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  * FIXME: remove all knowledge of the buffer layer from the core VM
  33  */
  34 #include <linux/buffer_head.h> /* for generic_osync_inode */
  35
  36 #include <asm/uaccess.h>
  37 #include <asm/mman.h>
  38
  39 /*
  40  * Shared mappings implemented 30.11.1994. It's not fully working yet,
  41  * though.
  42  *
  43  * Shared mappings now work. 15.8.1995  Bruno.
  44  *
  45  * finished 'unifying' the page and buffer cache and SMP-threaded the
  46  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
  47  *
  48  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
  49  */
  50
  51 /*
  52  * Lock ordering:
  53  *
  54  *  ->i_mmap_lock               (vmtruncate)
  55  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
  56  *      ->swap_list_lock
  57  *        ->swap_device_lock    (exclusive_swap_page, others)
  58  *          ->mapping->tree_lock
  59  *
  60  *  ->i_sem
  61  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
  62  *
  63  *  ->mmap_sem
  64  *    ->i_mmap_lock
  65  *      ->page_table_lock       (various places, mainly in mmap.c)
  66  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
  67  *
  68  *  ->mmap_sem
  69  *    ->lock_page               (access_process_vm)
  70  *
  71  *  ->mmap_sem
  72  *    ->i_sem                   (msync)
  73  *
  74  *  ->i_sem
  75  *    ->i_alloc_sem             (various)
  76  *
  77  *  ->inode_lock
  78  *    ->sb_lock                 (fs/fs-writeback.c)
  79  *    ->mapping->tree_lock      (__sync_single_inode)
  80  *
  81  *  ->i_mmap_lock
  82  *    ->anon_vma.lock           (vma_adjust)
  83  *
  84  *  ->anon_vma.lock
  85  *    ->page_table_lock         (anon_vma_prepare and various)
  86  *
  87  *  ->page_table_lock
  88  *    ->swap_device_lock        (try_to_unmap_one)
  89  *    ->private_lock            (try_to_unmap_one)
  90  *    ->tree_lock               (try_to_unmap_one)
  91  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
  92  *    ->private_lock            (page_remove_rmap->set_page_dirty)
  93  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
  94  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
  95  *    ->inode_lock              (zap_pte_range->set_page_dirty)
  96  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
  97  *
  98  *  ->task->proc_lock
  99  *    ->dcache_lock             (proc_pid_lookup)
 100  */
 101
 102 /*
 103  * Remove a page from the page cache and free it. Caller has to make
 104  * sure the page is locked and that nobody else uses it - or that usage
 105  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
 106  */
 107 void __remove_from_page_cache(struct page *page)
 108 {
 109         struct address_space *mapping = page->mapping;
 110
 111         radix_tree_delete(&mapping->page_tree, page->index);
 112         page->mapping = NULL;
 113         mapping->nrpages--;
 114         pagecache_acct(-1);
 115 }
 116
 117 void remove_from_page_cache(struct page *page)
 118 {
 119         struct address_space *mapping = page->mapping;
 120
 121         BUG_ON(!PageLocked(page));
 122
 123         write_lock_irq(&mapping->tree_lock);
 124         __remove_from_page_cache(page);
 125         write_unlock_irq(&mapping->tree_lock);
 126 }
 127
 128 static int sync_page(void *word)
 129 {
 130         struct address_space *mapping;
 131         struct page *page;
 132
 133         page = container_of((page_flags_t *)word, struct page, flags);
 134
 135         /*
 136          * page_mapping() is being called without PG_locked held.
 137          * Some knowledge of the state and use of the page is used to
 138          * reduce the requirements down to a memory barrier.
 139          * The danger here is of a stale page_mapping() return value
 140          * indicating a struct address_space different from the one it's
 141          * associated with when it is associated with one.
 142          * After smp_mb(), it's either the correct page_mapping() for
 143          * the page, or an old page_mapping() and the page's own
 144          * page_mapping() has gone NULL.
 145          * The ->sync_page() address_space operation must tolerate
 146          * page_mapping() going NULL. By an amazing coincidence,
 147          * this comes about because none of the users of the page
 148          * in the ->sync_page() methods make essential use of the
 149          * page_mapping(), merely passing the page down to the backing
 150          * device's unplug functions when it's non-NULL, which in turn
 151          * ignore it for all cases but swap, where only page->private is
 152          * of interest. When page_mapping() does go NULL, the entire
 153          * call stack gracefully ignores the page and returns.
