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
  39 #include <linux/kernel_stat.h>
  40 #include <linux/mm.h>
  41 #include <linux/hugetlb.h>
  42 #include <linux/mman.h>
  43 #include <linux/swap.h>
  44 #include <linux/highmem.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/rmap.h>
  47 #include <linux/module.h>
  48 #include <linux/init.h>
  49
  50 #include <asm/pgalloc.h>
  51 #include <asm/rmap.h>
  52 #include <asm/uaccess.h>
  53 #include <asm/tlb.h>
  54 #include <asm/tlbflush.h>
  55 #include <asm/pgtable.h>
  56
  57 #include <linux/swapops.h>
  58 #include <linux/elf.h>
  59
  60 #ifndef CONFIG_DISCONTIGMEM
  61 /* use the per-pgdat data instead for discontigmem - mbligh */
  62 unsigned long max_mapnr;
  63 struct page *mem_map;
  64
  65 EXPORT_SYMBOL(max_mapnr);
  66 EXPORT_SYMBOL(mem_map);
  67 #endif
  68
  69 unsigned long num_physpages;
  70 void * high_memory;
  71 struct page *highmem_start_page;
  72
  73 EXPORT_SYMBOL(num_physpages);
  74 EXPORT_SYMBOL(highmem_start_page);
  75 EXPORT_SYMBOL(high_memory);
  76
  77 /*
  78  * We special-case the C-O-W ZERO_PAGE, because it's such
  79  * a common occurrence (no need to read the page to know
  80  * that it's zero - better for the cache and memory subsystem).
  81  */
  82 static inline void copy_cow_page(struct page * from, struct page * to, unsigned long address)
  83 {
  84         if (from == ZERO_PAGE(address)) {
  85                 clear_user_highpage(to, address);
  86                 return;
  87         }
  88         copy_user_highpage(to, from, address);
  89 }
  90
  91 /*
  92  * Note: this doesn't free the actual pages themselves. That
  93  * has been handled earlier when unmapping all the memory regions.
  94  */
  95 static inline void free_one_pmd(struct mmu_gather *tlb, pmd_t * dir)
  96 {
  97         struct page *page;
  98
  99         if (pmd_none(*dir))
 100                 return;
 101         if (pmd_bad(*dir)) {
 102                 pmd_ERROR(*dir);
 103                 pmd_clear(dir);
 104                 return;
 105         }
 106         page = pmd_page(*dir);
 107         pmd_clear(dir);
 108         pgtable_remove_rmap(page);
 109         pte_free_tlb(tlb, page);
 110 }
 111
 112 static inline void free_one_pgd(struct mmu_gather *tlb, pgd_t * dir)
 113 {
 114         int j;
 115         pmd_t * pmd;
 116
 117         if (pgd_none(*dir))
 118                 return;
 119         if (pgd_bad(*dir)) {
 120                 pgd_ERROR(*dir);
 121                 pgd_clear(dir);
 122                 return;
 123         }
 124         pmd = pmd_offset(dir, 0);
 125         pgd_clear(dir);
 126         for (j = 0; j < PTRS_PER_PMD ; j++)
 127                 free_one_pmd(tlb, pmd+j);
 128         pmd_free_tlb(tlb, pmd);
 129 }
 130
 131 /*
 132  * This function clears all user-level page tables of a process - this
 133  * is needed by execve(), so that old pages aren't in the way.
 134  *
 135  * Must be called with pagetable lock held.
 136  */
 137 void clear_page_tables(struct mmu_gather *tlb, unsigned long first, int nr)
 138 {
 139         pgd_t * page_dir = tlb->mm->pgd;
 140
 141         page_dir += first;
 142         do {
 143                 free_one_pgd(tlb, page_dir);
 144                 page_dir++;
 145         } while (--nr);
 146 }
 147
 148 pte_t fastcall * pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 149 {
 150         if (!pmd_present(*pmd)) {
 151                 struct page *new;
 152
 153                 spin_unlock(&mm->page_table_lock);
 154                 new = pte_alloc_one(mm, address);
 155                 spin_lock(&mm->page_table_lock);
 156                 if (!new)
 157                         return NULL;
 158
 159                 /*
 160                  * Because we dropped the lock, we should re-check the
 161                  * entry, as somebody else could have populated it..
 162                  */
 163                 if (pmd_present(*pmd)) {
 164                         pte_free(new);
 165                         goto out;
 166                 }
 167                 pgtable_add_rmap(new, mm, address);
 168                 pmd_populate(mm, pmd, new);
 169         }
 170 out:
 171         return pte_offset_map(pmd, address);
 172 }
 173
 174 pte_t fastcall * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 175 {
 176         if (!pmd_present(*pmd)) {
 177                 pte_t *new;
 178
 179                 spin_unlock(&mm->page_table_lock);
 180                 new = pte_alloc_one_kernel(mm, address);
 181                 spin_lock(&mm->page_table_lock);
 182                 if (!new)
 183                         return NULL;
 184
 185                 /*
 186                  * Because we dropped the lock, we should re-check the
 187                  * entry, as somebody else could have populated it..
 188                  */
 189                 if (pmd_present(*pmd)) {
 190                         pte_free_kernel(new);
 191                         goto out;
 192                 }
 193                 pgtable_add_rmap(virt_to_page(new), mm, address);
 194                 pmd_populate_kernel(mm, pmd, new);
 195         }
 196 out:
 197         return pte_offset_kernel(pmd, address);
 198 }
 199 #define PTE_TABLE_MASK  ((PTRS_PER_PTE-1) * sizeof(pte_t))
 200 #define PMD_TABLE_MASK  ((PTRS_PER_PMD-1) * sizeof(pmd_t))
 201
 202 /*
 203  * copy one vm_area from one task to the other. Assumes the page tables
 204  * already present in the new task to be cleared in the whole range
 205  * covered by this vma.
 206  *
 207  * 08Jan98 Merged into one routine from several inline routines to reduce
 208  *         variable count and make things faster. -jj
 209  *
 210  * dst->page_table_lock is held on entry and exit,
 211  * but may be dropped within pmd_alloc() and pte_alloc_map().
