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/acct.h>
  50 #include <linux/module.h>
  51 #include <linux/init.h>
  52
  53 #include <asm/pgalloc.h>
  54 #include <asm/uaccess.h>
  55 #include <asm/tlb.h>
  56 #include <asm/tlbflush.h>
  57 #include <asm/pgtable.h>
  58
  59 #include <linux/swapops.h>
  60 #include <linux/elf.h>
  61
  62 #ifndef CONFIG_DISCONTIGMEM
  63 /* use the per-pgdat data instead for discontigmem - mbligh */
  64 unsigned long max_mapnr;
  65 struct page *mem_map;
  66
  67 EXPORT_SYMBOL(max_mapnr);
  68 EXPORT_SYMBOL(mem_map);
  69 #endif
  70
  71 unsigned long num_physpages;
  72 /*
  73  * A number of key systems in x86 including ioremap() rely on the assumption
  74  * that high_memory defines the upper bound on direct map memory, then end
  75  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  76  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  77  * and ZONE_HIGHMEM.
  78  */
  79 void * high_memory;
  80 unsigned long vmalloc_earlyreserve;
  81
  82 EXPORT_SYMBOL(num_physpages);
  83 EXPORT_SYMBOL(high_memory);
  84 EXPORT_SYMBOL(vmalloc_earlyreserve);
  85
  86 /*
  87  * Note: this doesn't free the actual pages themselves. That
  88  * has been handled earlier when unmapping all the memory regions.
  89  */
  90 static inline void clear_pmd_range(struct mmu_gather *tlb, pmd_t *pmd, unsigned long start, unsigned long end)
  91 {
  92         struct page *page;
  93
  94         if (pmd_none(*pmd))
  95                 return;
  96         if (unlikely(pmd_bad(*pmd))) {
  97                 pmd_ERROR(*pmd);
  98                 pmd_clear(pmd);
  99                 return;
 100         }
 101         if (!((start | end) & ~PMD_MASK)) {
 102                 /* Only clear full, aligned ranges */
 103                 page = pmd_page(*pmd);
 104                 pmd_clear(pmd);
 105                 dec_page_state(nr_page_table_pages);
 106                 tlb->mm->nr_ptes--;
 107                 pte_free_tlb(tlb, page);
 108         }
 109 }
 110
 111 static inline void clear_pud_range(struct mmu_gather *tlb, pud_t *pud, unsigned long start, unsigned long end)
 112 {
 113         unsigned long addr = start, next;
 114         pmd_t *pmd, *__pmd;
 115
 116         if (pud_none(*pud))
 117                 return;
 118         if (unlikely(pud_bad(*pud))) {
 119                 pud_ERROR(*pud);
 120                 pud_clear(pud);
 121                 return;
 122         }
 123
 124         pmd = __pmd = pmd_offset(pud, start);
 125         do {
 126                 next = (addr + PMD_SIZE) & PMD_MASK;
 127                 if (next > end || next <= addr)
 128                         next = end;
 129
 130                 clear_pmd_range(tlb, pmd, addr, next);
 131                 pmd++;
 132                 addr = next;
 133         } while (addr && (addr < end));
 134
 135         if (!((start | end) & ~PUD_MASK)) {
 136                 /* Only clear full, aligned ranges */
 137                 pud_clear(pud);
 138                 pmd_free_tlb(tlb, __pmd);
 139         }
 140 }
 141
 142
 143 static inline void clear_pgd_range(struct mmu_gather *tlb, pgd_t *pgd, unsigned long start, unsigned long end)
 144 {
 145         unsigned long addr = start, next;
 146         pud_t *pud, *__pud;
 147
 148         if (pgd_none(*pgd))
 149                 return;
 150         if (unlikely(pgd_bad(*pgd))) {
 151                 pgd_ERROR(*pgd);
 152                 pgd_clear(pgd);
 153                 return;
 154         }
 155
 156         pud = __pud = pud_offset(pgd, start);
 157         do {
 158                 next = (addr + PUD_SIZE) & PUD_MASK;
 159                 if (next > end || next <= addr)
 160                         next = end;
 161
 162                 clear_pud_range(tlb, pud, addr, next);
 163                 pud++;
 164                 addr = next;
 165         } while (addr && (addr < end));
 166
 167         if (!((start | end) & ~PGDIR_MASK)) {
 168                 /* Only clear full, aligned ranges */
 169                 pgd_clear(pgd);
 170                 pud_free_tlb(tlb, __pud);
 171         }
 172 }
 173
 174 /*
 175  * This function clears user-level page tables of a process.
 176  *
 177  * Must be called with pagetable lock held.
 178  */
 179 void clear_page_range(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 180 {
 181         unsigned long addr = start, next;
 182         pgd_t * pgd = pgd_offset(tlb->mm, start);
 183         unsigned long i;
 184
 185         for (i = pgd_index(start); i <= pgd_index(end-1); i++) {
 186                 next = (addr + PGDIR_SIZE) & PGDIR_MASK;
 187                 if (next > end || next <= addr)
 188                         next = end;
 189
 190                 clear_pgd_range(tlb, pgd, addr, next);
 191                 pgd++;
 192                 addr = next;
 193         }
 194 }
 195
 196 pte_t fastcall * pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 197 {
 198         if (!pmd_present(*pmd)) {
 199                 struct page *new;
 200
 201                 spin_unlock(&mm->page_table_lock);
 202                 new = pte_alloc_one(mm, address);
 203                 spin_lock(&mm->page_table_lock);
 204                 if (!new)
 205                         return NULL;
 206                 /*
 207                  * Because we dropped the lock, we should re-check the
 208                  * entry, as somebody else could have populated it..
 209                  */
 210                 if (pmd_present(*pmd)) {
 211                         pte_free(new);
 212                         goto out;
 213                 }
 214                 mm->nr_ptes++;
 215                 inc_page_state(nr_page_table_pages);
 216                 pmd_populate(mm, pmd, new);
 217         }
 218 out:
 219         return pte_offset_map(pmd, address);
 220 }
 221
 222 pte_t fastcall * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 223 {
 224         if (!pmd_present(*pmd)) {
 225                 pte_t *new;
 226
 227                 spin_unlock(&mm->page_table_lock);
 228                 new = pte_alloc_one_kernel(mm, address);
 229                 spin_lock(&mm->page_table_lock);
 230                 if (!new)
 231                         return NULL;
 232
 233                 /*
 234                  * Because we dropped the lock, we should re-check the
 235                  * entry, as somebody else could have populated it..
 236                  */
 237                 if (pmd_present(*pmd)) {
 238                         pte_free_kernel(new);
 239                         goto out;
 240                 }
 241                 pmd_populate_kernel(mm, pmd, new);
 242         }
 243 out:
 244         return pte_offset_kernel(pmd, address);
 245 }
 246
 247 /*
 248  * copy one vm_area from one task to the other. Assumes the page tables
 249  * already present in the new task to be cleared in the whole range
 250  * covered by this vma.
 251  *
 252  * dst->page_table_lock is held on entry and exit,
 253  * but may be dropped within p[mg]d_alloc() and pte_alloc_map().
 254  */
 255
 256 static inline void
 257 copy_swap_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, pte_t pte)
 258 {
 259         if (pte_file(pte))
 260                 return;
 261         swap_duplicate(pte_to_swp_entry(pte));
 262         if (list_empty(&dst_mm->mmlist)) {
 263                 spin_lock(&mmlist_lock);
 264                 list_add(&dst_mm->mmlist, &src_mm->mmlist);
 265                 spin_unlock(&mmlist_lock);
 266         }
 267 }
 268
 269 static inline void
 270 copy_one_pte(struct mm_struct *dst_mm,  struct mm_struct *src_mm,
 271                 pte_t *dst_pte, pte_t *src_pte, unsigned long vm_flags,
 272                 unsigned long addr)
 273 {
 274         pte_t pte = *src_pte;
 275         struct page *page;
 276         unsigned long pfn;
 277
 278         /* pte contains position in swap, so copy. */
 279         if (!pte_present(pte)) {
 280                 copy_swap_pte(dst_mm, src_mm, pte);
 281                 set_pte(dst_pte, pte);
 282                 return;
 283         }
 284         pfn = pte_pfn(pte);
 285         /* the pte points outside of valid memory, the
 286          * mapping is assumed to be good, meaningful
 287          * and not mapped via rmap - duplicate the
 288          * mapping as is.
