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, end_index, offset;
 692         loff_t isize;
 693         struct page *cached_page;
 694         int error;
 695         struct file_ra_state ra = *_ra;
 696
 697         cached_page = NULL;
 698         index = *ppos >> PAGE_CACHE_SHIFT;
 699         offset = *ppos & ~PAGE_CACHE_MASK;
 700
 701         isize = i_size_read(inode);
 702         if (!isize)
 703                 goto out;
 704
 705         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 706         for (;;) {
 707                 struct page *page;
 708                 unsigned long nr, ret;
 709
 710                 /* nr is the maximum number of bytes to copy from this page */
 711                 nr = PAGE_CACHE_SIZE;
 712                 if (index >= end_index) {
 713                         if (index > end_index)
 714                                 goto out;
 715                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 716                         if (nr <= offset) {
 717                                 goto out;
 718                         }
 719                 }
 720                 nr = nr - offset;
 721
 722                 cond_resched();
 723                 page_cache_readahead(mapping, &ra, filp, index);
 724
 725 find_page:
 726                 page = find_get_page(mapping, index);
 727                 if (unlikely(page == NULL)) {
 728                         handle_ra_miss(mapping, &ra, index);
 729                         goto no_cached_page;
 730                 }
 731                 if (!PageUptodate(page))
 732                         goto page_not_up_to_date;
 733 page_ok:
 734
 735                 /* If users can be writing to this page using arbitrary
 736                  * virtual addresses, take care about potential aliasing
 737                  * before reading the page on the kernel side.
 738                  */
 739                 if (mapping_writably_mapped(mapping))
 740                         flush_dcache_page(page);
 741
 742                 /*
 743                  * Mark the page accessed if we read the beginning.
 744                  */
 745                 if (!offset)
 746                         mark_page_accessed(page);
 747
 748                 /*
 749                  * Ok, we have the page, and it's up-to-date, so
 750                  * now we can copy it to user space...
 751                  *
 752                  * The actor routine returns how many bytes were actually used..
 753                  * NOTE! This may not be the same as how much of a user buffer
 754                  * we filled up (we may be padding etc), so we can only update
 755                  * "pos" here (the actor routine has to update the user buffer
 756                  * pointers and the remaining count).
 757                  */
 758                 ret = actor(desc, page, offset, nr);
 759                 offset += ret;
 760                 index += offset >> PAGE_CACHE_SHIFT;
 761                 offset &= ~PAGE_CACHE_MASK;
 762
 763                 page_cache_release(page);
 764                 if (ret == nr && desc->count)
 765                         continue;
 766                 goto out;
 767
 768 page_not_up_to_date:
 769                 /* Get exclusive access to the page ... */
 770                 lock_page(page);
 771
 772                 /* Did it get unhashed before we got the lock? */
 773                 if (!page->mapping) {
 774                         unlock_page(page);
 775                         page_cache_release(page);
 776                         continue;
 777                 }
 778
 779                 /* Did somebody else fill it already? */
 780                 if (PageUptodate(page)) {
 781                         unlock_page(page);
 782                         goto page_ok;
 783                 }
 784
 785 readpage:
 786                 /* Start the actual read. The read will unlock the page. */
 787                 error = mapping->a_ops->readpage(filp, page);
 788
 789                 if (unlikely(error))
 790                         goto readpage_error;
 791
 792                 if (!PageUptodate(page)) {
 793                         wait_on_page_locked(page);
 794                         if (!PageUptodate(page)) {
 795                                 error = -EIO;
 796                                 goto readpage_error;
 797                         }
 798                 }
 799
 800                 /*
 801                  * i_size must be checked after we have done ->readpage.
 802                  *
 803                  * Checking i_size after the readpage allows us to calculate
 804                  * the correct value for "nr", which means the zero-filled
 805                  * part of the page is not copied back to userspace (unless
 806                  * another truncate extends the file - this is desired though).
 807                  */
 808                 isize = i_size_read(inode);
 809                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 810                 if (unlikely(!isize || index > end_index)) {
 811                         page_cache_release(page);
 812                         goto out;
 813                 }
 814
 815                 /* nr is the maximum number of bytes to copy from this page */
 816                 nr = PAGE_CACHE_SIZE;
 817                 if (index == end_index) {
 818                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 819                         if (nr <= offset) {
 820                                 page_cache_release(page);
 821                                 goto out;
 822                         }
 823                 }
 824                 nr = nr - offset;
 825                 goto page_ok;
 826
 827 readpage_error:
 828                 /* UHHUH! A synchronous read error occurred. Report it */
 829                 desc->error = error;
 830                 page_cache_release(page);
 831                 goto out;
 832
 833 no_cached_page:
 834                 /*
 835                  * Ok, it wasn't cached, so we need to create a new
 836                  * page..
 837                  */
 838                 if (!cached_page) {
 839                         cached_page = page_cache_alloc_cold(mapping);
 840                         if (!cached_page) {
 841                                 desc->error = -ENOMEM;
 842                                 goto out;
 843                         }
 844                 }
 845                 error = add_to_page_cache_lru(cached_page, mapping,
 846                                                 index, GFP_KERNEL);
 847                 if (error) {
 848                         if (error == -EEXIST)
 849                                 goto find_page;
 850                         desc->error = error;
 851                         goto out;
 852                 }
 853                 page = cached_page;
 854                 cached_page = NULL;
 855                 goto readpage;
 856         }
 857
 858 out:
 859         *_ra = ra;
 860
 861         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 862         if (cached_page)
 863                 page_cache_release(cached_page);
 864         if (filp)
 865                 file_accessed(filp);
 866 }
 867
 868 EXPORT_SYMBOL(do_generic_mapping_read);
 869
 870 int file_read_actor(read_descriptor_t *desc, struct page *page,
 871                         unsigned long offset, unsigned long size)
 872 {
 873         char *kaddr;
 874         unsigned long left, count = desc->count;
 875
 876         if (size > count)
 877                 size = count;
 878
 879         /*
 880          * Faults on the destination of a read are common, so do it before
 881          * taking the kmap.
