mm/filemap.c

   1 /*
   2  *      linux/mm/filemap.c
   3  *
   4  * Copyright (C) 1994-1999  Linus Torvalds
   5  */
   6
   7 /*
   8  * This file handles the generic file mmap semantics used by
   9  * most "normal" filesystems (but you don't /have/ to use this:
  10  * the NFS filesystem used to do this differently, for example)
  11  */
  12 #include <linux/config.h>
  13 #include <linux/module.h>
  14 #include <linux/slab.h>
  15 #include <linux/compiler.h>
  16 #include <linux/fs.h>
  17 #include <linux/aio.h>
  18 #include <linux/kernel_stat.h>
  19 #include <linux/mm.h>
  20 #include <linux/swap.h>
  21 #include <linux/mman.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/file.h>
  24 #include <linux/uio.h>
  25 #include <linux/hash.h>
  26 #include <linux/writeback.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/security.h>
  30 #include <linux/syscalls.h>
  31 /*
  32  * This is needed for the following functions:
  33  *  - try_to_release_page
  34  *  - block_invalidatepage
  35  *  - generic_osync_inode
  36  *
  37  * FIXME: remove all knowledge of the buffer layer from the core VM
  38  */
  39 #include <linux/buffer_head.h> /* for generic_osync_inode */
  40
  41 #include <asm/uaccess.h>
  42 #include <asm/mman.h>
  43
  44 /*
  45  * Shared mappings implemented 30.11.1994. It's not fully working yet,
  46  * though.
  47  *
  48  * Shared mappings now work. 15.8.1995  Bruno.
  49  *
  50  * finished 'unifying' the page and buffer cache and SMP-threaded the
  51  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
  52  *
  53  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
  54  */
  55
  56 /*
  57  * Lock ordering:
  58  *
  59  *  ->i_mmap_lock               (vmtruncate)
  60  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
  61  *      ->swap_list_lock
  62  *        ->swap_device_lock    (exclusive_swap_page, others)
  63  *          ->mapping->tree_lock
  64  *
  65  *  ->i_sem
  66  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
  67  *
  68  *  ->mmap_sem
  69  *    ->i_mmap_lock
  70  *      ->page_table_lock       (various places, mainly in mmap.c)
  71  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
  72  *
  73  *  ->mmap_sem
  74  *    ->lock_page               (access_process_vm)
  75  *
  76  *  ->mmap_sem
  77  *    ->i_sem                   (msync)
  78  *
  79  *  ->i_sem
  80  *    ->i_alloc_sem             (various)
  81  *
  82  *  ->inode_lock
  83  *    ->sb_lock                 (fs/fs-writeback.c)
  84  *    ->mapping->tree_lock      (__sync_single_inode)
  85  *
  86  *  ->i_mmap_lock
  87  *    ->anon_vma.lock           (vma_adjust)
  88  *
  89  *  ->anon_vma.lock
  90  *    ->page_table_lock         (anon_vma_prepare and various)
  91  *
  92  *  ->page_table_lock
  93  *    ->swap_device_lock        (try_to_unmap_one)
  94  *    ->private_lock            (try_to_unmap_one)
  95  *    ->tree_lock               (try_to_unmap_one)
  96  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
  97  *    ->private_lock            (page_remove_rmap->set_page_dirty)
  98  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
  99  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
 100  *    ->inode_lock              (zap_pte_range->set_page_dirty)
 101  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
 102  *
 103  *  ->task->proc_lock
 104  *    ->dcache_lock             (proc_pid_lookup)
 105  */
 106
 107 /*
 108  * Remove a page from the page cache and free it. Caller has to make
 109  * sure the page is locked and that nobody else uses it - or that usage
 110  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
 111  */
 112 void __remove_from_page_cache(struct page *page)
 113 {
 114         struct address_space *mapping = page->mapping;
 115
 116         radix_tree_delete(&mapping->page_tree, page->index);
 117         page->mapping = NULL;
 118         mapping->nrpages--;
 119         pagecache_acct(-1);
 120 }
 121
 122 void remove_from_page_cache(struct page *page)
 123 {
 124         struct address_space *mapping = page->mapping;
 125
 126         if (unlikely(!PageLocked(page)))
 127                 PAGE_BUG(page);
 128
 129         spin_lock_irq(&mapping->tree_lock);
 130         __remove_from_page_cache(page);
 131         spin_unlock_irq(&mapping->tree_lock);
 132 }
 133
 134 static int sync_page(void *word)
 135 {
 136         struct address_space *mapping;
 137         struct page *page;
 138
 139         page = container_of((page_flags_t *)word, struct page, flags);
 140
 141         /*
 142          * FIXME, fercrissake.  What is this barrier here for?
 143          */
 144         smp_mb();
 145         mapping = page_mapping(page);
 146         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 147                 mapping->a_ops->sync_page(page);
 148         io_schedule();
 149         return 0;
 150 }
 151
 152 /**
 153  * filemap_fdatawrite_range - start writeback against all of a mapping's
 154  * dirty pages that lie within the byte offsets <start, end>
 155  * @mapping: address space structure to write
 156  * @start: offset in bytes where the range starts
 157  * @end : offset in bytes where the range ends
 158  *
 159  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 160  * opposed to a regular memory * cleansing writeback.  The difference between
 161  * these two operations is that if a dirty page/buffer is encountered, it must
 162  * be waited upon, and not just skipped over.
 163  */
 164 static int __filemap_fdatawrite_range(struct address_space *mapping,
 165         loff_t start, loff_t end, int sync_mode)
 166 {
 167         int ret;
 168         struct writeback_control wbc = {
 169                 .sync_mode = sync_mode,
 170                 .nr_to_write = mapping->nrpages * 2,
 171                 .start = start,
 172                 .end = end,
 173         };
 174
 175         if (mapping->backing_dev_info->memory_backed)
 176                 return 0;
 177
 178         ret = do_writepages(mapping, &wbc);
 179         return ret;
 180 }
 181
 182 static inline int __filemap_fdatawrite(struct address_space *mapping,
 183         int sync_mode)
 184 {
 185         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
 186 }
 187
 188 int filemap_fdatawrite(struct address_space *mapping)
 189 {
 190         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 191 }
 192 EXPORT_SYMBOL(filemap_fdatawrite);
 193
 194 static int filemap_fdatawrite_range(struct address_space *mapping,
 195         loff_t start, loff_t end)
 196 {
 197         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
 198 }
 199
 200 /*
 201  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 202  * purposes - I/O may not be started against all dirty pages.
 203  */
 204 int filemap_flush(struct address_space *mapping)
 205 {
 206         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 207 }
 208 EXPORT_SYMBOL(filemap_flush);
 209
 210 /*
 211  * Wait for writeback to complete against pages indexed by start->end
 212  * inclusive
 213  */
 214 static int wait_on_page_writeback_range(struct address_space *mapping,
 215                                 pgoff_t start, pgoff_t end)
 216 {
 217         struct pagevec pvec;
 218         int nr_pages;
 219         int ret = 0;
 220         pgoff_t index;
 221
 222         if (end < start)
 223                 return 0;
 224
 225         pagevec_init(&pvec, 0);
 226         index = start;
 227         while ((index <= end) &&
 228                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 229                         PAGECACHE_TAG_WRITEBACK,
 230                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 231                 unsigned i;
 232
 233                 for (i = 0; i < nr_pages; i++) {
 234                         struct page *page = pvec.pages[i];
 235
 236                         /* until radix tree lookup accepts end_index */
 237                         if (page->index > end)
 238                                 continue;
 239
 240                         wait_on_page_writeback(page);
 241                         if (PageError(page))
 242                                 ret = -EIO;
 243                 }
 244                 pagevec_release(&pvec);
 245                 cond_resched();
 246         }
 247
 248         /* Check for outstanding write errors */
 249         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 250                 ret = -ENOSPC;
 251         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 252                 ret = -EIO;
 253
 254         return ret;
 255 }
 256
 257 /*
 258  * Write and wait upon all the pages in the passed range.  This is a "data
 259  * integrity" operation.  It waits upon in-flight writeout before starting and
 260  * waiting upon new writeout.  If there was an IO error, return it.
 261  *
 262  * We need to re-take i_sem during the generic_osync_inode list walk because
 263  * it is otherwise livelockable.
 264  */
 265 int sync_page_range(struct inode *inode, struct address_space *mapping,
 266                         loff_t pos, size_t count)
 267 {
 268         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 269         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 270         int ret;
 271
 272         if (mapping->backing_dev_info->memory_backed || !count)
 273                 return 0;
 274         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 275         if (ret == 0) {
 276                 down(&inode->i_sem);
 277                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 278                 up(&inode->i_sem);
 279         }
 280         if (ret == 0)
 281                 ret = wait_on_page_writeback_range(mapping, start, end);
 282         return ret;
 283 }
 284 EXPORT_SYMBOL(sync_page_range);
 285
 286 /*
 287  * Note: Holding i_sem across sync_page_range_nolock is not a good idea
 288  * as it forces O_SYNC writers to different parts of the same file
 289  * to be serialised right until io completion.