 289          */
 290         page = NULL;
 291         if (pfn_valid(pfn))
 292                 page = pfn_to_page(pfn);
 293
 294         if (!page || PageReserved(page)) {
 295                 set_pte(dst_pte, pte);
 296                 return;
 297         }
 298
 299         /*
 300          * If it's a COW mapping, write protect it both
 301          * in the parent and the child
 302          */
 303         if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
 304                 ptep_set_wrprotect(src_pte);
 305                 pte = *src_pte;
 306         }
 307
 308         /*
 309          * If it's a shared mapping, mark it clean in
 310          * the child
 311          */
 312         if (vm_flags & VM_SHARED)
 313                 pte = pte_mkclean(pte);
 314         pte = pte_mkold(pte);
 315         get_page(page);
 316         vx_rsspages_inc(dst_mm);
 317         if (PageAnon(page))
 318                 vx_anonpages_inc(dst_mm);
 319         set_pte(dst_pte, pte);
 320         page_dup_rmap(page);
 321 }
 322
 323 static int copy_pte_range(struct mm_struct *dst_mm,  struct mm_struct *src_mm,
 324                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 325                 unsigned long addr, unsigned long end)
 326 {
 327         pte_t *src_pte, *dst_pte;
 328         pte_t *s, *d;
 329         unsigned long vm_flags = vma->vm_flags;
 330
 331         d = dst_pte = pte_alloc_map(dst_mm, dst_pmd, addr);
 332         if (!dst_pte)
 333                 return -ENOMEM;
 334
 335         spin_lock(&src_mm->page_table_lock);
 336         s = src_pte = pte_offset_map_nested(src_pmd, addr);
 337         for (; addr < end; addr += PAGE_SIZE, s++, d++) {
 338                 if (pte_none(*s))
 339                         continue;
 340                 copy_one_pte(dst_mm, src_mm, d, s, vm_flags, addr);
 341         }
 342         pte_unmap_nested(src_pte);
 343         pte_unmap(dst_pte);
 344         spin_unlock(&src_mm->page_table_lock);
 345         cond_resched_lock(&dst_mm->page_table_lock);
 346         return 0;
 347 }
 348
 349 static int copy_pmd_range(struct mm_struct *dst_mm,  struct mm_struct *src_mm,
 350                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 351                 unsigned long addr, unsigned long end)
 352 {
 353         pmd_t *src_pmd, *dst_pmd;
 354         int err = 0;
 355         unsigned long next;
 356
 357         src_pmd = pmd_offset(src_pud, addr);
 358         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 359         if (!dst_pmd)
 360                 return -ENOMEM;
 361
 362         for (; addr < end; addr = next, src_pmd++, dst_pmd++) {
 363                 next = (addr + PMD_SIZE) & PMD_MASK;
 364                 if (next > end || next <= addr)
 365                         next = end;
 366                 if (pmd_none(*src_pmd))
 367                         continue;
 368                 if (pmd_bad(*src_pmd)) {
 369                         pmd_ERROR(*src_pmd);
 370                         pmd_clear(src_pmd);
 371                         continue;
 372                 }
 373                 err = copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
 374                                                         vma, addr, next);
 375                 if (err)
 376                         break;
 377         }
 378         return err;
 379 }
 380
 381 static int copy_pud_range(struct mm_struct *dst_mm,  struct mm_struct *src_mm,
 382                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
 383                 unsigned long addr, unsigned long end)
 384 {
 385         pud_t *src_pud, *dst_pud;
 386         int err = 0;
 387         unsigned long next;
 388
 389         src_pud = pud_offset(src_pgd, addr);
 390         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
 391         if (!dst_pud)
 392                 return -ENOMEM;
 393
 394         for (; addr < end; addr = next, src_pud++, dst_pud++) {
 395                 next = (addr + PUD_SIZE) & PUD_MASK;
 396                 if (next > end || next <= addr)
 397                         next = end;
 398                 if (pud_none(*src_pud))
 399                         continue;
 400                 if (pud_bad(*src_pud)) {
 401                         pud_ERROR(*src_pud);
 402                         pud_clear(src_pud);
 403                         continue;
 404                 }
 405                 err = copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
 406                                                         vma, addr, next);
 407                 if (err)
 408                         break;
 409         }
 410         return err;
 411 }
 412
 413 int copy_page_range(struct mm_struct *dst, struct mm_struct *src,
 414                 struct vm_area_struct *vma)
 415 {
 416         pgd_t *src_pgd, *dst_pgd;
 417         unsigned long addr, start, end, next;
 418         int err = 0;
 419
 420         if (is_vm_hugetlb_page(vma))
 421                 return copy_hugetlb_page_range(dst, src, vma);
 422
 423         start = vma->vm_start;
 424         src_pgd = pgd_offset(src, start);
 425         dst_pgd = pgd_offset(dst, start);
 426
 427         end = vma->vm_end;
 428         addr = start;
 429         while (addr && (addr < end-1)) {
 430                 next = (addr + PGDIR_SIZE) & PGDIR_MASK;
 431                 if (next > end || next <= addr)
 432                         next = end;
 433                 if (pgd_none(*src_pgd))
 434                         goto next_pgd;
 435                 if (pgd_bad(*src_pgd)) {
 436                         pgd_ERROR(*src_pgd);
 437                         pgd_clear(src_pgd);
 438                         goto next_pgd;
 439                 }
 440                 err = copy_pud_range(dst, src, dst_pgd, src_pgd,
 441                                                         vma, addr, next);
 442                 if (err)
 443                         break;
 444
 445 next_pgd:
 446                 src_pgd++;
 447                 dst_pgd++;
 448                 addr = next;
 449         }
 450
 451         return err;
 452 }
 453
 454 static void zap_pte_range(struct mmu_gather *tlb,
 455                 pmd_t *pmd, unsigned long address,
 456                 unsigned long size, struct zap_details *details)
 457 {
 458         unsigned long offset;
 459         pte_t *ptep;
 460
 461         if (pmd_none(*pmd))
 462                 return;
 463         if (unlikely(pmd_bad(*pmd))) {
 464                 pmd_ERROR(*pmd);
 465                 pmd_clear(pmd);
 466                 return;
 467         }
 468         ptep = pte_offset_map(pmd, address);
 469         offset = address & ~PMD_MASK;
 470         if (offset + size > PMD_SIZE)
 471                 size = PMD_SIZE - offset;
 472         size &= PAGE_MASK;
 473         if (details && !details->check_mapping && !details->nonlinear_vma)
 474                 details = NULL;
 475         for (offset=0; offset < size; ptep++, offset += PAGE_SIZE) {
 476                 pte_t pte = *ptep;
 477                 if (pte_none(pte))
 478                         continue;
 479                 if (pte_present(pte)) {
 480                         struct page *page = NULL;
 481                         unsigned long pfn = pte_pfn(pte);
 482                         if (pfn_valid(pfn)) {
 483                                 page = pfn_to_page(pfn);
 484                                 if (PageReserved(page))
 485                                         page = NULL;
 486                         }
 487                         if (unlikely(details) && page) {
 488                                 /*
 489                                  * unmap_shared_mapping_pages() wants to
 490                                  * invalidate cache without truncating:
 491                                  * unmap shared but keep private pages.
 492                                  */
 493                                 if (details->check_mapping &&
 494                                     details->check_mapping != page->mapping)
 495                                         continue;
 496                                 /*
 497                                  * Each page->index must be checked when
 498                                  * invalidating or truncating nonlinear.
 499                                  */
 500                                 if (details->nonlinear_vma &&
 501                                     (page->index < details->first_index ||
 502                                      page->index > details->last_index))
 503                                         continue;
 504                         }
 505                         pte = ptep_get_and_clear(ptep);
 506                         tlb_remove_tlb_entry(tlb, ptep, address+offset);
 507                         if (unlikely(!page))
 508                                 continue;
 509                         if (unlikely(details) && details->nonlinear_vma
 510                             && linear_page_index(details->nonlinear_vma,
 511                                         address+offset) != page->index)
 512                                 set_pte(ptep, pgoff_to_pte(page->index));
 513                         if (pte_dirty(pte))
 514                                 set_page_dirty(page);
 515                         if (PageAnon(page))
 516                                 vx_anonpages_dec(tlb->mm);
 517                         else if (pte_young(pte))
 518                                 mark_page_accessed(page);
 519                         tlb->freed++;
 520                         page_remove_rmap(page);
 521                         tlb_remove_page(tlb, page);
 522                         continue;
 523                 }
 524                 /*
 525                  * If details->check_mapping, we leave swap entries;
 526                  * if details->nonlinear_vma, we leave file entries.
 527                  */
 528                 if (unlikely(details))
 529                         continue;
 530                 if (!pte_file(pte))
 531                         free_swap_and_cache(pte_to_swp_entry(pte));
 532                 pte_clear(ptep);
 533         }
 534         pte_unmap(ptep-1);
 535 }
 536
 537 static void zap_pmd_range(struct mmu_gather *tlb,
 538                 pud_t *pud, unsigned long address,
 539                 unsigned long size, struct zap_details *details)
 540 {
 541         pmd_t * pmd;
 542         unsigned long end;
 543
 544         if (pud_none(*pud))
 545                 return;
 546         if (unlikely(pud_bad(*pud))) {
 547                 pud_ERROR(*pud);
 548                 pud_clear(pud);
 549                 return;
 550         }
 551         pmd = pmd_offset(pud, address);
 552         end = address + size;
 553         if (end > ((address + PUD_SIZE) & PUD_MASK))
 554                 end = ((address + PUD_SIZE) & PUD_MASK);
 555         do {
 556                 zap_pte_range(tlb, pmd, address, end - address, details);
 557                 address = (address + PMD_SIZE) & PMD_MASK;
 558                 pmd++;
 559         } while (address && (address < end));
 560 }
 561
 562 static void zap_pud_range(struct mmu_gather *tlb,
 563                 pgd_t * pgd, unsigned long address,
 564                 unsigned long end, struct zap_details *details)
 565 {
 566         pud_t * pud;
 567
 568         if (pgd_none(*pgd))
 569                 return;
 570         if (unlikely(pgd_bad(*pgd))) {
 571                 pgd_ERROR(*pgd);
 572                 pgd_clear(pgd);
 573                 return;
 574         }
 575         pud = pud_offset(pgd, address);
 576         do {
 577                 zap_pmd_range(tlb, pud, address, end - address, details);
 578                 address = (address + PUD_SIZE) & PUD_MASK;
 579                 pud++;
 580         } while (address && (address < end));
 581 }
 582
 583 static void unmap_page_range(struct mmu_gather *tlb,
 584                 struct vm_area_struct *vma, unsigned long address,
 585                 unsigned long end, struct zap_details *details)
 586 {
 587         unsigned long next;
 588         pgd_t *pgd;
 589         int i;
 590
 591         BUG_ON(address >= end);
 592         pgd = pgd_offset(vma->vm_mm, address);
 593         tlb_start_vma(tlb, vma);
 594         for (i = pgd_index(address); i <= pgd_index(end-1); i++) {
 595                 next = (address + PGDIR_SIZE) & PGDIR_MASK;
 596                 if (next <= address || next > end)
 597                         next = end;
 598                 zap_pud_range(tlb, pgd, address, next, details);
 599                 address = next;
 600                 pgd++;
 601         }
 602         tlb_end_vma(tlb, vma);
 603 }
 604
 605 #ifdef CONFIG_PREEMPT
 606 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
 607 #else
 608 /* No preempt: go for improved straight-line efficiency */
 609 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
 610 #endif
 611
 612 /**
 613  * unmap_vmas - unmap a range of memory covered by a list of vma's
 614  * @tlbp: address of the caller's struct mmu_gather
 615  * @mm: the controlling mm_struct
 616  * @vma: the starting vma
 617  * @start_addr: virtual address at which to start unmapping
 618  * @end_addr: virtual address at which to end unmapping
 619  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 620  * @details: details of nonlinear truncation or shared cache invalidation
 621  *
 622  * Returns the number of vma's which were covered by the unmapping.