 212  */
 213 int copy_page_range(struct mm_struct *dst, struct mm_struct *src,
 214                         struct vm_area_struct *vma)
 215 {
 216         pgd_t * src_pgd, * dst_pgd;
 217         unsigned long address = vma->vm_start;
 218         unsigned long end = vma->vm_end;
 219         unsigned long cow;
 220         struct pte_chain *pte_chain = NULL;
 221
 222         if (is_vm_hugetlb_page(vma))
 223                 return copy_hugetlb_page_range(dst, src, vma);
 224
 225         pte_chain = pte_chain_alloc(GFP_ATOMIC | __GFP_NOWARN);
 226         if (!pte_chain) {
 227                 spin_unlock(&dst->page_table_lock);
 228                 pte_chain = pte_chain_alloc(GFP_KERNEL);
 229                 spin_lock(&dst->page_table_lock);
 230                 if (!pte_chain)
 231                         goto nomem;
 232         }
 233
 234         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 235         src_pgd = pgd_offset(src, address)-1;
 236         dst_pgd = pgd_offset(dst, address)-1;
 237
 238         for (;;) {
 239                 pmd_t * src_pmd, * dst_pmd;
 240
 241                 src_pgd++; dst_pgd++;
 242
 243                 /* copy_pmd_range */
 244
 245                 if (pgd_none(*src_pgd))
 246                         goto skip_copy_pmd_range;
 247                 if (pgd_bad(*src_pgd)) {
 248                         pgd_ERROR(*src_pgd);
 249                         pgd_clear(src_pgd);
 250 skip_copy_pmd_range:    address = (address + PGDIR_SIZE) & PGDIR_MASK;
 251                         if (!address || (address >= end))
 252                                 goto out;
 253                         continue;
 254                 }
 255
 256                 src_pmd = pmd_offset(src_pgd, address);
 257                 dst_pmd = pmd_alloc(dst, dst_pgd, address);
 258                 if (!dst_pmd)
 259                         goto nomem;
 260
 261                 do {
 262                         pte_t * src_pte, * dst_pte;
 263
 264                         /* copy_pte_range */
 265
 266                         if (pmd_none(*src_pmd))
 267                                 goto skip_copy_pte_range;
 268                         if (pmd_bad(*src_pmd)) {
 269                                 pmd_ERROR(*src_pmd);
 270                                 pmd_clear(src_pmd);
 271 skip_copy_pte_range:
 272                                 address = (address + PMD_SIZE) & PMD_MASK;
 273                                 if (address >= end)
 274                                         goto out;
 275                                 goto cont_copy_pmd_range;
 276                         }
 277
 278                         dst_pte = pte_alloc_map(dst, dst_pmd, address);
 279                         if (!dst_pte)
 280                                 goto nomem;
 281                         spin_lock(&src->page_table_lock);
 282                         src_pte = pte_offset_map_nested(src_pmd, address);
 283                         do {
 284                                 pte_t pte = *src_pte;
 285                                 struct page *page;
 286                                 unsigned long pfn;
 287
 288                                 /* copy_one_pte */
 289
 290                                 if (pte_none(pte))
 291                                         goto cont_copy_pte_range_noset;
 292                                 /* pte contains position in swap, so copy. */
 293                                 if (!pte_present(pte)) {
 294                                         if (!pte_file(pte))
 295                                                 swap_duplicate(pte_to_swp_entry(pte));
 296                                         set_pte(dst_pte, pte);
 297                                         goto cont_copy_pte_range_noset;
 298                                 }
 299                                 pfn = pte_pfn(pte);
 300                                 /* the pte points outside of valid memory, the
 301                                  * mapping is assumed to be good, meaningful
 302                                  * and not mapped via rmap - duplicate the
 303                                  * mapping as is.
 304                                  */
 305                                 page = NULL;
 306                                 if (pfn_valid(pfn))
 307                                         page = pfn_to_page(pfn);
 308
 309                                 if (!page || PageReserved(page)) {
 310                                         set_pte(dst_pte, pte);
 311                                         goto cont_copy_pte_range_noset;
 312                                 }
 313
 314                                 /*
 315                                  * If it's a COW mapping, write protect it both
 316                                  * in the parent and the child
 317                                  */
 318                                 if (cow) {
 319                                         ptep_set_wrprotect(src_pte);
 320                                         pte = *src_pte;
 321                                 }
 322
 323                                 /*
 324                                  * If it's a shared mapping, mark it clean in
 325                                  * the child
 326                                  */
 327                                 if (vma->vm_flags & VM_SHARED)
 328                                         pte = pte_mkclean(pte);
 329                                 pte = pte_mkold(pte);
 330                                 get_page(page);
 331                                 dst->rss++;
 332
 333                                 set_pte(dst_pte, pte);
 334                                 pte_chain = page_add_rmap(page, dst_pte,
 335                                                         pte_chain);
 336                                 if (pte_chain)
 337                                         goto cont_copy_pte_range_noset;
 338                                 pte_chain = pte_chain_alloc(GFP_ATOMIC | __GFP_NOWARN);
 339                                 if (pte_chain)
 340                                         goto cont_copy_pte_range_noset;
 341
 342                                 /*
 343                                  * pte_chain allocation failed, and we need to
 344                                  * run page reclaim.
 345                                  */
 346                                 pte_unmap_nested(src_pte);
 347                                 pte_unmap(dst_pte);
 348                                 spin_unlock(&src->page_table_lock);
 349                                 spin_unlock(&dst->page_table_lock);
 350                                 pte_chain = pte_chain_alloc(GFP_KERNEL);
 351                                 spin_lock(&dst->page_table_lock);
 352                                 if (!pte_chain)
 353                                         goto nomem;
 354                                 spin_lock(&src->page_table_lock);
 355                                 dst_pte = pte_offset_map(dst_pmd, address);
 356                                 src_pte = pte_offset_map_nested(src_pmd,
 357                                                                 address);
 358 cont_copy_pte_range_noset:
 359                                 address += PAGE_SIZE;
 360                                 if (address >= end) {
 361                                         pte_unmap_nested(src_pte);
 362                                         pte_unmap(dst_pte);
 363                                         goto out_unlock;
 364                                 }
 365                                 src_pte++;
 366                                 dst_pte++;
 367                         } while ((unsigned long)src_pte & PTE_TABLE_MASK);
 368                         pte_unmap_nested(src_pte-1);
 369                         pte_unmap(dst_pte-1);
 370                         spin_unlock(&src->page_table_lock);
 371
 372 cont_copy_pmd_range:
 373                         src_pmd++;
 374                         dst_pmd++;
 375                 } while ((unsigned long)src_pmd & PMD_TABLE_MASK);
 376         }
 377 out_unlock:
 378         spin_unlock(&src->page_table_lock);
 379 out:
 380         pte_chain_free(pte_chain);
 381         return 0;
 382 nomem:
 383         pte_chain_free(pte_chain);
 384         return -ENOMEM;
 385 }
 386
 387 static void zap_pte_range(struct mmu_gather *tlb,
 388                 pmd_t *pmd, unsigned long address,
 389                 unsigned long size, struct zap_details *details)
 390 {
 391         unsigned long offset;
 392         pte_t *ptep;
 393
 394         if (pmd_none(*pmd))
 395                 return;
 396         if (pmd_bad(*pmd)) {
 397                 pmd_ERROR(*pmd);
 398                 pmd_clear(pmd);
 399                 return;
 400         }
 401         ptep = pte_offset_map(pmd, address);
 402         offset = address & ~PMD_MASK;
 403         if (offset + size > PMD_SIZE)
 404                 size = PMD_SIZE - offset;
 405         size &= PAGE_MASK;
 406         for (offset=0; offset < size; ptep++, offset += PAGE_SIZE) {
 407                 pte_t pte = *ptep;
 408                 if (pte_none(pte))
 409                         continue;
 410                 if (pte_present(pte)) {
 411                         struct page *page = NULL;
 412                         unsigned long pfn = pte_pfn(pte);
 413                         if (pfn_valid(pfn)) {
 414                                 page = pfn_to_page(pfn);
 415                                 if (PageReserved(page))
 416                                         page = NULL;
 417                         }
 418                         if (unlikely(details) && page) {
 419                                 /*
 420                                  * unmap_shared_mapping_pages() wants to
 421                                  * invalidate cache without truncating:
 422                                  * unmap shared but keep private pages.
 423                                  */
 424                                 if (details->check_mapping &&
 425                                     details->check_mapping != page->mapping)
 426                                         continue;
 427                                 /*
 428                                  * Each page->index must be checked when
 429                                  * invalidating or truncating nonlinear.
 430                                  */
 431                                 if (details->nonlinear_vma &&
 432                                     (page->index < details->first_index ||
 433                                      page->index > details->last_index))
 434                                         continue;
 435                         }
 436                         pte = ptep_get_and_clear(ptep);
 437                         tlb_remove_tlb_entry(tlb, ptep, address+offset);
 438                         if (unlikely(!page))
 439                                 continue;
 440                         if (unlikely(details) && details->nonlinear_vma
 441                             && linear_page_index(details->nonlinear_vma,
 442                                         address+offset) != page->index)
 443                                 set_pte(ptep, pgoff_to_pte(page->index));
 444                         if (pte_dirty(pte))
 445                                 set_page_dirty(page);
 446                         if (pte_young(pte) && page_mapping(page))
 447                                 mark_page_accessed(page);
 448                         tlb->freed++;
 449                         page_remove_rmap(page, ptep);
 450                         tlb_remove_page(tlb, page);
 451                         continue;
 452                 }
 453                 /*
 454                  * If details->check_mapping, we leave swap entries;
 455                  * if details->nonlinear_vma, we leave file entries.