 154          * -- wli
 155          */
 156         smp_mb();
 157         mapping = page_mapping(page);
 158         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 159                 mapping->a_ops->sync_page(page);
 160         io_schedule();
 161         return 0;
 162 }
 163
 164 /**
 165  * filemap_fdatawrite_range - start writeback against all of a mapping's
 166  * dirty pages that lie within the byte offsets <start, end>
 167  * @mapping:    address space structure to write
 168  * @start:      offset in bytes where the range starts
 169  * @end:        offset in bytes where the range ends
 170  * @sync_mode:  enable synchronous operation
 171  *
 172  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 173  * opposed to a regular memory * cleansing writeback.  The difference between
 174  * these two operations is that if a dirty page/buffer is encountered, it must
 175  * be waited upon, and not just skipped over.
 176  */
 177 static int __filemap_fdatawrite_range(struct address_space *mapping,
 178         loff_t start, loff_t end, int sync_mode)
 179 {
 180         int ret;
 181         struct writeback_control wbc = {
 182                 .sync_mode = sync_mode,
 183                 .nr_to_write = mapping->nrpages * 2,
 184                 .start = start,
 185                 .end = end,
 186         };
 187
 188         if (!mapping_cap_writeback_dirty(mapping))
 189                 return 0;
 190
 191         ret = do_writepages(mapping, &wbc);
 192         return ret;
 193 }
 194
 195 static inline int __filemap_fdatawrite(struct address_space *mapping,
 196         int sync_mode)
 197 {
 198         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
 199 }
 200
 201 int filemap_fdatawrite(struct address_space *mapping)
 202 {
 203         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 204 }
 205 EXPORT_SYMBOL(filemap_fdatawrite);
 206
 207 static int filemap_fdatawrite_range(struct address_space *mapping,
 208         loff_t start, loff_t end)
 209 {
 210         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
 211 }
 212
 213 /*
 214  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 215  * purposes - I/O may not be started against all dirty pages.
 216  */
 217 int filemap_flush(struct address_space *mapping)
 218 {
 219         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 220 }
 221 EXPORT_SYMBOL(filemap_flush);
 222
 223 /*
 224  * Wait for writeback to complete against pages indexed by start->end
 225  * inclusive
 226  */
 227 static int wait_on_page_writeback_range(struct address_space *mapping,
 228                                 pgoff_t start, pgoff_t end)
 229 {
 230         struct pagevec pvec;
 231         int nr_pages;
 232         int ret = 0;
 233         pgoff_t index;
 234
 235         if (end < start)
 236                 return 0;
 237
 238         pagevec_init(&pvec, 0);
 239         index = start;
 240         while ((index <= end) &&
 241                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 242                         PAGECACHE_TAG_WRITEBACK,
 243                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 244                 unsigned i;
 245
 246                 for (i = 0; i < nr_pages; i++) {
 247                         struct page *page = pvec.pages[i];
 248
 249                         /* until radix tree lookup accepts end_index */
 250                         if (page->index > end)
 251                                 continue;
 252
 253                         wait_on_page_writeback(page);
 254                         if (PageError(page))
 255                                 ret = -EIO;
 256                 }
 257                 pagevec_release(&pvec);
 258                 cond_resched();
 259         }
 260
 261         /* Check for outstanding write errors */
 262         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 263                 ret = -ENOSPC;
 264         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 265                 ret = -EIO;
 266
 267         return ret;
 268 }
 269
 270 /*
 271  * Write and wait upon all the pages in the passed range.  This is a "data
 272  * integrity" operation.  It waits upon in-flight writeout before starting and
 273  * waiting upon new writeout.  If there was an IO error, return it.
 274  *
 275  * We need to re-take i_sem during the generic_osync_inode list walk because
 276  * it is otherwise livelockable.
 277  */
 278 int sync_page_range(struct inode *inode, struct address_space *mapping,
 279                         loff_t pos, size_t count)
 280 {
 281         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 282         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 283         int ret;
 284
 285         if (!mapping_cap_writeback_dirty(mapping) || !count)
 286                 return 0;
 287         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 288         if (ret == 0) {
 289                 down(&inode->i_sem);
 290                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 291                 up(&inode->i_sem);
 292         }
 293         if (ret == 0)
 294                 ret = wait_on_page_writeback_range(mapping, start, end);
 295         return ret;
 296 }
 297 EXPORT_SYMBOL(sync_page_range);
 298
 299 /*
 300  * Note: Holding i_sem across sync_page_range_nolock is not a good idea
 301  * as it forces O_SYNC writers to different parts of the same file
 302  * to be serialised right until io completion.