 623  *
 624  * Unmap all pages in the vma list.  Called under page_table_lock.
 625  *
 626  * We aim to not hold page_table_lock for too long (for scheduling latency
 627  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 628  * return the ending mmu_gather to the caller.
 629  *
 630  * Only addresses between `start' and `end' will be unmapped.
 631  *
 632  * The VMA list must be sorted in ascending virtual address order.
 633  *
 634  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 635  * range after unmap_vmas() returns.  So the only responsibility here is to
 636  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 637  * drops the lock and schedules.
 638  */
 639 int unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
 640                 struct vm_area_struct *vma, unsigned long start_addr,
 641                 unsigned long end_addr, unsigned long *nr_accounted,
 642                 struct zap_details *details)
 643 {
 644         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
 645         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 646         int tlb_start_valid = 0;
 647         int ret = 0;
 648         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
 649         int fullmm = tlb_is_full_mm(*tlbp);
 650
 651         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 652                 unsigned long start;
 653                 unsigned long end;
 654
 655                 start = max(vma->vm_start, start_addr);
 656                 if (start >= vma->vm_end)
 657                         continue;
 658                 end = min(vma->vm_end, end_addr);
 659                 if (end <= vma->vm_start)
 660                         continue;
 661
 662                 if (vma->vm_flags & VM_ACCOUNT)
 663                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 664
 665                 ret++;
 666                 while (start != end) {
 667                         unsigned long block;
 668
 669                         if (!tlb_start_valid) {
 670                                 tlb_start = start;
 671                                 tlb_start_valid = 1;
 672                         }
 673
 674                         if (is_vm_hugetlb_page(vma)) {
 675                                 block = end - start;
 676                                 unmap_hugepage_range(vma, start, end);
 677                         } else {
 678                                 block = min(zap_bytes, end - start);
 679                                 unmap_page_range(*tlbp, vma, start,
 680                                                 start + block, details);
 681                         }
 682
 683                         start += block;
 684                         zap_bytes -= block;
 685                         if ((long)zap_bytes > 0)
 686                                 continue;
 687
 688                         tlb_finish_mmu(*tlbp, tlb_start, start);
 689
 690                         if (need_resched() ||
 691                                 need_lockbreak(&mm->page_table_lock) ||
 692                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
 693                                 if (i_mmap_lock) {
 694                                         /* must reset count of rss freed */
 695                                         *tlbp = tlb_gather_mmu(mm, fullmm);
 696                                         details->break_addr = start;
 697                                         goto out;
 698                                 }
 699                                 spin_unlock(&mm->page_table_lock);
 700                                 cond_resched();
 701                                 spin_lock(&mm->page_table_lock);
 702                         }
 703
 704                         *tlbp = tlb_gather_mmu(mm, fullmm);
 705                         tlb_start_valid = 0;
 706                         zap_bytes = ZAP_BLOCK_SIZE;
 707                 }
 708         }
 709 out:
 710         return ret;
 711 }
 712
 713 /**
 714  * zap_page_range - remove user pages in a given range
 715  * @vma: vm_area_struct holding the applicable pages
 716  * @address: starting address of pages to zap
 717  * @size: number of bytes to zap
 718  * @details: details of nonlinear truncation or shared cache invalidation
 719  */
 720 void zap_page_range(struct vm_area_struct *vma, unsigned long address,
 721                 unsigned long size, struct zap_details *details)
 722 {
 723         struct mm_struct *mm = vma->vm_mm;
 724         struct mmu_gather *tlb;
 725         unsigned long end = address + size;
 726         unsigned long nr_accounted = 0;
 727
 728         if (is_vm_hugetlb_page(vma)) {
 729                 zap_hugepage_range(vma, address, size);
 730                 return;
 731         }
 732
 733         lru_add_drain();
 734         spin_lock(&mm->page_table_lock);
 735         tlb = tlb_gather_mmu(mm, 0);
 736         unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
 737         tlb_finish_mmu(tlb, address, end);
 738         acct_update_integrals();
 739         spin_unlock(&mm->page_table_lock);
 740 }
 741
 742 /*
 743  * Do a quick page-table lookup for a single page.
 744  * mm->page_table_lock must be held.
 745  */
 746 static struct page *
 747 __follow_page(struct mm_struct *mm, unsigned long address, int read, int write)
 748 {
 749         pgd_t *pgd;
 750         pud_t *pud;
 751         pmd_t *pmd;
 752         pte_t *ptep, pte;
 753         unsigned long pfn;
 754         struct page *page;
 755
 756         page = follow_huge_addr(mm, address, write);
 757         if (! IS_ERR(page))
 758                 return page;
 759
 760         pgd = pgd_offset(mm, address);
 761         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 762                 goto out;
 763
 764         pud = pud_offset(pgd, address);
 765         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
 766                 goto out;
 767
 768         pmd = pmd_offset(pud, address);
 769         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 770                 goto out;
 771         if (pmd_huge(*pmd))
 772                 return follow_huge_pmd(mm, address, pmd, write);
 773
 774         ptep = pte_offset_map(pmd, address);
 775         if (!ptep)
 776                 goto out;
 777
 778         pte = *ptep;
 779         pte_unmap(ptep);
 780         if (pte_present(pte)) {
 781                 if (write && !pte_write(pte))
 782                         goto out;
 783                 if (read && !pte_read(pte))
 784                         goto out;
 785                 pfn = pte_pfn(pte);
 786                 if (pfn_valid(pfn)) {
 787                         page = pfn_to_page(pfn);
 788                         if (write && !pte_dirty(pte) && !PageDirty(page))
 789                                 set_page_dirty(page);
 790                         mark_page_accessed(page);
 791                         return page;
 792                 }
 793         }
 794
 795 out:
 796         return NULL;
 797 }
 798
 799 struct page *
 800 follow_page(struct mm_struct *mm, unsigned long address, int write)
 801 {
 802         return __follow_page(mm, address, /*read*/0, write);
 803 }
 804
 805 int
 806 check_user_page_readable(struct mm_struct *mm, unsigned long address)
 807 {
 808         return __follow_page(mm, address, /*read*/1, /*write*/0) != NULL;
 809 }
 810
 811 EXPORT_SYMBOL(check_user_page_readable);
 812
 813 /*
 814  * Given a physical address, is there a useful struct page pointing to
 815  * it?  This may become more complex in the future if we start dealing
 816  * with IO-aperture pages for direct-IO.
 817  */
 818
 819 static inline struct page *get_page_map(struct page *page)
 820 {
 821         if (!pfn_valid(page_to_pfn(page)))
 822                 return NULL;
 823         return page;
 824 }
 825
 826
 827 static inline int
 828 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
 829                          unsigned long address)
 830 {
 831         pgd_t *pgd;
 832         pud_t *pud;
 833         pmd_t *pmd;
 834
 835         /* Check if the vma is for an anonymous mapping. */
 836         if (vma->vm_ops && vma->vm_ops->nopage)
 837                 return 0;
 838
 839         /* Check if page directory entry exists. */
 840         pgd = pgd_offset(mm, address);
 841         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 842                 return 1;
 843
 844         pud = pud_offset(pgd, address);
 845         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
 846                 return 1;
 847
 848         /* Check if page middle directory entry exists. */
 849         pmd = pmd_offset(pud, address);
 850         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 851                 return 1;
 852
 853         /* There is a pte slot for 'address' in 'mm'. */
 854         return 0;
 855 }
 856
 857
 858 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 859                 unsigned long start, int len, int write, int force,
 860                 struct page **pages, struct vm_area_struct **vmas)
 861 {
 862         int i;
 863         unsigned int flags;
 864
 865         /*
 866          * Require read or write permissions.
 867          * If 'force' is set, we only require the "MAY" flags.