 456                  */
 457                 if (unlikely(details))
 458                         continue;
 459                 if (!pte_file(pte))
 460                         free_swap_and_cache(pte_to_swp_entry(pte));
 461                 pte_clear(ptep);
 462         }
 463         pte_unmap(ptep-1);
 464 }
 465
 466 static void zap_pmd_range(struct mmu_gather *tlb,
 467                 pgd_t * dir, unsigned long address,
 468                 unsigned long size, struct zap_details *details)
 469 {
 470         pmd_t * pmd;
 471         unsigned long end;
 472
 473         if (pgd_none(*dir))
 474                 return;
 475         if (pgd_bad(*dir)) {
 476                 pgd_ERROR(*dir);
 477                 pgd_clear(dir);
 478                 return;
 479         }
 480         pmd = pmd_offset(dir, address);
 481         end = address + size;
 482         if (end > ((address + PGDIR_SIZE) & PGDIR_MASK))
 483                 end = ((address + PGDIR_SIZE) & PGDIR_MASK);
 484         do {
 485                 zap_pte_range(tlb, pmd, address, end - address, details);
 486                 address = (address + PMD_SIZE) & PMD_MASK;
 487                 pmd++;
 488         } while (address < end);
 489 }
 490
 491 static void unmap_page_range(struct mmu_gather *tlb,
 492                 struct vm_area_struct *vma, unsigned long address,
 493                 unsigned long end, struct zap_details *details)
 494 {
 495         pgd_t * dir;
 496
 497         BUG_ON(address >= end);
 498         dir = pgd_offset(vma->vm_mm, address);
 499         tlb_start_vma(tlb, vma);
 500         do {
 501                 zap_pmd_range(tlb, dir, address, end - address, details);
 502                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 503                 dir++;
 504         } while (address && (address < end));
 505         tlb_end_vma(tlb, vma);
 506 }
 507
 508 /* Dispose of an entire struct mmu_gather per rescheduling point */
 509 #if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 510 #define ZAP_BLOCK_SIZE  (FREE_PTE_NR * PAGE_SIZE)
 511 #endif
 512
 513 /* For UP, 256 pages at a time gives nice low latency */
 514 #if !defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 515 #define ZAP_BLOCK_SIZE  (256 * PAGE_SIZE)
 516 #endif
 517
 518 /* No preempt: go for improved straight-line efficiency */
 519 #if !defined(CONFIG_PREEMPT)
 520 #define ZAP_BLOCK_SIZE  (1024 * PAGE_SIZE)
 521 #endif
 522
 523 /**
 524  * unmap_vmas - unmap a range of memory covered by a list of vma's
 525  * @tlbp: address of the caller's struct mmu_gather
 526  * @mm: the controlling mm_struct
 527  * @vma: the starting vma
 528  * @start_addr: virtual address at which to start unmapping
 529  * @end_addr: virtual address at which to end unmapping
 530  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 531  * @details: details of nonlinear truncation or shared cache invalidation
 532  *
 533  * Returns the number of vma's which were covered by the unmapping.
 534  *
 535  * Unmap all pages in the vma list.  Called under page_table_lock.
 536  *
 537  * We aim to not hold page_table_lock for too long (for scheduling latency
 538  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 539  * return the ending mmu_gather to the caller.
 540  *
 541  * Only addresses between `start' and `end' will be unmapped.
 542  *
 543  * The VMA list must be sorted in ascending virtual address order.
 544  *
 545  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 546  * range after unmap_vmas() returns.  So the only responsibility here is to
 547  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 548  * drops the lock and schedules.
 549  */
 550 int unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
 551                 struct vm_area_struct *vma, unsigned long start_addr,
 552                 unsigned long end_addr, unsigned long *nr_accounted,
 553                 struct zap_details *details)
 554 {
 555         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
 556         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 557         int tlb_start_valid = 0;
 558         int ret = 0;
 559
 560         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 561                 unsigned long start;
 562                 unsigned long end;
 563
 564                 start = max(vma->vm_start, start_addr);
 565                 if (start >= vma->vm_end)
 566                         continue;
 567                 end = min(vma->vm_end, end_addr);
 568                 if (end <= vma->vm_start)
 569                         continue;
 570
 571                 if (vma->vm_flags & VM_ACCOUNT)
 572                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 573
 574                 ret++;
 575                 while (start != end) {
 576                         unsigned long block;
 577
 578                         if (!tlb_start_valid) {
 579                                 tlb_start = start;
 580                                 tlb_start_valid = 1;
 581                         }
 582
 583                         if (is_vm_hugetlb_page(vma)) {
 584                                 block = end - start;
 585                                 unmap_hugepage_range(vma, start, end);
 586                         } else {
 587                                 block = min(zap_bytes, end - start);
 588                                 unmap_page_range(*tlbp, vma, start,
 589                                                 start + block, details);
 590                         }
 591
 592                         start += block;
 593                         zap_bytes -= block;
 594                         if ((long)zap_bytes > 0)
 595                                 continue;
 596                         if (need_resched()) {
 597                                 int fullmm = tlb_is_full_mm(*tlbp);
 598                                 tlb_finish_mmu(*tlbp, tlb_start, start);
 599                                 cond_resched_lock(&mm->page_table_lock);
 600                                 *tlbp = tlb_gather_mmu(mm, fullmm);
 601                                 tlb_start_valid = 0;
 602                         }
 603                         zap_bytes = ZAP_BLOCK_SIZE;
 604                 }
 605         }
 606         return ret;
 607 }
 608
 609 /**
 610  * zap_page_range - remove user pages in a given range
 611  * @vma: vm_area_struct holding the applicable pages
 612  * @address: starting address of pages to zap
 613  * @size: number of bytes to zap
 614  * @details: details of nonlinear truncation or shared cache invalidation
 615  */
 616 void zap_page_range(struct vm_area_struct *vma, unsigned long address,
 617                 unsigned long size, struct zap_details *details)
 618 {
 619         struct mm_struct *mm = vma->vm_mm;
 620         struct mmu_gather *tlb;
 621         unsigned long end = address + size;
 622         unsigned long nr_accounted = 0;
 623
 624         might_sleep();
 625
 626         if (is_vm_hugetlb_page(vma)) {
 627                 zap_hugepage_range(vma, address, size);
 628                 return;
 629         }
 630
 631         lru_add_drain();
 632         spin_lock(&mm->page_table_lock);
 633         tlb = tlb_gather_mmu(mm, 0);
 634         unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
 635         tlb_finish_mmu(tlb, address, end);
 636         spin_unlock(&mm->page_table_lock);
 637 }
 638
 639 /*
 640  * Do a quick page-table lookup for a single page.
 641  * mm->page_table_lock must be held.
 642  */
 643 struct page *
 644 follow_page(struct mm_struct *mm, unsigned long address, int write)
 645 {
 646         pgd_t *pgd;
 647         pmd_t *pmd;
 648         pte_t *ptep, pte;
 649         unsigned long pfn;
 650         struct page *page;
 651
 652         page = follow_huge_addr(mm, address, write);
 653         if (! IS_ERR(page))
 654                 return page;
 655
 656         pgd = pgd_offset(mm, address);
 657         if (pgd_none(*pgd) || pgd_bad(*pgd))
 658                 goto out;
 659
 660         pmd = pmd_offset(pgd, address);
 661         if (pmd_none(*pmd))
 662                 goto out;
 663         if (pmd_huge(*pmd))
 664                 return follow_huge_pmd(mm, address, pmd, write);
 665         if (pmd_bad(*pmd))
 666                 goto out;
 667
 668         ptep = pte_offset_map(pmd, address);
 669         if (!ptep)
 670                 goto out;
 671
 672         pte = *ptep;
 673         pte_unmap(ptep);
 674         if (pte_present(pte)) {
 675                 if (write && !pte_write(pte))
 676                         goto out;
 677                 if (write && !pte_dirty(pte)) {
 678                         struct page *page = pte_page(pte);
 679                         if (!PageDirty(page))
 680                                 set_page_dirty(page);
 681                 }
 682                 pfn = pte_pfn(pte);
 683                 if (pfn_valid(pfn)) {
 684                         struct page *page = pfn_to_page(pfn);
 685
 686                         mark_page_accessed(page);
 687                         return page;
 688                 }
 689         }
 690
 691 out:
 692         return NULL;
 693 }
 694
 695 /*
 696  * Given a physical address, is there a useful struct page pointing to
 697  * it?  This may become more complex in the future if we start dealing
 698  * with IO-aperture pages for direct-IO.