 303  */
 304 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
 305                         loff_t pos, size_t count)
 306 {
 307         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 308         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 309         int ret;
 310
 311         if (!mapping_cap_writeback_dirty(mapping) || !count)
 312                 return 0;
 313         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 314         if (ret == 0)
 315                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 316         if (ret == 0)
 317                 ret = wait_on_page_writeback_range(mapping, start, end);
 318         return ret;
 319 }
 320 EXPORT_SYMBOL(sync_page_range_nolock);
 321
 322 /**
 323  * filemap_fdatawait - walk the list of under-writeback pages of the given
 324  *     address space and wait for all of them.
 325  *
 326  * @mapping: address space structure to wait for
 327  */
 328 int filemap_fdatawait(struct address_space *mapping)
 329 {
 330         loff_t i_size = i_size_read(mapping->host);
 331
 332         if (i_size == 0)
 333                 return 0;
 334
 335         return wait_on_page_writeback_range(mapping, 0,
 336                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 337 }
 338 EXPORT_SYMBOL(filemap_fdatawait);
 339
 340 int filemap_write_and_wait(struct address_space *mapping)
 341 {
 342         int retval = 0;
 343
 344         if (mapping->nrpages) {
 345                 retval = filemap_fdatawrite(mapping);
 346                 if (retval == 0)
 347                         retval = filemap_fdatawait(mapping);
 348         }
 349         return retval;
 350 }
 351
 352 int filemap_write_and_wait_range(struct address_space *mapping,
 353                                  loff_t lstart, loff_t lend)
 354 {
 355         int retval = 0;
 356
 357         if (mapping->nrpages) {
 358                 retval = __filemap_fdatawrite_range(mapping, lstart, lend,
 359                                                     WB_SYNC_ALL);
 360                 if (retval == 0)
 361                         retval = wait_on_page_writeback_range(mapping,
 362                                                     lstart >> PAGE_CACHE_SHIFT,
 363                                                     lend >> PAGE_CACHE_SHIFT);
 364         }
 365         return retval;
 366 }
 367
 368 /*
 369  * This function is used to add newly allocated pagecache pages:
 370  * the page is new, so we can just run SetPageLocked() against it.
 371  * The other page state flags were set by rmqueue().
 372  *
 373  * This function does not add the page to the LRU.  The caller must do that.
 374  */
 375 int add_to_page_cache(struct page *page, struct address_space *mapping,
 376                 pgoff_t offset, int gfp_mask)
 377 {
 378         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 379
 380         if (error == 0) {
 381                 write_lock_irq(&mapping->tree_lock);
 382                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 383                 if (!error) {
 384                         page_cache_get(page);
 385                         SetPageLocked(page);
 386                         page->mapping = mapping;
 387                         page->index = offset;
 388                         mapping->nrpages++;
 389                         pagecache_acct(1);
 390                 }
 391                 write_unlock_irq(&mapping->tree_lock);
 392                 radix_tree_preload_end();
 393         }
 394         return error;
 395 }
 396
 397 EXPORT_SYMBOL(add_to_page_cache);
 398
 399 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 400                                 pgoff_t offset, int gfp_mask)
 401 {
 402         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 403         if (ret == 0)
 404                 lru_cache_add(page);
 405         return ret;
 406 }
 407
 408 /*
 409  * In order to wait for pages to become available there must be
 410  * waitqueues associated with pages. By using a hash table of
 411  * waitqueues where the bucket discipline is to maintain all
 412  * waiters on the same queue and wake all when any of the pages
 413  * become available, and for the woken contexts to check to be
 414  * sure the appropriate page became available, this saves space
 415  * at a cost of "thundering herd" phenomena during rare hash
 416  * collisions.
 417  */
 418 static wait_queue_head_t *page_waitqueue(struct page *page)
 419 {
 420         const struct zone *zone = page_zone(page);
 421
 422         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 423 }
 424
 425 static inline void wake_up_page(struct page *page, int bit)
 426 {
 427         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 428 }
 429
 430 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 431 {
 432         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 433
 434         if (test_bit(bit_nr, &page->flags))
 435                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 436                                                         TASK_UNINTERRUPTIBLE);
 437 }
 438 EXPORT_SYMBOL(wait_on_page_bit);
 439
 440 /**
 441  * unlock_page() - unlock a locked page
 442  *
 443  * @page: the page
 444  *
 445  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 446  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 447  * mechananism between PageLocked pages and PageWriteback pages is shared.