 868          */
 869         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 870         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 871         i = 0;
 872
 873         do {
 874                 struct vm_area_struct * vma;
 875
 876                 vma = find_extend_vma(mm, start);
 877                 if (!vma && in_gate_area(tsk, start)) {
 878                         unsigned long pg = start & PAGE_MASK;
 879                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
 880                         pgd_t *pgd;
 881                         pud_t *pud;
 882                         pmd_t *pmd;
 883                         pte_t *pte;
 884                         if (write) /* user gate pages are read-only */
 885                                 return i ? : -EFAULT;
 886                         if (pg > TASK_SIZE)
 887                                 pgd = pgd_offset_k(pg);
 888                         else
 889                                 pgd = pgd_offset_gate(mm, pg);
 890                         BUG_ON(pgd_none(*pgd));
 891                         pud = pud_offset(pgd, pg);
 892                         BUG_ON(pud_none(*pud));
 893                         pmd = pmd_offset(pud, pg);
 894                         BUG_ON(pmd_none(*pmd));
 895                         pte = pte_offset_map(pmd, pg);
 896                         BUG_ON(pte_none(*pte));
 897                         if (pages) {
 898                                 pages[i] = pte_page(*pte);
 899                                 get_page(pages[i]);
 900                         }
 901                         pte_unmap(pte);
 902                         if (vmas)
 903                                 vmas[i] = gate_vma;
 904                         i++;
 905                         start += PAGE_SIZE;
 906                         len--;
 907                         continue;
 908                 }
 909
 910                 if (!vma || (vma->vm_flags & VM_IO)
 911                                 || !(flags & vma->vm_flags))
 912                         return i ? : -EFAULT;
 913
 914                 if (is_vm_hugetlb_page(vma)) {
 915                         i = follow_hugetlb_page(mm, vma, pages, vmas,
 916                                                 &start, &len, i);
 917                         continue;
 918                 }
 919                 spin_lock(&mm->page_table_lock);
 920                 do {
 921                         struct page *map;
 922                         int lookup_write = write;
 923
 924                         cond_resched_lock(&mm->page_table_lock);
 925                         while (!(map = follow_page(mm, start, lookup_write))) {
 926                                 /*
 927                                  * Shortcut for anonymous pages. We don't want
 928                                  * to force the creation of pages tables for
 929                                  * insanly big anonymously mapped areas that
 930                                  * nobody touched so far. This is important
 931                                  * for doing a core dump for these mappings.
 932                                  */
 933                                 if (!lookup_write &&
 934                                     untouched_anonymous_page(mm,vma,start)) {
 935                                         map = ZERO_PAGE(start);
 936                                         break;
 937                                 }
 938                                 spin_unlock(&mm->page_table_lock);
 939                                 switch (handle_mm_fault(mm,vma,start,write)) {
 940                                 case VM_FAULT_MINOR:
 941                                         tsk->min_flt++;
 942                                         break;
 943                                 case VM_FAULT_MAJOR:
 944                                         tsk->maj_flt++;
 945                                         break;
 946                                 case VM_FAULT_SIGBUS:
 947                                         return i ? i : -EFAULT;
 948                                 case VM_FAULT_OOM:
 949                                         return i ? i : -ENOMEM;
 950                                 default:
 951                                         BUG();
 952                                 }
 953                                 /*
 954                                  * Now that we have performed a write fault
 955                                  * and surely no longer have a shared page we
 956                                  * shouldn't write, we shouldn't ignore an
 957                                  * unwritable page in the page table if
 958                                  * we are forcing write access.
 959                                  */
 960                                 lookup_write = write && !force;
 961                                 spin_lock(&mm->page_table_lock);
 962                         }
 963                         if (pages) {
 964                                 pages[i] = get_page_map(map);
 965                                 if (!pages[i]) {
 966                                         spin_unlock(&mm->page_table_lock);
 967                                         while (i--)
 968                                                 page_cache_release(pages[i]);
 969                                         i = -EFAULT;
 970                                         goto out;
 971                                 }
 972                                 flush_dcache_page(pages[i]);
 973                                 if (!PageReserved(pages[i]))
 974                                         page_cache_get(pages[i]);
 975                         }
 976                         if (vmas)
 977                                 vmas[i] = vma;
 978                         i++;
 979                         start += PAGE_SIZE;
 980                         len--;
 981                 } while(len && start < vma->vm_end);
 982                 spin_unlock(&mm->page_table_lock);
 983         } while(len);
 984 out:
 985         return i;
 986 }
 987
 988 EXPORT_SYMBOL(get_user_pages);
 989
 990 static void zeromap_pte_range(pte_t * pte, unsigned long address,
 991                                      unsigned long size, pgprot_t prot)
 992 {
 993         unsigned long end;
 994
 995         address &= ~PMD_MASK;
 996         end = address + size;
 997         if (end > PMD_SIZE)
 998                 end = PMD_SIZE;
 999         do {
1000                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
1001                 BUG_ON(!pte_none(*pte));
1002                 set_pte(pte, zero_pte);
1003                 address += PAGE_SIZE;
1004                 pte++;
1005         } while (address && (address < end));
1006 }
1007
1008 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd,
1009                 unsigned long address, unsigned long size, pgprot_t prot)
1010 {
1011         unsigned long base, end;
1012
1013         base = address & PUD_MASK;
1014         address &= ~PUD_MASK;
1015         end = address + size;
1016         if (end > PUD_SIZE)
1017                 end = PUD_SIZE;
1018         do {
1019                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
1020                 if (!pte)
1021                         return -ENOMEM;
1022                 zeromap_pte_range(pte, base + address, end - address, prot);
1023                 pte_unmap(pte);
1024                 address = (address + PMD_SIZE) & PMD_MASK;
1025                 pmd++;
1026         } while (address && (address < end));
1027         return 0;
1028 }
1029
1030 static inline int zeromap_pud_range(struct mm_struct *mm, pud_t * pud,
1031                                     unsigned long address,
1032                                     unsigned long size, pgprot_t prot)
1033 {
1034         unsigned long base, end;
1035         int error = 0;
1036
1037         base = address & PGDIR_MASK;
1038         address &= ~PGDIR_MASK;
1039         end = address + size;
1040         if (end > PGDIR_SIZE)
1041                 end = PGDIR_SIZE;
1042         do {
1043                 pmd_t * pmd = pmd_alloc(mm, pud, base + address);
1044                 error = -ENOMEM;
1045                 if (!pmd)
1046                         break;
1047                 error = zeromap_pmd_range(mm, pmd, base + address,
1048                                           end - address, prot);
1049                 if (error)
1050                         break;
1051                 address = (address + PUD_SIZE) & PUD_MASK;
1052                 pud++;
1053         } while (address && (address < end));
1054         return 0;
1055 }
1056
1057 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address,
1058                                         unsigned long size, pgprot_t prot)
1059 {
1060         int i;
1061         int error = 0;
1062         pgd_t * pgd;
1063         unsigned long beg = address;
1064         unsigned long end = address + size;
1065         unsigned long next;
1066         struct mm_struct *mm = vma->vm_mm;
1067
1068         pgd = pgd_offset(mm, address);
1069         flush_cache_range(vma, beg, end);
1070         BUG_ON(address >= end);
1071         BUG_ON(end > vma->vm_end);
1072
1073         spin_lock(&mm->page_table_lock);
1074         for (i = pgd_index(address); i <= pgd_index(end-1); i++) {
1075                 pud_t *pud = pud_alloc(mm, pgd, address);
1076                 error = -ENOMEM;
1077                 if (!pud)
1078                         break;
1079                 next = (address + PGDIR_SIZE) & PGDIR_MASK;
1080                 if (next <= beg || next > end)
1081                         next = end;
1082                 error = zeromap_pud_range(mm, pud, address,
1083                                                 next - address, prot);
1084                 if (error)
1085                         break;
1086                 address = next;
1087                 pgd++;
1088         }
1089         /*
1090          * Why flush? zeromap_pte_range has a BUG_ON for !pte_none()
1091          */
1092         flush_tlb_range(vma, beg, end);
1093         spin_unlock(&mm->page_table_lock);
1094         return error;
1095 }
1096
1097 /*
1098  * maps a range of physical memory into the requested pages. the old
1099  * mappings are removed. any references to nonexistent pages results
1100  * in null mappings (currently treated as "copy-on-access")
1101  */
1102 static inline void
1103 remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
1104                 unsigned long pfn, pgprot_t prot)
1105 {
1106         unsigned long end;
1107
1108         address &= ~PMD_MASK;
1109         end = address + size;
1110         if (end > PMD_SIZE)
1111                 end = PMD_SIZE;
1112         do {
1113                 BUG_ON(!pte_none(*pte));
1114                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1115                         set_pte(pte, pfn_pte(pfn, prot));
1116                 address += PAGE_SIZE;
1117                 pfn++;
1118                 pte++;
1119         } while (address && (address < end));
1120 }
1121
1122 static inline int
1123 remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
1124                 unsigned long size, unsigned long pfn, pgprot_t prot)
1125 {
1126         unsigned long base, end;
1127
1128         base = address & PUD_MASK;
1129         address &= ~PUD_MASK;
1130         end = address + size;
1131         if (end > PUD_SIZE)
1132                 end = PUD_SIZE;
1133         pfn -= (address >> PAGE_SHIFT);
1134         do {
1135                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
1136                 if (!pte)
1137                         return -ENOMEM;
1138                 remap_pte_range(pte, base + address, end - address,
1139                                 (address >> PAGE_SHIFT) + pfn, prot);
1140                 pte_unmap(pte);
1141                 address = (address + PMD_SIZE) & PMD_MASK;
1142                 pmd++;
1143         } while (address && (address < end));
1144         return 0;
1145 }
1146
1147 static inline int remap_pud_range(struct mm_struct *mm, pud_t * pud,
1148                                   unsigned long address, unsigned long size,
1149                                   unsigned long pfn, pgprot_t prot)
1150 {
1151         unsigned long base, end;
1152         int error;
1153
1154         base = address & PGDIR_MASK;
1155         address &= ~PGDIR_MASK;
1156         end = address + size;
1157         if (end > PGDIR_SIZE)
1158                 end = PGDIR_SIZE;
1159         pfn -= address >> PAGE_SHIFT;
1160         do {
1161                 pmd_t *pmd = pmd_alloc(mm, pud, base+address);
1162                 error = -ENOMEM;
1163                 if (!pmd)
1164                         break;
1165                 error = remap_pmd_range(mm, pmd, base + address, end - address,
1166                                 (address >> PAGE_SHIFT) + pfn, prot);
1167                 if (error)
1168                         break;
1169                 address = (address + PUD_SIZE) & PUD_MASK;
1170                 pud++;
1171         } while (address && (address < end));
1172         return error;
1173 }
1174
1175 /*  Note: this is only safe if the mm semaphore is held when called. */
1176 int remap_pfn_range(struct vm_area_struct *vma, unsigned long from,
1177                     unsigned long pfn, unsigned long size, pgprot_t prot)
1178 {
1179         int error = 0;
1180         pgd_t *pgd;
1181         unsigned long beg = from;
1182         unsigned long end = from + size;
1183         unsigned long next;
1184         struct mm_struct *mm = vma->vm_mm;
1185         int i;
1186
1187         pfn -= from >> PAGE_SHIFT;
1188         pgd = pgd_offset(mm, from);
1189         flush_cache_range(vma, beg, end);
1190         BUG_ON(from >= end);
1191
1192         /*
1193          * Physically remapped pages are special. Tell the
1194          * rest of the world about it:
1195          *   VM_IO tells people not to look at these pages
1196          *      (accesses can have side effects).