 699  */
 700
 701 static inline struct page *get_page_map(struct page *page)
 702 {
 703         if (!pfn_valid(page_to_pfn(page)))
 704                 return 0;
 705         return page;
 706 }
 707
 708
 709 static inline int
 710 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
 711                          unsigned long address)
 712 {
 713         pgd_t *pgd;
 714         pmd_t *pmd;
 715
 716         /* Check if the vma is for an anonymous mapping. */
 717         if (vma->vm_ops && vma->vm_ops->nopage)
 718                 return 0;
 719
 720         /* Check if page directory entry exists. */
 721         pgd = pgd_offset(mm, address);
 722         if (pgd_none(*pgd) || pgd_bad(*pgd))
 723                 return 1;
 724
 725         /* Check if page middle directory entry exists. */
 726         pmd = pmd_offset(pgd, address);
 727         if (pmd_none(*pmd) || pmd_bad(*pmd))
 728                 return 1;
 729
 730         /* There is a pte slot for 'address' in 'mm'. */
 731         return 0;
 732 }
 733
 734
 735 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 736                 unsigned long start, int len, int write, int force,
 737                 struct page **pages, struct vm_area_struct **vmas)
 738 {
 739         int i;
 740         unsigned int flags;
 741
 742         /*
 743          * Require read or write permissions.
 744          * If 'force' is set, we only require the "MAY" flags.
 745          */
 746         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 747         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 748         i = 0;
 749
 750         do {
 751                 struct vm_area_struct * vma;
 752
 753                 vma = find_extend_vma(mm, start);
 754                 if (!vma && in_gate_area(tsk, start)) {
 755                         unsigned long pg = start & PAGE_MASK;
 756                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
 757                         pgd_t *pgd;
 758                         pmd_t *pmd;
 759                         pte_t *pte;
 760                         if (write) /* user gate pages are read-only */
 761                                 return i ? : -EFAULT;
 762                         pgd = pgd_offset_k(pg);
 763                         if (!pgd)
 764                                 return i ? : -EFAULT;
 765                         pmd = pmd_offset(pgd, pg);
 766                         if (!pmd)
 767                                 return i ? : -EFAULT;
 768                         pte = pte_offset_kernel(pmd, pg);
 769                         if (!pte || !pte_present(*pte))
 770                                 return i ? : -EFAULT;
 771                         if (pages) {
 772                                 pages[i] = pte_page(*pte);
 773                                 get_page(pages[i]);
 774                         }
 775                         if (vmas)
 776                                 vmas[i] = gate_vma;
 777                         i++;
 778                         start += PAGE_SIZE;
 779                         len--;
 780                         continue;
 781                 }
 782
 783                 if (!vma || (pages && (vma->vm_flags & VM_IO))
 784                                 || !(flags & vma->vm_flags))
 785                         return i ? : -EFAULT;
 786
 787                 if (is_vm_hugetlb_page(vma)) {
 788                         i = follow_hugetlb_page(mm, vma, pages, vmas,
 789                                                 &start, &len, i);
 790                         continue;
 791                 }
 792                 spin_lock(&mm->page_table_lock);
 793                 do {
 794                         struct page *map;
 795                         int lookup_write = write;
 796                         while (!(map = follow_page(mm, start, lookup_write))) {
 797                                 /*
 798                                  * Shortcut for anonymous pages. We don't want
 799                                  * to force the creation of pages tables for
 800                                  * insanly big anonymously mapped areas that
 801                                  * nobody touched so far. This is important
 802                                  * for doing a core dump for these mappings.
 803                                  */
 804                                 if (!lookup_write &&
 805                                     untouched_anonymous_page(mm,vma,start)) {
 806                                         map = ZERO_PAGE(start);
 807                                         break;
 808                                 }
 809                                 spin_unlock(&mm->page_table_lock);
 810                                 switch (handle_mm_fault(mm,vma,start,write)) {
 811                                 case VM_FAULT_MINOR:
 812                                         tsk->min_flt++;
 813                                         break;
 814                                 case VM_FAULT_MAJOR:
 815                                         tsk->maj_flt++;
 816                                         break;
 817                                 case VM_FAULT_SIGBUS:
 818                                         return i ? i : -EFAULT;
 819                                 case VM_FAULT_OOM:
 820                                         return i ? i : -ENOMEM;
 821                                 default:
 822                                         BUG();
 823                                 }
 824                                 /*
 825                                  * Now that we have performed a write fault
 826                                  * and surely no longer have a shared page we
 827                                  * shouldn't write, we shouldn't ignore an
 828                                  * unwritable page in the page table if
 829                                  * we are forcing write access.
 830                                  */
 831                                 lookup_write = write && !force;
 832                                 spin_lock(&mm->page_table_lock);
 833                         }
 834                         if (pages) {
 835                                 pages[i] = get_page_map(map);
 836                                 if (!pages[i]) {
 837                                         spin_unlock(&mm->page_table_lock);
 838                                         while (i--)
 839                                                 page_cache_release(pages[i]);
 840                                         i = -EFAULT;
 841                                         goto out;
 842                                 }
 843                                 flush_dcache_page(pages[i]);
 844                                 if (!PageReserved(pages[i]))
 845                                         page_cache_get(pages[i]);
 846                         }
 847                         if (vmas)
 848                                 vmas[i] = vma;
 849                         i++;
 850                         start += PAGE_SIZE;
 851                         len--;
 852                 } while(len && start < vma->vm_end);
 853                 spin_unlock(&mm->page_table_lock);
 854         } while(len);
 855 out:
 856         return i;
 857 }
 858
 859 EXPORT_SYMBOL(get_user_pages);
 860
 861 static void zeromap_pte_range(pte_t * pte, unsigned long address,
 862                                      unsigned long size, pgprot_t prot)
 863 {
 864         unsigned long end;
 865
 866         address &= ~PMD_MASK;
 867         end = address + size;
 868         if (end > PMD_SIZE)
 869                 end = PMD_SIZE;
 870         do {
 871                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
 872                 BUG_ON(!pte_none(*pte));
 873                 set_pte(pte, zero_pte);
 874                 address += PAGE_SIZE;
 875                 pte++;
 876         } while (address && (address < end));
 877 }
 878
 879 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
 880                                     unsigned long size, pgprot_t prot)
 881 {
 882         unsigned long base, end;
 883
 884         base = address & PGDIR_MASK;
 885         address &= ~PGDIR_MASK;
 886         end = address + size;
 887         if (end > PGDIR_SIZE)
 888                 end = PGDIR_SIZE;
 889         do {
 890                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
 891                 if (!pte)
 892                         return -ENOMEM;
 893                 zeromap_pte_range(pte, base + address, end - address, prot);
 894                 pte_unmap(pte);
 895                 address = (address + PMD_SIZE) & PMD_MASK;
 896                 pmd++;
 897         } while (address && (address < end));
 898         return 0;
 899 }
 900
 901 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address, unsigned long size, pgprot_t prot)
 902 {
 903         int error = 0;
 904         pgd_t * dir;
 905         unsigned long beg = address;
 906         unsigned long end = address + size;
 907         struct mm_struct *mm = vma->vm_mm;
 908
 909         dir = pgd_offset(mm, address);
 910         flush_cache_range(vma, beg, end);
 911         if (address >= end)
 912                 BUG();
