at(dst_mm, addr, dst_pte, pte);
 365                 return;
 366         }
 367
 368         pfn = pte_pfn(pte);
 369         /* the pte points outside of valid memory, the
 370          * mapping is assumed to be good, meaningful
 371          * and not mapped via rmap - duplicate the
 372          * mapping as is.
 373          */
 374         page = NULL;
 375         if (pfn_valid(pfn))
 376                 page = pfn_to_page(pfn);
 377
 378         if (!page || PageReserved(page)) {
 379                 set_pte_at(dst_mm, addr, dst_pte, pte);
 380                 return;
 381         }
 382
 383         /*
 384          * If it's a COW mapping, write protect it both
 385          * in the parent and the child
 386          */
 387         if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
 388                 ptep_set_wrprotect(src_mm, addr, src_pte);
 389                 pte = *src_pte;
 390         }
 391
 392         /*
 393          * If it's a shared mapping, mark it clean in
 394          * the child
 395          */
 396         if (vm_flags & VM_SHARED)
 397                 pte = pte_mkclean(pte);
 398         pte = pte_mkold(pte);
 399         get_page(page);
 400         inc_mm_counter(dst_mm, rss);
 401         if (PageAnon(page))
 402                 inc_mm_counter(dst_mm, anon_rss);
 403         set_pte_at(dst_mm, addr, dst_pte, pte);
 404         page_dup_rmap(page);
 405 }
 406
 407 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 408                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 409                 unsigned long addr, unsigned long end)
 410 {
 411         pte_t *src_pte, *dst_pte;
 412         unsigned long vm_flags = vma->vm_flags;
 413         int progress = 0;
 414
 415 again:
 416         dst_pte = pte_alloc_map(dst_mm, dst_pmd, addr);
 417         if (!dst_pte)
 418                 return -ENOMEM;
 419         src_pte = pte_offset_map_nested(src_pmd, addr);
 420
 421         spin_lock(&src_mm->page_table_lock);
 422         do {
 423                 /*
 424                  * We are holding two locks at this point - either of them
 425                  * could generate latencies in another task on another CPU.
 426                  */
 427                 if (progress >= 32) {
 428                         progress = 0;
 429                         if (need_resched() ||
 430                             need_lockbreak(&src_mm->page_table_lock) ||
 431                             need_lockbreak(&dst_mm->page_table_lock))
 432                                 break;
 433                 }
 434                 if (pte_none(*src_pte)) {
 435                         progress++;
 436                         continue;
 437                 }
 438                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vm_flags, addr);
 439                 progress += 8;
 440         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 441         spin_unlock(&src_mm->page_table_lock);
 442
 443         pte_unmap_nested(src_pte - 1);
 444         pte_unmap(dst_pte - 1);
 445         cond_resched_lock(&dst_mm->page_table_lock);
 446         if (addr != end)
 447                 goto again;
 448         return 0;
 449 }
 450
 451 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 452                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 453                 unsigned long addr, unsigned long end)
 454 {
 455         pmd_t *src_pmd, *dst_pmd;
 456         unsigned long next;
 457
 458         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 459         if (!dst_pmd)
 460                 return -ENOMEM;
 461         src_pmd = pmd_offset(src_pud, addr);
 462         do {
 463                 next = pmd_addr_end(addr, end);
 464                 if (pmd_none_or_clear_bad(src_pmd))
 465                         continue;
 466                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
 467                                                 vma, addr, next))
 468                         return -ENOMEM;
 469         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
 470         return 0;
 471 }
 472
 473 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 474                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
 475                 unsigned long addr, unsigned long end)
 476 {
 477         pud_t *src_pud, *dst_pud;
 478         unsigned long next;
 479
 480         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
 481         if (!dst_pud)
 482                 return -ENOMEM;
 483         src_pud = pud_offset(src_pgd, addr);
 484         do {
 485                 next = pud_addr_end(addr, end);
 486                 if (pud_none_or_clear_bad(src_pud))
 487                         continue;
 488                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
 489                                                 vma, addr, next))
 490                         return -ENOMEM;
 491         } while (dst_pud++, src_pud++, addr = next, addr != end);
 492         return 0;
 493 }
 494
 495 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 496                 struct vm_area_struct *vma)
 497 {
 498         pgd_t *src_pgd, *dst_pgd;
 499         unsigned long next;
 500         unsigned long addr = vma->vm_start;
 501         unsigned long end = vma->vm_end;
 502
 503         /*
 504          * Don't copy ptes where a page fault will fill them correctly.
 505          * Fork becomes much lighter when there are big shared or private
 506          * readonly mappings. The tradeoff is that copy_page_range is more
 507          * efficient than faulting.
 508          */
 509         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_RESERVED))) {
 510                 if (!vma->anon_vma)
 511                         return 0;
 512         }
 513
 514         if (is_vm_hugetlb_page(vma))
 515                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
 516
 517         dst_pgd = pgd_offset(dst_mm, addr);
 518         src_pgd = pgd_offset(src_mm, addr);
 519         do {
 520                 next = pgd_addr_end(addr, end);
 521                 if (pgd_none_or_clear_bad(src_pgd))
 522                         continue;
 523                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
 524                                                 vma, addr, next))
 525                         return -ENOMEM;
 526         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
 527         return 0;
 528 }
 529
 530 static void zap_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 531                                 unsigned long addr, unsigned long end,
 532                                 struct zap_details *details)
 533 {
 534         pte_t *pte;
 535
 536         pte = pte_offset_map(pmd, addr);
 537         do {
 538                 pte_t ptent = *pte;
 539                 if (pte_none(ptent))
 540                         continue;
 541                 if (pte_present(ptent)) {
 542                         struct page *page = NULL;
 543                         unsigned long pfn = pte_pfn(ptent);
 544                         if (pfn_valid(pfn)) {
 545                                 page = pfn_to_page(pfn);
 546                                 if (PageReserved(page))
 547                                         page = NULL;
 548                         }
 549                         if (unlikely(details) && page) {
 550                                 /*
 551                                  * unmap_shared_mapping_pages() wants to
 552                                  * invalidate cache without truncating:
 553                                  * unmap shared but keep private pages.
 554                                  */
 555                                 if (details->check_mapping &&
 556                                     details->check_mapping != page->mapping)
 557                                         continue;
 558                                 /*
 559                                  * Each page->index must be checked when
 560                                  * invalidating or truncating nonlinear.
 561                                  */
 562                                 if (details->nonlinear_vma &&
 563                                     (page->index < details->first_index ||
 564                                      page->index > details->last_index))
 565                                         continue;
 566                         }
 567                         ptent = ptep_get_and_clear_full(tlb->mm, addr, pte,
 568                                                         tlb->fullmm);
 569                         tlb_remove_tlb_entry(tlb, pte, addr);
 570                         if (unlikely(!page))
 571                                 continue;
 572                         if (unlikely(details) && details->nonlinear_vma
 573                             && linear_page_index(details->nonlinear_vma,
 574                                                 addr) != page->index)
 575                                 set_pte_at(tlb->mm, addr, pte,
 576                                            pgoff_to_pte(page->index));
 577                         if (PageAnon(page))
 578                                 dec_mm_counter(tlb->mm, anon_rss);
 579                         else {
 580                                 if (pte_dirty(ptent))
 581                                         set_page_dirty(page);
 582                                 if (pte_young(ptent))
 583                                         mark_page_accessed(page);
 584                         }
 585                         dec_mm_counter(tlb->mm, rss);
 586                         page_remove_rmap(page);
 587                         tlb_remove_page(tlb, page);
 588                         continue;
 589                 }
 590                 /*
 591                  * If details->check_mapping, we leave swap entries;
 592                  * if details->nonlinear_vma, we leave file entries.