 448  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 449  *
 450  * The first mb is necessary to safely close the critical section opened by the
 451  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 452  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 453  * parallel wait_on_page_locked()).
 454  */
 455 void fastcall unlock_page(struct page *page)
 456 {
 457         smp_mb__before_clear_bit();
 458         if (!TestClearPageLocked(page))
 459                 BUG();
 460         smp_mb__after_clear_bit();
 461         wake_up_page(page, PG_locked);
 462 }
 463 EXPORT_SYMBOL(unlock_page);
 464
 465 /*
 466  * End writeback against a page.
 467  */
 468 void end_page_writeback(struct page *page)
 469 {
 470         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 471                 if (!test_clear_page_writeback(page))
 472                         BUG();
 473         }
 474         smp_mb__after_clear_bit();
 475         wake_up_page(page, PG_writeback);
 476 }
 477 EXPORT_SYMBOL(end_page_writeback);
 478
 479 /*
 480  * Get a lock on the page, assuming we need to sleep to get it.
 481  *
 482  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 483  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 484  * chances are that on the second loop, the block layer's plug list is empty,
 485  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 486  */
 487 void fastcall __lock_page(struct page *page)
 488 {
 489         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 490
 491         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 492                                                         TASK_UNINTERRUPTIBLE);
 493 }
 494 EXPORT_SYMBOL(__lock_page);
 495
 496 /*
 497  * a rather lightweight function, finding and getting a reference to a
 498  * hashed page atomically.
 499  */
 500 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 501 {
 502         struct page *page;
 503
 504         read_lock_irq(&mapping->tree_lock);
 505         page = radix_tree_lookup(&mapping->page_tree, offset);
 506         if (page)
 507                 page_cache_get(page);
 508         read_unlock_irq(&mapping->tree_lock);
 509         return page;
 510 }
 511
 512 EXPORT_SYMBOL(find_get_page);
 513
 514 /*
 515  * Same as above, but trylock it instead of incrementing the count.
 516  */
 517 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 518 {
 519         struct page *page;
 520
 521         read_lock_irq(&mapping->tree_lock);
 522         page = radix_tree_lookup(&mapping->page_tree, offset);
 523         if (page && TestSetPageLocked(page))
 524                 page = NULL;
 525         read_unlock_irq(&mapping->tree_lock);
 526         return page;
 527 }
 528
 529 EXPORT_SYMBOL(find_trylock_page);
 530
 531 /**
 532  * find_lock_page - locate, pin and lock a pagecache page
 533  *
 534  * @mapping: the address_space to search
 535  * @offset: the page index
 536  *
 537  * Locates the desired pagecache page, locks it, increments its reference
 538  * count and returns its address.
 539  *
 540  * Returns zero if the page was not present. find_lock_page() may sleep.
 541  */
 542 struct page *find_lock_page(struct address_space *mapping,
 543                                 unsigned long offset)
 544 {
 545         struct page *page;
 546
 547         read_lock_irq(&mapping->tree_lock);
 548 repeat:
 549         page = radix_tree_lookup(&mapping->page_tree, offset);
 550         if (page) {
 551                 page_cache_get(page);
 552                 if (TestSetPageLocked(page)) {
 553                         read_unlock_irq(&mapping->tree_lock);
 554                         lock_page(page);
 555                         read_lock_irq(&mapping->tree_lock);
 556
 557                         /* Has the page been truncated while we slept? */
 558                         if (page->mapping != mapping || page->index != offset) {
 559                                 unlock_page(page);
 560                                 page_cache_release(page);
 561                                 goto repeat;
 562                         }
 563                 }
 564         }
 565         read_unlock_irq(&mapping->tree_lock);
 566         return page;
 567 }
 568
 569 EXPORT_SYMBOL(find_lock_page);
 570
 571 /**
 572  * find_or_create_page - locate or add a pagecache page
 573  *
 574  * @mapping: the page's address_space
 575  * @index: the page's index into the mapping
 576  * @gfp_mask: page allocation mode
 577  *
 578  * Locates a page in the pagecache.  If the page is not present, a new page
 579  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 580  * LRU list.  The returned page is locked and has its reference count
 581  * incremented.
 582  *
 583  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 584  * allocation!