 882          */
 883         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 884                 kaddr = kmap_atomic(page, KM_USER0);
 885                 left = __copy_to_user_inatomic(desc->arg.buf,
 886                                                 kaddr + offset, size);
 887                 kunmap_atomic(kaddr, KM_USER0);
 888                 if (left == 0)
 889                         goto success;
 890         }
 891
 892         /* Do it the slow way */
 893         kaddr = kmap(page);
 894         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 895         kunmap(page);
 896
 897         if (left) {
 898                 size -= left;
 899                 desc->error = -EFAULT;
 900         }
 901 success:
 902         desc->count = count - size;
 903         desc->written += size;
 904         desc->arg.buf += size;
 905         return size;
 906 }
 907
 908 /*
 909  * This is the "read()" routine for all filesystems
 910  * that can use the page cache directly.
 911  */
 912 ssize_t
 913 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 914                 unsigned long nr_segs, loff_t *ppos)
 915 {
 916         struct file *filp = iocb->ki_filp;
 917         ssize_t retval;
 918         unsigned long seg;
 919         size_t count;
 920
 921         count = 0;
 922         for (seg = 0; seg < nr_segs; seg++) {
 923                 const struct iovec *iv = &iov[seg];
 924
 925                 /*
 926                  * If any segment has a negative length, or the cumulative
 927                  * length ever wraps negative then return -EINVAL.
 928                  */
 929                 count += iv->iov_len;
 930                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
 931                         return -EINVAL;
 932                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
 933                         continue;
 934                 if (seg == 0)
 935                         return -EFAULT;
 936                 nr_segs = seg;
 937                 count -= iv->iov_len;   /* This segment is no good */
 938                 break;
 939         }
 940
 941         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
 942         if (filp->f_flags & O_DIRECT) {
 943                 loff_t pos = *ppos, size;
 944                 struct address_space *mapping;
 945                 struct inode *inode;
 946
 947                 mapping = filp->f_mapping;
 948                 inode = mapping->host;
 949                 retval = 0;
 950                 if (!count)
 951                         goto out; /* skip atime */
 952                 size = i_size_read(inode);
 953                 if (pos < size) {
 954                         retval = generic_file_direct_IO(READ, iocb,
 955                                                 iov, pos, nr_segs);
 956                         if (retval >= 0 && !is_sync_kiocb(iocb))
 957                                 retval = -EIOCBQUEUED;
 958                         if (retval > 0)
 959                                 *ppos = pos + retval;
 960                 }
 961                 file_accessed(filp);
 962                 goto out;
 963         }
 964
 965         retval = 0;
 966         if (count) {
 967                 for (seg = 0; seg < nr_segs; seg++) {
 968                         read_descriptor_t desc;
 969
 970                         desc.written = 0;
 971                         desc.arg.buf = iov[seg].iov_base;
 972                         desc.count = iov[seg].iov_len;
 973                         if (desc.count == 0)
 974                                 continue;
 975                         desc.error = 0;
 976                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
 977                         retval += desc.written;
 978                         if (!retval) {
 979                                 retval = desc.error;
 980                                 break;
 981                         }
 982                 }
 983         }
 984 out:
 985         return retval;
 986 }
 987
 988 EXPORT_SYMBOL(__generic_file_aio_read);
 989
 990 ssize_t
 991 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 992 {
 993         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
 994
 995         BUG_ON(iocb->ki_pos != pos);
 996         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
 997 }
 998
 999 EXPORT_SYMBOL(generic_file_aio_read);
1000
1001 ssize_t
1002 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1003 {
1004         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1005         struct kiocb kiocb;
1006         ssize_t ret;
1007
1008         init_sync_kiocb(&kiocb, filp);
1009         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1010         if (-EIOCBQUEUED == ret)
1011                 ret = wait_on_sync_kiocb(&kiocb);
1012         return ret;
1013 }
1014
1015 EXPORT_SYMBOL(generic_file_read);
1016
1017 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1018 {
1019         ssize_t written;
1020         unsigned long count = desc->count;
1021         struct file *file = desc->arg.data;
1022
1023         if (size > count)
1024                 size = count;
1025
1026         written = file->f_op->sendpage(file, page, offset,
1027                                        size, &file->f_pos, size<count);
1028         if (written < 0) {
1029                 desc->error = written;
1030                 written = 0;
1031         }
1032         desc->count = count - written;
1033         desc->written += written;
1034         return written;
1035 }
1036
1037 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1038                          size_t count, read_actor_t actor, void *target)
1039 {
1040         read_descriptor_t desc;
1041
1042         if (!count)
1043                 return 0;
1044
1045         desc.written = 0;
1046         desc.count = count;
1047         desc.arg.data = target;
1048         desc.error = 0;
1049
1050         do_generic_file_read(in_file, ppos, &desc, actor);
1051         if (desc.written)
1052                 return desc.written;
1053         return desc.error;
1054 }
1055
1056 EXPORT_SYMBOL(generic_file_sendfile);
1057
1058 static ssize_t
1059 do_readahead(struct address_space *mapping, struct file *filp,
1060              unsigned long index, unsigned long nr)
1061 {
1062         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1063                 return -EINVAL;
1064
1065         force_page_cache_readahead(mapping, filp, index,
1066                                         max_sane_readahead(nr));
1067         return 0;
1068 }
1069
1070 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1071 {
1072         ssize_t ret;
1073         struct file *file;
1074
1075         ret = -EBADF;
1076         file = fget(fd);
1077         if (file) {
1078                 if (file->f_mode & FMODE_READ) {
1079                         struct address_space *mapping = file->f_mapping;
1080                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1081                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1082                         unsigned long len = end - start + 1;
1083                         ret = do_readahead(mapping, file, start, len);
1084                 }
1085                 fput(file);
1086         }
1087         return ret;
1088 }
1089
1090 #ifdef CONFIG_MMU
1091 /*
1092  * This adds the requested page to the page cache if it isn't already there,
1093  * and schedules an I/O to read in its contents from disk.