1197          *   VM_RESERVED tells swapout not to try to touch
1198          *      this region.
1199          */
1200         vma->vm_flags |= VM_IO | VM_RESERVED;
1201
1202         spin_lock(&mm->page_table_lock);
1203         for (i = pgd_index(beg); i <= pgd_index(end-1); i++) {
1204                 pud_t *pud = pud_alloc(mm, pgd, from);
1205                 error = -ENOMEM;
1206                 if (!pud)
1207                         break;
1208                 next = (from + PGDIR_SIZE) & PGDIR_MASK;
1209                 if (next > end || next <= from)
1210                         next = end;
1211                 error = remap_pud_range(mm, pud, from, end - from,
1212                                         pfn + (from >> PAGE_SHIFT), prot);
1213                 if (error)
1214                         break;
1215                 from = next;
1216                 pgd++;
1217         }
1218         /*
1219          * Why flush? remap_pte_range has a BUG_ON for !pte_none()
1220          */
1221         flush_tlb_range(vma, beg, end);
1222         spin_unlock(&mm->page_table_lock);
1223
1224         return error;
1225 }
1226
1227 EXPORT_SYMBOL(remap_pfn_range);
1228
1229 /*
1230  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1231  * servicing faults for write access.  In the normal case, do always want
1232  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1233  * that do not have writing enabled, when used by access_process_vm.
1234  */
1235 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1236 {
1237         if (likely(vma->vm_flags & VM_WRITE))
1238                 pte = pte_mkwrite(pte);
1239         return pte;
1240 }
1241
1242 /*
1243  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1244  */
1245 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
1246                 pte_t *page_table)
1247 {
1248         pte_t entry;
1249
1250         flush_cache_page(vma, address);
1251         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1252                               vma);
1253         ptep_establish(vma, address, page_table, entry);
1254         update_mmu_cache(vma, address, entry);
1255 }
1256
1257 /*
1258  * This routine handles present pages, when users try to write
1259  * to a shared page. It is done by copying the page to a new address
1260  * and decrementing the shared-page counter for the old page.
1261  *
1262  * Goto-purists beware: the only reason for goto's here is that it results
1263  * in better assembly code.. The "default" path will see no jumps at all.
1264  *
1265  * Note that this routine assumes that the protection checks have been
1266  * done by the caller (the low-level page fault routine in most cases).
1267  * Thus we can safely just mark it writable once we've done any necessary
1268  * COW.
1269  *
1270  * We also mark the page dirty at this point even though the page will
1271  * change only once the write actually happens. This avoids a few races,
1272  * and potentially makes it more efficient.
1273  *
1274  * We hold the mm semaphore and the page_table_lock on entry and exit
1275  * with the page_table_lock released.
1276  */
1277 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1278         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1279 {
1280         struct page *old_page, *new_page;
1281         unsigned long pfn = pte_pfn(pte);
1282         pte_t entry;
1283
1284         if (unlikely(!pfn_valid(pfn))) {
1285                 /*
1286                  * This should really halt the system so it can be debugged or
1287                  * at least the kernel stops what it's doing before it corrupts
1288                  * data, but for the moment just pretend this is OOM.
1289                  */
1290                 pte_unmap(page_table);
1291                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1292                                 address);
1293                 spin_unlock(&mm->page_table_lock);
1294                 return VM_FAULT_OOM;
1295         }
1296         old_page = pfn_to_page(pfn);
1297
1298         if (!TestSetPageLocked(old_page)) {
1299                 int reuse = can_share_swap_page(old_page);
1300                 unlock_page(old_page);
1301                 if (reuse) {
1302                         flush_cache_page(vma, address);
1303                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1304                                               vma);
1305                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1306                         update_mmu_cache(vma, address, entry);
1307                         pte_unmap(page_table);
1308                         spin_unlock(&mm->page_table_lock);
1309                         return VM_FAULT_MINOR;
1310                 }
1311         }
1312         pte_unmap(page_table);
1313
1314         /*
1315          * Ok, we need to copy. Oh, well..
1316          */
1317         if (!PageReserved(old_page))
1318                 page_cache_get(old_page);
1319         spin_unlock(&mm->page_table_lock);
1320
1321         if (unlikely(anon_vma_prepare(vma)))
1322                 goto no_new_page;
1323         if (old_page == ZERO_PAGE(address)) {
1324                 new_page = alloc_zeroed_user_highpage(vma, address);
1325                 if (!new_page)
1326                         goto no_new_page;
1327         } else {
1328                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1329                 if (!new_page)
1330                         goto no_new_page;
1331                 copy_user_highpage(new_page, old_page, address);
1332         }
1333         /*
1334          * Re-check the pte - we dropped the lock
1335          */
1336         spin_lock(&mm->page_table_lock);
1337         page_table = pte_offset_map(pmd, address);
1338         if (likely(pte_same(*page_table, pte))) {
1339                 if (PageAnon(old_page))
1340                         vx_anonpages_dec(mm);
1341                 if (PageReserved(old_page)) {
1342                         vx_rsspages_inc(mm);
1343                         acct_update_integrals();
1344                         update_mem_hiwater();
1345                 } else
1346                         page_remove_rmap(old_page);
1347                 break_cow(vma, new_page, address, page_table);
1348                 lru_cache_add_active(new_page);
1349                 page_add_anon_rmap(new_page, vma, address);
1350
1351                 /* Free the old page.. */
1352                 new_page = old_page;
1353         }
1354         pte_unmap(page_table);
1355         page_cache_release(new_page);
1356         page_cache_release(old_page);
1357         spin_unlock(&mm->page_table_lock);
1358         return VM_FAULT_MINOR;
1359
1360 no_new_page:
1361         page_cache_release(old_page);
1362         return VM_FAULT_OOM;
1363 }
1364
1365 /*
1366  * Helper functions for unmap_mapping_range().
1367  *
1368  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1369  *
1370  * We have to restart searching the prio_tree whenever we drop the lock,
1371  * since the iterator is only valid while the lock is held, and anyway
1372  * a later vma might be split and reinserted earlier while lock dropped.
1373  *
1374  * The list of nonlinear vmas could be handled more efficiently, using
1375  * a placeholder, but handle it in the same way until a need is shown.
1376  * It is important to search the prio_tree before nonlinear list: a vma
1377  * may become nonlinear and be shifted from prio_tree to nonlinear list
1378  * while the lock is dropped; but never shifted from list to prio_tree.
1379  *
1380  * In order to make forward progress despite restarting the search,
1381  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1382  * quickly skip it next time around.  Since the prio_tree search only
1383  * shows us those vmas affected by unmapping the range in question, we
1384  * can't efficiently keep all vmas in step with mapping->truncate_count:
1385  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1386  * mapping->truncate_count and vma->vm_truncate_count are protected by
1387  * i_mmap_lock.
1388  *
1389  * In order to make forward progress despite repeatedly restarting some
1390  * large vma, note the break_addr set by unmap_vmas when it breaks out:
1391  * and restart from that address when we reach that vma again.  It might
1392  * have been split or merged, shrunk or extended, but never shifted: so
1393  * restart_addr remains valid so long as it remains in the vma's range.
1394  * unmap_mapping_range forces truncate_count to leap over page-aligned
1395  * values so we can save vma's restart_addr in its truncate_count field.
1396  */
1397 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1398
1399 static void reset_vma_truncate_counts(struct address_space *mapping)
1400 {
1401         struct vm_area_struct *vma;
1402         struct prio_tree_iter iter;
1403
1404         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1405                 vma->vm_truncate_count = 0;
1406         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1407                 vma->vm_truncate_count = 0;
1408 }
1409
1410 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1411                 unsigned long start_addr, unsigned long end_addr,
1412                 struct zap_details *details)
1413 {
1414         unsigned long restart_addr;
1415         int need_break;
1416
1417 again:
1418         restart_addr = vma->vm_truncate_count;
1419         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1420                 start_addr = restart_addr;
1421                 if (start_addr >= end_addr) {
1422                         /* Top of vma has been split off since last time */
1423                         vma->vm_truncate_count = details->truncate_count;
1424                         return 0;
1425                 }
1426         }
1427
1428         details->break_addr = end_addr;
1429         zap_page_range(vma, start_addr, end_addr - start_addr, details);
1430
1431         /*
1432          * We cannot rely on the break test in unmap_vmas:
1433          * on the one hand, we don't want to restart our loop
1434          * just because that broke out for the page_table_lock;
1435          * on the other hand, it does no test when vma is small.