 593                  */
 594                 if (unlikely(details))
 595                         continue;
 596                 if (!pte_file(ptent))
 597                         free_swap_and_cache(pte_to_swp_entry(ptent));
 598                 pte_clear_full(tlb->mm, addr, pte, tlb->fullmm);
 599         } while (pte++, addr += PAGE_SIZE, addr != end);
 600         pte_unmap(pte - 1);
 601 }
 602
 603 static inline void zap_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 604                                 unsigned long addr, unsigned long end,
 605                                 struct zap_details *details)
 606 {
 607         pmd_t *pmd;
 608         unsigned long next;
 609
 610         pmd = pmd_offset(pud, addr);
 611         do {
 612                 next = pmd_addr_end(addr, end);
 613                 if (pmd_none_or_clear_bad(pmd))
 614                         continue;
 615                 zap_pte_range(tlb, pmd, addr, next, details);
 616         } while (pmd++, addr = next, addr != end);
 617 }
 618
 619 static inline void zap_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 620                                 unsigned long addr, unsigned long end,
 621                                 struct zap_details *details)
 622 {
 623         pud_t *pud;
 624         unsigned long next;
 625
 626         pud = pud_offset(pgd, addr);
 627         do {
 628                 next = pud_addr_end(addr, end);
 629                 if (pud_none_or_clear_bad(pud))
 630                         continue;
 631                 zap_pmd_range(tlb, pud, addr, next, details);
 632         } while (pud++, addr = next, addr != end);
 633 }
 634
 635 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
 636                                 unsigned long addr, unsigned long end,
 637                                 struct zap_details *details)
 638 {
 639         pgd_t *pgd;
 640         unsigned long next;
 641
 642         if (details && !details->check_mapping && !details->nonlinear_vma)
 643                 details = NULL;
 644
 645         BUG_ON(addr >= end);
 646         tlb_start_vma(tlb, vma);
 647         pgd = pgd_offset(vma->vm_mm, addr);
 648         do {
 649                 next = pgd_addr_end(addr, end);
 650                 if (pgd_none_or_clear_bad(pgd))
 651                         continue;
 652                 zap_pud_range(tlb, pgd, addr, next, details);
 653         } while (pgd++, addr = next, addr != end);
 654         tlb_end_vma(tlb, vma);
 655 }
 656
 657 #ifdef CONFIG_PREEMPT
 658 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
 659 #else
 660 /* No preempt: go for improved straight-line efficiency */
 661 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
 662 #endif
 663
 664 /**
 665  * unmap_vmas - unmap a range of memory covered by a list of vma's
 666  * @tlbp: address of the caller's struct mmu_gather
 667  * @mm: the controlling mm_struct
 668  * @vma: the starting vma
 669  * @start_addr: virtual address at which to start unmapping
 670  * @end_addr: virtual address at which to end unmapping
 671  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 672  * @details: details of nonlinear truncation or shared cache invalidation
 673  *
 674  * Returns the end address of the unmapping (restart addr if interrupted).
 675  *
 676  * Unmap all pages in the vma list.  Called under page_table_lock.
 677  *
 678  * We aim to not hold page_table_lock for too long (for scheduling latency
 679  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 680  * return the ending mmu_gather to the caller.
 681  *
 682  * Only addresses between `start' and `end' will be unmapped.
 683  *
 684  * The VMA list must be sorted in ascending virtual address order.
 685  *
 686  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 687  * range after unmap_vmas() returns.  So the only responsibility here is to
 688  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 689  * drops the lock and schedules.
 690  */
 691 unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
 692                 struct vm_area_struct *vma, unsigned long start_addr,
 693                 unsigned long end_addr, unsigned long *nr_accounted,
 694                 struct zap_details *details)
 695 {
 696         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
 697         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 698         int tlb_start_valid = 0;
 699         unsigned long start = start_addr;
 700         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
 701         int fullmm = (*tlbp)->fullmm;
 702
 703         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 704                 unsigned long end;
 705
 706                 start = max(vma->vm_start, start_addr);
 707                 if (start >= vma->vm_end)
 708                         continue;
 709                 end = min(vma->vm_end, end_addr);
 710                 if (end <= vma->vm_start)
 711                         continue;
 712
 713                 if (vma->vm_flags & VM_ACCOUNT)
 714                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 715
 716                 while (start != end) {
 717                         unsigned long block;
 718
 719                         if (!tlb_start_valid) {
 720                                 tlb_start = start;
 721                                 tlb_start_valid = 1;
 722                         }
 723
 724                         if (is_vm_hugetlb_page(vma)) {
 725                                 block = end - start;
 726                                 unmap_hugepage_range(vma, start, end);
 727                         } else {
 728                                 block = min(zap_bytes, end - start);
 729                                 unmap_page_range(*tlbp, vma, start,
 730                                                 start + block, details);
 731                         }
 732
 733                         start += block;
 734                         zap_bytes -= block;
 735                         if ((long)zap_bytes > 0)
 736                                 continue;
 737
 738                         tlb_finish_mmu(*tlbp, tlb_start, start);
 739
 740                         if (need_resched() ||
 741                                 need_lockbreak(&mm->page_table_lock) ||
 742                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
 743                                 if (i_mmap_lock) {
 744                                         /* must reset count of rss freed */
 745                                         *tlbp = tlb_gather_mmu(mm, fullmm);
 746                                         goto out;
 747                                 }
 748                                 spin_unlock(&mm->page_table_lock);
 749                                 cond_resched();
 750                                 spin_lock(&mm->page_table_lock);
 751                         }
 752
 753                         *tlbp = tlb_gather_mmu(mm, fullmm);
 754                         tlb_start_valid = 0;
 755                         zap_bytes = ZAP_BLOCK_SIZE;
 756                 }
 757         }
 758 out:
 759         return start;   /* which is now the end (or restart) address */
 760 }
 761
 762 /**
 763  * zap_page_range - remove user pages in a given range
 764  * @vma: vm_area_struct holding the applicable pages
 765  * @address: starting address of pages to zap
 766  * @size: number of bytes to zap
 767  * @details: details of nonlinear truncation or shared cache invalidation
 768  */
 769 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
 770                 unsigned long size, struct zap_details *details)
 771 {
 772         struct mm_struct *mm = vma->vm_mm;
 773         struct mmu_gather *tlb;
 774         unsigned long end = address + size;
 775         unsigned long nr_accounted = 0;
 776
 777         if (is_vm_hugetlb_page(vma)) {
 778                 zap_hugepage_range(vma, address, size);
 779                 return end;
 780         }
 781
 782         lru_add_drain();
 783         spin_lock(&mm->page_table_lock);
 784         tlb = tlb_gather_mmu(mm, 0);
 785         end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
 786         tlb_finish_mmu(tlb, address, end);
 787         spin_unlock(&mm->page_table_lock);
 788         return end;
 789 }
 790
 791 /*
 792  * Do a quick page-table lookup for a single page.
 793  * mm->page_table_lock must be held.