1094  */
1095 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1096 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1097 {
1098         struct address_space *mapping = file->f_mapping;
1099         struct page *page;
1100         int error;
1101
1102         page = page_cache_alloc_cold(mapping);
1103         if (!page)
1104                 return -ENOMEM;
1105
1106         error = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1107         if (!error) {
1108                 error = mapping->a_ops->readpage(file, page);
1109                 page_cache_release(page);
1110                 return error;
1111         }
1112
1113         /*
1114          * We arrive here in the unlikely event that someone
1115          * raced with us and added our page to the cache first
1116          * or we are out of memory for radix-tree nodes.
1117          */
1118         page_cache_release(page);
1119         return error == -EEXIST ? 0 : error;
1120 }
1121
1122 #define MMAP_LOTSAMISS  (100)
1123
1124 /*
1125  * filemap_nopage() is invoked via the vma operations vector for a
1126  * mapped memory region to read in file data during a page fault.
1127  *
1128  * The goto's are kind of ugly, but this streamlines the normal case of having
1129  * it in the page cache, and handles the special cases reasonably without
1130  * having a lot of duplicated code.
1131  */
1132 struct page * filemap_nopage(struct vm_area_struct * area, unsigned long address, int *type)
1133 {
1134         int error;
1135         struct file *file = area->vm_file;
1136         struct address_space *mapping = file->f_mapping;
1137         struct file_ra_state *ra = &file->f_ra;
1138         struct inode *inode = mapping->host;
1139         struct page *page;
1140         unsigned long size, pgoff, endoff;
1141         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1142
1143         pgoff = ((address - area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1144         endoff = ((area->vm_end - area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1145
1146 retry_all:
1147         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1148         if (pgoff >= size)
1149                 goto outside_data_content;
1150
1151         /* If we don't want any read-ahead, don't bother */
1152         if (VM_RandomReadHint(area))
1153                 goto no_cached_page;
1154
1155         /*
1156          * The "size" of the file, as far as mmap is concerned, isn't bigger
1157          * than the mapping
1158          */
1159         if (size > endoff)
1160                 size = endoff;
1161
1162         /*
1163          * The readahead code wants to be told about each and every page
1164          * so it can build and shrink its windows appropriately
1165          *
1166          * For sequential accesses, we use the generic readahead logic.
1167          */
1168         if (VM_SequentialReadHint(area))
1169                 page_cache_readahead(mapping, ra, file, pgoff);
1170
1171         /*
1172          * Do we have something in the page cache already?
1173          */
1174 retry_find:
1175         page = find_get_page(mapping, pgoff);
1176         if (!page) {
1177                 unsigned long ra_pages;
1178
1179                 if (VM_SequentialReadHint(area)) {
1180                         handle_ra_miss(mapping, ra, pgoff);
1181                         goto no_cached_page;
1182                 }
1183                 ra->mmap_miss++;
1184
1185                 /*
1186                  * Do we miss much more than hit in this file? If so,
1187                  * stop bothering with read-ahead. It will only hurt.
1188                  */
1189                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1190                         goto no_cached_page;
1191
1192                 /*
1193                  * To keep the pgmajfault counter straight, we need to
1194                  * check did_readaround, as this is an inner loop.
1195                  */
1196                 if (!did_readaround) {
1197                         majmin = VM_FAULT_MAJOR;
1198                         inc_page_state(pgmajfault);
1199                 }
1200                 did_readaround = 1;
1201                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1202                 if (ra_pages) {
1203                         pgoff_t start = 0;
1204
1205                         if (pgoff > ra_pages / 2)
1206                                 start = pgoff - ra_pages / 2;
1207                         do_page_cache_readahead(mapping, file, start, ra_pages);
1208                 }
1209                 page = find_get_page(mapping, pgoff);
1210                 if (!page)
1211                         goto no_cached_page;
1212         }
1213
1214         if (!did_readaround)
1215                 ra->mmap_hit++;
1216
1217         /*
1218          * Ok, found a page in the page cache, now we need to check
1219          * that it's up-to-date.
1220          */
1221         if (!PageUptodate(page))
1222                 goto page_not_uptodate;
1223
1224 success:
1225         /*
1226          * Found the page and have a reference on it.
1227          */
1228         mark_page_accessed(page);
1229         if (type)
1230                 *type = majmin;
1231         return page;
1232
1233 outside_data_content:
1234         /*
1235          * An external ptracer can access pages that normally aren't
1236          * accessible..
1237          */
1238         if (area->vm_mm == current->mm)
1239                 return NULL;
1240         /* Fall through to the non-read-ahead case */
1241 no_cached_page:
1242         /*
1243          * We're only likely to ever get here if MADV_RANDOM is in
1244          * effect.
1245          */
1246         error = page_cache_read(file, pgoff);
1247         grab_swap_token();
1248
1249         /*
1250          * The page we want has now been added to the page cache.
1251          * In the unlikely event that someone removed it in the
1252          * meantime, we'll just come back here and read it again.
1253          */
1254         if (error >= 0)
1255                 goto retry_find;
1256
1257         /*
1258          * An error return from page_cache_read can result if the
1259          * system is low on memory, or a problem occurs while trying
1260          * to schedule I/O.
1261          */
1262         if (error == -ENOMEM)
1263                 return NOPAGE_OOM;
1264         return NULL;
1265
1266 page_not_uptodate:
1267         if (!did_readaround) {
1268                 majmin = VM_FAULT_MAJOR;
1269                 inc_page_state(pgmajfault);
1270         }
1271         lock_page(page);
1272
1273         /* Did it get unhashed while we waited for it? */
1274         if (!page->mapping) {
1275                 unlock_page(page);
1276                 page_cache_release(page);
1277                 goto retry_all;
1278         }
1279
1280         /* Did somebody else get it up-to-date? */
1281         if (PageUptodate(page)) {
1282                 unlock_page(page);
1283                 goto success;
1284         }
1285
1286         if (!mapping->a_ops->readpage(file, page)) {
1287                 wait_on_page_locked(page);
1288                 if (PageUptodate(page))
1289                         goto success;
1290         }
1291
1292         /*
1293          * Umm, take care of errors if the page isn't up-to-date.