 290  */
 291 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
 292                         loff_t pos, size_t count)
 293 {
 294         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 295         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 296         int ret;
 297
 298         if (mapping->backing_dev_info->memory_backed || !count)
 299                 return 0;
 300         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 301         if (ret == 0)
 302                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 303         if (ret == 0)
 304                 ret = wait_on_page_writeback_range(mapping, start, end);
 305         return ret;
 306 }
 307 EXPORT_SYMBOL(sync_page_range_nolock);
 308
 309 /**
 310  * filemap_fdatawait - walk the list of under-writeback pages of the given
 311  *     address space and wait for all of them.
 312  *
 313  * @mapping: address space structure to wait for
 314  */
 315 int filemap_fdatawait(struct address_space *mapping)
 316 {
 317         loff_t i_size = i_size_read(mapping->host);
 318
 319         if (i_size == 0)
 320                 return 0;
 321
 322         return wait_on_page_writeback_range(mapping, 0,
 323                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 324 }
 325 EXPORT_SYMBOL(filemap_fdatawait);
 326
 327 int filemap_write_and_wait(struct address_space *mapping)
 328 {
 329         int retval = 0;
 330
 331         if (mapping->nrpages) {
 332                 retval = filemap_fdatawrite(mapping);
 333                 if (retval == 0)
 334                         retval = filemap_fdatawait(mapping);
 335         }
 336         return retval;
 337 }
 338
 339 /*
 340  * This function is used to add newly allocated pagecache pages:
 341  * the page is new, so we can just run SetPageLocked() against it.
 342  * The other page state flags were set by rmqueue().
 343  *
 344  * This function does not add the page to the LRU.  The caller must do that.
 345  */
 346 int add_to_page_cache(struct page *page, struct address_space *mapping,
 347                 pgoff_t offset, int gfp_mask)
 348 {
 349         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 350
 351         if (error == 0) {
 352                 spin_lock_irq(&mapping->tree_lock);
 353                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 354                 if (!error) {
 355                         page_cache_get(page);
 356                         SetPageLocked(page);
 357                         page->mapping = mapping;
 358                         page->index = offset;
 359                         mapping->nrpages++;
 360                         pagecache_acct(1);
 361                 }
 362                 spin_unlock_irq(&mapping->tree_lock);
 363                 radix_tree_preload_end();
 364         }
 365         return error;
 366 }
 367
 368 EXPORT_SYMBOL(add_to_page_cache);
 369
 370 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 371                                 pgoff_t offset, int gfp_mask)
 372 {
 373         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 374         if (ret == 0)
 375                 lru_cache_add(page);
 376         return ret;
 377 }
 378
 379 /*
 380  * In order to wait for pages to become available there must be
 381  * waitqueues associated with pages. By using a hash table of
 382  * waitqueues where the bucket discipline is to maintain all
 383  * waiters on the same queue and wake all when any of the pages
 384  * become available, and for the woken contexts to check to be
 385  * sure the appropriate page became available, this saves space
 386  * at a cost of "thundering herd" phenomena during rare hash
 387  * collisions.
 388  */
 389 static wait_queue_head_t *page_waitqueue(struct page *page)
 390 {
 391         const struct zone *zone = page_zone(page);
 392
 393         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 394 }
 395
 396 static inline void wake_up_page(struct page *page, int bit)
 397 {
 398         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 399 }
 400
 401 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 402 {
 403         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 404
 405         if (test_bit(bit_nr, &page->flags))
 406                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 407                                                         TASK_UNINTERRUPTIBLE);
 408 }
 409 EXPORT_SYMBOL(wait_on_page_bit);
 410
 411 /**
 412  * unlock_page() - unlock a locked page
 413  *
 414  * @page: the page
 415  *
 416  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 417  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 418  * mechananism between PageLocked pages and PageWriteback pages is shared.
 419  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 420  *
 421  * The first mb is necessary to safely close the critical section opened by the
 422  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 423  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 424  * parallel wait_on_page_locked()).
 425  */
 426 void fastcall unlock_page(struct page *page)
 427 {
 428         smp_mb__before_clear_bit();
 429         if (!TestClearPageLocked(page))
 430                 BUG();
 431         smp_mb__after_clear_bit();
 432         wake_up_page(page, PG_locked);
 433 }
 434 EXPORT_SYMBOL(unlock_page);
 435
 436 /*
 437  * End writeback against a page.
 438  */
 439 void end_page_writeback(struct page *page)
 440 {
 441         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 442                 if (!test_clear_page_writeback(page))
 443                         BUG();
 444         }
 445         smp_mb__after_clear_bit();
 446         wake_up_page(page, PG_writeback);
 447 }
 448 EXPORT_SYMBOL(end_page_writeback);
 449
 450 /*
 451  * Get a lock on the page, assuming we need to sleep to get it.
 452  *
 453  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 454  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 455  * chances are that on the second loop, the block layer's plug list is empty,
 456  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 457  */
 458 void fastcall __lock_page(struct page *page)
 459 {
 460         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 461
 462         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 463                                                         TASK_UNINTERRUPTIBLE);
 464 }
 465 EXPORT_SYMBOL(__lock_page);
 466
 467 /*
 468  * a rather lightweight function, finding and getting a reference to a
 469  * hashed page atomically.
 470  */
 471 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 472 {
 473         struct page *page;
 474
 475         spin_lock_irq(&mapping->tree_lock);
 476         page = radix_tree_lookup(&mapping->page_tree, offset);
 477         if (page)
 478                 page_cache_get(page);
 479         spin_unlock_irq(&mapping->tree_lock);
 480         return page;
 481 }
 482
 483 EXPORT_SYMBOL(find_get_page);
 484
 485 /*
 486  * Same as above, but trylock it instead of incrementing the count.
 487  */
 488 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 489 {
 490         struct page *page;
 491
 492         spin_lock_irq(&mapping->tree_lock);
 493         page = radix_tree_lookup(&mapping->page_tree, offset);
 494         if (page && TestSetPageLocked(page))
 495                 page = NULL;
 496         spin_unlock_irq(&mapping->tree_lock);
 497         return page;
 498 }
 499
 500 EXPORT_SYMBOL(find_trylock_page);
 501
 502 /**
 503  * find_lock_page - locate, pin and lock a pagecache page
 504  *
 505  * @mapping - the address_space to search
 506  * @offset - the page index
 507  *
 508  * Locates the desired pagecache page, locks it, increments its reference
 509  * count and returns its address.
 510  *
 511  * Returns zero if the page was not present. find_lock_page() may sleep.
 512  */
 513 struct page *find_lock_page(struct address_space *mapping,
 514                                 unsigned long offset)
 515 {
 516         struct page *page;
 517
 518         spin_lock_irq(&mapping->tree_lock);
 519 repeat:
 520         page = radix_tree_lookup(&mapping->page_tree, offset);
 521         if (page) {
 522                 page_cache_get(page);
 523                 if (TestSetPageLocked(page)) {
 524                         spin_unlock_irq(&mapping->tree_lock);
 525                         lock_page(page);
 526                         spin_lock_irq(&mapping->tree_lock);
 527
 528                         /* Has the page been truncated while we slept? */
 529                         if (page->mapping != mapping || page->index != offset) {
 530                                 unlock_page(page);
 531                                 page_cache_release(page);
 532                                 goto repeat;
 533                         }
 534                 }
 535         }
 536         spin_unlock_irq(&mapping->tree_lock);
 537         return page;
 538 }
 539
 540 EXPORT_SYMBOL(find_lock_page);
 541
 542 /**
 543  * find_or_create_page - locate or add a pagecache page
 544  *
 545  * @mapping - the page's address_space
 546  * @index - the page's index into the mapping
 547  * @gfp_mask - page allocation mode
 548  *
 549  * Locates a page in the pagecache.  If the page is not present, a new page
 550  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 551  * LRU list.  The returned page is locked and has its reference count
 552  * incremented.
 553  *
 554  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 555  * allocation!
 556  *
 557  * find_or_create_page() returns the desired page's address, or zero on
 558  * memory exhaustion.