 585  *
 586  * find_or_create_page() returns the desired page's address, or zero on
 587  * memory exhaustion.
 588  */
 589 struct page *find_or_create_page(struct address_space *mapping,
 590                 unsigned long index, unsigned int gfp_mask)
 591 {
 592         struct page *page, *cached_page = NULL;
 593         int err;
 594 repeat:
 595         page = find_lock_page(mapping, index);
 596         if (!page) {
 597                 if (!cached_page) {
 598                         cached_page = alloc_page(gfp_mask);
 599                         if (!cached_page)
 600                                 return NULL;
 601                 }
 602                 err = add_to_page_cache_lru(cached_page, mapping,
 603                                         index, gfp_mask);
 604                 if (!err) {
 605                         page = cached_page;
 606                         cached_page = NULL;
 607                 } else if (err == -EEXIST)
 608                         goto repeat;
 609         }
 610         if (cached_page)
 611                 page_cache_release(cached_page);
 612         return page;
 613 }
 614
 615 EXPORT_SYMBOL(find_or_create_page);
 616
 617 /**
 618  * find_get_pages - gang pagecache lookup
 619  * @mapping:    The address_space to search
 620  * @start:      The starting page index
 621  * @nr_pages:   The maximum number of pages
 622  * @pages:      Where the resulting pages are placed
 623  *
 624  * find_get_pages() will search for and return a group of up to
 625  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 626  * find_get_pages() takes a reference against the returned pages.
 627  *
 628  * The search returns a group of mapping-contiguous pages with ascending
 629  * indexes.  There may be holes in the indices due to not-present pages.
 630  *
 631  * find_get_pages() returns the number of pages which were found.
 632  */
 633 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 634                             unsigned int nr_pages, struct page **pages)
 635 {
 636         unsigned int i;
 637         unsigned int ret;
 638
 639         read_lock_irq(&mapping->tree_lock);
 640         ret = radix_tree_gang_lookup(&mapping->page_tree,
 641                                 (void **)pages, start, nr_pages);
 642         for (i = 0; i < ret; i++)
 643                 page_cache_get(pages[i]);
 644         read_unlock_irq(&mapping->tree_lock);
 645         return ret;
 646 }
 647
 648 /*
 649  * Like find_get_pages, except we only return pages which are tagged with
 650  * `tag'.   We update *index to index the next page for the traversal.
 651  */
 652 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 653                         int tag, unsigned int nr_pages, struct page **pages)
 654 {
 655         unsigned int i;
 656         unsigned int ret;
 657
 658         read_lock_irq(&mapping->tree_lock);
 659         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 660                                 (void **)pages, *index, nr_pages, tag);
 661         for (i = 0; i < ret; i++)
 662                 page_cache_get(pages[i]);
 663         if (ret)
 664                 *index = pages[ret - 1]->index + 1;
 665         read_unlock_irq(&mapping->tree_lock);
 666         return ret;
 667 }
 668
 669 /*
 670  * Same as grab_cache_page, but do not wait if the page is unavailable.
 671  * This is intended for speculative data generators, where the data can
 672  * be regenerated if the page couldn't be grabbed.  This routine should
 673  * be safe to call while holding the lock for another page.
 674  *
 675  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 676  * and deadlock against the caller's locked page.
 677  */
 678 struct page *
 679 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 680 {
 681         struct page *page = find_get_page(mapping, index);
 682         unsigned int gfp_mask;
 683
 684         if (page) {
 685                 if (!TestSetPageLocked(page))
 686                         return page;
 687                 page_cache_release(page);
 688                 return NULL;
 689         }
 690         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 691         page = alloc_pages(gfp_mask, 0);
 692         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 693                 page_cache_release(page);
 694                 page = NULL;
 695         }
 696         return page;
 697 }
 698
 699 EXPORT_SYMBOL(grab_cache_page_nowait);
 700
 701 /*
 702  * This is a generic file read routine, and uses the
 703  * mapping->a_ops->readpage() function for the actual low-level
 704  * stuff.
 705  *
 706  * This is really ugly. But the goto's actually try to clarify some
 707  * of the logic when it comes to error handling etc.
 708  *
 709  * Note the struct file* is only passed for the use of readpage.  It may be
 710  * NULL.