1436          */
1437         need_break = need_resched() ||
1438                         need_lockbreak(details->i_mmap_lock);
1439
1440         if (details->break_addr >= end_addr) {
1441                 /* We have now completed this vma: mark it so */
1442                 vma->vm_truncate_count = details->truncate_count;
1443                 if (!need_break)
1444                         return 0;
1445         } else {
1446                 /* Note restart_addr in vma's truncate_count field */
1447                 vma->vm_truncate_count = details->break_addr;
1448                 if (!need_break)
1449                         goto again;
1450         }
1451
1452         spin_unlock(details->i_mmap_lock);
1453         cond_resched();
1454         spin_lock(details->i_mmap_lock);
1455         return -EINTR;
1456 }
1457
1458 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1459                                             struct zap_details *details)
1460 {
1461         struct vm_area_struct *vma;
1462         struct prio_tree_iter iter;
1463         pgoff_t vba, vea, zba, zea;
1464
1465 restart:
1466         vma_prio_tree_foreach(vma, &iter, root,
1467                         details->first_index, details->last_index) {
1468                 /* Skip quickly over those we have already dealt with */
1469                 if (vma->vm_truncate_count == details->truncate_count)
1470                         continue;
1471
1472                 vba = vma->vm_pgoff;
1473                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1474                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1475                 zba = details->first_index;
1476                 if (zba < vba)
1477                         zba = vba;
1478                 zea = details->last_index;
1479                 if (zea > vea)
1480                         zea = vea;
1481
1482                 if (unmap_mapping_range_vma(vma,
1483                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1484                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1485                                 details) < 0)
1486                         goto restart;
1487         }
1488 }
1489
1490 static inline void unmap_mapping_range_list(struct list_head *head,
1491                                             struct zap_details *details)
1492 {
1493         struct vm_area_struct *vma;
1494
1495         /*
1496          * In nonlinear VMAs there is no correspondence between virtual address
1497          * offset and file offset.  So we must perform an exhaustive search
1498          * across *all* the pages in each nonlinear VMA, not just the pages
1499          * whose virtual address lies outside the file truncation point.
1500          */
1501 restart:
1502         list_for_each_entry(vma, head, shared.vm_set.list) {
1503                 /* Skip quickly over those we have already dealt with */
1504                 if (vma->vm_truncate_count == details->truncate_count)
1505                         continue;
1506                 details->nonlinear_vma = vma;
1507                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1508                                         vma->vm_end, details) < 0)
1509                         goto restart;
1510         }
1511 }
1512
1513 /**
1514  * unmap_mapping_range - unmap the portion of all mmaps
1515  * in the specified address_space corresponding to the specified
1516  * page range in the underlying file.
1517  * @address_space: the address space containing mmaps to be unmapped.
1518  * @holebegin: byte in first page to unmap, relative to the start of
1519  * the underlying file.  This will be rounded down to a PAGE_SIZE
1520  * boundary.  Note that this is different from vmtruncate(), which
1521  * must keep the partial page.  In contrast, we must get rid of
1522  * partial pages.
1523  * @holelen: size of prospective hole in bytes.  This will be rounded
1524  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1525  * end of the file.
1526  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1527  * but 0 when invalidating pagecache, don't throw away private data.
1528  */
1529 void unmap_mapping_range(struct address_space *mapping,
1530                 loff_t const holebegin, loff_t const holelen, int even_cows)
1531 {
1532         struct zap_details details;
1533         pgoff_t hba = holebegin >> PAGE_SHIFT;
1534         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1535
1536         /* Check for overflow. */
1537         if (sizeof(holelen) > sizeof(hlen)) {
1538                 long long holeend =
1539                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1540                 if (holeend & ~(long long)ULONG_MAX)
1541                         hlen = ULONG_MAX - hba + 1;
1542         }
1543
1544         details.check_mapping = even_cows? NULL: mapping;
1545         details.nonlinear_vma = NULL;
1546         details.first_index = hba;
1547         details.last_index = hba + hlen - 1;
1548         if (details.last_index < details.first_index)
1549                 details.last_index = ULONG_MAX;
1550         details.i_mmap_lock = &mapping->i_mmap_lock;
1551
1552         spin_lock(&mapping->i_mmap_lock);
1553
1554         /* serialize i_size write against truncate_count write */
1555         smp_wmb();
1556         /* Protect against page faults, and endless unmapping loops */
1557         mapping->truncate_count++;
1558         /*
1559          * For archs where spin_lock has inclusive semantics like ia64
1560          * this smp_mb() will prevent to read pagetable contents
1561          * before the truncate_count increment is visible to
1562          * other cpus.
1563          */
1564         smp_mb();
1565         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1566                 if (mapping->truncate_count == 0)
1567                         reset_vma_truncate_counts(mapping);
1568                 mapping->truncate_count++;
1569         }
1570         details.truncate_count = mapping->truncate_count;
1571
1572         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1573                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1574         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1575                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1576         spin_unlock(&mapping->i_mmap_lock);
1577 }
1578 EXPORT_SYMBOL(unmap_mapping_range);
1579
1580 /*
1581  * Handle all mappings that got truncated by a "truncate()"
1582  * system call.
1583  *
1584  * NOTE! We have to be ready to update the memory sharing
1585  * between the file and the memory map for a potential last
1586  * incomplete page.  Ugly, but necessary.
1587  */
1588 int vmtruncate(struct inode * inode, loff_t offset)
1589 {
1590         struct address_space *mapping = inode->i_mapping;
1591         unsigned long limit;
1592
1593         if (inode->i_size < offset)
1594                 goto do_expand;
1595         /*
1596          * truncation of in-use swapfiles is disallowed - it would cause
1597          * subsequent swapout to scribble on the now-freed blocks.
1598          */
1599         if (IS_SWAPFILE(inode))
1600                 goto out_busy;
1601         i_size_write(inode, offset);
1602         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1603         truncate_inode_pages(mapping, offset);
1604         goto out_truncate;
1605
1606 do_expand:
1607         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1608         if (limit != RLIM_INFINITY && offset > limit)
1609                 goto out_sig;
1610         if (offset > inode->i_sb->s_maxbytes)
1611                 goto out_big;
1612         i_size_write(inode, offset);
1613
1614 out_truncate:
1615         if (inode->i_op && inode->i_op->truncate)
1616                 inode->i_op->truncate(inode);
1617         return 0;
1618 out_sig:
1619         send_sig(SIGXFSZ, current, 0);
1620 out_big:
1621         return -EFBIG;
1622 out_busy:
1623         return -ETXTBSY;
1624 }
1625
1626 EXPORT_SYMBOL(vmtruncate);
1627
1628 /*
1629  * Primitive swap readahead code. We simply read an aligned block of
1630  * (1 << page_cluster) entries in the swap area. This method is chosen
1631  * because it doesn't cost us any seek time.  We also make sure to queue
1632  * the 'original' request together with the readahead ones...
1633  *
1634  * This has been extended to use the NUMA policies from the mm triggering
1635  * the readahead.
1636  *
1637  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1638  */
1639 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1640 {
1641 #ifdef CONFIG_NUMA
1642         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1643 #endif
1644         int i, num;
1645         struct page *new_page;
1646         unsigned long offset;
1647
1648         /*
1649          * Get the number of handles we should do readahead io to.
1650          */
1651         num = valid_swaphandles(entry, &offset);
1652         for (i = 0; i < num; offset++, i++) {
1653                 /* Ok, do the async read-ahead now */
1654                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1655                                                            offset), vma, addr);
1656                 if (!new_page)
1657                         break;
1658                 page_cache_release(new_page);
1659 #ifdef CONFIG_NUMA
1660                 /*
1661                  * Find the next applicable VMA for the NUMA policy.
1662                  */
1663                 addr += PAGE_SIZE;
1664                 if (addr == 0)
1665                         vma = NULL;
1666                 if (vma) {
1667                         if (addr >= vma->vm_end) {
1668                                 vma = next_vma;
1669                                 next_vma = vma ? vma->vm_next : NULL;
1670                         }
1671                         if (vma && addr < vma->vm_start)
1672                                 vma = NULL;
1673                 } else {
1674                         if (next_vma && addr >= next_vma->vm_start) {
1675                                 vma = next_vma;
1676                                 next_vma = vma->vm_next;
1677                         }
1678                 }
1679 #endif
1680         }
1681         lru_add_drain();        /* Push any new pages onto the LRU now */
1682 }
1683
1684 /*
1685  * We hold the mm semaphore and the page_table_lock on entry and
1686  * should release the pagetable lock on exit..
1687  */
1688 static int do_swap_page(struct mm_struct * mm,
1689         struct vm_area_struct * vma, unsigned long address,
1690         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1691 {
1692         struct page *page;
1693         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1694         pte_t pte;
1695         int ret = VM_FAULT_MINOR;
1696
1697         pte_unmap(page_table);
1698         spin_unlock(&mm->page_table_lock);
1699         page = lookup_swap_cache(entry);
1700         if (!page) {
1701                 swapin_readahead(entry, address, vma);
1702                 page = read_swap_cache_async(entry, vma, address);
1703                 if (!page) {
1704                         /*
1705                          * Back out if somebody else faulted in this pte while
1706                          * we released the page table lock.