 794  */
 795 static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
 796                         int read, int write, int accessed)
 797 {
 798         pgd_t *pgd;
 799         pud_t *pud;
 800         pmd_t *pmd;
 801         pte_t *ptep, pte;
 802         unsigned long pfn;
 803         struct page *page;
 804
 805         page = follow_huge_addr(mm, address, write);
 806         if (! IS_ERR(page))
 807                 return page;
 808
 809         pgd = pgd_offset(mm, address);
 810         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 811                 goto out;
 812
 813         pud = pud_offset(pgd, address);
 814         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
 815                 goto out;
 816
 817         pmd = pmd_offset(pud, address);
 818         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 819                 goto out;
 820         if (pmd_huge(*pmd))
 821                 return follow_huge_pmd(mm, address, pmd, write);
 822
 823         ptep = pte_offset_map(pmd, address);
 824         if (!ptep)
 825                 goto out;
 826
 827         pte = *ptep;
 828         pte_unmap(ptep);
 829         if (pte_present(pte)) {
 830                 if (write && !pte_write(pte))
 831                         goto out;
 832                 if (read && !pte_read(pte))
 833                         goto out;
 834                 pfn = pte_pfn(pte);
 835                 if (pfn_valid(pfn)) {
 836                         page = pfn_to_page(pfn);
 837                         if (accessed) {
 838                                 if (write && !pte_dirty(pte) &&!PageDirty(page))
 839                                         set_page_dirty(page);
 840                                 mark_page_accessed(page);
 841                         }
 842                         return page;
 843                 }
 844         }
 845
 846 out:
 847         return NULL;
 848 }
 849
 850 inline struct page *
 851 follow_page(struct mm_struct *mm, unsigned long address, int write)
 852 {
 853         return __follow_page(mm, address, 0, write, 1);
 854 }
 855
 856 /*
 857  * check_user_page_readable() can be called frm niterrupt context by oprofile,
 858  * so we need to avoid taking any non-irq-safe locks
 859  */
 860 int check_user_page_readable(struct mm_struct *mm, unsigned long address)
 861 {
 862         return __follow_page(mm, address, 1, 0, 0) != NULL;
 863 }
 864 EXPORT_SYMBOL(check_user_page_readable);
 865
 866 static inline int
 867 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
 868                          unsigned long address)
 869 {
 870         pgd_t *pgd;
 871         pud_t *pud;
 872         pmd_t *pmd;
 873
 874         /* Check if the vma is for an anonymous mapping. */
 875         if (vma->vm_ops && vma->vm_ops->nopage)
 876                 return 0;
 877
 878         /* Check if page directory entry exists. */
 879         pgd = pgd_offset(mm, address);
 880         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 881                 return 1;
 882
 883         pud = pud_offset(pgd, address);
 884         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
 885                 return 1;
 886
 887         /* Check if page middle directory entry exists. */
 888         pmd = pmd_offset(pud, address);
 889         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 890                 return 1;
 891
 892         /* There is a pte slot for 'address' in 'mm'. */
 893         return 0;
 894 }
 895
 896 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 897                 unsigned long start, int len, int write, int force,
 898                 struct page **pages, struct vm_area_struct **vmas)
 899 {
 900         int i;
 901         unsigned int flags;
 902
 903         /*
 904          * Require read or write permissions.
 905          * If 'force' is set, we only require the "MAY" flags.
 906          */
 907         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 908         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 909         i = 0;
 910
 911         do {
 912                 struct vm_area_struct * vma;
 913
 914                 vma = find_extend_vma(mm, start);
 915                 if (!vma && in_gate_area(tsk, start)) {
 916                         unsigned long pg = start & PAGE_MASK;
 917                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
 918                         pgd_t *pgd;
 919                         pud_t *pud;
 920                         pmd_t *pmd;
 921                         pte_t *pte;
 922                         if (write) /* user gate pages are read-only */
 923                                 return i ? : -EFAULT;
 924                         if (pg > TASK_SIZE)
 925                                 pgd = pgd_offset_k(pg);
 926                         else
 927                                 pgd = pgd_offset_gate(mm, pg);
 928                         BUG_ON(pgd_none(*pgd));
 929                         pud = pud_offset(pgd, pg);
 930                         BUG_ON(pud_none(*pud));
 931                         pmd = pmd_offset(pud, pg);
 932                         if (pmd_none(*pmd))
 933                                 return i ? : -EFAULT;
 934                         pte = pte_offset_map(pmd, pg);
 935                         if (pte_none(*pte)) {
 936                                 pte_unmap(pte);
 937                                 return i ? : -EFAULT;
 938                         }
 939                         if (pages) {
 940                                 pages[i] = pte_page(*pte);
 941                                 get_page(pages[i]);
 942                         }
 943                         pte_unmap(pte);
 944                         if (vmas)
 945                                 vmas[i] = gate_vma;
 946                         i++;
 947                         start += PAGE_SIZE;
 948                         len--;
 949                         continue;
 950                 }
 951
 952                 if (!vma || (vma->vm_flags & VM_IO)
 953                                 || !(flags & vma->vm_flags))
 954                         return i ? : -EFAULT;
 955
 956                 if (is_vm_hugetlb_page(vma)) {
 957                         i = follow_hugetlb_page(mm, vma, pages, vmas,
 958                                                 &start, &len, i);
 959                         continue;
 960                 }
 961                 spin_lock(&mm->page_table_lock);
 962                 do {
 963                         int write_access = write;
 964                         struct page *page;
 965
 966                         cond_resched_lock(&mm->page_table_lock);
 967                         while (!(page = follow_page(mm, start, write_access))) {
 968                                 int ret;
 969
 970                                 /*
 971                                  * Shortcut for anonymous pages. We don't want
 972                                  * to force the creation of pages tables for
 973                                  * insanely big anonymously mapped areas that
 974                                  * nobody touched so far. This is important
 975                                  * for doing a core dump for these mappings.
 976                                  */
 977                                 if (!write && untouched_anonymous_page(mm,vma,start)) {
 978                                         page = ZERO_PAGE(start);
 979                                         break;
 980                                 }
 981                                 spin_unlock(&mm->page_table_lock);
 982                                 ret = __handle_mm_fault(mm, vma, start, write_access);
 983
 984                                 /*
 985                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
 986                                  * broken COW when necessary, even if maybe_mkwrite
 987                                  * decided not to set pte_write. We can thus safely do
 988                                  * subsequent page lookups as if they were reads.
 989                                  */
 990                                 if (ret & VM_FAULT_WRITE)
 991                                         write_access = 0;
 992
 993                                 switch (ret & ~VM_FAULT_WRITE) {
 994                                 case VM_FAULT_MINOR:
 995                                         tsk->min_flt++;
 996                                         break;
 997                                 case VM_FAULT_MAJOR:
 998                                         tsk->maj_flt++;
 999                                         break;
1000                                 case VM_FAULT_SIGBUS:
1001                                         return i ? i : -EFAULT;
1002                                 case VM_FAULT_OOM:
1003                                         return i ? i : -ENOMEM;
1004                                 default:
1005                                         BUG();
1006                                 }
1007                                 spin_lock(&mm->page_table_lock);
1008                         }
1009                         if (pages) {
1010                                 pages[i] = page;
1011                                 flush_dcache_page(page);
1012                                 if (!PageReserved(page))
1013                                         page_cache_get(page);
1014                         }
1015                         if (vmas)
1016                                 vmas[i] = vma;
1017                         i++;
1018                         start += PAGE_SIZE;
1019                         len--;
1020                 } while (len && start < vma->vm_end);
1021                 spin_unlock(&mm->page_table_lock);
1022         } while (len);
1023         return i;
1024 }
1025 EXPORT_SYMBOL(get_user_pages);
1026
1027 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1028                         unsigned long addr, unsigned long end, pgprot_t prot)
1029 {
1030         pte_t *pte;
1031
1032         pte = pte_alloc_map(mm, pmd, addr);
1033         if (!pte)
1034                 return -ENOMEM;
1035         do {
1036                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(addr), prot));
1037                 BUG_ON(!pte_none(*pte));
1038                 set_pte_at(mm, addr, pte, zero_pte);
1039         } while (pte++, addr += PAGE_SIZE, addr != end);
1040         pte_unmap(pte - 1);
1041         return 0;
1042 }
1043
1044 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1045                         unsigned long addr, unsigned long end, pgprot_t prot)
1046 {
1047         pmd_t *pmd;
1048         unsigned long next;
1049