 559  */
 560 struct page *find_or_create_page(struct address_space *mapping,
 561                 unsigned long index, unsigned int gfp_mask)
 562 {
 563         struct page *page, *cached_page = NULL;
 564         int err;
 565 repeat:
 566         page = find_lock_page(mapping, index);
 567         if (!page) {
 568                 if (!cached_page) {
 569                         cached_page = alloc_page(gfp_mask);
 570                         if (!cached_page)
 571                                 return NULL;
 572                 }
 573                 err = add_to_page_cache_lru(cached_page, mapping,
 574                                         index, gfp_mask);
 575                 if (!err) {
 576                         page = cached_page;
 577                         cached_page = NULL;
 578                 } else if (err == -EEXIST)
 579                         goto repeat;
 580         }
 581         if (cached_page)
 582                 page_cache_release(cached_page);
 583         return page;
 584 }
 585
 586 EXPORT_SYMBOL(find_or_create_page);
 587
 588 /**
 589  * find_get_pages - gang pagecache lookup
 590  * @mapping:    The address_space to search
 591  * @start:      The starting page index
 592  * @nr_pages:   The maximum number of pages
 593  * @pages:      Where the resulting pages are placed
 594  *
 595  * find_get_pages() will search for and return a group of up to
 596  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 597  * find_get_pages() takes a reference against the returned pages.
 598  *
 599  * The search returns a group of mapping-contiguous pages with ascending
 600  * indexes.  There may be holes in the indices due to not-present pages.
 601  *
 602  * find_get_pages() returns the number of pages which were found.
 603  */
 604 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 605                             unsigned int nr_pages, struct page **pages)
 606 {
 607         unsigned int i;
 608         unsigned int ret;
 609
 610         spin_lock_irq(&mapping->tree_lock);
 611         ret = radix_tree_gang_lookup(&mapping->page_tree,
 612                                 (void **)pages, start, nr_pages);
 613         for (i = 0; i < ret; i++)
 614                 page_cache_get(pages[i]);
 615         spin_unlock_irq(&mapping->tree_lock);
 616         return ret;
 617 }
 618
 619 /*
 620  * Like find_get_pages, except we only return pages which are tagged with
 621  * `tag'.   We update *index to index the next page for the traversal.
 622  */
 623 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 624                         int tag, unsigned int nr_pages, struct page **pages)
 625 {
 626         unsigned int i;
 627         unsigned int ret;
 628
 629         spin_lock_irq(&mapping->tree_lock);
 630         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 631                                 (void **)pages, *index, nr_pages, tag);
 632         for (i = 0; i < ret; i++)
 633                 page_cache_get(pages[i]);
 634         if (ret)
 635                 *index = pages[ret - 1]->index + 1;
 636         spin_unlock_irq(&mapping->tree_lock);
 637         return ret;
 638 }
 639
 640 /*
 641  * Same as grab_cache_page, but do not wait if the page is unavailable.
 642  * This is intended for speculative data generators, where the data can
 643  * be regenerated if the page couldn't be grabbed.  This routine should
 644  * be safe to call while holding the lock for another page.
 645  *
 646  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 647  * and deadlock against the caller's locked page.
 648  */
 649 struct page *
 650 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 651 {
 652         struct page *page = find_get_page(mapping, index);
 653         int gfp_mask;
 654
 655         if (page) {
 656                 if (!TestSetPageLocked(page))
 657                         return page;
 658                 page_cache_release(page);
 659                 return NULL;
 660         }
 661         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 662         page = alloc_pages(gfp_mask, 0);
 663         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 664                 page_cache_release(page);
 665                 page = NULL;
 666         }
 667         return page;
 668 }
 669
 670 EXPORT_SYMBOL(grab_cache_page_nowait);
 671
 672 /*
 673  * This is a generic file read routine, and uses the
 674  * mapping->a_ops->readpage() function for the actual low-level
 675  * stuff.
 676  *
 677  * This is really ugly. But the goto's actually try to clarify some
 678  * of the logic when it comes to error handling etc.
 679  *
 680  * Note the struct file* is only passed for the use of readpage.  It may be
 681  * NULL.
 682  */
 683 void do_generic_mapping_read(struct address_space *mapping,
 684                              struct file_ra_state *_ra,
 685                              struct file *filp,
 686                              loff_t *ppos,
 687                              read_descriptor_t *desc,
 688                              read_actor_t actor)
 689 {
 690         struct inode *inode = mapping->host;
 691         unsigned long index;
 692         unsigned long end_index;
 693         unsigned long offset;
 694         unsigned long req_size;
 695         unsigned long next_index;
 696         unsigned long prev_index;
 697         loff_t isize;
 698         struct page *cached_page;
 699         int error;
 700         struct file_ra_state ra = *_ra;
 701
 702         cached_page = NULL;
 703         index = *ppos >> PAGE_CACHE_SHIFT;
 704         next_index = index;
 705         prev_index = ra.prev_page;
 706         req_size = (desc->count + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 707         offset = *ppos & ~PAGE_CACHE_MASK;
 708
 709         isize = i_size_read(inode);
 710         if (!isize)
 711                 goto out;
 712
 713         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 714         for (;;) {
 715                 struct page *page;
 716                 unsigned long ret_size, nr, ret;
 717
 718                 /* nr is the maximum number of bytes to copy from this page */
 719                 nr = PAGE_CACHE_SIZE;
 720                 if (index >= end_index) {
 721                         if (index > end_index)
 722                                 goto out;
 723                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 724                         if (nr <= offset) {
 725                                 goto out;
 726                         }
 727                 }
 728                 nr = nr - offset;
 729
 730                 cond_resched();
 731                 if (index == next_index && req_size) {
 732                         ret_size = page_cache_readahead(mapping, &ra,
 733                                         filp, index, req_size);
 734                         next_index += ret_size;
 735                         req_size -= ret_size;
 736                 }
 737
 738 find_page:
 739                 page = find_get_page(mapping, index);
 740                 if (unlikely(page == NULL)) {
 741                         handle_ra_miss(mapping, &ra, index);
 742                         goto no_cached_page;
 743                 }
 744                 if (!PageUptodate(page))
 745                         goto page_not_up_to_date;
 746 page_ok:
 747
 748                 /* If users can be writing to this page using arbitrary
 749                  * virtual addresses, take care about potential aliasing
 750                  * before reading the page on the kernel side.
 751                  */
 752                 if (mapping_writably_mapped(mapping))
 753                         flush_dcache_page(page);
 754
 755                 /*
 756                  * When (part of) the same page is read multiple times
 757                  * in succession, only mark it as accessed the first time.
 758                  */
 759                 if (prev_index != index)
 760                         mark_page_accessed(page);
 761                 prev_index = index;
 762
 763                 /*
 764                  * Ok, we have the page, and it's up-to-date, so
 765                  * now we can copy it to user space...
 766                  *
 767                  * The actor routine returns how many bytes were actually used..
 768                  * NOTE! This may not be the same as how much of a user buffer
 769                  * we filled up (we may be padding etc), so we can only update
 770                  * "pos" here (the actor routine has to update the user buffer
 771                  * pointers and the remaining count).
 772                  */
 773                 ret = actor(desc, page, offset, nr);
 774                 offset += ret;
 775                 index += offset >> PAGE_CACHE_SHIFT;
 776                 offset &= ~PAGE_CACHE_MASK;
 777
 778                 page_cache_release(page);
 779                 if (ret == nr && desc->count)
 780                         continue;
 781                 goto out;
 782
 783 page_not_up_to_date:
 784                 /* Get exclusive access to the page ... */
 785                 lock_page(page);
 786
 787                 /* Did it get unhashed before we got the lock? */
 788                 if (!page->mapping) {
 789                         unlock_page(page);
 790                         page_cache_release(page);
 791                         continue;
 792                 }
 793
 794                 /* Did somebody else fill it already? */
 795                 if (PageUptodate(page)) {
 796                         unlock_page(page);
 797                         goto page_ok;
 798                 }
 799
 800 readpage:
 801                 /* Start the actual read. The read will unlock the page. */
 802                 error = mapping->a_ops->readpage(filp, page);
 803
 804                 if (unlikely(error))
 805                         goto readpage_error;
 806
 807                 if (!PageUptodate(page)) {
 808                         lock_page(page);
 809                         if (!PageUptodate(page)) {
 810                                 if (page->mapping == NULL) {
 811                                         /*
 812                                          * invalidate_inode_pages got it
 813                                          */
 814                                         unlock_page(page);
 815                                         page_cache_release(page);
 816                                         goto find_page;
 817                                 }
 818                                 unlock_page(page);
 819                                 error = -EIO;
 820                                 goto readpage_error;
 821                         }
 822                         unlock_page(page);
 823                 }
 824
 825                 /*
 826                  * i_size must be checked after we have done ->readpage.
 827                  *
 828                  * Checking i_size after the readpage allows us to calculate
 829                  * the correct value for "nr", which means the zero-filled
 830                  * part of the page is not copied back to userspace (unless
 831                  * another truncate extends the file - this is desired though).