 913
 914         spin_lock(&mm->page_table_lock);
 915         do {
 916                 pmd_t *pmd = pmd_alloc(mm, dir, address);
 917                 error = -ENOMEM;
 918                 if (!pmd)
 919                         break;
 920                 error = zeromap_pmd_range(mm, pmd, address, end - address, prot);
 921                 if (error)
 922                         break;
 923                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 924                 dir++;
 925         } while (address && (address < end));
 926         /*
 927          * Why flush? zeromap_pte_range has a BUG_ON for !pte_none()
 928          */
 929         flush_tlb_range(vma, beg, end);
 930         spin_unlock(&mm->page_table_lock);
 931         return error;
 932 }
 933
 934 /*
 935  * maps a range of physical memory into the requested pages. the old
 936  * mappings are removed. any references to nonexistent pages results
 937  * in null mappings (currently treated as "copy-on-access")
 938  */
 939 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
 940         unsigned long phys_addr, pgprot_t prot)
 941 {
 942         unsigned long end;
 943         unsigned long pfn;
 944
 945         address &= ~PMD_MASK;
 946         end = address + size;
 947         if (end > PMD_SIZE)
 948                 end = PMD_SIZE;
 949         pfn = phys_addr >> PAGE_SHIFT;
 950         do {
 951                 BUG_ON(!pte_none(*pte));
 952                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
 953                         set_pte(pte, pfn_pte(pfn, prot));
 954                 address += PAGE_SIZE;
 955                 pfn++;
 956                 pte++;
 957         } while (address && (address < end));
 958 }
 959
 960 static inline int remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address, unsigned long size,
 961         unsigned long phys_addr, pgprot_t prot)
 962 {
 963         unsigned long base, end;
 964
 965         base = address & PGDIR_MASK;
 966         address &= ~PGDIR_MASK;
 967         end = address + size;
 968         if (end > PGDIR_SIZE)
 969                 end = PGDIR_SIZE;
 970         phys_addr -= address;
 971         do {
 972                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
 973                 if (!pte)
 974                         return -ENOMEM;
 975                 remap_pte_range(pte, base + address, end - address, address + phys_addr, prot);
 976                 pte_unmap(pte);
 977                 address = (address + PMD_SIZE) & PMD_MASK;
 978                 pmd++;
 979         } while (address && (address < end));
 980         return 0;
 981 }
 982
 983 /*  Note: this is only safe if the mm semaphore is held when called. */
 984 int remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
 985 {
 986         int error = 0;
 987         pgd_t * dir;
 988         unsigned long beg = from;
 989         unsigned long end = from + size;
 990         struct mm_struct *mm = vma->vm_mm;
 991
 992         phys_addr -= from;
 993         dir = pgd_offset(mm, from);
 994         flush_cache_range(vma, beg, end);
 995         if (from >= end)
 996                 BUG();
 997
 998         spin_lock(&mm->page_table_lock);
 999         do {
1000                 pmd_t *pmd = pmd_alloc(mm, dir, from);
1001                 error = -ENOMEM;
1002                 if (!pmd)
1003                         break;
1004                 error = remap_pmd_range(mm, pmd, from, end - from, phys_addr + from, prot);
1005                 if (error)
1006                         break;
1007                 from = (from + PGDIR_SIZE) & PGDIR_MASK;
1008                 dir++;
1009         } while (from && (from < end));
1010         /*
1011          * Why flush? remap_pte_range has a BUG_ON for !pte_none()
1012          */
1013         flush_tlb_range(vma, beg, end);
1014         spin_unlock(&mm->page_table_lock);
1015         return error;
1016 }
1017
1018 EXPORT_SYMBOL(remap_page_range);
1019
1020 /*
1021  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1022  * servicing faults for write access.  In the normal case, do always want
1023  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1024  * that do not have writing enabled, when used by access_process_vm.
1025  */
1026 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1027 {
1028         if (likely(vma->vm_flags & VM_WRITE))
1029                 pte = pte_mkwrite(pte);
1030         return pte;
1031 }
1032
1033 /*
1034  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1035  */
1036 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
1037                 pte_t *page_table)
1038 {
1039         pte_t entry;
1040
1041         flush_cache_page(vma, address);
1042         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1043                               vma);
1044         ptep_establish(vma, address, page_table, entry);
1045         update_mmu_cache(vma, address, entry);
1046 }
1047
1048 /*
1049  * This routine handles present pages, when users try to write
1050  * to a shared page. It is done by copying the page to a new address
1051  * and decrementing the shared-page counter for the old page.
1052  *
1053  * Goto-purists beware: the only reason for goto's here is that it results
1054  * in better assembly code.. The "default" path will see no jumps at all.
1055  *
1056  * Note that this routine assumes that the protection checks have been
1057  * done by the caller (the low-level page fault routine in most cases).
1058  * Thus we can safely just mark it writable once we've done any necessary
1059  * COW.
1060  *
1061  * We also mark the page dirty at this point even though the page will
1062  * change only once the write actually happens. This avoids a few races,
1063  * and potentially makes it more efficient.
1064  *
1065  * We hold the mm semaphore and the page_table_lock on entry and exit
1066  * with the page_table_lock released.
1067  */
1068 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1069         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1070 {
1071         struct page *old_page, *new_page;
1072         unsigned long pfn = pte_pfn(pte);
1073         struct pte_chain *pte_chain;
1074         pte_t entry;
1075
1076         if (unlikely(!pfn_valid(pfn))) {
1077                 /*
1078                  * This should really halt the system so it can be debugged or
1079                  * at least the kernel stops what it's doing before it corrupts
1080                  * data, but for the moment just pretend this is OOM.
1081                  */
1082                 pte_unmap(page_table);
1083                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1084                                 address);
1085                 spin_unlock(&mm->page_table_lock);
1086                 return VM_FAULT_OOM;
1087         }
1088         old_page = pfn_to_page(pfn);
1089
1090         if (!TestSetPageLocked(old_page)) {
1091                 int reuse = can_share_swap_page(old_page);
1092                 unlock_page(old_page);
1093                 if (reuse) {
1094                         flush_cache_page(vma, address);
1095                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1096                                               vma);
1097                         ptep_establish(vma, address, page_table, entry);
1098                         update_mmu_cache(vma, address, entry);
1099                         pte_unmap(page_table);
1100                         spin_unlock(&mm->page_table_lock);
1101                         return VM_FAULT_MINOR;
1102                 }
1103         }
1104         pte_unmap(page_table);
1105
1106         /*
1107          * Ok, we need to copy. Oh, well..
1108          */
1109         page_cache_get(old_page);
1110         spin_unlock(&mm->page_table_lock);
1111
1112         pte_chain = pte_chain_alloc(GFP_KERNEL);
1113         if (!pte_chain)
1114                 goto no_pte_chain;
1115         new_page = alloc_page(GFP_HIGHUSER);
1116         if (!new_page)
1117                 goto no_new_page;
1118         copy_cow_page(old_page,new_page,address);
1119
1120         /*
1121          * Re-check the pte - we dropped the lock
1122          */
1123         spin_lock(&mm->page_table_lock);
1124         page_table = pte_offset_map(pmd, address);
1125         if (pte_same(*page_table, pte)) {
1126                 if (PageReserved(old_page))
1127                         ++mm->rss;
1128                 page_remove_rmap(old_page, page_table);
1129                 break_cow(vma, new_page, address, page_table);
1130                 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1131                 lru_cache_add_active(new_page);
1132
1133                 /* Free the old page.. */
1134                 new_page = old_page;
1135         }
1136         pte_unmap(page_table);
1137         page_cache_release(new_page);
1138         page_cache_release(old_page);
1139         spin_unlock(&mm->page_table_lock);
1140         pte_chain_free(pte_chain);
1141         return VM_FAULT_MINOR;
1142
1143 no_new_page:
1144         pte_chain_free(pte_chain);
1145 no_pte_chain:
1146         page_cache_release(old_page);
1147         return VM_FAULT_OOM;
1148 }
1149
1150 /*
1151  * Helper function for unmap_mapping_range().