1050         pmd = pmd_alloc(mm, pud, addr);
1051         if (!pmd)
1052                 return -ENOMEM;
1053         do {
1054                 next = pmd_addr_end(addr, end);
1055                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1056                         return -ENOMEM;
1057         } while (pmd++, addr = next, addr != end);
1058         return 0;
1059 }
1060
1061 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1062                         unsigned long addr, unsigned long end, pgprot_t prot)
1063 {
1064         pud_t *pud;
1065         unsigned long next;
1066
1067         pud = pud_alloc(mm, pgd, addr);
1068         if (!pud)
1069                 return -ENOMEM;
1070         do {
1071                 next = pud_addr_end(addr, end);
1072                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1073                         return -ENOMEM;
1074         } while (pud++, addr = next, addr != end);
1075         return 0;
1076 }
1077
1078 int zeromap_page_range(struct vm_area_struct *vma,
1079                         unsigned long addr, unsigned long size, pgprot_t prot)
1080 {
1081         pgd_t *pgd;
1082         unsigned long next;
1083         unsigned long end = addr + size;
1084         struct mm_struct *mm = vma->vm_mm;
1085         int err;
1086
1087         BUG_ON(addr >= end);
1088         pgd = pgd_offset(mm, addr);
1089         flush_cache_range(vma, addr, end);
1090         spin_lock(&mm->page_table_lock);
1091         do {
1092                 next = pgd_addr_end(addr, end);
1093                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1094                 if (err)
1095                         break;
1096         } while (pgd++, addr = next, addr != end);
1097         spin_unlock(&mm->page_table_lock);
1098         return err;
1099 }
1100
1101 /*
1102  * maps a range of physical memory into the requested pages. the old
1103  * mappings are removed. any references to nonexistent pages results
1104  * in null mappings (currently treated as "copy-on-access")
1105  */
1106 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1107                         unsigned long addr, unsigned long end,
1108                         unsigned long pfn, pgprot_t prot)
1109 {
1110         pte_t *pte;
1111
1112         pte = pte_alloc_map(mm, pmd, addr);
1113         if (!pte)
1114                 return -ENOMEM;
1115         do {
1116                 BUG_ON(!pte_none(*pte));
1117                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1118                         set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1119                 pfn++;
1120         } while (pte++, addr += PAGE_SIZE, addr != end);
1121         pte_unmap(pte - 1);
1122         return 0;
1123 }
1124
1125 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1126                         unsigned long addr, unsigned long end,
1127                         unsigned long pfn, pgprot_t prot)
1128 {
1129         pmd_t *pmd;
1130         unsigned long next;
1131
1132         pfn -= addr >> PAGE_SHIFT;
1133         pmd = pmd_alloc(mm, pud, addr);
1134         if (!pmd)
1135                 return -ENOMEM;
1136         do {
1137                 next = pmd_addr_end(addr, end);
1138                 if (remap_pte_range(mm, pmd, addr, next,
1139                                 pfn + (addr >> PAGE_SHIFT), prot))
1140                         return -ENOMEM;
1141         } while (pmd++, addr = next, addr != end);
1142         return 0;
1143 }
1144
1145 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1146                         unsigned long addr, unsigned long end,
1147                         unsigned long pfn, pgprot_t prot)
1148 {
1149         pud_t *pud;
1150         unsigned long next;
1151
1152         pfn -= addr >> PAGE_SHIFT;
1153         pud = pud_alloc(mm, pgd, addr);
1154         if (!pud)
1155                 return -ENOMEM;
1156         do {
1157                 next = pud_addr_end(addr, end);
1158                 if (remap_pmd_range(mm, pud, addr, next,
1159                                 pfn + (addr >> PAGE_SHIFT), prot))
1160                         return -ENOMEM;
1161         } while (pud++, addr = next, addr != end);
1162         return 0;
1163 }
1164
1165 /*  Note: this is only safe if the mm semaphore is held when called. */
1166 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1167                     unsigned long pfn, unsigned long size, pgprot_t prot)
1168 {
1169         pgd_t *pgd;
1170         unsigned long next;
1171         unsigned long end = addr + PAGE_ALIGN(size);
1172         struct mm_struct *mm = vma->vm_mm;
1173         int err;
1174
1175         /*
1176          * Physically remapped pages are special. Tell the
1177          * rest of the world about it:
1178          *   VM_IO tells people not to look at these pages
1179          *      (accesses can have side effects).
1180          *   VM_RESERVED tells swapout not to try to touch
1181          *      this region.
1182          */
1183         vma->vm_flags |= VM_IO | VM_RESERVED;
1184
1185         BUG_ON(addr >= end);
1186         pfn -= addr >> PAGE_SHIFT;
1187         pgd = pgd_offset(mm, addr);
1188         flush_cache_range(vma, addr, end);
1189         spin_lock(&mm->page_table_lock);
1190         do {
1191                 next = pgd_addr_end(addr, end);
1192                 err = remap_pud_range(mm, pgd, addr, next,
1193                                 pfn + (addr >> PAGE_SHIFT), prot);
1194                 if (err)
1195                         break;
1196         } while (pgd++, addr = next, addr != end);
1197         spin_unlock(&mm->page_table_lock);
1198         return err;
1199 }
1200 EXPORT_SYMBOL(remap_pfn_range);
1201
1202 /*
1203  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1204  * servicing faults for write access.  In the normal case, do always want
1205  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1206  * that do not have writing enabled, when used by access_process_vm.
1207  */
1208 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1209 {
1210         if (likely(vma->vm_flags & VM_WRITE))
1211                 pte = pte_mkwrite(pte);
1212         return pte;
1213 }
1214
1215 /*
1216  * This routine handles present pages, when users try to write
1217  * to a shared page. It is done by copying the page to a new address
1218  * and decrementing the shared-page counter for the old page.
1219  *
1220  * Note that this routine assumes that the protection checks have been
1221  * done by the caller (the low-level page fault routine in most cases).
1222  * Thus we can safely just mark it writable once we've done any necessary
1223  * COW.
1224  *
1225  * We also mark the page dirty at this point even though the page will
1226  * change only once the write actually happens. This avoids a few races,
1227  * and potentially makes it more efficient.
1228  *
1229  * We hold the mm semaphore and the page_table_lock on entry and exit
1230  * with the page_table_lock released.
1231  */
1232 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1233                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1234                 pte_t orig_pte)
1235 {
1236         struct page *old_page, *new_page;
1237         unsigned long pfn = pte_pfn(orig_pte);
1238         pte_t entry;
1239         int ret = VM_FAULT_MINOR;
1240
1241         if (unlikely(!pfn_valid(pfn))) {
1242                 /*
1243                  * Page table corrupted: show pte and kill process.
1244                  */
1245                 pte_ERROR(orig_pte);
1246                 ret = VM_FAULT_OOM;
1247                 goto unlock;
1248         }
1249         old_page = pfn_to_page(pfn);
1250
1251         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1252                 int reuse = can_share_swap_page(old_page);
1253                 unlock_page(old_page);
1254                 if (reuse) {
1255                         flush_cache_page(vma, address, pfn);
1256                         entry = pte_mkyoung(orig_pte);
1257                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1258                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1259                         update_mmu_cache(vma, address, entry);
1260                         lazy_mmu_prot_update(entry);
1261                         ret |= VM_FAULT_WRITE;
1262                         goto unlock;
1263                 }
1264         }
1265
1266         /*
1267          * Ok, we need to copy. Oh, well..
1268          */
1269         if (!PageReserved(old_page))
1270                 page_cache_get(old_page);
1271         pte_unmap(page_table);
1272         spin_unlock(&mm->page_table_lock);
1273
1274         if (unlikely(anon_vma_prepare(vma)))
1275                 goto oom;
1276         if (old_page == ZERO_PAGE(address)) {
1277                 new_page = alloc_zeroed_user_highpage(vma, address);
1278                 if (!new_page)
1279                         goto oom;
1280         } else {
1281                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1282                 if (!new_page)
1283                         goto oom;
1284                 copy_user_highpage(new_page, old_page, address);
1285         }
1286
1287         /*
1288          * Re-check the pte - we dropped the lock
1289          */
1290         spin_lock(&mm->page_table_lock);
1291         page_table = pte_offset_map(pmd, address);
1292         if (likely(pte_same(*page_table, orig_pte))) {
1293                 if (PageAnon(old_page))
1294                         dec_mm_counter(mm, anon_rss);
1295                 if (PageReserved(old_page))
1296                         inc_mm_counter(mm, rss);
1297                 else
1298                         page_remove_rmap(old_page);
1299
1300                 flush_cache_page(vma, address, pfn);
1301                 entry = mk_pte(new_page, vma->vm_page_prot);
1302                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1303                 ptep_establish(vma, address, page_table, entry);
1304                 update_mmu_cache(vma, address, entry);
1305                 lazy_mmu_prot_update(entry);
1306
1307                 lru_cache_add_active(new_page);
1308                 page_add_anon_rmap(new_page, vma, address);
1309
1310                 /* Free the old page.. */
1311                 new_page = old_page;
1312                 ret |= VM_FAULT_WRITE;
1313         }
1314         page_cache_release(new_page);
1315         page_cache_release(old_page);
1316 unlock:
1317         pte_unmap(page_table);
1318         spin_unlock(&mm->page_table_lock);
1319         return ret;
1320 oom:
1321         page_cache_release(old_page);
1322         return VM_FAULT_OOM;
1323 }
1324
1325 /*
1326  * Helper functions for unmap_mapping_range().