1294          * Try to re-read it _once_. We do this synchronously,
1295          * because there really aren't any performance issues here
1296          * and we need to check for errors.
1297          */
1298         lock_page(page);
1299
1300         /* Somebody truncated the page on us? */
1301         if (!page->mapping) {
1302                 unlock_page(page);
1303                 page_cache_release(page);
1304                 goto retry_all;
1305         }
1306
1307         /* Somebody else successfully read it in? */
1308         if (PageUptodate(page)) {
1309                 unlock_page(page);
1310                 goto success;
1311         }
1312         ClearPageError(page);
1313         if (!mapping->a_ops->readpage(file, page)) {
1314                 wait_on_page_locked(page);
1315                 if (PageUptodate(page))
1316                         goto success;
1317         }
1318
1319         /*
1320          * Things didn't work out. Return zero to tell the
1321          * mm layer so, possibly freeing the page cache page first.
1322          */
1323         page_cache_release(page);
1324         return NULL;
1325 }
1326
1327 EXPORT_SYMBOL(filemap_nopage);
1328
1329 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1330                                         int nonblock)
1331 {
1332         struct address_space *mapping = file->f_mapping;
1333         struct page *page;
1334         int error;
1335
1336         /*
1337          * Do we have something in the page cache already?
1338          */
1339 retry_find:
1340         page = find_get_page(mapping, pgoff);
1341         if (!page) {
1342                 if (nonblock)
1343                         return NULL;
1344                 goto no_cached_page;
1345         }
1346
1347         /*
1348          * Ok, found a page in the page cache, now we need to check
1349          * that it's up-to-date.
1350          */
1351         if (!PageUptodate(page))
1352                 goto page_not_uptodate;
1353
1354 success:
1355         /*
1356          * Found the page and have a reference on it.
1357          */
1358         mark_page_accessed(page);
1359         return page;
1360
1361 no_cached_page:
1362         error = page_cache_read(file, pgoff);
1363
1364         /*
1365          * The page we want has now been added to the page cache.
1366          * In the unlikely event that someone removed it in the
1367          * meantime, we'll just come back here and read it again.
1368          */
1369         if (error >= 0)
1370                 goto retry_find;
1371
1372         /*
1373          * An error return from page_cache_read can result if the
1374          * system is low on memory, or a problem occurs while trying
1375          * to schedule I/O.
1376          */
1377         return NULL;
1378
1379 page_not_uptodate:
1380         lock_page(page);
1381
1382         /* Did it get unhashed while we waited for it? */
1383         if (!page->mapping) {
1384                 unlock_page(page);
1385                 goto err;
1386         }
1387
1388         /* Did somebody else get it up-to-date? */
1389         if (PageUptodate(page)) {
1390                 unlock_page(page);
1391                 goto success;
1392         }
1393
1394         if (!mapping->a_ops->readpage(file, page)) {
1395                 wait_on_page_locked(page);
1396                 if (PageUptodate(page))
1397                         goto success;
1398         }
1399
1400         /*
1401          * Umm, take care of errors if the page isn't up-to-date.
1402          * Try to re-read it _once_. We do this synchronously,
1403          * because there really aren't any performance issues here
1404          * and we need to check for errors.
1405          */
1406         lock_page(page);
1407
1408         /* Somebody truncated the page on us? */
1409         if (!page->mapping) {
1410                 unlock_page(page);
1411                 goto err;
1412         }
1413         /* Somebody else successfully read it in? */
1414         if (PageUptodate(page)) {
1415                 unlock_page(page);
1416                 goto success;
1417         }
1418
1419         ClearPageError(page);
1420         if (!mapping->a_ops->readpage(file, page)) {
1421                 wait_on_page_locked(page);
1422                 if (PageUptodate(page))
1423                         goto success;
1424         }
1425
1426         /*
1427          * Things didn't work out. Return zero to tell the
1428          * mm layer so, possibly freeing the page cache page first.
1429          */
1430 err:
1431         page_cache_release(page);
1432
1433         return NULL;
1434 }
1435
1436 static int filemap_populate(struct vm_area_struct *vma,
1437                         unsigned long addr,
1438                         unsigned long len,
1439                         pgprot_t prot,
1440                         unsigned long pgoff,
1441                         int nonblock)
1442 {
1443         struct file *file = vma->vm_file;
1444         struct address_space *mapping = file->f_mapping;
1445         struct inode *inode = mapping->host;
1446         unsigned long size;
1447         struct mm_struct *mm = vma->vm_mm;
1448         struct page *page;
1449         int err;
1450
1451         if (!nonblock)
1452                 force_page_cache_readahead(mapping, vma->vm_file,
1453                                         pgoff, len >> PAGE_CACHE_SHIFT);
1454
1455 repeat:
1456         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1457         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1458                 return -EINVAL;
1459
1460         page = filemap_getpage(file, pgoff, nonblock);
1461         if (!page && !nonblock)
1462                 return -ENOMEM;
1463         if (page) {
1464                 err = install_page(mm, vma, addr, page, prot);
1465                 if (err) {
1466                         page_cache_release(page);
1467                         return err;
1468                 }
1469         } else {
1470                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1471                 if (err)
1472                         return err;
1473         }
1474
1475         len -= PAGE_SIZE;
1476         addr += PAGE_SIZE;
1477         pgoff++;
1478         if (len)
1479                 goto repeat;
1480
1481         return 0;
1482 }
1483
1484 struct vm_operations_struct generic_file_vm_ops = {
1485         .nopage         = filemap_nopage,
1486         .populate       = filemap_populate,
1487 };
1488
1489 /* This is used for a general mmap of a disk file */
1490
1491 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1492 {
1493         struct address_space *mapping = file->f_mapping;
1494
1495         if (!mapping->a_ops->readpage)
1496                 return -ENOEXEC;
1497         file_accessed(file);
1498         vma->vm_ops = &generic_file_vm_ops;
1499         return 0;
1500 }
1501
1502 /*
1503  * This is for filesystems which do not implement ->writepage.