1152  */
1153 static void unmap_mapping_range_list(struct list_head *head,
1154                                      struct zap_details *details)
1155 {
1156         struct vm_area_struct *vma;
1157         pgoff_t vba, vea, zba, zea;
1158
1159         list_for_each_entry(vma, head, shared) {
1160                 if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
1161                         details->nonlinear_vma = vma;
1162                         zap_page_range(vma, vma->vm_start,
1163                                 vma->vm_end - vma->vm_start, details);
1164                         details->nonlinear_vma = NULL;
1165                         continue;
1166                 }
1167                 vba = vma->vm_pgoff;
1168                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1169                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1170                 if (vba > details->last_index || vea < details->first_index)
1171                         continue;       /* Mapping disjoint from hole. */
1172                 zba = details->first_index;
1173                 if (zba < vba)
1174                         zba = vba;
1175                 zea = details->last_index;
1176                 if (zea > vea)
1177                         zea = vea;
1178                 zap_page_range(vma,
1179                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1180                         (zea - zba + 1) << PAGE_SHIFT,
1181                         details->check_mapping? details: NULL);
1182         }
1183 }
1184
1185 /**
1186  * unmap_mapping_range - unmap the portion of all mmaps
1187  * in the specified address_space corresponding to the specified
1188  * page range in the underlying file.
1189  * @address_space: the address space containing mmaps to be unmapped.
1190  * @holebegin: byte in first page to unmap, relative to the start of
1191  * the underlying file.  This will be rounded down to a PAGE_SIZE
1192  * boundary.  Note that this is different from vmtruncate(), which
1193  * must keep the partial page.  In contrast, we must get rid of
1194  * partial pages.
1195  * @holelen: size of prospective hole in bytes.  This will be rounded
1196  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1197  * end of the file.
1198  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1199  * but 0 when invalidating pagecache, don't throw away private data.
1200  */
1201 void unmap_mapping_range(struct address_space *mapping,
1202         loff_t const holebegin, loff_t const holelen, int even_cows)
1203 {
1204         struct zap_details details;
1205         pgoff_t hba = holebegin >> PAGE_SHIFT;
1206         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1207
1208         /* Check for overflow. */
1209         if (sizeof(holelen) > sizeof(hlen)) {
1210                 long long holeend =
1211                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1212                 if (holeend & ~(long long)ULONG_MAX)
1213                         hlen = ULONG_MAX - hba + 1;
1214         }
1215
1216         details.check_mapping = even_cows? NULL: mapping;
1217         details.nonlinear_vma = NULL;
1218         details.first_index = hba;
1219         details.last_index = hba + hlen - 1;
1220         if (details.last_index < details.first_index)
1221                 details.last_index = ULONG_MAX;
1222
1223         down(&mapping->i_shared_sem);
1224         /* Protect against page fault */
1225         atomic_inc(&mapping->truncate_count);
1226         if (unlikely(!list_empty(&mapping->i_mmap)))
1227                 unmap_mapping_range_list(&mapping->i_mmap, &details);
1228
1229         /* Don't waste time to check mapping on fully shared vmas */
1230         details.check_mapping = NULL;
1231
1232         if (unlikely(!list_empty(&mapping->i_mmap_shared)))
1233                 unmap_mapping_range_list(&mapping->i_mmap_shared, &details);
1234         up(&mapping->i_shared_sem);
1235 }
1236 EXPORT_SYMBOL(unmap_mapping_range);
1237
1238 /*
1239  * Handle all mappings that got truncated by a "truncate()"
1240  * system call.
1241  *
1242  * NOTE! We have to be ready to update the memory sharing
1243  * between the file and the memory map for a potential last
1244  * incomplete page.  Ugly, but necessary.
1245  */
1246 int vmtruncate(struct inode * inode, loff_t offset)
1247 {
1248         struct address_space *mapping = inode->i_mapping;
1249         unsigned long limit;
1250
1251         if (inode->i_size < offset)
1252                 goto do_expand;
1253         i_size_write(inode, offset);
1254         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1255         truncate_inode_pages(mapping, offset);
1256         goto out_truncate;
1257
1258 do_expand:
1259         limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1260         if (limit != RLIM_INFINITY && offset > limit)
1261                 goto out_sig;
1262         if (offset > inode->i_sb->s_maxbytes)
1263                 goto out;
1264         i_size_write(inode, offset);
1265
1266 out_truncate:
1267         if (inode->i_op && inode->i_op->truncate)
1268                 inode->i_op->truncate(inode);
1269         return 0;
1270 out_sig:
1271         send_sig(SIGXFSZ, current, 0);
1272 out:
1273         return -EFBIG;
1274 }
1275
1276 EXPORT_SYMBOL(vmtruncate);
1277
1278 /*
1279  * Primitive swap readahead code. We simply read an aligned block of
1280  * (1 << page_cluster) entries in the swap area. This method is chosen
1281  * because it doesn't cost us any seek time.  We also make sure to queue
1282  * the 'original' request together with the readahead ones...
1283  */
1284 void swapin_readahead(swp_entry_t entry)
1285 {
1286         int i, num;
1287         struct page *new_page;
1288         unsigned long offset;
1289
1290         /*
1291          * Get the number of handles we should do readahead io to.
1292          */
1293         num = valid_swaphandles(entry, &offset);
1294         for (i = 0; i < num; offset++, i++) {
1295                 /* Ok, do the async read-ahead now */
1296                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1297                                                 offset));
1298                 if (!new_page)
1299                         break;
1300                 page_cache_release(new_page);
1301         }
1302         lru_add_drain();        /* Push any new pages onto the LRU now */
1303 }
1304
1305 /*
1306  * We hold the mm semaphore and the page_table_lock on entry and
1307  * should release the pagetable lock on exit..
1308  */
1309 static int do_swap_page(struct mm_struct * mm,
1310         struct vm_area_struct * vma, unsigned long address,
1311         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1312 {
1313         struct page *page;
1314         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1315         pte_t pte;
1316         int ret = VM_FAULT_MINOR;
1317         struct pte_chain *pte_chain = NULL;
1318
1319         pte_unmap(page_table);
1320         spin_unlock(&mm->page_table_lock);
1321         page = lookup_swap_cache(entry);
1322         if (!page) {
1323                 swapin_readahead(entry);
1324                 page = read_swap_cache_async(entry);
1325                 if (!page) {
1326                         /*
1327                          * Back out if somebody else faulted in this pte while
1328                          * we released the page table lock.
1329                          */
1330                         spin_lock(&mm->page_table_lock);
1331                         page_table = pte_offset_map(pmd, address);
1332                         if (pte_same(*page_table, orig_pte))
1333                                 ret = VM_FAULT_OOM;
1334                         else
1335                                 ret = VM_FAULT_MINOR;
1336                         pte_unmap(page_table);
1337                         spin_unlock(&mm->page_table_lock);
1338                         goto out;
1339                 }
1340
1341                 /* Had to read the page from swap area: Major fault */
1342                 ret = VM_FAULT_MAJOR;
1343                 inc_page_state(pgmajfault);
1344         }
1345
1346         mark_page_accessed(page);
1347         pte_chain = pte_chain_alloc(GFP_KERNEL);
1348         if (!pte_chain) {
1349                 ret = VM_FAULT_OOM;
1350                 goto out;
1351         }
1352         lock_page(page);
1353
1354         /*
1355          * Back out if somebody else faulted in this pte while we
1356          * released the page table lock.