1327  *
1328  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1329  *
1330  * We have to restart searching the prio_tree whenever we drop the lock,
1331  * since the iterator is only valid while the lock is held, and anyway
1332  * a later vma might be split and reinserted earlier while lock dropped.
1333  *
1334  * The list of nonlinear vmas could be handled more efficiently, using
1335  * a placeholder, but handle it in the same way until a need is shown.
1336  * It is important to search the prio_tree before nonlinear list: a vma
1337  * may become nonlinear and be shifted from prio_tree to nonlinear list
1338  * while the lock is dropped; but never shifted from list to prio_tree.
1339  *
1340  * In order to make forward progress despite restarting the search,
1341  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1342  * quickly skip it next time around.  Since the prio_tree search only
1343  * shows us those vmas affected by unmapping the range in question, we
1344  * can't efficiently keep all vmas in step with mapping->truncate_count:
1345  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1346  * mapping->truncate_count and vma->vm_truncate_count are protected by
1347  * i_mmap_lock.
1348  *
1349  * In order to make forward progress despite repeatedly restarting some
1350  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1351  * and restart from that address when we reach that vma again.  It might
1352  * have been split or merged, shrunk or extended, but never shifted: so
1353  * restart_addr remains valid so long as it remains in the vma's range.
1354  * unmap_mapping_range forces truncate_count to leap over page-aligned
1355  * values so we can save vma's restart_addr in its truncate_count field.
1356  */
1357 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1358
1359 static void reset_vma_truncate_counts(struct address_space *mapping)
1360 {
1361         struct vm_area_struct *vma;
1362         struct prio_tree_iter iter;
1363
1364         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1365                 vma->vm_truncate_count = 0;
1366         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1367                 vma->vm_truncate_count = 0;
1368 }
1369
1370 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1371                 unsigned long start_addr, unsigned long end_addr,
1372                 struct zap_details *details)
1373 {
1374         unsigned long restart_addr;
1375         int need_break;
1376
1377 again:
1378         restart_addr = vma->vm_truncate_count;
1379         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1380                 start_addr = restart_addr;
1381                 if (start_addr >= end_addr) {
1382                         /* Top of vma has been split off since last time */
1383                         vma->vm_truncate_count = details->truncate_count;
1384                         return 0;
1385                 }
1386         }
1387
1388         restart_addr = zap_page_range(vma, start_addr,
1389                                         end_addr - start_addr, details);
1390
1391         /*
1392          * We cannot rely on the break test in unmap_vmas:
1393          * on the one hand, we don't want to restart our loop
1394          * just because that broke out for the page_table_lock;
1395          * on the other hand, it does no test when vma is small.
1396          */
1397         need_break = need_resched() ||
1398                         need_lockbreak(details->i_mmap_lock);
1399
1400         if (restart_addr >= end_addr) {
1401                 /* We have now completed this vma: mark it so */
1402                 vma->vm_truncate_count = details->truncate_count;
1403                 if (!need_break)
1404                         return 0;
1405         } else {
1406                 /* Note restart_addr in vma's truncate_count field */
1407                 vma->vm_truncate_count = restart_addr;
1408                 if (!need_break)
1409                         goto again;
1410         }
1411
1412         spin_unlock(details->i_mmap_lock);
1413         cond_resched();
1414         spin_lock(details->i_mmap_lock);
1415         return -EINTR;
1416 }
1417
1418 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1419                                             struct zap_details *details)
1420 {
1421         struct vm_area_struct *vma;
1422         struct prio_tree_iter iter;
1423         pgoff_t vba, vea, zba, zea;
1424
1425 restart:
1426         vma_prio_tree_foreach(vma, &iter, root,
1427                         details->first_index, details->last_index) {
1428                 /* Skip quickly over those we have already dealt with */
1429                 if (vma->vm_truncate_count == details->truncate_count)
1430                         continue;
1431
1432                 vba = vma->vm_pgoff;
1433                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1434                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1435                 zba = details->first_index;
1436                 if (zba < vba)
1437                         zba = vba;
1438                 zea = details->last_index;
1439                 if (zea > vea)
1440                         zea = vea;
1441
1442                 if (unmap_mapping_range_vma(vma,
1443                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1444                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1445                                 details) < 0)
1446                         goto restart;
1447         }
1448 }
1449
1450 static inline void unmap_mapping_range_list(struct list_head *head,
1451                                             struct zap_details *details)
1452 {
1453         struct vm_area_struct *vma;
1454
1455         /*
1456          * In nonlinear VMAs there is no correspondence between virtual address
1457          * offset and file offset.  So we must perform an exhaustive search
1458          * across *all* the pages in each nonlinear VMA, not just the pages
1459          * whose virtual address lies outside the file truncation point.
1460          */
1461 restart:
1462         list_for_each_entry(vma, head, shared.vm_set.list) {
1463                 /* Skip quickly over those we have already dealt with */
1464                 if (vma->vm_truncate_count == details->truncate_count)
1465                         continue;
1466                 details->nonlinear_vma = vma;
1467                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1468                                         vma->vm_end, details) < 0)
1469                         goto restart;
1470         }
1471 }
1472
1473 /**
1474  * unmap_mapping_range - unmap the portion of all mmaps
1475  * in the specified address_space corresponding to the specified
1476  * page range in the underlying file.
1477  * @mapping: the address space containing mmaps to be unmapped.
1478  * @holebegin: byte in first page to unmap, relative to the start of
1479  * the underlying file.  This will be rounded down to a PAGE_SIZE
1480  * boundary.  Note that this is different from vmtruncate(), which
1481  * must keep the partial page.  In contrast, we must get rid of
1482  * partial pages.
1483  * @holelen: size of prospective hole in bytes.  This will be rounded
1484  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1485  * end of the file.
1486  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1487  * but 0 when invalidating pagecache, don't throw away private data.
1488  */
1489 void unmap_mapping_range(struct address_space *mapping,
1490                 loff_t const holebegin, loff_t const holelen, int even_cows)
1491 {
1492         struct zap_details details;
1493         pgoff_t hba = holebegin >> PAGE_SHIFT;
1494         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1495
1496         /* Check for overflow. */
1497         if (sizeof(holelen) > sizeof(hlen)) {
1498                 long long holeend =
1499                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1500                 if (holeend & ~(long long)ULONG_MAX)
1501                         hlen = ULONG_MAX - hba + 1;
1502         }
1503
1504         details.check_mapping = even_cows? NULL: mapping;
1505         details.nonlinear_vma = NULL;
1506         details.first_index = hba;
1507         details.last_index = hba + hlen - 1;
1508         if (details.last_index < details.first_index)
1509                 details.last_index = ULONG_MAX;
1510         details.i_mmap_lock = &mapping->i_mmap_lock;
1511
1512         spin_lock(&mapping->i_mmap_lock);
1513
1514         /* serialize i_size write against truncate_count write */
1515         smp_wmb();
1516         /* Protect against page faults, and endless unmapping loops */
1517         mapping->truncate_count++;
1518         /*
1519          * For archs where spin_lock has inclusive semantics like ia64
1520          * this smp_mb() will prevent to read pagetable contents
1521          * before the truncate_count increment is visible to
1522          * other cpus.