1707                          */
1708                         spin_lock(&mm->page_table_lock);
1709                         page_table = pte_offset_map(pmd, address);
1710                         if (likely(pte_same(*page_table, orig_pte)))
1711                                 ret = VM_FAULT_OOM;
1712                         else
1713                                 ret = VM_FAULT_MINOR;
1714                         pte_unmap(page_table);
1715                         spin_unlock(&mm->page_table_lock);
1716                         goto out;
1717                 }
1718
1719                 /* Had to read the page from swap area: Major fault */
1720                 ret = VM_FAULT_MAJOR;
1721                 inc_page_state(pgmajfault);
1722                 grab_swap_token();
1723         }
1724
1725         if (!vx_rsspages_avail(mm, 1)) {
1726                 ret = VM_FAULT_OOM;
1727                 goto out;
1728         }
1729         mark_page_accessed(page);
1730         lock_page(page);
1731
1732         /*
1733          * Back out if somebody else faulted in this pte while we
1734          * released the page table lock.
1735          */
1736         spin_lock(&mm->page_table_lock);
1737         page_table = pte_offset_map(pmd, address);
1738         if (unlikely(!pte_same(*page_table, orig_pte))) {
1739                 pte_unmap(page_table);
1740                 spin_unlock(&mm->page_table_lock);
1741                 unlock_page(page);
1742                 page_cache_release(page);
1743                 ret = VM_FAULT_MINOR;
1744                 goto out;
1745         }
1746
1747         /* The page isn't present yet, go ahead with the fault. */
1748
1749         swap_free(entry);
1750         if (vm_swap_full())
1751                 remove_exclusive_swap_page(page);
1752
1753         vx_rsspages_inc(mm);
1754         acct_update_integrals();
1755         update_mem_hiwater();
1756
1757         pte = mk_pte(page, vma->vm_page_prot);
1758         if (write_access && can_share_swap_page(page)) {
1759                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1760                 write_access = 0;
1761         }
1762         unlock_page(page);
1763
1764         flush_icache_page(vma, page);
1765         set_pte(page_table, pte);
1766         page_add_anon_rmap(page, vma, address);
1767
1768         if (write_access) {
1769                 if (do_wp_page(mm, vma, address,
1770                                 page_table, pmd, pte) == VM_FAULT_OOM)
1771                         ret = VM_FAULT_OOM;
1772                 goto out;
1773         }
1774
1775         /* No need to invalidate - it was non-present before */
1776         update_mmu_cache(vma, address, pte);
1777         pte_unmap(page_table);
1778         spin_unlock(&mm->page_table_lock);
1779 out:
1780         return ret;
1781 }
1782
1783 /*
1784  * We are called with the MM semaphore and page_table_lock
1785  * spinlock held to protect against concurrent faults in
1786  * multithreaded programs.
1787  */
1788 static int
1789 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1790                 pte_t *page_table, pmd_t *pmd, int write_access,
1791                 unsigned long addr)
1792 {
1793         pte_t entry;
1794         struct page * page = ZERO_PAGE(addr);
1795
1796         /* Read-only mapping of ZERO_PAGE. */
1797         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1798
1799         /* ..except if it's a write access */
1800         if (write_access) {
1801                 /* Allocate our own private page. */
1802                 pte_unmap(page_table);
1803                 spin_unlock(&mm->page_table_lock);
1804
1805                 if (!vx_rsspages_avail(mm, 1))
1806                         goto no_mem;
1807                 if (unlikely(anon_vma_prepare(vma)))
1808                         goto no_mem;
1809                 page = alloc_zeroed_user_highpage(vma, addr);
1810                 if (!page)
1811                         goto no_mem;
1812
1813                 spin_lock(&mm->page_table_lock);
1814                 page_table = pte_offset_map(pmd, addr);
1815
1816                 if (!pte_none(*page_table)) {
1817                         pte_unmap(page_table);
1818                         page_cache_release(page);
1819                         spin_unlock(&mm->page_table_lock);
1820                         goto out;
1821                 }
1822                 vx_rsspages_inc(mm);
1823                 acct_update_integrals();
1824                 update_mem_hiwater();
1825                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1826                                                          vma->vm_page_prot)),
1827                                       vma);
1828                 lru_cache_add_active(page);
1829                 SetPageReferenced(page);
1830                 page_add_anon_rmap(page, vma, addr);
1831         }
1832
1833         set_pte(page_table, entry);
1834         pte_unmap(page_table);
1835
1836         /* No need to invalidate - it was non-present before */
1837         update_mmu_cache(vma, addr, entry);
1838         spin_unlock(&mm->page_table_lock);
1839 out:
1840         return VM_FAULT_MINOR;
1841 no_mem:
1842         return VM_FAULT_OOM;
1843 }
1844
1845 /*
1846  * do_no_page() tries to create a new page mapping. It aggressively
1847  * tries to share with existing pages, but makes a separate copy if
1848  * the "write_access" parameter is true in order to avoid the next
1849  * page fault.
1850  *
1851  * As this is called only for pages that do not currently exist, we
1852  * do not need to flush old virtual caches or the TLB.
1853  *
1854  * This is called with the MM semaphore held and the page table
1855  * spinlock held. Exit with the spinlock released.
1856  */
1857 static int
1858 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1859         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1860 {
1861         struct page * new_page;
1862         struct address_space *mapping = NULL;
1863         pte_t entry;
1864         unsigned int sequence = 0;
1865         int ret = VM_FAULT_MINOR;
1866         int anon = 0;
1867
1868         if (!vma->vm_ops || !vma->vm_ops->nopage)
1869                 return do_anonymous_page(mm, vma, page_table,
1870                                         pmd, write_access, address);
1871         pte_unmap(page_table);
1872         spin_unlock(&mm->page_table_lock);
1873
1874         if (vma->vm_file) {
1875                 mapping = vma->vm_file->f_mapping;
1876                 sequence = mapping->truncate_count;
1877                 smp_rmb(); /* serializes i_size against truncate_count */
1878         }
1879 retry:
1880         cond_resched();
1881         /* FIXME: is that check useful here? */
1882         if (!vx_rsspages_avail(mm, 1))
1883                 return VM_FAULT_OOM;
1884         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1885         /*
1886          * No smp_rmb is needed here as long as there's a full
1887          * spin_lock/unlock sequence inside the ->nopage callback
1888          * (for the pagecache lookup) that acts as an implicit
1889          * smp_mb() and prevents the i_size read to happen
1890          * after the next truncate_count read.
1891          */
1892
1893         /* no page was available -- either SIGBUS or OOM */
1894         if (new_page == NOPAGE_SIGBUS)
1895                 return VM_FAULT_SIGBUS;
1896         if (new_page == NOPAGE_OOM)
1897                 return VM_FAULT_OOM;
1898
1899         /*
1900          * Should we do an early C-O-W break?
1901          */
1902         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1903                 struct page *page;
1904
1905                 if (unlikely(anon_vma_prepare(vma)))
1906                         goto oom;
1907                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1908                 if (!page)
1909                         goto oom;
1910                 copy_user_highpage(page, new_page, address);
1911                 page_cache_release(new_page);
1912                 new_page = page;
1913                 anon = 1;
1914         }
1915
1916         spin_lock(&mm->page_table_lock);
1917         /*
1918          * For a file-backed vma, someone could have truncated or otherwise
1919          * invalidated this page.  If unmap_mapping_range got called,
1920          * retry getting the page.
1921          */
1922         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1923                 sequence = mapping->truncate_count;
1924                 spin_unlock(&mm->page_table_lock);
1925                 page_cache_release(new_page);
1926                 goto retry;
1927         }
1928         page_table = pte_offset_map(pmd, address);
1929
1930         /*
1931          * This silly early PAGE_DIRTY setting removes a race
1932          * due to the bad i386 page protection. But it's valid
1933          * for other architectures too.
1934          *
1935          * Note that if write_access is true, we either now have
1936          * an exclusive copy of the page, or this is a shared mapping,
1937          * so we can make it writable and dirty to avoid having to
1938          * handle that later.
1939          */
1940         /* Only go through if we didn't race with anybody else... */
1941         if (pte_none(*page_table)) {
1942                 if (!PageReserved(new_page))
1943                         vx_rsspages_inc(mm);
1944                 acct_update_integrals();
1945                 update_mem_hiwater();
1946
1947                 flush_icache_page(vma, new_page);
1948                 entry = mk_pte(new_page, vma->vm_page_prot);
1949                 if (write_access)
1950                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1951                 set_pte(page_table, entry);
1952                 if (anon) {
1953                         lru_cache_add_active(new_page);
1954                         page_add_anon_rmap(new_page, vma, address);
1955                 } else
1956                         page_add_file_rmap(new_page);
1957                 pte_unmap(page_table);
1958         } else {
1959                 /* One of our sibling threads was faster, back out. */
1960                 pte_unmap(page_table);
1961                 page_cache_release(new_page);
1962                 spin_unlock(&mm->page_table_lock);
1963                 goto out;
1964         }
1965
1966         /* no need to invalidate: a not-present page shouldn't be cached */
1967         update_mmu_cache(vma, address, entry);
1968         spin_unlock(&mm->page_table_lock);
1969 out:
1970         return ret;
1971 oom:
1972         page_cache_release(new_page);
1973         ret = VM_FAULT_OOM;
1974         goto out;
1975 }
1976
1977 /*
1978  * Fault of a previously existing named mapping. Repopulate the pte
1979  * from the encoded file_pte if possible. This enables swappable
1980  * nonlinear vmas.
1981  */
1982 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1983         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1984 {
1985         unsigned long pgoff;
1986         int err;
1987
1988         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1989         /*
1990          * Fall back to the linear mapping if the fs does not support
1991          * ->populate:
1992          */
1993         if (!vma->vm_ops || !vma->vm_ops->populate ||
1994                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1995                 pte_clear(pte);
1996                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1997         }
1998
1999         pgoff = pte_to_pgoff(*pte);
2000
2001         pte_unmap(pte);
2002         spin_unlock(&mm->page_table_lock);
2003
2004         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
2005         if (err == -ENOMEM)
2006                 return VM_FAULT_OOM;
2007         if (err)
2008                 return VM_FAULT_SIGBUS;
2009         return VM_FAULT_MAJOR;
2010 }
2011
2012 /*
2013  * These routines also need to handle stuff like marking pages dirty
2014  * and/or accessed for architectures that don't do it in hardware (most
2015  * RISC architectures).  The early dirtying is also good on the i386.