1357          */
1358         spin_lock(&mm->page_table_lock);
1359         page_table = pte_offset_map(pmd, address);
1360         if (!pte_same(*page_table, orig_pte)) {
1361                 pte_unmap(page_table);
1362                 spin_unlock(&mm->page_table_lock);
1363                 unlock_page(page);
1364                 page_cache_release(page);
1365                 ret = VM_FAULT_MINOR;
1366                 goto out;
1367         }
1368
1369         /* The page isn't present yet, go ahead with the fault. */
1370
1371         swap_free(entry);
1372         if (vm_swap_full())
1373                 remove_exclusive_swap_page(page);
1374
1375         mm->rss++;
1376         pte = mk_pte(page, vma->vm_page_prot);
1377         if (write_access && can_share_swap_page(page))
1378                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1379         unlock_page(page);
1380
1381         flush_icache_page(vma, page);
1382         set_pte(page_table, pte);
1383         pte_chain = page_add_rmap(page, page_table, pte_chain);
1384
1385         /* No need to invalidate - it was non-present before */
1386         update_mmu_cache(vma, address, pte);
1387         pte_unmap(page_table);
1388         spin_unlock(&mm->page_table_lock);
1389 out:
1390         pte_chain_free(pte_chain);
1391         return ret;
1392 }
1393
1394 /*
1395  * We are called with the MM semaphore and page_table_lock
1396  * spinlock held to protect against concurrent faults in
1397  * multithreaded programs.
1398  */
1399 static int
1400 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1401                 pte_t *page_table, pmd_t *pmd, int write_access,
1402                 unsigned long addr)
1403 {
1404         pte_t entry;
1405         struct page * page = ZERO_PAGE(addr);
1406         struct pte_chain *pte_chain;
1407         int ret;
1408
1409         pte_chain = pte_chain_alloc(GFP_ATOMIC | __GFP_NOWARN);
1410         if (!pte_chain) {
1411                 pte_unmap(page_table);
1412                 spin_unlock(&mm->page_table_lock);
1413                 pte_chain = pte_chain_alloc(GFP_KERNEL);
1414                 if (!pte_chain)
1415                         goto no_mem;
1416                 spin_lock(&mm->page_table_lock);
1417                 page_table = pte_offset_map(pmd, addr);
1418         }
1419
1420         /* Read-only mapping of ZERO_PAGE. */
1421         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1422
1423         /* ..except if it's a write access */
1424         if (write_access) {
1425                 /* Allocate our own private page. */
1426                 pte_unmap(page_table);
1427                 spin_unlock(&mm->page_table_lock);
1428
1429                 page = alloc_page(GFP_HIGHUSER);
1430                 if (!page)
1431                         goto no_mem;
1432                 clear_user_highpage(page, addr);
1433
1434                 spin_lock(&mm->page_table_lock);
1435                 page_table = pte_offset_map(pmd, addr);
1436
1437                 if (!pte_none(*page_table)) {
1438                         pte_unmap(page_table);
1439                         page_cache_release(page);
1440                         spin_unlock(&mm->page_table_lock);
1441                         ret = VM_FAULT_MINOR;
1442                         goto out;
1443                 }
1444                 mm->rss++;
1445                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1446                                                          vma->vm_page_prot)),
1447                                       vma);
1448                 lru_cache_add_active(page);
1449                 mark_page_accessed(page);
1450         }
1451
1452         set_pte(page_table, entry);
1453         /* ignores ZERO_PAGE */
1454         pte_chain = page_add_rmap(page, page_table, pte_chain);
1455         pte_unmap(page_table);
1456
1457         /* No need to invalidate - it was non-present before */
1458         update_mmu_cache(vma, addr, entry);
1459         spin_unlock(&mm->page_table_lock);
1460         ret = VM_FAULT_MINOR;
1461         goto out;
1462
1463 no_mem:
1464         ret = VM_FAULT_OOM;
1465 out:
1466         pte_chain_free(pte_chain);
1467         return ret;
1468 }
1469
1470 /*
1471  * do_no_page() tries to create a new page mapping. It aggressively
1472  * tries to share with existing pages, but makes a separate copy if
1473  * the "write_access" parameter is true in order to avoid the next
1474  * page fault.
1475  *
1476  * As this is called only for pages that do not currently exist, we
1477  * do not need to flush old virtual caches or the TLB.
1478  *
1479  * This is called with the MM semaphore held and the page table
1480  * spinlock held. Exit with the spinlock released.
1481  */
1482 static int
1483 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1484         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1485 {
1486         struct page * new_page;
1487         struct address_space *mapping = NULL;
1488         pte_t entry;
1489         struct pte_chain *pte_chain;
1490         int sequence = 0;
1491         int ret = VM_FAULT_MINOR;
1492
1493         if (!vma->vm_ops || !vma->vm_ops->nopage)
1494                 return do_anonymous_page(mm, vma, page_table,
1495                                         pmd, write_access, address);
1496         pte_unmap(page_table);
1497         spin_unlock(&mm->page_table_lock);
1498
1499         if (vma->vm_file) {
1500                 mapping = vma->vm_file->f_mapping;
1501                 sequence = atomic_read(&mapping->truncate_count);
1502         }
1503         smp_rmb();  /* Prevent CPU from reordering lock-free ->nopage() */
1504 retry:
1505         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1506
1507         /* no page was available -- either SIGBUS or OOM */
1508         if (new_page == NOPAGE_SIGBUS)
1509                 return VM_FAULT_SIGBUS;
1510         if (new_page == NOPAGE_OOM)
1511                 return VM_FAULT_OOM;
1512
1513         pte_chain = pte_chain_alloc(GFP_KERNEL);
1514         if (!pte_chain)
1515                 goto oom;
1516
1517         /*
1518          * Should we do an early C-O-W break?
1519          */
1520         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1521                 struct page * page = alloc_page(GFP_HIGHUSER);
1522                 if (!page)
1523                         goto oom;
1524                 copy_user_highpage(page, new_page, address);
1525                 page_cache_release(new_page);
1526                 lru_cache_add_active(page);
1527                 new_page = page;
1528         }
1529
1530         spin_lock(&mm->page_table_lock);
1531         /*
1532          * For a file-backed vma, someone could have truncated or otherwise
1533          * invalidated this page.  If unmap_mapping_range got called,
1534          * retry getting the page.
1535          */
1536         if (mapping &&
1537               (unlikely(sequence != atomic_read(&mapping->truncate_count)))) {
1538                 sequence = atomic_read(&mapping->truncate_count);
1539                 spin_unlock(&mm->page_table_lock);
1540                 page_cache_release(new_page);
1541                 pte_chain_free(pte_chain);
1542                 goto retry;
1543         }
1544         page_table = pte_offset_map(pmd, address);
1545
1546         /*
1547          * This silly early PAGE_DIRTY setting removes a race
1548          * due to the bad i386 page protection. But it's valid
1549          * for other architectures too.
1550          *
1551          * Note that if write_access is true, we either now have
1552          * an exclusive copy of the page, or this is a shared mapping,
1553          * so we can make it writable and dirty to avoid having to
1554          * handle that later.
1555          */
1556         /* Only go through if we didn't race with anybody else... */
1557         if (pte_none(*page_table)) {
1558                 if (!PageReserved(new_page))
1559                         ++mm->rss;
1560                 flush_icache_page(vma, new_page);
1561                 entry = mk_pte(new_page, vma->vm_page_prot);
1562                 if (write_access)
1563                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1564                 set_pte(page_table, entry);
1565                 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1566                 pte_unmap(page_table);
1567         } else {
1568                 /* One of our sibling threads was faster, back out. */
1569                 pte_unmap(page_table);
1570                 page_cache_release(new_page);
1571                 spin_unlock(&mm->page_table_lock);
1572                 goto out;
1573         }
1574
1575         /* no need to invalidate: a not-present page shouldn't be cached */
1576         update_mmu_cache(vma, address, entry);
1577         spin_unlock(&mm->page_table_lock);
1578         goto out;
1579 oom:
1580         page_cache_release(new_page);
1581         ret = VM_FAULT_OOM;
1582 out:
1583         pte_chain_free(pte_chain);
1584         return ret;
1585 }
1586
1587 /*
1588  * Fault of a previously existing named mapping. Repopulate the pte
1589  * from the encoded file_pte if possible. This enables swappable
1590  * nonlinear vmas.