1523          */
1524         smp_mb();
1525         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1526                 if (mapping->truncate_count == 0)
1527                         reset_vma_truncate_counts(mapping);
1528                 mapping->truncate_count++;
1529         }
1530         details.truncate_count = mapping->truncate_count;
1531
1532         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1533                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1534         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1535                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1536         spin_unlock(&mapping->i_mmap_lock);
1537 }
1538 EXPORT_SYMBOL(unmap_mapping_range);
1539
1540 /*
1541  * Handle all mappings that got truncated by a "truncate()"
1542  * system call.
1543  *
1544  * NOTE! We have to be ready to update the memory sharing
1545  * between the file and the memory map for a potential last
1546  * incomplete page.  Ugly, but necessary.
1547  */
1548 int vmtruncate(struct inode * inode, loff_t offset)
1549 {
1550         struct address_space *mapping = inode->i_mapping;
1551         unsigned long limit;
1552
1553         if (inode->i_size < offset)
1554                 goto do_expand;
1555         /*
1556          * truncation of in-use swapfiles is disallowed - it would cause
1557          * subsequent swapout to scribble on the now-freed blocks.
1558          */
1559         if (IS_SWAPFILE(inode))
1560                 goto out_busy;
1561         i_size_write(inode, offset);
1562         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1563         truncate_inode_pages(mapping, offset);
1564         goto out_truncate;
1565
1566 do_expand:
1567         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1568         if (limit != RLIM_INFINITY && offset > limit)
1569                 goto out_sig;
1570         if (offset > inode->i_sb->s_maxbytes)
1571                 goto out_big;
1572         i_size_write(inode, offset);
1573
1574 out_truncate:
1575         if (inode->i_op && inode->i_op->truncate)
1576                 inode->i_op->truncate(inode);
1577         return 0;
1578 out_sig:
1579         send_sig(SIGXFSZ, current, 0);
1580 out_big:
1581         return -EFBIG;
1582 out_busy:
1583         return -ETXTBSY;
1584 }
1585
1586 EXPORT_SYMBOL(vmtruncate);
1587
1588 /*
1589  * Primitive swap readahead code. We simply read an aligned block of
1590  * (1 << page_cluster) entries in the swap area. This method is chosen
1591  * because it doesn't cost us any seek time.  We also make sure to queue
1592  * the 'original' request together with the readahead ones...
1593  *
1594  * This has been extended to use the NUMA policies from the mm triggering
1595  * the readahead.
1596  *
1597  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1598  */
1599 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1600 {
1601 #ifdef CONFIG_NUMA
1602         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1603 #endif
1604         int i, num;
1605         struct page *new_page;
1606         unsigned long offset;
1607
1608         /*
1609          * Get the number of handles we should do readahead io to.
1610          */
1611         num = valid_swaphandles(entry, &offset);
1612         for (i = 0; i < num; offset++, i++) {
1613                 /* Ok, do the async read-ahead now */
1614                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1615                                                            offset), vma, addr);
1616                 if (!new_page)
1617                         break;
1618                 page_cache_release(new_page);
1619 #ifdef CONFIG_NUMA
1620                 /*
1621                  * Find the next applicable VMA for the NUMA policy.
1622                  */
1623                 addr += PAGE_SIZE;
1624                 if (addr == 0)
1625                         vma = NULL;
1626                 if (vma) {
1627                         if (addr >= vma->vm_end) {
1628                                 vma = next_vma;
1629                                 next_vma = vma ? vma->vm_next : NULL;
1630                         }
1631                         if (vma && addr < vma->vm_start)
1632                                 vma = NULL;
1633                 } else {
1634                         if (next_vma && addr >= next_vma->vm_start) {
1635                                 vma = next_vma;
1636                                 next_vma = vma->vm_next;
1637                         }
1638                 }
1639 #endif
1640         }
1641         lru_add_drain();        /* Push any new pages onto the LRU now */
1642 }
1643
1644 /*
1645  * We hold the mm semaphore and the page_table_lock on entry and
1646  * should release the pagetable lock on exit..
1647  */
1648 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1649                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1650                 int write_access, pte_t orig_pte)
1651 {
1652         struct page *page;
1653         swp_entry_t entry;
1654         pte_t pte;
1655         int ret = VM_FAULT_MINOR;
1656
1657         pte_unmap(page_table);
1658         spin_unlock(&mm->page_table_lock);
1659
1660         entry = pte_to_swp_entry(orig_pte);
1661         page = lookup_swap_cache(entry);
1662         if (!page) {
1663                 swapin_readahead(entry, address, vma);
1664                 page = read_swap_cache_async(entry, vma, address);
1665                 if (!page) {
1666                         /*
1667                          * Back out if somebody else faulted in this pte while
1668                          * we released the page table lock.
1669                          */
1670                         spin_lock(&mm->page_table_lock);
1671                         page_table = pte_offset_map(pmd, address);
1672                         if (likely(pte_same(*page_table, orig_pte)))
1673                                 ret = VM_FAULT_OOM;
1674                         goto unlock;
1675                 }
1676
1677                 /* Had to read the page from swap area: Major fault */
1678                 ret = VM_FAULT_MAJOR;
1679                 inc_page_state(pgmajfault);
1680                 grab_swap_token();
1681         }
1682
1683         mark_page_accessed(page);
1684         lock_page(page);
1685
1686         /*
1687          * Back out if somebody else faulted in this pte while we
1688          * released the page table lock.
1689          */
1690         spin_lock(&mm->page_table_lock);
1691         page_table = pte_offset_map(pmd, address);
1692         if (unlikely(!pte_same(*page_table, orig_pte))) {
1693                 ret = VM_FAULT_MINOR;
1694                 goto out_nomap;
1695         }
1696
1697         if (unlikely(!PageUptodate(page))) {
1698                 ret = VM_FAULT_SIGBUS;
1699                 goto out_nomap;
1700         }
1701
1702         /* The page isn't present yet, go ahead with the fault. */
1703
1704         inc_mm_counter(mm, rss);
1705         pte = mk_pte(page, vma->vm_page_prot);
1706         if (write_access && can_share_swap_page(page)) {
1707                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1708                 write_access = 0;
1709         }
1710
1711         flush_icache_page(vma, page);
1712         set_pte_at(mm, address, page_table, pte);
1713         page_add_anon_rmap(page, vma, address);
1714
1715         swap_free(entry);
1716         if (vm_swap_full())
1717                 remove_exclusive_swap_page(page);
1718         unlock_page(page);
1719
1720         if (write_access) {
1721                 if (do_wp_page(mm, vma, address,
1722                                 page_table, pmd, pte) == VM_FAULT_OOM)
1723                         ret = VM_FAULT_OOM;
1724                 goto out;
1725         }
1726
1727         /* No need to invalidate - it was non-present before */
1728         update_mmu_cache(vma, address, pte);
1729         lazy_mmu_prot_update(pte);
1730 unlock:
1731         pte_unmap(page_table);
1732         spin_unlock(&mm->page_table_lock);
1733 out:
1734         return ret;
1735 out_nomap:
1736         pte_unmap(page_table);
1737         spin_unlock(&mm->page_table_lock);
1738         unlock_page(page);
1739         page_cache_release(page);
1740         return ret;
1741 }
1742
1743 /*
1744  * We are called with the MM semaphore and page_table_lock
1745  * spinlock held to protect against concurrent faults in
1746  * multithreaded programs.