 711  */
 712 void do_generic_mapping_read(struct address_space *mapping,
 713                              struct file_ra_state *_ra,
 714                              struct file *filp,
 715                              loff_t *ppos,
 716                              read_descriptor_t *desc,
 717                              read_actor_t actor)
 718 {
 719         struct inode *inode = mapping->host;
 720         unsigned long index;
 721         unsigned long end_index;
 722         unsigned long offset;
 723         unsigned long last_index;
 724         unsigned long next_index;
 725         unsigned long prev_index;
 726         loff_t isize;
 727         struct page *cached_page;
 728         int error;
 729         struct file_ra_state ra = *_ra;
 730
 731         cached_page = NULL;
 732         index = *ppos >> PAGE_CACHE_SHIFT;
 733         next_index = index;
 734         prev_index = ra.prev_page;
 735         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 736         offset = *ppos & ~PAGE_CACHE_MASK;
 737
 738         isize = i_size_read(inode);
 739         if (!isize)
 740                 goto out;
 741
 742         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 743         for (;;) {
 744                 struct page *page;
 745                 unsigned long nr, ret;
 746
 747                 /* nr is the maximum number of bytes to copy from this page */
 748                 nr = PAGE_CACHE_SIZE;
 749                 if (index >= end_index) {
 750                         if (index > end_index)
 751                                 goto out;
 752                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 753                         if (nr <= offset) {
 754                                 goto out;
 755                         }
 756                 }
 757                 nr = nr - offset;
 758
 759                 cond_resched();
 760                 if (index == next_index)
 761                         next_index = page_cache_readahead(mapping, &ra, filp,
 762                                         index, last_index - index);
 763
 764 find_page:
 765                 page = find_get_page(mapping, index);
 766                 if (unlikely(page == NULL)) {
 767                         handle_ra_miss(mapping, &ra, index);
 768                         goto no_cached_page;
 769                 }
 770                 if (!PageUptodate(page))
 771                         goto page_not_up_to_date;
 772 page_ok:
 773
 774                 /* If users can be writing to this page using arbitrary
 775                  * virtual addresses, take care about potential aliasing
 776                  * before reading the page on the kernel side.
 777                  */
 778                 if (mapping_writably_mapped(mapping))
 779                         flush_dcache_page(page);
 780
 781                 /*
 782                  * When (part of) the same page is read multiple times
 783                  * in succession, only mark it as accessed the first time.
 784                  */
 785                 if (prev_index != index)
 786                         mark_page_accessed(page);
 787                 prev_index = index;
 788
 789                 /*
 790                  * Ok, we have the page, and it's up-to-date, so
 791                  * now we can copy it to user space...
 792                  *
 793                  * The actor routine returns how many bytes were actually used..
 794                  * NOTE! This may not be the same as how much of a user buffer
 795                  * we filled up (we may be padding etc), so we can only update
 796                  * "pos" here (the actor routine has to update the user buffer
 797                  * pointers and the remaining count).
 798                  */
 799                 ret = actor(desc, page, offset, nr);
 800                 offset += ret;
 801                 index += offset >> PAGE_CACHE_SHIFT;
 802                 offset &= ~PAGE_CACHE_MASK;
 803
 804                 page_cache_release(page);
 805                 if (ret == nr && desc->count)
 806                         continue;
 807                 goto out;
 808
 809 page_not_up_to_date:
 810                 /* Get exclusive access to the page ... */
 811                 lock_page(page);
 812
 813                 /* Did it get unhashed before we got the lock? */
 814                 if (!page->mapping) {
 815                         unlock_page(page);
 816                         page_cache_release(page);
 817                         continue;
 818                 }
 819
 820                 /* Did somebody else fill it already? */
 821                 if (PageUptodate(page)) {
 822                         unlock_page(page);
 823                         goto page_ok;
 824                 }
 825
 826 readpage:
 827                 /* Start the actual read. The read will unlock the page. */
 828                 error = mapping->a_ops->readpage(filp, page);
 829
 830                 if (unlikely(error))
 831                         goto readpage_error;
 832
 833                 if (!PageUptodate(page)) {
 834                         lock_page(page);
 835                         if (!PageUptodate(page)) {
 836                                 if (page->mapping == NULL) {
 837                                         /*
 838                                          * invalidate_inode_pages got it
 839                                          */
 840                                         unlock_page(page);
 841                                         page_cache_release(page);
 842                                         goto find_page;
 843                                 }
 844                                 unlock_page(page);
 845                                 error = -EIO;
 846                                 goto readpage_error;
 847                         }
 848                         unlock_page(page);
 849                 }
 850
 851                 /*
 852                  * i_size must be checked after we have done ->readpage.