1504  */
1505 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1506 {
1507         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1508                 return -EINVAL;
1509         return generic_file_mmap(file, vma);
1510 }
1511 #else
1512 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1513 {
1514         return -ENOSYS;
1515 }
1516 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1517 {
1518         return -ENOSYS;
1519 }
1520 #endif /* CONFIG_MMU */
1521
1522 EXPORT_SYMBOL(generic_file_mmap);
1523 EXPORT_SYMBOL(generic_file_readonly_mmap);
1524
1525 static inline struct page *__read_cache_page(struct address_space *mapping,
1526                                 unsigned long index,
1527                                 int (*filler)(void *,struct page*),
1528                                 void *data)
1529 {
1530         struct page *page, *cached_page = NULL;
1531         int err;
1532 repeat:
1533         page = find_get_page(mapping, index);
1534         if (!page) {
1535                 if (!cached_page) {
1536                         cached_page = page_cache_alloc_cold(mapping);
1537                         if (!cached_page)
1538                                 return ERR_PTR(-ENOMEM);
1539                 }
1540                 err = add_to_page_cache_lru(cached_page, mapping,
1541                                         index, GFP_KERNEL);
1542                 if (err == -EEXIST)
1543                         goto repeat;
1544                 if (err < 0) {
1545                         /* Presumably ENOMEM for radix tree node */
1546                         page_cache_release(cached_page);
1547                         return ERR_PTR(err);
1548                 }
1549                 page = cached_page;
1550                 cached_page = NULL;
1551                 err = filler(data, page);
1552                 if (err < 0) {
1553                         page_cache_release(page);
1554                         page = ERR_PTR(err);
1555                 }
1556         }
1557         if (cached_page)
1558                 page_cache_release(cached_page);
1559         return page;
1560 }
1561
1562 /*
1563  * Read into the page cache. If a page already exists,
1564  * and PageUptodate() is not set, try to fill the page.
1565  */
1566 struct page *read_cache_page(struct address_space *mapping,
1567                                 unsigned long index,
1568                                 int (*filler)(void *,struct page*),
1569                                 void *data)
1570 {
1571         struct page *page;
1572         int err;
1573
1574 retry:
1575         page = __read_cache_page(mapping, index, filler, data);
1576         if (IS_ERR(page))
1577                 goto out;
1578         mark_page_accessed(page);
1579         if (PageUptodate(page))
1580                 goto out;
1581
1582         lock_page(page);
1583         if (!page->mapping) {
1584                 unlock_page(page);
1585                 page_cache_release(page);
1586                 goto retry;
1587         }
1588         if (PageUptodate(page)) {
1589                 unlock_page(page);
1590                 goto out;
1591         }
1592         err = filler(data, page);
1593         if (err < 0) {
1594                 page_cache_release(page);
1595                 page = ERR_PTR(err);
1596         }
1597  out:
1598         return page;
1599 }
1600
1601 EXPORT_SYMBOL(read_cache_page);
1602
1603 /*
1604  * If the page was newly created, increment its refcount and add it to the
1605  * caller's lru-buffering pagevec.  This function is specifically for
1606  * generic_file_write().
1607  */
1608 static inline struct page *
1609 __grab_cache_page(struct address_space *mapping, unsigned long index,
1610                         struct page **cached_page, struct pagevec *lru_pvec)
1611 {
1612         int err;
1613         struct page *page;
1614 repeat:
1615         page = find_lock_page(mapping, index);
1616         if (!page) {
1617                 if (!*cached_page) {
1618                         *cached_page = page_cache_alloc(mapping);
1619                         if (!*cached_page)
1620                                 return NULL;
1621                 }
1622                 err = add_to_page_cache(*cached_page, mapping,
1623                                         index, GFP_KERNEL);
1624                 if (err == -EEXIST)
1625                         goto repeat;
1626                 if (err == 0) {
1627                         page = *cached_page;
1628                         page_cache_get(page);
1629                         if (!pagevec_add(lru_pvec, page))
1630                                 __pagevec_lru_add(lru_pvec);
1631                         *cached_page = NULL;
1632                 }
1633         }
1634         return page;
1635 }
1636
1637 /*
1638  * The logic we want is
1639  *
1640  *      if suid or (sgid and xgrp)
1641  *              remove privs
1642  */
1643 int remove_suid(struct dentry *dentry)
1644 {
1645         mode_t mode = dentry->d_inode->i_mode;
1646         int kill = 0;
1647         int result = 0;
1648
1649         /* suid always must be killed */
1650         if (unlikely(mode & S_ISUID))
1651                 kill = ATTR_KILL_SUID;
1652
1653         /*
1654          * sgid without any exec bits is just a mandatory locking mark; leave
1655          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1656          */
1657         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1658                 kill |= ATTR_KILL_SGID;
1659
1660         if (unlikely(kill && !capable(CAP_FSETID))) {
1661                 struct iattr newattrs;
1662
1663                 newattrs.ia_valid = ATTR_FORCE | kill;
1664                 result = notify_change(dentry, &newattrs);
1665         }
1666         return result;
1667 }
1668 EXPORT_SYMBOL(remove_suid);
1669
1670 /*
1671  * Copy as much as we can into the page and return the number of bytes which
1672  * were sucessfully copied.  If a fault is encountered then clear the page
1673  * out to (offset+bytes) and return the number of bytes which were copied.