 832                  */
 833                 isize = i_size_read(inode);
 834                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 835                 if (unlikely(!isize || index > end_index)) {
 836                         page_cache_release(page);
 837                         goto out;
 838                 }
 839
 840                 /* nr is the maximum number of bytes to copy from this page */
 841                 nr = PAGE_CACHE_SIZE;
 842                 if (index == end_index) {
 843                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 844                         if (nr <= offset) {
 845                                 page_cache_release(page);
 846                                 goto out;
 847                         }
 848                 }
 849                 nr = nr - offset;
 850                 goto page_ok;
 851
 852 readpage_error:
 853                 /* UHHUH! A synchronous read error occurred. Report it */
 854                 desc->error = error;
 855                 page_cache_release(page);
 856                 goto out;
 857
 858 no_cached_page:
 859                 /*
 860                  * Ok, it wasn't cached, so we need to create a new
 861                  * page..
 862                  */
 863                 if (!cached_page) {
 864                         cached_page = page_cache_alloc_cold(mapping);
 865                         if (!cached_page) {
 866                                 desc->error = -ENOMEM;
 867                                 goto out;
 868                         }
 869                 }
 870                 error = add_to_page_cache_lru(cached_page, mapping,
 871                                                 index, GFP_KERNEL);
 872                 if (error) {
 873                         if (error == -EEXIST)
 874                                 goto find_page;
 875                         desc->error = error;
 876                         goto out;
 877                 }
 878                 page = cached_page;
 879                 cached_page = NULL;
 880                 goto readpage;
 881         }
 882
 883 out:
 884         *_ra = ra;
 885
 886         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 887         if (cached_page)
 888                 page_cache_release(cached_page);
 889         if (filp)
 890                 file_accessed(filp);
 891 }
 892
 893 EXPORT_SYMBOL(do_generic_mapping_read);
 894
 895 int file_read_actor(read_descriptor_t *desc, struct page *page,
 896                         unsigned long offset, unsigned long size)
 897 {
 898         char *kaddr;
 899         unsigned long left, count = desc->count;
 900
 901         if (size > count)
 902                 size = count;
 903
 904         /*
 905          * Faults on the destination of a read are common, so do it before
 906          * taking the kmap.
 907          */
 908         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 909                 kaddr = kmap_atomic(page, KM_USER0);
 910                 left = __copy_to_user_inatomic(desc->arg.buf,
 911                                                 kaddr + offset, size);
 912                 kunmap_atomic(kaddr, KM_USER0);
 913                 if (left == 0)
 914                         goto success;
 915         }
 916
 917         /* Do it the slow way */
 918         kaddr = kmap(page);
 919         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 920         kunmap(page);
 921
 922         if (left) {
 923                 size -= left;
 924                 desc->error = -EFAULT;
 925         }
 926 success:
 927         desc->count = count - size;
 928         desc->written += size;
 929         desc->arg.buf += size;
 930         return size;
 931 }
 932
 933 /*
 934  * This is the "read()" routine for all filesystems
 935  * that can use the page cache directly.
 936  */
 937 ssize_t
 938 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 939                 unsigned long nr_segs, loff_t *ppos)
 940 {
 941         struct file *filp = iocb->ki_filp;
 942         ssize_t retval;
 943         unsigned long seg;
 944         size_t count;
 945
 946         count = 0;
 947         for (seg = 0; seg < nr_segs; seg++) {
 948                 const struct iovec *iv = &iov[seg];
 949
 950                 /*
 951                  * If any segment has a negative length, or the cumulative
 952                  * length ever wraps negative then return -EINVAL.
 953                  */
 954                 count += iv->iov_len;
 955                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
 956                         return -EINVAL;
 957                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
 958                         continue;
 959                 if (seg == 0)
 960                         return -EFAULT;
 961                 nr_segs = seg;
 962                 count -= iv->iov_len;   /* This segment is no good */
 963                 break;
 964         }
 965
 966         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
 967         if (filp->f_flags & O_DIRECT) {
 968                 loff_t pos = *ppos, size;
 969                 struct address_space *mapping;
 970                 struct inode *inode;
 971
 972                 mapping = filp->f_mapping;
 973                 inode = mapping->host;
 974                 retval = 0;
 975                 if (!count)
 976                         goto out; /* skip atime */
 977                 size = i_size_read(inode);
 978                 if (pos < size) {
 979                         retval = generic_file_direct_IO(READ, iocb,
 980                                                 iov, pos, nr_segs);
 981                         if (retval >= 0 && !is_sync_kiocb(iocb))
 982                                 retval = -EIOCBQUEUED;
 983                         if (retval > 0)
 984                                 *ppos = pos + retval;
 985                 }
 986                 file_accessed(filp);
 987                 goto out;
 988         }
 989
 990         retval = 0;
 991         if (count) {
 992                 for (seg = 0; seg < nr_segs; seg++) {
 993                         read_descriptor_t desc;
 994
 995                         desc.written = 0;
 996                         desc.arg.buf = iov[seg].iov_base;
 997                         desc.count = iov[seg].iov_len;
 998                         if (desc.count == 0)
 999                                 continue;
1000                         desc.error = 0;
1001                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1002                         retval += desc.written;
1003                         if (!retval) {
1004                                 retval = desc.error;
1005                                 break;
1006                         }
1007                 }
1008         }
1009 out:
1010         return retval;
1011 }
1012
1013 EXPORT_SYMBOL(__generic_file_aio_read);
1014
1015 ssize_t
1016 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1017 {
1018         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1019
1020         BUG_ON(iocb->ki_pos != pos);
1021         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1022 }
1023
1024 EXPORT_SYMBOL(generic_file_aio_read);
1025
1026 ssize_t
1027 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1028 {
1029         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1030         struct kiocb kiocb;
1031         ssize_t ret;
1032
1033         init_sync_kiocb(&kiocb, filp);
1034         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1035         if (-EIOCBQUEUED == ret)
1036                 ret = wait_on_sync_kiocb(&kiocb);
1037         return ret;
1038 }
1039
1040 EXPORT_SYMBOL(generic_file_read);
1041
1042 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1043 {
1044         ssize_t written;
1045         unsigned long count = desc->count;
1046         struct file *file = desc->arg.data;
1047
1048         if (size > count)
1049                 size = count;
1050
1051         written = file->f_op->sendpage(file, page, offset,
1052                                        size, &file->f_pos, size<count);
1053         if (written < 0) {
1054                 desc->error = written;
1055                 written = 0;
1056         }
1057         desc->count = count - written;
1058         desc->written += written;
1059         return written;
1060 }
1061
1062 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1063                          size_t count, read_actor_t actor, void *target)
1064 {
1065         read_descriptor_t desc;
1066
1067         if (!count)
1068                 return 0;
1069
1070         desc.written = 0;
1071         desc.count = count;
1072         desc.arg.data = target;
1073         desc.error = 0;
1074
1075         do_generic_file_read(in_file, ppos, &desc, actor);
1076         if (desc.written)
1077                 return desc.written;
1078         return desc.error;
1079 }
1080
1081 EXPORT_SYMBOL(generic_file_sendfile);
1082
1083 static ssize_t
1084 do_readahead(struct address_space *mapping, struct file *filp,
1085              unsigned long index, unsigned long nr)
1086 {
1087         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1088                 return -EINVAL;
1089
1090         force_page_cache_readahead(mapping, filp, index,
1091                                         max_sane_readahead(nr));
1092         return 0;
1093 }
1094
1095 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1096 {
1097         ssize_t ret;
1098         struct file *file;
1099
1100         ret = -EBADF;
1101         file = fget(fd);
1102         if (file) {
1103                 if (file->f_mode & FMODE_READ) {
1104                         struct address_space *mapping = file->f_mapping;
1105                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1106                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1107                         unsigned long len = end - start + 1;
1108                         ret = do_readahead(mapping, file, start, len);
1109                 }
1110                 fput(file);
1111         }
1112         return ret;
1113 }
1114
1115 #ifdef CONFIG_MMU
1116 /*
1117  * This adds the requested page to the page cache if it isn't already there,
1118  * and schedules an I/O to read in its contents from disk.
1119  */
1120 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1121 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1122 {
1123         struct address_space *mapping = file->f_mapping;
1124         struct page *page;
1125         int error;
1126
1127         page = page_cache_alloc_cold(mapping);
1128         if (!page)
1129                 return -ENOMEM;
1130
1131         error = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1132         if (!error) {
1133                 error = mapping->a_ops->readpage(file, page);
1134                 page_cache_release(page);
1135                 return error;
1136         }
1137
1138         /*
1139          * We arrive here in the unlikely event that someone
1140          * raced with us and added our page to the cache first
1141          * or we are out of memory for radix-tree nodes.