 853                  *
 854                  * Checking i_size after the readpage allows us to calculate
 855                  * the correct value for "nr", which means the zero-filled
 856                  * part of the page is not copied back to userspace (unless
 857                  * another truncate extends the file - this is desired though).
 858                  */
 859                 isize = i_size_read(inode);
 860                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 861                 if (unlikely(!isize || index > end_index)) {
 862                         page_cache_release(page);
 863                         goto out;
 864                 }
 865
 866                 /* nr is the maximum number of bytes to copy from this page */
 867                 nr = PAGE_CACHE_SIZE;
 868                 if (index == end_index) {
 869                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 870                         if (nr <= offset) {
 871                                 page_cache_release(page);
 872                                 goto out;
 873                         }
 874                 }
 875                 nr = nr - offset;
 876                 goto page_ok;
 877
 878 readpage_error:
 879                 /* UHHUH! A synchronous read error occurred. Report it */
 880                 desc->error = error;
 881                 page_cache_release(page);
 882                 goto out;
 883
 884 no_cached_page:
 885                 /*
 886                  * Ok, it wasn't cached, so we need to create a new
 887                  * page..
 888                  */
 889                 if (!cached_page) {
 890                         cached_page = page_cache_alloc_cold(mapping);
 891                         if (!cached_page) {
 892                                 desc->error = -ENOMEM;
 893                                 goto out;
 894                         }
 895                 }
 896                 error = add_to_page_cache_lru(cached_page, mapping,
 897                                                 index, GFP_KERNEL);
 898                 if (error) {
 899                         if (error == -EEXIST)
 900                                 goto find_page;
 901                         desc->error = error;
 902                         goto out;
 903                 }
 904                 page = cached_page;
 905                 cached_page = NULL;
 906                 goto readpage;
 907         }
 908
 909 out:
 910         *_ra = ra;
 911
 912         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 913         if (cached_page)
 914                 page_cache_release(cached_page);
 915         if (filp)
 916                 file_accessed(filp);
 917 }
 918
 919 EXPORT_SYMBOL(do_generic_mapping_read);
 920
 921 int file_read_actor(read_descriptor_t *desc, struct page *page,
 922                         unsigned long offset, unsigned long size)
 923 {
 924         char *kaddr;
 925         unsigned long left, count = desc->count;
 926
 927         if (size > count)
 928                 size = count;
 929
 930         /*
 931          * Faults on the destination of a read are common, so do it before
 932          * taking the kmap.
 933          */
 934         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 935                 kaddr = kmap_atomic(page, KM_USER0);
 936                 left = __copy_to_user_inatomic(desc->arg.buf,
 937                                                 kaddr + offset, size);
 938                 kunmap_atomic(kaddr, KM_USER0);
 939                 if (left == 0)
 940                         goto success;
 941         }
 942
 943         /* Do it the slow way */
 944         kaddr = kmap(page);
 945         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 946         kunmap(page);
 947
 948         if (left) {
 949                 size -= left;
 950                 desc->error = -EFAULT;
 951         }
 952 success:
 953         desc->count = count - size;
 954         desc->written += size;
 955         desc->arg.buf += size;
 956         return size;
 957 }
 958
 959 /*
 960  * This is the "read()" routine for all filesystems
 961  * that can use the page cache directly.
 962  */
 963 ssize_t
 964 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 965                 unsigned long nr_segs, loff_t *ppos)
 966 {
 967         struct file *filp = iocb->ki_filp;
 968         ssize_t retval;
 969         unsigned long seg;
 970         size_t count;
 971
 972         count = 0;
 973         for (seg = 0; seg < nr_segs; seg++) {
 974                 const struct iovec *iv = &iov[seg];
 975
 976                 /*
 977                  * If any segment has a negative length, or the cumulative
 978                  * length ever wraps negative then return -EINVAL.