1747  */
1748 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1749                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1750                 int write_access)
1751 {
1752         pte_t entry;
1753
1754         /* Mapping of ZERO_PAGE - vm_page_prot is readonly */
1755         entry = mk_pte(ZERO_PAGE(addr), vma->vm_page_prot);
1756
1757         if (write_access) {
1758                 struct page *page;
1759
1760                 /* Allocate our own private page. */
1761                 pte_unmap(page_table);
1762                 spin_unlock(&mm->page_table_lock);
1763
1764                 if (unlikely(anon_vma_prepare(vma)))
1765                         goto oom;
1766                 page = alloc_zeroed_user_highpage(vma, address);
1767                 if (!page)
1768                         goto oom;
1769
1770                 spin_lock(&mm->page_table_lock);
1771                 page_table = pte_offset_map(pmd, address);
1772
1773                 if (!pte_none(*page_table)) {
1774                         page_cache_release(page);
1775                         goto unlock;
1776                 }
1777                 inc_mm_counter(mm, rss);
1778                 entry = mk_pte(page, vma->vm_page_prot);
1779                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1780                 lru_cache_add_active(page);
1781                 SetPageReferenced(page);
1782                 page_add_anon_rmap(page, vma, address);
1783         }
1784
1785         set_pte_at(mm, address, page_table, entry);
1786
1787         /* No need to invalidate - it was non-present before */
1788         update_mmu_cache(vma, address, entry);
1789         lazy_mmu_prot_update(entry);
1790 unlock:
1791         pte_unmap(page_table);
1792         spin_unlock(&mm->page_table_lock);
1793         return VM_FAULT_MINOR;
1794 oom:
1795         return VM_FAULT_OOM;
1796 }
1797
1798 /*
1799  * do_no_page() tries to create a new page mapping. It aggressively
1800  * tries to share with existing pages, but makes a separate copy if
1801  * the "write_access" parameter is true in order to avoid the next
1802  * page fault.
1803  *
1804  * As this is called only for pages that do not currently exist, we
1805  * do not need to flush old virtual caches or the TLB.
1806  *
1807  * This is called with the MM semaphore held and the page table
1808  * spinlock held. Exit with the spinlock released.
1809  */
1810 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1811                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1812                 int write_access)
1813 {
1814         struct page *new_page;
1815         struct address_space *mapping = NULL;
1816         pte_t entry;
1817         unsigned int sequence = 0;
1818         int ret = VM_FAULT_MINOR;
1819         int anon = 0;
1820
1821         pte_unmap(page_table);
1822         spin_unlock(&mm->page_table_lock);
1823
1824         if (vma->vm_file) {
1825                 mapping = vma->vm_file->f_mapping;
1826                 sequence = mapping->truncate_count;
1827                 smp_rmb(); /* serializes i_size against truncate_count */
1828         }
1829 retry:
1830         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1831         /*
1832          * No smp_rmb is needed here as long as there's a full
1833          * spin_lock/unlock sequence inside the ->nopage callback
1834          * (for the pagecache lookup) that acts as an implicit
1835          * smp_mb() and prevents the i_size read to happen
1836          * after the next truncate_count read.
1837          */
1838
1839         /* no page was available -- either SIGBUS or OOM */
1840         if (new_page == NOPAGE_SIGBUS)
1841                 return VM_FAULT_SIGBUS;
1842         if (new_page == NOPAGE_OOM)
1843                 return VM_FAULT_OOM;
1844
1845         /*
1846          * Should we do an early C-O-W break?
1847          */
1848         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1849                 struct page *page;
1850
1851                 if (unlikely(anon_vma_prepare(vma)))
1852                         goto oom;
1853                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1854                 if (!page)
1855                         goto oom;
1856                 copy_user_highpage(page, new_page, address);
1857                 page_cache_release(new_page);
1858                 new_page = page;
1859                 anon = 1;
1860         }
1861
1862         spin_lock(&mm->page_table_lock);
1863         /*
1864          * For a file-backed vma, someone could have truncated or otherwise
1865          * invalidated this page.  If unmap_mapping_range got called,
1866          * retry getting the page.
1867          */
1868         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1869                 spin_unlock(&mm->page_table_lock);
1870                 page_cache_release(new_page);
1871                 cond_resched();
1872                 sequence = mapping->truncate_count;
1873                 smp_rmb();
1874                 goto retry;
1875         }
1876         page_table = pte_offset_map(pmd, address);
1877
1878         /*
1879          * This silly early PAGE_DIRTY setting removes a race
1880          * due to the bad i386 page protection. But it's valid
1881          * for other architectures too.
1882          *
1883          * Note that if write_access is true, we either now have
1884          * an exclusive copy of the page, or this is a shared mapping,
1885          * so we can make it writable and dirty to avoid having to
1886          * handle that later.
1887          */
1888         /* Only go through if we didn't race with anybody else... */
1889         if (pte_none(*page_table)) {
1890                 if (!PageReserved(new_page))
1891                         inc_mm_counter(mm, rss);
1892
1893                 flush_icache_page(vma, new_page);
1894                 entry = mk_pte(new_page, vma->vm_page_prot);
1895                 if (write_access)
1896                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1897                 set_pte_at(mm, address, page_table, entry);
1898                 if (anon) {
1899                         lru_cache_add_active(new_page);
1900                         page_add_anon_rmap(new_page, vma, address);
1901                 } else
1902                         page_add_file_rmap(new_page);
1903         } else {
1904                 /* One of our sibling threads was faster, back out. */
1905                 page_cache_release(new_page);
1906                 goto unlock;
1907         }
1908
1909         /* no need to invalidate: a not-present page shouldn't be cached */
1910         update_mmu_cache(vma, address, entry);
1911         lazy_mmu_prot_update(entry);
1912 unlock:
1913         pte_unmap(page_table);
1914         spin_unlock(&mm->page_table_lock);
1915         return ret;
1916 oom:
1917         page_cache_release(new_page);
1918         return VM_FAULT_OOM;
1919 }
1920
1921 /*
1922  * Fault of a previously existing named mapping. Repopulate the pte
1923  * from the encoded file_pte if possible. This enables swappable
1924  * nonlinear vmas.
1925  */
1926 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
1927                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1928                 int write_access, pte_t orig_pte)
1929 {
1930         pgoff_t pgoff;
1931         int err;
1932
1933         pte_unmap(page_table);
1934         spin_unlock(&mm->page_table_lock);
1935
1936         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
1937                 /*
1938                  * Page table corrupted: show pte and kill process.
1939                  */
1940                 pte_ERROR(orig_pte);
1941                 return VM_FAULT_OOM;
1942         }
1943         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1944
1945         pgoff = pte_to_pgoff(orig_pte);
1946         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
1947                                         vma->vm_page_prot, pgoff, 0);
1948         if (err == -ENOMEM)
1949                 return VM_FAULT_OOM;
1950         if (err)
1951                 return VM_FAULT_SIGBUS;
1952         return VM_FAULT_MAJOR;
1953 }
1954
1955 /*
1956  * These routines also need to handle stuff like marking pages dirty
1957  * and/or accessed for architectures that don't do it in hardware (most
1958  * RISC architectures).  The early dirtying is also good on the i386.
1959  *
1960  * There is also a hook called "update_mmu_cache()" that architectures
1961  * with external mmu caches can use to update those (ie the Sparc or
1962  * PowerPC hashed page tables that act as extended TLBs).
1963  *
1964  * Note the "page_table_lock". It is to protect against kswapd removing
1965  * pages from under us. Note that kswapd only ever _removes_ pages, never
1966  * adds them. As such, once we have noticed that the page is not present,
1967  * we can drop the lock early.
1968  *
1969  * The adding of pages is protected by the MM semaphore (which we hold),
1970  * so we don't need to worry about a page being suddenly been added into
1971  * our VM.
1972  *
1973  * We enter with the pagetable spinlock held, we are supposed to
1974  * release it when done.