1142          */
1143         page_cache_release(page);
1144         return error == -EEXIST ? 0 : error;
1145 }
1146
1147 #define MMAP_LOTSAMISS  (100)
1148
1149 /*
1150  * filemap_nopage() is invoked via the vma operations vector for a
1151  * mapped memory region to read in file data during a page fault.
1152  *
1153  * The goto's are kind of ugly, but this streamlines the normal case of having
1154  * it in the page cache, and handles the special cases reasonably without
1155  * having a lot of duplicated code.
1156  */
1157 struct page * filemap_nopage(struct vm_area_struct * area, unsigned long address, int *type)
1158 {
1159         int error;
1160         struct file *file = area->vm_file;
1161         struct address_space *mapping = file->f_mapping;
1162         struct file_ra_state *ra = &file->f_ra;
1163         struct inode *inode = mapping->host;
1164         struct page *page;
1165         unsigned long size, pgoff, endoff;
1166         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1167
1168         pgoff = ((address - area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1169         endoff = ((area->vm_end - area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1170
1171 retry_all:
1172         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1173         if (pgoff >= size)
1174                 goto outside_data_content;
1175
1176         /* If we don't want any read-ahead, don't bother */
1177         if (VM_RandomReadHint(area))
1178                 goto no_cached_page;
1179
1180         /*
1181          * The "size" of the file, as far as mmap is concerned, isn't bigger
1182          * than the mapping
1183          */
1184         if (size > endoff)
1185                 size = endoff;
1186
1187         /*
1188          * The readahead code wants to be told about each and every page
1189          * so it can build and shrink its windows appropriately
1190          *
1191          * For sequential accesses, we use the generic readahead logic.
1192          */
1193         if (VM_SequentialReadHint(area))
1194                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1195
1196         /*
1197          * Do we have something in the page cache already?
1198          */
1199 retry_find:
1200         page = find_get_page(mapping, pgoff);
1201         if (!page) {
1202                 unsigned long ra_pages;
1203
1204                 if (VM_SequentialReadHint(area)) {
1205                         handle_ra_miss(mapping, ra, pgoff);
1206                         goto no_cached_page;
1207                 }
1208                 ra->mmap_miss++;
1209
1210                 /*
1211                  * Do we miss much more than hit in this file? If so,
1212                  * stop bothering with read-ahead. It will only hurt.
1213                  */
1214                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1215                         goto no_cached_page;
1216
1217                 /*
1218                  * To keep the pgmajfault counter straight, we need to
1219                  * check did_readaround, as this is an inner loop.
1220                  */
1221                 if (!did_readaround) {
1222                         majmin = VM_FAULT_MAJOR;
1223                         inc_page_state(pgmajfault);
1224                 }
1225                 did_readaround = 1;
1226                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1227                 if (ra_pages) {
1228                         pgoff_t start = 0;
1229
1230                         if (pgoff > ra_pages / 2)
1231                                 start = pgoff - ra_pages / 2;
1232                         do_page_cache_readahead(mapping, file, start, ra_pages);
1233                 }
1234                 page = find_get_page(mapping, pgoff);
1235                 if (!page)
1236                         goto no_cached_page;
1237         }
1238
1239         if (!did_readaround)
1240                 ra->mmap_hit++;
1241
1242         /*
1243          * Ok, found a page in the page cache, now we need to check
1244          * that it's up-to-date.
1245          */
1246         if (!PageUptodate(page))
1247                 goto page_not_uptodate;
1248
1249 success:
1250         /*
1251          * Found the page and have a reference on it.
1252          */
1253         mark_page_accessed(page);
1254         if (type)
1255                 *type = majmin;
1256         return page;
1257
1258 outside_data_content:
1259         /*
1260          * An external ptracer can access pages that normally aren't
1261          * accessible..
1262          */
1263         if (area->vm_mm == current->mm)
1264                 return NULL;
1265         /* Fall through to the non-read-ahead case */
1266 no_cached_page:
1267         /*
1268          * We're only likely to ever get here if MADV_RANDOM is in
1269          * effect.
1270          */
1271         error = page_cache_read(file, pgoff);
1272         grab_swap_token();
1273
1274         /*
1275          * The page we want has now been added to the page cache.
1276          * In the unlikely event that someone removed it in the
1277          * meantime, we'll just come back here and read it again.
1278          */
1279         if (error >= 0)
1280                 goto retry_find;
1281
1282         /*
1283          * An error return from page_cache_read can result if the
1284          * system is low on memory, or a problem occurs while trying
1285          * to schedule I/O.
1286          */
1287         if (error == -ENOMEM)
1288                 return NOPAGE_OOM;
1289         return NULL;
1290
1291 page_not_uptodate:
1292         if (!did_readaround) {
1293                 majmin = VM_FAULT_MAJOR;
1294                 inc_page_state(pgmajfault);
1295         }
1296         lock_page(page);
1297
1298         /* Did it get unhashed while we waited for it? */
1299         if (!page->mapping) {
1300                 unlock_page(page);
1301                 page_cache_release(page);
1302                 goto retry_all;
1303         }
1304
1305         /* Did somebody else get it up-to-date? */
1306         if (PageUptodate(page)) {
1307                 unlock_page(page);
1308                 goto success;
1309         }
1310
1311         if (!mapping->a_ops->readpage(file, page)) {
1312                 wait_on_page_locked(page);
1313                 if (PageUptodate(page))
1314                         goto success;
1315         }
1316
1317         /*
1318          * Umm, take care of errors if the page isn't up-to-date.
1319          * Try to re-read it _once_. We do this synchronously,
1320          * because there really aren't any performance issues here
1321          * and we need to check for errors.
1322          */
1323         lock_page(page);
1324
1325         /* Somebody truncated the page on us? */
1326         if (!page->mapping) {
1327                 unlock_page(page);
1328                 page_cache_release(page);
1329                 goto retry_all;
1330         }
1331
1332         /* Somebody else successfully read it in? */
1333         if (PageUptodate(page)) {
1334                 unlock_page(page);
1335                 goto success;
1336         }
1337         ClearPageError(page);
1338         if (!mapping->a_ops->readpage(file, page)) {
1339                 wait_on_page_locked(page);
1340                 if (PageUptodate(page))
1341                         goto success;
1342         }
1343
1344         /*
1345          * Things didn't work out. Return zero to tell the
1346          * mm layer so, possibly freeing the page cache page first.
1347          */
1348         page_cache_release(page);
1349         return NULL;
1350 }
1351
1352 EXPORT_SYMBOL(filemap_nopage);
1353
1354 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1355                                         int nonblock)
1356 {
1357         struct address_space *mapping = file->f_mapping;
1358         struct page *page;
1359         int error;
1360
1361         /*
1362          * Do we have something in the page cache already?
1363          */
1364 retry_find:
1365         page = find_get_page(mapping, pgoff);
1366         if (!page) {
1367                 if (nonblock)
1368                         return NULL;
1369                 goto no_cached_page;
1370         }
1371
1372         /*
1373          * Ok, found a page in the page cache, now we need to check
1374          * that it's up-to-date.
1375          */
1376         if (!PageUptodate(page))
1377                 goto page_not_uptodate;
1378
1379 success:
1380         /*
1381          * Found the page and have a reference on it.
1382          */
1383         mark_page_accessed(page);
1384         return page;
1385
1386 no_cached_page:
1387         error = page_cache_read(file, pgoff);
1388
1389         /*
1390          * The page we want has now been added to the page cache.
1391          * In the unlikely event that someone removed it in the
1392          * meantime, we'll just come back here and read it again.
1393          */
1394         if (error >= 0)
1395                 goto retry_find;
1396
1397         /*
1398          * An error return from page_cache_read can result if the
1399          * system is low on memory, or a problem occurs while trying
1400          * to schedule I/O.
1401          */
1402         return NULL;
1403
1404 page_not_uptodate:
1405         lock_page(page);
1406
1407         /* Did it get unhashed while we waited for it? */
1408         if (!page->mapping) {
1409                 unlock_page(page);
1410                 goto err;
1411         }
1412
1413         /* Did somebody else get it up-to-date? */
1414         if (PageUptodate(page)) {
1415                 unlock_page(page);
1416                 goto success;
1417         }
1418
1419         if (!mapping->a_ops->readpage(file, page)) {
1420                 wait_on_page_locked(page);
1421                 if (PageUptodate(page))
1422                         goto success;
1423         }
1424
1425         /*
1426          * Umm, take care of errors if the page isn't up-to-date.
1427          * Try to re-read it _once_. We do this synchronously,
1428          * because there really aren't any performance issues here
1429          * and we need to check for errors.