1591  */
1592 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1593         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1594 {
1595         unsigned long pgoff;
1596         int err;
1597
1598         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1599         /*
1600          * Fall back to the linear mapping if the fs does not support
1601          * ->populate:
1602          */
1603         if (!vma->vm_ops || !vma->vm_ops->populate ||
1604                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1605                 pte_clear(pte);
1606                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1607         }
1608
1609         pgoff = pte_to_pgoff(*pte);
1610
1611         pte_unmap(pte);
1612         spin_unlock(&mm->page_table_lock);
1613
1614         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1615         if (err == -ENOMEM)
1616                 return VM_FAULT_OOM;
1617         if (err)
1618                 return VM_FAULT_SIGBUS;
1619         return VM_FAULT_MAJOR;
1620 }
1621
1622 /*
1623  * These routines also need to handle stuff like marking pages dirty
1624  * and/or accessed for architectures that don't do it in hardware (most
1625  * RISC architectures).  The early dirtying is also good on the i386.
1626  *
1627  * There is also a hook called "update_mmu_cache()" that architectures
1628  * with external mmu caches can use to update those (ie the Sparc or
1629  * PowerPC hashed page tables that act as extended TLBs).
1630  *
1631  * Note the "page_table_lock". It is to protect against kswapd removing
1632  * pages from under us. Note that kswapd only ever _removes_ pages, never
1633  * adds them. As such, once we have noticed that the page is not present,
1634  * we can drop the lock early.
1635  *
1636  * The adding of pages is protected by the MM semaphore (which we hold),
1637  * so we don't need to worry about a page being suddenly been added into
1638  * our VM.
1639  *
1640  * We enter with the pagetable spinlock held, we are supposed to
1641  * release it when done.
1642  */
1643 static inline int handle_pte_fault(struct mm_struct *mm,
1644         struct vm_area_struct * vma, unsigned long address,
1645         int write_access, pte_t *pte, pmd_t *pmd)
1646 {
1647         pte_t entry;
1648
1649         entry = *pte;
1650         if (!pte_present(entry)) {
1651                 /*
1652                  * If it truly wasn't present, we know that kswapd
1653                  * and the PTE updates will not touch it later. So
1654                  * drop the lock.
1655                  */
1656                 if (pte_none(entry))
1657                         return do_no_page(mm, vma, address, write_access, pte, pmd);
1658                 if (pte_file(entry))
1659                         return do_file_page(mm, vma, address, write_access, pte, pmd);
1660                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1661         }
1662
1663         if (write_access) {
1664                 if (!pte_write(entry))
1665                         return do_wp_page(mm, vma, address, pte, pmd, entry);
1666
1667                 entry = pte_mkdirty(entry);
1668         }
1669         entry = pte_mkyoung(entry);
1670         ptep_establish(vma, address, pte, entry);
1671         update_mmu_cache(vma, address, entry);
1672         pte_unmap(pte);
1673         spin_unlock(&mm->page_table_lock);
1674         return VM_FAULT_MINOR;
1675 }
1676
1677 /*
1678  * By the time we get here, we already hold the mm semaphore
1679  */
1680 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1681         unsigned long address, int write_access)
1682 {
1683         pgd_t *pgd;
1684         pmd_t *pmd;
1685
1686         __set_current_state(TASK_RUNNING);
1687         pgd = pgd_offset(mm, address);
1688
1689         inc_page_state(pgfault);
1690
1691         if (is_vm_hugetlb_page(vma))
1692                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1693
1694         /*
1695          * We need the page table lock to synchronize with kswapd
1696          * and the SMP-safe atomic PTE updates.
1697          */
1698         spin_lock(&mm->page_table_lock);
1699         pmd = pmd_alloc(mm, pgd, address);
1700
1701         if (pmd) {
1702                 pte_t * pte = pte_alloc_map(mm, pmd, address);
1703                 if (pte)
1704                         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1705         }
1706         spin_unlock(&mm->page_table_lock);
1707         return VM_FAULT_OOM;
1708 }
1709
1710 /*
1711  * Allocate page middle directory.
1712  *
1713  * We've already handled the fast-path in-line, and we own the
1714  * page table lock.
1715  *
1716  * On a two-level page table, this ends up actually being entirely
1717  * optimized away.
1718  */
1719 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1720 {
1721         pmd_t *new;
1722
1723         spin_unlock(&mm->page_table_lock);
1724         new = pmd_alloc_one(mm, address);
1725         spin_lock(&mm->page_table_lock);
1726         if (!new)
1727                 return NULL;
1728
1729         /*
1730          * Because we dropped the lock, we should re-check the
1731          * entry, as somebody else could have populated it..
1732          */
1733         if (pgd_present(*pgd)) {
1734                 pmd_free(new);
1735                 goto out;
1736         }
1737         pgd_populate(mm, pgd, new);
1738 out:
1739         return pmd_offset(pgd, address);
1740 }
1741
1742 int make_pages_present(unsigned long addr, unsigned long end)
1743 {
1744         int ret, len, write;
1745         struct vm_area_struct * vma;
1746
1747         vma = find_vma(current->mm, addr);
1748         write = (vma->vm_flags & VM_WRITE) != 0;
1749         if (addr >= end)
1750                 BUG();
1751         if (end > vma->vm_end)
1752                 BUG();
1753         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1754         ret = get_user_pages(current, current->mm, addr,
1755                         len, write, 0, NULL, NULL);
1756         if (ret < 0)
1757                 return ret;
1758         return ret == len ? 0 : -1;
1759 }
1760
1761 /*
1762  * Map a vmalloc()-space virtual address to the physical page.
1763  */
1764 struct page * vmalloc_to_page(void * vmalloc_addr)
1765 {
1766         unsigned long addr = (unsigned long) vmalloc_addr;
1767         struct page *page = NULL;
1768         pgd_t *pgd = pgd_offset_k(addr);
1769         pmd_t *pmd;
1770         pte_t *ptep, pte;
1771
1772         if (!pgd_none(*pgd)) {
1773                 pmd = pmd_offset(pgd, addr);
1774                 if (!pmd_none(*pmd)) {
1775                         preempt_disable();
1776                         ptep = pte_offset_map(pmd, addr);
1777                         pte = *ptep;
1778                         if (pte_present(pte))
1779                                 page = pte_page(pte);
1780                         pte_unmap(ptep);
1781                         preempt_enable();
1782                 }
1783         }
1784         return page;
1785 }
1786
1787 EXPORT_SYMBOL(vmalloc_to_page);
1788
1789 #if !defined(CONFIG_ARCH_GATE_AREA)
1790
1791 #if defined(AT_SYSINFO_EHDR)
1792 struct vm_area_struct gate_vma;
1793
1794 static int __init gate_vma_init(void)
1795 {
1796         gate_vma.vm_mm = NULL;
1797         gate_vma.vm_start = FIXADDR_USER_START;
1798         gate_vma.vm_end = FIXADDR_USER_END;
1799         gate_vma.vm_page_prot = PAGE_READONLY;
1800         gate_vma.vm_flags = 0;
1801         return 0;
1802 }
1803 __initcall(gate_vma_init);
1804 #endif
1805
1806 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1807 {
1808 #ifdef AT_SYSINFO_EHDR
1809         return &gate_vma;
1810 #else
1811         return 0;
1812 #endif
1813 }
1814
1815 int in_gate_area(struct task_struct *task, unsigned long addr)
1816 {
1817 #ifdef AT_SYSINFO_EHDR
1818         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
1819                 return 1;
1820 #endif
1821         return 0;
1822 }
1823
1824 #endif