1975  */
1976 static inline int handle_pte_fault(struct mm_struct *mm,
1977                 struct vm_area_struct *vma, unsigned long address,
1978                 pte_t *pte, pmd_t *pmd, int write_access)
1979 {
1980         pte_t entry;
1981
1982         entry = *pte;
1983         if (!pte_present(entry)) {
1984                 if (pte_none(entry)) {
1985                         if (!vma->vm_ops || !vma->vm_ops->nopage)
1986                                 return do_anonymous_page(mm, vma, address,
1987                                         pte, pmd, write_access);
1988                         return do_no_page(mm, vma, address,
1989                                         pte, pmd, write_access);
1990                 }
1991                 if (pte_file(entry))
1992                         return do_file_page(mm, vma, address,
1993                                         pte, pmd, write_access, entry);
1994                 return do_swap_page(mm, vma, address,
1995                                         pte, pmd, write_access, entry);
1996         }
1997
1998         if (write_access) {
1999                 if (!pte_write(entry))
2000                         return do_wp_page(mm, vma, address, pte, pmd, entry);
2001                 entry = pte_mkdirty(entry);
2002         }
2003         entry = pte_mkyoung(entry);
2004         ptep_set_access_flags(vma, address, pte, entry, write_access);
2005         update_mmu_cache(vma, address, entry);
2006         lazy_mmu_prot_update(entry);
2007         pte_unmap(pte);
2008         spin_unlock(&mm->page_table_lock);
2009         return VM_FAULT_MINOR;
2010 }
2011
2012 /*
2013  * By the time we get here, we already hold the mm semaphore
2014  */
2015 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2016                 unsigned long address, int write_access)
2017 {
2018         pgd_t *pgd;
2019         pud_t *pud;
2020         pmd_t *pmd;
2021         pte_t *pte;
2022
2023         __set_current_state(TASK_RUNNING);
2024
2025         inc_page_state(pgfault);
2026
2027         if (unlikely(is_vm_hugetlb_page(vma)))
2028                 return hugetlb_fault(mm, vma, address, write_access);
2029
2030         /*
2031          * We need the page table lock to synchronize with kswapd
2032          * and the SMP-safe atomic PTE updates.
2033          */
2034         pgd = pgd_offset(mm, address);
2035         spin_lock(&mm->page_table_lock);
2036
2037         pud = pud_alloc(mm, pgd, address);
2038         if (!pud)
2039                 goto oom;
2040
2041         pmd = pmd_alloc(mm, pud, address);
2042         if (!pmd)
2043                 goto oom;
2044
2045         pte = pte_alloc_map(mm, pmd, address);
2046         if (!pte)
2047                 goto oom;
2048
2049         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2050
2051  oom:
2052         spin_unlock(&mm->page_table_lock);
2053         return VM_FAULT_OOM;
2054 }
2055
2056 #ifndef __PAGETABLE_PUD_FOLDED
2057 /*
2058  * Allocate page upper directory.
2059  *
2060  * We've already handled the fast-path in-line, and we own the
2061  * page table lock.
2062  */
2063 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2064 {
2065         pud_t *new;
2066
2067         spin_unlock(&mm->page_table_lock);
2068         new = pud_alloc_one(mm, address);
2069         spin_lock(&mm->page_table_lock);
2070         if (!new)
2071                 return NULL;
2072
2073         /*
2074          * Because we dropped the lock, we should re-check the
2075          * entry, as somebody else could have populated it..
2076          */
2077         if (pgd_present(*pgd)) {
2078                 pud_free(new);
2079                 goto out;
2080         }
2081         pgd_populate(mm, pgd, new);
2082  out:
2083         return pud_offset(pgd, address);
2084 }
2085 #endif /* __PAGETABLE_PUD_FOLDED */
2086
2087 #ifndef __PAGETABLE_PMD_FOLDED
2088 /*
2089  * Allocate page middle directory.
2090  *
2091  * We've already handled the fast-path in-line, and we own the
2092  * page table lock.
2093  */
2094 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2095 {
2096         pmd_t *new;
2097
2098         spin_unlock(&mm->page_table_lock);
2099         new = pmd_alloc_one(mm, address);
2100         spin_lock(&mm->page_table_lock);
2101         if (!new)
2102                 return NULL;
2103
2104         /*
2105          * Because we dropped the lock, we should re-check the
2106          * entry, as somebody else could have populated it..
2107          */
2108 #ifndef __ARCH_HAS_4LEVEL_HACK
2109         if (pud_present(*pud)) {
2110                 pmd_free(new);
2111                 goto out;
2112         }
2113         pud_populate(mm, pud, new);
2114 #else
2115         if (pgd_present(*pud)) {
2116                 pmd_free(new);
2117                 goto out;
2118         }
2119         pgd_populate(mm, pud, new);
2120 #endif /* __ARCH_HAS_4LEVEL_HACK */
2121
2122  out:
2123         return pmd_offset(pud, address);
2124 }
2125 #endif /* __PAGETABLE_PMD_FOLDED */
2126
2127 int make_pages_present(unsigned long addr, unsigned long end)
2128 {
2129         int ret, len, write;
2130         struct vm_area_struct * vma;
2131
2132         vma = find_vma(current->mm, addr);
2133         if (!vma)
2134                 return -1;
2135         write = (vma->vm_flags & VM_WRITE) != 0;
2136         if (addr >= end)
2137                 BUG();
2138         if (end > vma->vm_end)
2139                 BUG();
2140         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2141         ret = get_user_pages(current, current->mm, addr,
2142                         len, write, 0, NULL, NULL);
2143         if (ret < 0)
2144                 return ret;
2145         return ret == len ? 0 : -1;
2146 }
2147
2148 /*
2149  * Map a vmalloc()-space virtual address to the physical page.
2150  */
2151 struct page * vmalloc_to_page(void * vmalloc_addr)
2152 {
2153         unsigned long addr = (unsigned long) vmalloc_addr;
2154         struct page *page = NULL;
2155         pgd_t *pgd = pgd_offset_k(addr);
2156         pud_t *pud;
2157         pmd_t *pmd;
2158         pte_t *ptep, pte;
2159
2160         if (!pgd_none(*pgd)) {
2161                 pud = pud_offset(pgd, addr);
2162                 if (!pud_none(*pud)) {
2163                         pmd = pmd_offset(pud, addr);
2164                         if (!pmd_none(*pmd)) {
2165                                 ptep = pte_offset_map(pmd, addr);
2166                                 pte = *ptep;
2167                                 if (pte_present(pte))
2168                                         page = pte_page(pte);
2169                                 pte_unmap(ptep);
2170                         }
2171                 }
2172         }
2173         return page;
2174 }
2175
2176 EXPORT_SYMBOL(vmalloc_to_page);
2177
2178 /*
2179  * Map a vmalloc()-space virtual address to the physical page frame number.
2180  */
2181 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2182 {
2183         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2184 }
2185
2186 EXPORT_SYMBOL(vmalloc_to_pfn);
2187
2188 /*
2189  * update_mem_hiwater
2190  *      - update per process rss and vm high water data
2191  */
2192 void update_mem_hiwater(struct task_struct *tsk)
2193 {
2194         if (tsk->mm) {
2195                 unsigned long rss = get_mm_counter(tsk->mm, rss);
2196
2197                 if (tsk->mm->hiwater_rss < rss)
2198                         tsk->mm->hiwater_rss = rss;
2199                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2200                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2201         }
2202 }
2203
2204 #if !defined(__HAVE_ARCH_GATE_AREA)
2205
2206 #if defined(AT_SYSINFO_EHDR)
2207 static struct vm_area_struct gate_vma;
2208
2209 static int __init gate_vma_init(void)
2210 {
2211         gate_vma.vm_mm = NULL;
2212         gate_vma.vm_start = FIXADDR_USER_START;
2213         gate_vma.vm_end = FIXADDR_USER_END;
2214         gate_vma.vm_page_prot = PAGE_READONLY;
2215         gate_vma.vm_flags = 0;
2216         return 0;
2217 }
2218 __initcall(gate_vma_init);
2219 #endif
2220
2221 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2222 {
2223 #ifdef AT_SYSINFO_EHDR
2224         return &gate_vma;
2225 #else
2226         return NULL;
2227 #endif
2228 }
2229
2230 int in_gate_area_no_task(unsigned long addr)
2231 {
2232 #ifdef AT_SYSINFO_EHDR
2233         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2234                 return 1;
2235 #endif
2236         return 0;
2237 }
2238
2239 #endif  /* __HAVE_ARCH_GATE_AREA */