1430          */
1431         lock_page(page);
1432
1433         /* Somebody truncated the page on us? */
1434         if (!page->mapping) {
1435                 unlock_page(page);
1436                 goto err;
1437         }
1438         /* Somebody else successfully read it in? */
1439         if (PageUptodate(page)) {
1440                 unlock_page(page);
1441                 goto success;
1442         }
1443
1444         ClearPageError(page);
1445         if (!mapping->a_ops->readpage(file, page)) {
1446                 wait_on_page_locked(page);
1447                 if (PageUptodate(page))
1448                         goto success;
1449         }
1450
1451         /*
1452          * Things didn't work out. Return zero to tell the
1453          * mm layer so, possibly freeing the page cache page first.
1454          */
1455 err:
1456         page_cache_release(page);
1457
1458         return NULL;
1459 }
1460
1461 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1462                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1463                 int nonblock)
1464 {
1465         struct file *file = vma->vm_file;
1466         struct address_space *mapping = file->f_mapping;
1467         struct inode *inode = mapping->host;
1468         unsigned long size;
1469         struct mm_struct *mm = vma->vm_mm;
1470         struct page *page;
1471         int err;
1472
1473         if (!nonblock)
1474                 force_page_cache_readahead(mapping, vma->vm_file,
1475                                         pgoff, len >> PAGE_CACHE_SHIFT);
1476
1477 repeat:
1478         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1479         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1480                 return -EINVAL;
1481
1482         page = filemap_getpage(file, pgoff, nonblock);
1483         if (!page && !nonblock)
1484                 return -ENOMEM;
1485         if (page) {
1486                 err = install_page(mm, vma, addr, page, prot);
1487                 if (err) {
1488                         page_cache_release(page);
1489                         return err;
1490                 }
1491         } else {
1492                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1493                 if (err)
1494                         return err;
1495         }
1496
1497         len -= PAGE_SIZE;
1498         addr += PAGE_SIZE;
1499         pgoff++;
1500         if (len)
1501                 goto repeat;
1502
1503         return 0;
1504 }
1505
1506 struct vm_operations_struct generic_file_vm_ops = {
1507         .nopage         = filemap_nopage,
1508         .populate       = filemap_populate,
1509 };
1510
1511 /* This is used for a general mmap of a disk file */
1512
1513 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1514 {
1515         struct address_space *mapping = file->f_mapping;
1516
1517         if (!mapping->a_ops->readpage)
1518                 return -ENOEXEC;
1519         file_accessed(file);
1520         vma->vm_ops = &generic_file_vm_ops;
1521         return 0;
1522 }
1523 EXPORT_SYMBOL(filemap_populate);
1524
1525 /*
1526  * This is for filesystems which do not implement ->writepage.
1527  */
1528 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1529 {
1530         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1531                 return -EINVAL;
1532         return generic_file_mmap(file, vma);
1533 }
1534 #else
1535 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1536 {
1537         return -ENOSYS;
1538 }
1539 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1540 {
1541         return -ENOSYS;
1542 }
1543 #endif /* CONFIG_MMU */
1544
1545 EXPORT_SYMBOL(generic_file_mmap);
1546 EXPORT_SYMBOL(generic_file_readonly_mmap);
1547
1548 static inline struct page *__read_cache_page(struct address_space *mapping,
1549                                 unsigned long index,
1550                                 int (*filler)(void *,struct page*),
1551                                 void *data)
1552 {
1553         struct page *page, *cached_page = NULL;
1554         int err;
1555 repeat:
1556         page = find_get_page(mapping, index);
1557         if (!page) {
1558                 if (!cached_page) {
1559                         cached_page = page_cache_alloc_cold(mapping);
1560                         if (!cached_page)
1561                                 return ERR_PTR(-ENOMEM);
1562                 }
1563                 err = add_to_page_cache_lru(cached_page, mapping,
1564                                         index, GFP_KERNEL);
1565                 if (err == -EEXIST)
1566                         goto repeat;
1567                 if (err < 0) {
1568                         /* Presumably ENOMEM for radix tree node */
1569                         page_cache_release(cached_page);
1570                         return ERR_PTR(err);
1571                 }
1572                 page = cached_page;
1573                 cached_page = NULL;
1574                 err = filler(data, page);
1575                 if (err < 0) {
1576                         page_cache_release(page);
1577                         page = ERR_PTR(err);
1578                 }
1579         }
1580         if (cached_page)
1581                 page_cache_release(cached_page);
1582         return page;
1583 }
1584
1585 /*
1586  * Read into the page cache. If a page already exists,
1587  * and PageUptodate() is not set, try to fill the page.
1588  */
1589 struct page *read_cache_page(struct address_space *mapping,
1590                                 unsigned long index,
1591                                 int (*filler)(void *,struct page*),
1592                                 void *data)
1593 {
1594         struct page *page;
1595         int err;
1596
1597 retry:
1598         page = __read_cache_page(mapping, index, filler, data);
1599         if (IS_ERR(page))
1600                 goto out;
1601         mark_page_accessed(page);
1602         if (PageUptodate(page))
1603                 goto out;
1604
1605         lock_page(page);
1606         if (!page->mapping) {
1607                 unlock_page(page);
1608                 page_cache_release(page);
1609                 goto retry;
1610         }
1611         if (PageUptodate(page)) {
1612                 unlock_page(page);
1613                 goto out;
1614         }
1615         err = filler(data, page);
1616         if (err < 0) {
1617                 page_cache_release(page);
1618                 page = ERR_PTR(err);
1619         }
1620  out:
1621         return page;
1622 }
1623
1624 EXPORT_SYMBOL(read_cache_page);
1625
1626 /*
1627  * If the page was newly created, increment its refcount and add it to the
1628  * caller's lru-buffering pagevec.  This function is specifically for
1629  * generic_file_write().
1630  */
1631 static inline struct page *
1632 __grab_cache_page(struct address_space *mapping, unsigned long index,
1633                         struct page **cached_page, struct pagevec *lru_pvec)
1634 {
1635         int err;
1636         struct page *page;
1637 repeat:
1638         page = find_lock_page(mapping, index);
1639         if (!page) {
1640                 if (!*cached_page) {
1641                         *cached_page = page_cache_alloc(mapping);
1642                         if (!*cached_page)
1643                                 return NULL;
1644                 }
1645                 err = add_to_page_cache(*cached_page, mapping,
1646                                         index, GFP_KERNEL);
1647                 if (err == -EEXIST)
1648                         goto repeat;
1649                 if (err == 0) {
1650                         page = *cached_page;
1651                         page_cache_get(page);
1652                         if (!pagevec_add(lru_pvec, page))
1653                                 __pagevec_lru_add(lru_pvec);
1654                         *cached_page = NULL;
1655                 }
1656         }
1657         return page;
1658 }
1659
1660 /*
1661  * The logic we want is
1662  *
1663  *      if suid or (sgid and xgrp)
1664  *              remove privs
1665  */
1666 int remove_suid(struct dentry *dentry)
1667 {
1668         mode_t mode = dentry->d_inode->i_mode;
1669         int kill = 0;
1670         int result = 0;
1671
1672         /* suid always must be killed */
1673         if (unlikely(mode & S_ISUID))
1674                 kill = ATTR_KILL_SUID;
1675
1676         /*
1677          * sgid without any exec bits is just a mandatory locking mark; leave
1678          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1679          */
1680         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1681                 kill |= ATTR_KILL_SGID;
1682
1683         if (unlikely(kill && !capable(CAP_FSETID))) {
1684                 struct iattr newattrs;
1685
1686                 newattrs.ia_valid = ATTR_FORCE | kill;
1687                 result = notify_change(dentry, &newattrs);
1688         }
1689         return result;
1690 }
1691 EXPORT_SYMBOL(remove_suid);
1692
1693 /*
1694  * Copy as much as we can into the page and return the number of bytes which
1695  * were sucessfully copied.  If a fault is encountered then clear the page
1696  * out to (offset+bytes) and return the number of bytes which were copied.
1697  */
1698 static inline size_t
1699 filemap_copy_from_user(struct page *page, unsigned long offset,
1700                         const char __user *buf, unsigned bytes)
1701 {
1702         char *kaddr;
1703         int left;
1704
1705         kaddr = kmap_atomic(page, KM_USER0);
1706         left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1707         kunmap_atomic(kaddr, KM_USER0);
1708
1709         if (left != 0) {
1710                 /* Do it the slow way */
1711                 kaddr = kmap(page);
1712                 left = __copy_from_user(kaddr + offset, buf, bytes);
1713                 kunmap(page);
1714         }
1715         return bytes - left;
1716 }
1717
1718 static size_t
1719 __filemap_copy_from_user_iovec(char *vaddr,
1720                         const struct iovec *iov, size_t base, size_t bytes)
1721 {
1722         size_t copied = 0, left = 0;
1723
1724         while (bytes) {
1725                 char __user *buf = iov->iov_base + base;
1726                 int copy = min(bytes, iov->iov_len - base);
1727
1728                 base = 0;
1729                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1730                 copied += copy;
1731                 bytes -= copy;
1732                 vaddr += copy;
1733                 iov++;
1734
1735                 if (unlikely(left)) {
1736                         /* zero the rest of the target like __copy_from_user */
1737                         if (bytes)
1738                                 memset(vaddr, 0, bytes);
1739                         break;
1740                 }
1741         }
1742         return copied - left;
1743 }
1744
1745 /*
1746  * This has the same sideeffects and return value as filemap_copy_from_user().