1674  */
1675 static inline size_t
1676 filemap_copy_from_user(struct page *page, unsigned long offset,
1677                         const char __user *buf, unsigned bytes)
1678 {
1679         char *kaddr;
1680         int left;
1681
1682         kaddr = kmap_atomic(page, KM_USER0);
1683         left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1684         kunmap_atomic(kaddr, KM_USER0);
1685
1686         if (left != 0) {
1687                 /* Do it the slow way */
1688                 kaddr = kmap(page);
1689                 left = __copy_from_user(kaddr + offset, buf, bytes);
1690                 kunmap(page);
1691         }
1692         return bytes - left;
1693 }
1694
1695 static size_t
1696 __filemap_copy_from_user_iovec(char *vaddr,
1697                         const struct iovec *iov, size_t base, size_t bytes)
1698 {
1699         size_t copied = 0, left = 0;
1700
1701         while (bytes) {
1702                 char __user *buf = iov->iov_base + base;
1703                 int copy = min(bytes, iov->iov_len - base);
1704
1705                 base = 0;
1706                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1707                 copied += copy;
1708                 bytes -= copy;
1709                 vaddr += copy;
1710                 iov++;
1711
1712                 if (unlikely(left)) {
1713                         /* zero the rest of the target like __copy_from_user */
1714                         if (bytes)
1715                                 memset(vaddr, 0, bytes);
1716                         break;
1717                 }
1718         }
1719         return copied - left;
1720 }
1721
1722 /*
1723  * This has the same sideeffects and return value as filemap_copy_from_user().
1724  * The difference is that on a fault we need to memset the remainder of the
1725  * page (out to offset+bytes), to emulate filemap_copy_from_user()'s
1726  * single-segment behaviour.
1727  */
1728 static inline size_t
1729 filemap_copy_from_user_iovec(struct page *page, unsigned long offset,
1730                         const struct iovec *iov, size_t base, size_t bytes)
1731 {
1732         char *kaddr;
1733         size_t copied;
1734
1735         kaddr = kmap_atomic(page, KM_USER0);
1736         copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1737                                                 base, bytes);
1738         kunmap_atomic(kaddr, KM_USER0);
1739         if (copied != bytes) {
1740                 kaddr = kmap(page);
1741                 copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1742                                                         base, bytes);
1743                 kunmap(page);
1744         }
1745         return copied;
1746 }
1747
1748 static inline void
1749 filemap_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1750 {
1751         const struct iovec *iov = *iovp;
1752         size_t base = *basep;
1753
1754         while (bytes) {
1755                 int copy = min(bytes, iov->iov_len - base);
1756
1757                 bytes -= copy;
1758                 base += copy;
1759                 if (iov->iov_len == base) {
1760                         iov++;
1761                         base = 0;
1762                 }
1763         }
1764         *iovp = iov;
1765         *basep = base;
1766 }
1767
1768 /*
1769  * Performs necessary checks before doing a write
1770  *
1771  * Can adjust writing position aor amount of bytes to write.
1772  * Returns appropriate error code that caller should return or
1773  * zero in case that write should be allowed.
1774  */
1775 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1776 {
1777         struct inode *inode = file->f_mapping->host;
1778         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1779
1780         if (unlikely(*pos < 0))
1781                 return -EINVAL;
1782
1783         if (unlikely(file->f_error)) {
1784                 int err = file->f_error;
1785                 file->f_error = 0;
1786                 return err;
1787         }
1788
1789         if (!isblk) {
1790                 /* FIXME: this is for backwards compatibility with 2.4 */
1791                 if (file->f_flags & O_APPEND)
1792                         *pos = i_size_read(inode);
1793
1794                 if (limit != RLIM_INFINITY) {
1795                         if (*pos >= limit) {
1796                                 send_sig(SIGXFSZ, current, 0);
1797                                 return -EFBIG;
1798                         }
1799                         if (*count > limit - (typeof(limit))*pos) {
1800                                 *count = limit - (typeof(limit))*pos;
1801                         }
1802                 }
1803         }
1804
1805         /*
1806          * LFS rule
1807          */
1808         if (unlikely(*pos + *count > MAX_NON_LFS &&
1809                                 !(file->f_flags & O_LARGEFILE))) {
1810                 if (*pos >= MAX_NON_LFS) {
1811                         send_sig(SIGXFSZ, current, 0);
1812                         return -EFBIG;
1813                 }
1814                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1815                         *count = MAX_NON_LFS - (unsigned long)*pos;
1816                 }
1817         }
1818
1819         /*
1820          * Are we about to exceed the fs block limit ?
1821          *
1822          * If we have written data it becomes a short write.  If we have
1823          * exceeded without writing data we send a signal and return EFBIG.
1824          * Linus frestrict idea will clean these up nicely..
1825          */
1826         if (likely(!isblk)) {
1827                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1828                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1829                                 send_sig(SIGXFSZ, current, 0);
1830                                 return -EFBIG;
1831                         }
1832                         /* zero-length writes at ->s_maxbytes are OK */
1833                 }
1834
1835                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1836                         *count = inode->i_sb->s_maxbytes - *pos;
1837         } else {
1838                 loff_t isize;
1839                 if (bdev_read_only(I_BDEV(inode)))
1840                         return -EPERM;
1841                 isize = i_size_read(inode);
1842                 if (*pos >= isize) {
1843                         if (*count || *pos > isize)
1844                                 return -ENOSPC;
1845                 }
1846
1847                 if (*pos + *count > isize)
1848                         *count = isize - *pos;
1849         }
1850         return 0;
1851 }
1852 EXPORT_SYMBOL(generic_write_checks);
1853
1854 ssize_t
1855 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1856                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1857                 size_t count, size_t ocount)
1858 {
1859         struct file     *file = iocb->ki_filp;
1860         struct address_space *mapping = file->f_mapping;
1861         struct inode    *inode = mapping->host;
1862         ssize_t         written;
1863
1864         if (count != ocount)
1865                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1866
1867         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1868         if (written > 0) {
1869                 loff_t end = pos + written;
1870                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1871                         i_size_write(inode,  end);
1872                         mark_inode_dirty(inode);
1873                 }
1874                 *ppos = end;
1875         }
1876
1877         /*
1878          * Sync the fs metadata but not the minor inode changes and
1879          * of course not the data as we did direct DMA for the IO.