2016  *
2017  * There is also a hook called "update_mmu_cache()" that architectures
2018  * with external mmu caches can use to update those (ie the Sparc or
2019  * PowerPC hashed page tables that act as extended TLBs).
2020  *
2021  * Note the "page_table_lock". It is to protect against kswapd removing
2022  * pages from under us. Note that kswapd only ever _removes_ pages, never
2023  * adds them. As such, once we have noticed that the page is not present,
2024  * we can drop the lock early.
2025  *
2026  * The adding of pages is protected by the MM semaphore (which we hold),
2027  * so we don't need to worry about a page being suddenly been added into
2028  * our VM.
2029  *
2030  * We enter with the pagetable spinlock held, we are supposed to
2031  * release it when done.
2032  */
2033 static inline int handle_pte_fault(struct mm_struct *mm,
2034         struct vm_area_struct * vma, unsigned long address,
2035         int write_access, pte_t *pte, pmd_t *pmd)
2036 {
2037         pte_t entry;
2038
2039         entry = *pte;
2040         if (!pte_present(entry)) {
2041                 /*
2042                  * If it truly wasn't present, we know that kswapd
2043                  * and the PTE updates will not touch it later. So
2044                  * drop the lock.
2045                  */
2046                 if (pte_none(entry))
2047                         return do_no_page(mm, vma, address, write_access, pte, pmd);
2048                 if (pte_file(entry))
2049                         return do_file_page(mm, vma, address, write_access, pte, pmd);
2050                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
2051         }
2052
2053         if (write_access) {
2054                 if (!pte_write(entry))
2055                         return do_wp_page(mm, vma, address, pte, pmd, entry);
2056
2057                 entry = pte_mkdirty(entry);
2058         }
2059         entry = pte_mkyoung(entry);
2060         ptep_set_access_flags(vma, address, pte, entry, write_access);
2061         update_mmu_cache(vma, address, entry);
2062         pte_unmap(pte);
2063         spin_unlock(&mm->page_table_lock);
2064         return VM_FAULT_MINOR;
2065 }
2066
2067 /*
2068  * By the time we get here, we already hold the mm semaphore
2069  */
2070 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
2071                 unsigned long address, int write_access)
2072 {
2073         pgd_t *pgd;
2074         pud_t *pud;
2075         pmd_t *pmd;
2076         pte_t *pte;
2077
2078         __set_current_state(TASK_RUNNING);
2079
2080         inc_page_state(pgfault);
2081
2082         if (is_vm_hugetlb_page(vma))
2083                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
2084
2085         /*
2086          * We need the page table lock to synchronize with kswapd
2087          * and the SMP-safe atomic PTE updates.
2088          */
2089         pgd = pgd_offset(mm, address);
2090         spin_lock(&mm->page_table_lock);
2091
2092         pud = pud_alloc(mm, pgd, address);
2093         if (!pud)
2094                 goto oom;
2095
2096         pmd = pmd_alloc(mm, pud, address);
2097         if (!pmd)
2098                 goto oom;
2099
2100         pte = pte_alloc_map(mm, pmd, address);
2101         if (!pte)
2102                 goto oom;
2103
2104         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
2105
2106  oom:
2107         spin_unlock(&mm->page_table_lock);
2108         return VM_FAULT_OOM;
2109 }
2110
2111 #ifndef __ARCH_HAS_4LEVEL_HACK
2112 /*
2113  * Allocate page upper directory.
2114  *
2115  * We've already handled the fast-path in-line, and we own the
2116  * page table lock.
2117  *
2118  * On a two-level or three-level page table, this ends up actually being
2119  * entirely optimized away.
2120  */
2121 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2122 {
2123         pud_t *new;
2124
2125         spin_unlock(&mm->page_table_lock);
2126         new = pud_alloc_one(mm, address);
2127         spin_lock(&mm->page_table_lock);
2128         if (!new)
2129                 return NULL;
2130
2131         /*
2132          * Because we dropped the lock, we should re-check the
2133          * entry, as somebody else could have populated it..
2134          */
2135         if (pgd_present(*pgd)) {
2136                 pud_free(new);
2137                 goto out;
2138         }
2139         pgd_populate(mm, pgd, new);
2140  out:
2141         return pud_offset(pgd, address);
2142 }
2143
2144 /*
2145  * Allocate page middle directory.
2146  *
2147  * We've already handled the fast-path in-line, and we own the
2148  * page table lock.
2149  *
2150  * On a two-level page table, this ends up actually being entirely
2151  * optimized away.
2152  */
2153 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2154 {
2155         pmd_t *new;
2156
2157         spin_unlock(&mm->page_table_lock);
2158         new = pmd_alloc_one(mm, address);
2159         spin_lock(&mm->page_table_lock);
2160         if (!new)
2161                 return NULL;
2162
2163         /*
2164          * Because we dropped the lock, we should re-check the
2165          * entry, as somebody else could have populated it..
2166          */
2167         if (pud_present(*pud)) {
2168                 pmd_free(new);
2169                 goto out;
2170         }
2171         pud_populate(mm, pud, new);
2172  out:
2173         return pmd_offset(pud, address);
2174 }
2175 #else
2176 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2177 {
2178         pmd_t *new;
2179
2180         spin_unlock(&mm->page_table_lock);
2181         new = pmd_alloc_one(mm, address);
2182         spin_lock(&mm->page_table_lock);
2183         if (!new)
2184                 return NULL;
2185
2186         /*
2187          * Because we dropped the lock, we should re-check the
2188          * entry, as somebody else could have populated it..
2189          */
2190         if (pgd_present(*pud)) {
2191                 pmd_free(new);
2192                 goto out;
2193         }
2194         pgd_populate(mm, pud, new);
2195 out:
2196         return pmd_offset(pud, address);
2197 }
2198 #endif
2199
2200 int make_pages_present(unsigned long addr, unsigned long end)
2201 {
2202         int ret, len, write;
2203         struct vm_area_struct * vma;
2204
2205         vma = find_vma(current->mm, addr);
2206         if (!vma)
2207                 return -1;
2208         write = (vma->vm_flags & VM_WRITE) != 0;
2209         if (addr >= end)
2210                 BUG();
2211         if (end > vma->vm_end)
2212                 BUG();
2213         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2214         ret = get_user_pages(current, current->mm, addr,
2215                         len, write, 0, NULL, NULL);
2216         if (ret < 0)
2217                 return ret;
2218         return ret == len ? 0 : -1;
2219 }
2220
2221 /*
2222  * Map a vmalloc()-space virtual address to the physical page.
2223  */
2224 struct page * vmalloc_to_page(void * vmalloc_addr)
2225 {
2226         unsigned long addr = (unsigned long) vmalloc_addr;
2227         struct page *page = NULL;
2228         pgd_t *pgd = pgd_offset_k(addr);
2229         pud_t *pud;
2230         pmd_t *pmd;
2231         pte_t *ptep, pte;
2232
2233         if (!pgd_none(*pgd)) {
2234                 pud = pud_offset(pgd, addr);
2235                 if (!pud_none(*pud)) {
2236                         pmd = pmd_offset(pud, addr);
2237                         if (!pmd_none(*pmd)) {
2238                                 ptep = pte_offset_map(pmd, addr);
2239                                 pte = *ptep;
2240                                 if (pte_present(pte))
2241                                         page = pte_page(pte);
2242                                 pte_unmap(ptep);
2243                         }
2244                 }
2245         }
2246         return page;
2247 }
2248
2249 EXPORT_SYMBOL(vmalloc_to_page);
2250
2251 /*
2252  * Map a vmalloc()-space virtual address to the physical page frame number.
2253  */
2254 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2255 {
2256         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2257 }
2258
2259 EXPORT_SYMBOL(vmalloc_to_pfn);
2260
2261 /*
2262  * update_mem_hiwater
2263  *      - update per process rss and vm high water data
2264  */
2265 void update_mem_hiwater(void)
2266 {
2267         struct task_struct *tsk = current;
2268
2269         if (tsk->mm) {
2270                 if (tsk->mm->hiwater_rss < tsk->mm->rss)
2271                         tsk->mm->hiwater_rss = tsk->mm->rss;
2272                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2273                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2274         }
2275 }
2276
2277 #if !defined(__HAVE_ARCH_GATE_AREA)
2278
2279 #if defined(AT_SYSINFO_EHDR)
2280 struct vm_area_struct gate_vma;
2281
2282 static int __init gate_vma_init(void)
2283 {
2284         gate_vma.vm_mm = NULL;
2285         gate_vma.vm_start = FIXADDR_USER_START;
2286         gate_vma.vm_end = FIXADDR_USER_END;
2287         gate_vma.vm_page_prot = PAGE_READONLY;
2288         gate_vma.vm_flags = 0;
2289         return 0;
2290 }
2291 __initcall(gate_vma_init);
2292 #endif
2293
2294 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2295 {
2296 #ifdef AT_SYSINFO_EHDR
2297         return &gate_vma;
2298 #else
2299         return NULL;
2300 #endif
2301 }
2302
2303 int in_gate_area_no_task(unsigned long addr)
2304 {
2305 #ifdef AT_SYSINFO_EHDR
2306         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2307                 return 1;
2308 #endif
2309         return 0;
2310 }
2311
2312 #endif  /* __HAVE_ARCH_GATE_AREA */