1747  * The difference is that on a fault we need to memset the remainder of the
1748  * page (out to offset+bytes), to emulate filemap_copy_from_user()'s
1749  * single-segment behaviour.
1750  */
1751 static inline size_t
1752 filemap_copy_from_user_iovec(struct page *page, unsigned long offset,
1753                         const struct iovec *iov, size_t base, size_t bytes)
1754 {
1755         char *kaddr;
1756         size_t copied;
1757
1758         kaddr = kmap_atomic(page, KM_USER0);
1759         copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1760                                                 base, bytes);
1761         kunmap_atomic(kaddr, KM_USER0);
1762         if (copied != bytes) {
1763                 kaddr = kmap(page);
1764                 copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1765                                                         base, bytes);
1766                 kunmap(page);
1767         }
1768         return copied;
1769 }
1770
1771 static inline void
1772 filemap_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1773 {
1774         const struct iovec *iov = *iovp;
1775         size_t base = *basep;
1776
1777         while (bytes) {
1778                 int copy = min(bytes, iov->iov_len - base);
1779
1780                 bytes -= copy;
1781                 base += copy;
1782                 if (iov->iov_len == base) {
1783                         iov++;
1784                         base = 0;
1785                 }
1786         }
1787         *iovp = iov;
1788         *basep = base;
1789 }
1790
1791 /*
1792  * Performs necessary checks before doing a write
1793  *
1794  * Can adjust writing position aor amount of bytes to write.
1795  * Returns appropriate error code that caller should return or
1796  * zero in case that write should be allowed.
1797  */
1798 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1799 {
1800         struct inode *inode = file->f_mapping->host;
1801         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1802
1803         if (unlikely(*pos < 0))
1804                 return -EINVAL;
1805
1806         if (unlikely(file->f_error)) {
1807                 int err = file->f_error;
1808                 file->f_error = 0;
1809                 return err;
1810         }
1811
1812         if (!isblk) {
1813                 /* FIXME: this is for backwards compatibility with 2.4 */
1814                 if (file->f_flags & O_APPEND)
1815                         *pos = i_size_read(inode);
1816
1817                 if (limit != RLIM_INFINITY) {
1818                         if (*pos >= limit) {
1819                                 send_sig(SIGXFSZ, current, 0);
1820                                 return -EFBIG;
1821                         }
1822                         if (*count > limit - (typeof(limit))*pos) {
1823                                 *count = limit - (typeof(limit))*pos;
1824                         }
1825                 }
1826         }
1827
1828         /*
1829          * LFS rule
1830          */
1831         if (unlikely(*pos + *count > MAX_NON_LFS &&
1832                                 !(file->f_flags & O_LARGEFILE))) {
1833                 if (*pos >= MAX_NON_LFS) {
1834                         send_sig(SIGXFSZ, current, 0);
1835                         return -EFBIG;
1836                 }
1837                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1838                         *count = MAX_NON_LFS - (unsigned long)*pos;
1839                 }
1840         }
1841
1842         /*
1843          * Are we about to exceed the fs block limit ?
1844          *
1845          * If we have written data it becomes a short write.  If we have
1846          * exceeded without writing data we send a signal and return EFBIG.
1847          * Linus frestrict idea will clean these up nicely..
1848          */
1849         if (likely(!isblk)) {
1850                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1851                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1852                                 send_sig(SIGXFSZ, current, 0);
1853                                 return -EFBIG;
1854                         }
1855                         /* zero-length writes at ->s_maxbytes are OK */
1856                 }
1857
1858                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1859                         *count = inode->i_sb->s_maxbytes - *pos;
1860         } else {
1861                 loff_t isize;
1862                 if (bdev_read_only(I_BDEV(inode)))
1863                         return -EPERM;
1864                 isize = i_size_read(inode);
1865                 if (*pos >= isize) {
1866                         if (*count || *pos > isize)
1867                                 return -ENOSPC;
1868                 }
1869
1870                 if (*pos + *count > isize)
1871                         *count = isize - *pos;
1872         }
1873         return 0;
1874 }
1875 EXPORT_SYMBOL(generic_write_checks);
1876
1877 ssize_t
1878 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1879                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1880                 size_t count, size_t ocount)
1881 {
1882         struct file     *file = iocb->ki_filp;
1883         struct address_space *mapping = file->f_mapping;
1884         struct inode    *inode = mapping->host;
1885         ssize_t         written;
1886
1887         if (count != ocount)
1888                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1889
1890         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1891         if (written > 0) {
1892                 loff_t end = pos + written;
1893                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1894                         i_size_write(inode,  end);
1895                         mark_inode_dirty(inode);
1896                 }
1897                 *ppos = end;
1898         }
1899
1900         /*
1901          * Sync the fs metadata but not the minor inode changes and
1902          * of course not the data as we did direct DMA for the IO.
1903          * i_sem is held, which protects generic_osync_inode() from
1904          * livelocking.
1905          */
1906         if (written >= 0 && file->f_flags & O_SYNC)
1907                 generic_osync_inode(inode, mapping, OSYNC_METADATA);
1908         if (written == count && !is_sync_kiocb(iocb))
1909                 written = -EIOCBQUEUED;
1910         return written;
1911 }
1912 EXPORT_SYMBOL(generic_file_direct_write);
1913
1914 ssize_t
1915 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1916                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1917                 size_t count, ssize_t written)
1918 {
1919         struct file *file = iocb->ki_filp;
1920         struct address_space * mapping = file->f_mapping;
1921         struct address_space_operations *a_ops = mapping->a_ops;
1922         struct inode    *inode = mapping->host;
1923         long            status = 0;
1924         struct page     *page;
1925         struct page     *cached_page = NULL;
1926         size_t          bytes;
1927         struct pagevec  lru_pvec;
1928         const struct iovec *cur_iov = iov; /* current iovec */
1929         size_t          iov_base = 0;      /* offset in the current iovec */
1930         char __user     *buf;
1931
1932         pagevec_init(&lru_pvec, 0);
1933
1934         /*
1935          * handle partial DIO write.  Adjust cur_iov if needed.
1936          */
1937         if (likely(nr_segs == 1))
1938                 buf = iov->iov_base + written;
1939         else {
1940                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1941                 buf = iov->iov_base + iov_base;
1942         }
1943
1944         do {
1945                 unsigned long index;
1946                 unsigned long offset;
1947                 size_t copied;
1948
1949                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1950                 index = pos >> PAGE_CACHE_SHIFT;
1951                 bytes = PAGE_CACHE_SIZE - offset;
1952                 if (bytes > count)
1953                         bytes = count;
1954
1955                 /*
1956                  * Bring in the user page that we will copy from _first_.
1957                  * Otherwise there's a nasty deadlock on copying from the
1958                  * same page as we're writing to, without it being marked
1959                  * up-to-date.
1960                  */
1961                 fault_in_pages_readable(buf, bytes);
1962
1963                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1964                 if (!page) {
1965                         status = -ENOMEM;
1966                         break;
1967                 }
1968
1969                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1970                 if (unlikely(status)) {
1971                         loff_t isize = i_size_read(inode);
1972                         /*
1973                          * prepare_write() may have instantiated a few blocks
1974                          * outside i_size.  Trim these off again.