1880          * i_sem is held, which protects generic_osync_inode() from
1881          * livelocking.
1882          */
1883         if (written >= 0 && file->f_flags & O_SYNC)
1884                 generic_osync_inode(inode, mapping, OSYNC_METADATA);
1885         if (written == count && !is_sync_kiocb(iocb))
1886                 written = -EIOCBQUEUED;
1887         return written;
1888 }
1889 EXPORT_SYMBOL(generic_file_direct_write);
1890
1891 ssize_t
1892 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1893                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1894                 size_t count, ssize_t written)
1895 {
1896         struct file *file = iocb->ki_filp;
1897         struct address_space * mapping = file->f_mapping;
1898         struct address_space_operations *a_ops = mapping->a_ops;
1899         struct inode    *inode = mapping->host;
1900         long            status = 0;
1901         struct page     *page;
1902         struct page     *cached_page = NULL;
1903         size_t          bytes;
1904         struct pagevec  lru_pvec;
1905         const struct iovec *cur_iov = iov; /* current iovec */
1906         size_t          iov_base = 0;      /* offset in the current iovec */
1907         char __user     *buf;
1908
1909         pagevec_init(&lru_pvec, 0);
1910
1911         buf = iov->iov_base + written;  /* handle partial DIO write */
1912         do {
1913                 unsigned long index;
1914                 unsigned long offset;
1915                 size_t copied;
1916
1917                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1918                 index = pos >> PAGE_CACHE_SHIFT;
1919                 bytes = PAGE_CACHE_SIZE - offset;
1920                 if (bytes > count)
1921                         bytes = count;
1922
1923                 /*
1924                  * Bring in the user page that we will copy from _first_.
1925                  * Otherwise there's a nasty deadlock on copying from the
1926                  * same page as we're writing to, without it being marked
1927                  * up-to-date.
1928                  */
1929                 fault_in_pages_readable(buf, bytes);
1930
1931                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1932                 if (!page) {
1933                         status = -ENOMEM;
1934                         break;
1935                 }
1936
1937                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1938                 if (unlikely(status)) {
1939                         loff_t isize = i_size_read(inode);
1940                         /*
1941                          * prepare_write() may have instantiated a few blocks
1942                          * outside i_size.  Trim these off again.
1943                          */
1944                         unlock_page(page);
1945                         page_cache_release(page);
1946                         if (pos + bytes > isize)
1947                                 vmtruncate(inode, isize);
1948                         break;
1949                 }
1950                 if (likely(nr_segs == 1))
1951                         copied = filemap_copy_from_user(page, offset,
1952                                                         buf, bytes);
1953                 else
1954                         copied = filemap_copy_from_user_iovec(page, offset,
1955                                                 cur_iov, iov_base, bytes);
1956                 flush_dcache_page(page);
1957                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1958                 if (likely(copied > 0)) {
1959                         if (!status)
1960                                 status = copied;
1961
1962                         if (status >= 0) {
1963                                 written += status;
1964                                 count -= status;
1965                                 pos += status;
1966                                 buf += status;
1967                                 if (unlikely(nr_segs > 1))
1968                                         filemap_set_next_iovec(&cur_iov,
1969                                                         &iov_base, status);
1970                         }
1971                 }
1972                 if (unlikely(copied != bytes))
1973                         if (status >= 0)
1974                                 status = -EFAULT;
1975                 unlock_page(page);
1976                 mark_page_accessed(page);
1977                 page_cache_release(page);
1978                 if (status < 0)
1979                         break;
1980                 balance_dirty_pages_ratelimited(mapping);
1981                 cond_resched();
1982         } while (count);
1983         *ppos = pos;
1984
1985         if (cached_page)
1986                 page_cache_release(cached_page);
1987
1988         /*
1989          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
1990          */
1991         if (likely(status >= 0)) {
1992                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1993                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
1994                                 status = generic_osync_inode(inode, mapping,
1995                                                 OSYNC_METADATA|OSYNC_DATA);
1996                 }
1997         }
1998
1999         /*
2000          * If we get here for O_DIRECT writes then we must have fallen through
2001          * to buffered writes (block instantiation inside i_size).  So we sync
2002          * the file data here, to try to honour O_DIRECT expectations.
2003          */
2004         if (unlikely(file->f_flags & O_DIRECT) && written)
2005                 status = filemap_write_and_wait(mapping);
2006
2007         pagevec_lru_add(&lru_pvec);
2008         return written ? written : status;
2009 }
2010 EXPORT_SYMBOL(generic_file_buffered_write);
2011
2012 ssize_t
2013 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2014                                 unsigned long nr_segs, loff_t *ppos)
2015 {
2016         struct file *file = iocb->ki_filp;
2017         struct address_space * mapping = file->f_mapping;
2018         size_t ocount;          /* original count */
2019         size_t count;           /* after file limit checks */
2020         struct inode    *inode = mapping->host;
2021         unsigned long   seg;
2022         loff_t          pos;
2023         ssize_t         written;
2024         ssize_t         err;
2025
2026         ocount = 0;
2027         for (seg = 0; seg < nr_segs; seg++) {
2028                 const struct iovec *iv = &iov[seg];
2029
2030                 /*
2031                  * If any segment has a negative length, or the cumulative
2032                  * length ever wraps negative then return -EINVAL.