1975                          */
1976                         unlock_page(page);
1977                         page_cache_release(page);
1978                         if (pos + bytes > isize)
1979                                 vmtruncate(inode, isize);
1980                         break;
1981                 }
1982                 if (likely(nr_segs == 1))
1983                         copied = filemap_copy_from_user(page, offset,
1984                                                         buf, bytes);
1985                 else
1986                         copied = filemap_copy_from_user_iovec(page, offset,
1987                                                 cur_iov, iov_base, bytes);
1988                 flush_dcache_page(page);
1989                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1990                 if (likely(copied > 0)) {
1991                         if (!status)
1992                                 status = copied;
1993
1994                         if (status >= 0) {
1995                                 written += status;
1996                                 count -= status;
1997                                 pos += status;
1998                                 buf += status;
1999                                 if (unlikely(nr_segs > 1))
2000                                         filemap_set_next_iovec(&cur_iov,
2001                                                         &iov_base, status);
2002                         }
2003                 }
2004                 if (unlikely(copied != bytes))
2005                         if (status >= 0)
2006                                 status = -EFAULT;
2007                 unlock_page(page);
2008                 mark_page_accessed(page);
2009                 page_cache_release(page);
2010                 if (status < 0)
2011                         break;
2012                 balance_dirty_pages_ratelimited(mapping);
2013                 cond_resched();
2014         } while (count);
2015         *ppos = pos;
2016
2017         if (cached_page)
2018                 page_cache_release(cached_page);
2019
2020         /*
2021          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2022          */
2023         if (likely(status >= 0)) {
2024                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2025                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2026                                 status = generic_osync_inode(inode, mapping,
2027                                                 OSYNC_METADATA|OSYNC_DATA);
2028                 }
2029         }
2030
2031         /*
2032          * If we get here for O_DIRECT writes then we must have fallen through
2033          * to buffered writes (block instantiation inside i_size).  So we sync
2034          * the file data here, to try to honour O_DIRECT expectations.
2035          */
2036         if (unlikely(file->f_flags & O_DIRECT) && written)
2037                 status = filemap_write_and_wait(mapping);
2038
2039         pagevec_lru_add(&lru_pvec);
2040         return written ? written : status;
2041 }
2042 EXPORT_SYMBOL(generic_file_buffered_write);
2043
2044 ssize_t
2045 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2046                                 unsigned long nr_segs, loff_t *ppos)
2047 {
2048         struct file *file = iocb->ki_filp;
2049         struct address_space * mapping = file->f_mapping;
2050         size_t ocount;          /* original count */
2051         size_t count;           /* after file limit checks */
2052         struct inode    *inode = mapping->host;
2053         unsigned long   seg;
2054         loff_t          pos;
2055         ssize_t         written;
2056         ssize_t         err;
2057
2058         ocount = 0;
2059         for (seg = 0; seg < nr_segs; seg++) {
2060                 const struct iovec *iv = &iov[seg];
2061
2062                 /*
2063                  * If any segment has a negative length, or the cumulative
2064                  * length ever wraps negative then return -EINVAL.
2065                  */
2066                 ocount += iv->iov_len;
2067                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2068                         return -EINVAL;
2069                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2070                         continue;
2071                 if (seg == 0)
2072                         return -EFAULT;
2073                 nr_segs = seg;
2074                 ocount -= iv->iov_len;  /* This segment is no good */
2075                 break;
2076         }
2077
2078         count = ocount;
2079         pos = *ppos;
2080
2081         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2082
2083         /* We can write back this queue in page reclaim */
2084         current->backing_dev_info = mapping->backing_dev_info;
2085         written = 0;
2086
2087         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2088         if (err)
2089                 goto out;
2090
2091         if (count == 0)
2092                 goto out;
2093
2094         err = remove_suid(file->f_dentry);
2095         if (err)
2096                 goto out;
2097
2098         inode_update_time(inode, 1);
2099
2100         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2101         if (unlikely(file->f_flags & O_DIRECT)) {
2102                 written = generic_file_direct_write(iocb, iov,
2103                                 &nr_segs, pos, ppos, count, ocount);
2104                 if (written < 0 || written == count)
2105                         goto out;
2106                 /*
2107                  * direct-io write to a hole: fall through to buffered I/O
2108                  * for completing the rest of the request.
2109                  */
2110                 pos += written;
2111                 count -= written;
2112         }
2113
2114         written = generic_file_buffered_write(iocb, iov, nr_segs,
2115                         pos, ppos, count, written);
2116 out:
2117         current->backing_dev_info = NULL;
2118         return written ? written : err;
2119 }
2120 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2121
2122 ssize_t
2123 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2124                                 unsigned long nr_segs, loff_t *ppos)
2125 {
2126         struct file *file = iocb->ki_filp;
2127         struct address_space *mapping = file->f_mapping;
2128         struct inode *inode = mapping->host;
2129         ssize_t ret;
2130         loff_t pos = *ppos;
2131
2132         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2133
2134         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2135                 int err;
2136
2137                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2138                 if (err < 0)
2139                         ret = err;
2140         }
2141         return ret;
2142 }
2143
2144 ssize_t
2145 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2146                                 unsigned long nr_segs, loff_t *ppos)
2147 {
2148         struct kiocb kiocb;
2149         ssize_t ret;
2150
2151         init_sync_kiocb(&kiocb, file);
2152         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2153         if (ret == -EIOCBQUEUED)
2154                 ret = wait_on_sync_kiocb(&kiocb);
2155         return ret;
2156 }
2157
2158 ssize_t
2159 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2160                                 unsigned long nr_segs, loff_t *ppos)
2161 {
2162         struct kiocb kiocb;
2163         ssize_t ret;
2164
2165         init_sync_kiocb(&kiocb, file);
2166         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2167         if (-EIOCBQUEUED == ret)
2168                 ret = wait_on_sync_kiocb(&kiocb);
2169         return ret;
2170 }
2171 EXPORT_SYMBOL(generic_file_write_nolock);
2172
2173 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2174                                size_t count, loff_t pos)
2175 {
2176         struct file *file = iocb->ki_filp;
2177         struct address_space *mapping = file->f_mapping;
2178         struct inode *inode = mapping->host;
2179         ssize_t ret;
2180         struct iovec local_iov = { .iov_base = (void __user *)buf,
2181                                         .iov_len = count };
2182
2183         BUG_ON(iocb->ki_pos != pos);
2184
2185         down(&inode->i_sem);
2186         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2187                                                 &iocb->ki_pos);
2188         up(&inode->i_sem);
2189
2190         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2191                 ssize_t err;
2192
2193                 err = sync_page_range(inode, mapping, pos, ret);
2194                 if (err < 0)
2195                         ret = err;
2196         }
2197         return ret;
2198 }
2199 EXPORT_SYMBOL(generic_file_aio_write);
2200
2201 ssize_t generic_file_write(struct file *file, const char __user *buf,
2202                            size_t count, loff_t *ppos)
2203 {
2204         struct address_space *mapping = file->f_mapping;
2205         struct inode *inode = mapping->host;
2206         ssize_t ret;
2207         struct iovec local_iov = { .iov_base = (void __user *)buf,
2208                                         .iov_len = count };
2209
2210         down(&inode->i_sem);
2211         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2212         up(&inode->i_sem);
2213
2214         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2215                 ssize_t err;
2216
2217                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2218                 if (err < 0)
2219                         ret = err;
2220         }
2221         return ret;
2222 }
2223 EXPORT_SYMBOL(generic_file_write);
2224
2225 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2226                         unsigned long nr_segs, loff_t *ppos)
2227 {
2228         struct kiocb kiocb;
2229         ssize_t ret;
2230
2231         init_sync_kiocb(&kiocb, filp);
2232         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2233         if (-EIOCBQUEUED == ret)
2234                 ret = wait_on_sync_kiocb(&kiocb);
2235         return ret;
2236 }
2237 EXPORT_SYMBOL(generic_file_readv);
2238
2239 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2240                         unsigned long nr_segs, loff_t *ppos)
2241 {
2242         struct address_space *mapping = file->f_mapping;
2243         struct inode *inode = mapping->host;
2244         ssize_t ret;
2245
2246         down(&inode->i_sem);
2247         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2248         up(&inode->i_sem);
2249
2250         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2251                 int err;
2252
2253                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2254                 if (err < 0)
2255                         ret = err;
2256         }
2257         return ret;
2258 }
2259 EXPORT_SYMBOL(generic_file_writev);
2260
2261 /*
2262  * Called under i_sem for writes to S_ISREG files.   Returns -EIO if something
2263  * went wrong during pagecache shootdown.
2264  */
2265 ssize_t
2266 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2267         loff_t offset, unsigned long nr_segs)
2268 {
2269         struct file *file = iocb->ki_filp;
2270         struct address_space *mapping = file->f_mapping;
2271         ssize_t retval;
2272
2273         /*
2274          * If it's a write, unmap all mmappings of the file up-front.  This
2275          * will cause any pte dirty bits to be propagated into the pageframes
2276          * for the subsequent filemap_write_and_wait().
2277          */
2278         if (rw == WRITE && mapping_mapped(mapping))
2279                 unmap_mapping_range(mapping, 0, -1, 0);
2280
2281         retval = filemap_write_and_wait(mapping);
2282         if (retval == 0) {
2283                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2284                                                 offset, nr_segs);
2285                 if (rw == WRITE && mapping->nrpages) {
2286                         int err = invalidate_inode_pages2(mapping);
2287                         if (err)
2288                                 retval = err;
2289                 }
2290         }
2291         return retval;
2292 }
2293 EXPORT_SYMBOL_GPL(generic_file_direct_IO);