2033                  */
2034                 ocount += iv->iov_len;
2035                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2036                         return -EINVAL;
2037                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2038                         continue;
2039                 if (seg == 0)
2040                         return -EFAULT;
2041                 nr_segs = seg;
2042                 ocount -= iv->iov_len;  /* This segment is no good */
2043                 break;
2044         }
2045
2046         count = ocount;
2047         pos = *ppos;
2048
2049         /* We can write back this queue in page reclaim */
2050         current->backing_dev_info = mapping->backing_dev_info;
2051         written = 0;
2052
2053         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2054         if (err)
2055                 goto out;
2056
2057         if (count == 0)
2058                 goto out;
2059
2060         err = remove_suid(file->f_dentry);
2061         if (err)
2062                 goto out;
2063
2064         inode_update_time(inode, 1);
2065
2066         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2067         if (unlikely(file->f_flags & O_DIRECT)) {
2068                 written = generic_file_direct_write(iocb, iov,
2069                                 &nr_segs, pos, ppos, count, ocount);
2070                 if (written < 0 || written == count)
2071                         goto out;
2072                 /*
2073                  * direct-io write to a hole: fall through to buffered I/O
2074                  * for completing the rest of the request.
2075                  */
2076                 pos += written;
2077                 count -= written;
2078         }
2079
2080         written = generic_file_buffered_write(iocb, iov, nr_segs,
2081                         pos, ppos, count, written);
2082 out:
2083         current->backing_dev_info = NULL;
2084         return written ? written : err;
2085 }
2086 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2087
2088 ssize_t
2089 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2090                                 unsigned long nr_segs, loff_t *ppos)
2091 {
2092         struct file *file = iocb->ki_filp;
2093         struct address_space *mapping = file->f_mapping;
2094         struct inode *inode = mapping->host;
2095         ssize_t ret;
2096         loff_t pos = *ppos;
2097
2098         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2099
2100         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2101                 int err;
2102
2103                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2104                 if (err < 0)
2105                         ret = err;
2106         }
2107         return ret;
2108 }
2109
2110 ssize_t
2111 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2112                                 unsigned long nr_segs, loff_t *ppos)
2113 {
2114         struct kiocb kiocb;
2115         ssize_t ret;
2116
2117         init_sync_kiocb(&kiocb, file);
2118         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2119         if (ret == -EIOCBQUEUED)
2120                 ret = wait_on_sync_kiocb(&kiocb);
2121         return ret;
2122 }
2123
2124 ssize_t
2125 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2126                                 unsigned long nr_segs, loff_t *ppos)
2127 {
2128         struct kiocb kiocb;
2129         ssize_t ret;
2130
2131         init_sync_kiocb(&kiocb, file);
2132         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2133         if (-EIOCBQUEUED == ret)
2134                 ret = wait_on_sync_kiocb(&kiocb);
2135         return ret;
2136 }
2137 EXPORT_SYMBOL(generic_file_write_nolock);
2138
2139 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2140                                size_t count, loff_t pos)
2141 {
2142         struct file *file = iocb->ki_filp;
2143         struct address_space *mapping = file->f_mapping;
2144         struct inode *inode = mapping->host;
2145         ssize_t ret;
2146         struct iovec local_iov = { .iov_base = (void __user *)buf,
2147                                         .iov_len = count };
2148
2149         BUG_ON(iocb->ki_pos != pos);
2150
2151         down(&inode->i_sem);
2152         ret = generic_file_aio_write_nolock(iocb, &local_iov, 1,
2153                                                 &iocb->ki_pos);
2154         up(&inode->i_sem);
2155
2156         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2157                 ssize_t err;
2158
2159                 err = sync_page_range(inode, mapping, pos, ret);
2160                 if (err < 0)
2161                         ret = err;
2162         }
2163         return ret;
2164 }
2165 EXPORT_SYMBOL(generic_file_aio_write);
2166
2167 ssize_t generic_file_write(struct file *file, const char __user *buf,
2168                            size_t count, loff_t *ppos)
2169 {
2170         struct address_space *mapping = file->f_mapping;
2171         struct inode *inode = mapping->host;
2172         ssize_t ret;
2173         struct iovec local_iov = { .iov_base = (void __user *)buf,
2174                                         .iov_len = count };
2175
2176         down(&inode->i_sem);
2177         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2178         up(&inode->i_sem);
2179
2180         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2181                 ssize_t err;
2182
2183                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2184                 if (err < 0)
2185                         ret = err;
2186         }
2187         return ret;
2188 }
2189 EXPORT_SYMBOL(generic_file_write);
2190
2191 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2192                         unsigned long nr_segs, loff_t *ppos)
2193 {
2194         struct kiocb kiocb;
2195         ssize_t ret;
2196
2197         init_sync_kiocb(&kiocb, filp);
2198         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2199         if (-EIOCBQUEUED == ret)
2200                 ret = wait_on_sync_kiocb(&kiocb);
2201         return ret;
2202 }
2203 EXPORT_SYMBOL(generic_file_readv);
2204
2205 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2206                         unsigned long nr_segs, loff_t *ppos)
2207 {
2208         struct address_space *mapping = file->f_mapping;
2209         struct inode *inode = mapping->host;
2210         ssize_t ret;
2211
2212         down(&inode->i_sem);
2213         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2214         up(&inode->i_sem);
2215
2216         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2217                 int err;
2218
2219                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2220                 if (err < 0)
2221                         ret = err;
2222         }
2223         return ret;
2224 }
2225 EXPORT_SYMBOL(generic_file_writev);
2226
2227 /*
2228  * Called under i_sem for writes to S_ISREG files
2229  */
2230 ssize_t
2231 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2232         loff_t offset, unsigned long nr_segs)
2233 {
2234         struct file *file = iocb->ki_filp;
2235         struct address_space *mapping = file->f_mapping;
2236         ssize_t retval;
2237
2238         retval = filemap_write_and_wait(mapping);
2239         if (retval == 0) {
2240                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2241                                                 offset, nr_segs);
2242                 if (rw == WRITE && mapping->nrpages)
2243                         invalidate_inode_pages2(mapping);
2244         }
2245         return retval;
2246 }
2247 EXPORT_SYMBOL_GPL(generic_file_direct_IO);