 979                  */
 980                 count += iv->iov_len;
 981                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
 982                         return -EINVAL;
 983                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
 984                         continue;
 985                 if (seg == 0)
 986                         return -EFAULT;
 987                 nr_segs = seg;
 988                 count -= iv->iov_len;   /* This segment is no good */
 989                 break;
 990         }
 991
 992         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
 993         if (filp->f_flags & O_DIRECT) {
 994                 loff_t pos = *ppos, size;
 995                 struct address_space *mapping;
 996                 struct inode *inode;
 997
 998                 mapping = filp->f_mapping;
 999                 inode = mapping->host;
1000                 retval = 0;
1001                 if (!count)
1002                         goto out; /* skip atime */
1003                 size = i_size_read(inode);
1004                 if (pos < size) {
1005                         retval = generic_file_direct_IO(READ, iocb,
1006                                                 iov, pos, nr_segs);
1007                         if (retval > 0 && !is_sync_kiocb(iocb))
1008                                 retval = -EIOCBQUEUED;
1009                         if (retval > 0)
1010                                 *ppos = pos + retval;
1011                 }
1012                 file_accessed(filp);
1013                 goto out;
1014         }
1015
1016         retval = 0;
1017         if (count) {
1018                 for (seg = 0; seg < nr_segs; seg++) {
1019                         read_descriptor_t desc;
1020
1021                         desc.written = 0;
1022                         desc.arg.buf = iov[seg].iov_base;
1023                         desc.count = iov[seg].iov_len;
1024                         if (desc.count == 0)
1025                                 continue;
1026                         desc.error = 0;
1027                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1028                         retval += desc.written;
1029                         if (!retval) {
1030                                 retval = desc.error;
1031                                 break;
1032                         }
1033                 }
1034         }
1035 out:
1036         return retval;
1037 }
1038
1039 EXPORT_SYMBOL(__generic_file_aio_read);
1040
1041 ssize_t
1042 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1043 {
1044         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1045
1046         BUG_ON(iocb->ki_pos != pos);
1047         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1048 }
1049
1050 EXPORT_SYMBOL(generic_file_aio_read);
1051
1052 ssize_t
1053 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1054 {
1055         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1056         struct kiocb kiocb;
1057         ssize_t ret;
1058
1059         init_sync_kiocb(&kiocb, filp);
1060         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1061         if (-EIOCBQUEUED == ret)
1062                 ret = wait_on_sync_kiocb(&kiocb);
1063         return ret;
1064 }
1065
1066 EXPORT_SYMBOL(generic_file_read);
1067
1068 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1069 {
1070         ssize_t written;
1071         unsigned long count = desc->count;
1072         struct file *file = desc->arg.data;
1073
1074         if (size > count)
1075                 size = count;
1076
1077         written = file->f_op->sendpage(file, page, offset,
1078                                        size, &file->f_pos, size<count);
1079         if (written < 0) {
1080                 desc->error = written;
1081                 written = 0;
1082         }
1083         desc->count = count - written;
1084         desc->written += written;
1085         return written;
1086 }
1087
1088 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1089                          size_t count, read_actor_t actor, void *target)
1090 {
1091         read_descriptor_t desc;
1092
1093         if (!count)
1094                 return 0;
1095
1096         desc.written = 0;
1097         desc.count = count;
1098         desc.arg.data = target;
1099         desc.error = 0;
1100
1101         do_generic_file_read(in_file, ppos, &desc, actor);
1102         if (desc.written)
1103                 return desc.written;
1104         return desc.error;
1105 }
1106
1107 EXPORT_SYMBOL(generic_file_sendfile);
1108
1109 static ssize_t
1110 do_readahead(struct address_space *mapping, struct file *filp,
1111              unsigned long index, unsigned long nr)
1112 {
1113         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1114                 return -EINVAL;
1115
1116         force_page_cache_readahead(mapping, filp, index,
1117                                         max_sane_readahead(nr));
1118         return 0;
1119 }
1120
1121 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1122 {
1123         ssize_t ret;
1124         struct file *file;
1125
1126         ret = -EBADF;
1127         file = fget(fd);
1128         if (file) {
1129                 if (file->f_mode & FMODE_READ) {
1130                         struct address_space *mapping = file->f_mapping;
1131                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1132                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1133                         unsigned long len = end - start + 1;
1134                         ret = do_readahead(mapping, file, start, len);
1135                 }
1136                 fput(file);
1137         }
1138         return ret;
1139 }
1140
1141 #ifdef CONFIG_MMU
1142 /*
1143  * This adds the requested page to the page cache if it isn't already there,
1144  * and schedules an I/O to read in its contents from disk.
1145  */
1146 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1147 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1148 {
1149         struct address_space *mapping = file->f_mapping;
1150         struct page *page;
1151         int error;
1152
1153         page = page_cache_alloc_cold(mapping);
1154         if (!page)
1155                 return -ENOMEM;
1156
1157         error = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1158         if (!error) {
1159                 error = mapping->a_ops->readpage(file, page);
1160                 page_cache_release(page);
1161                 return error;
1162         }
1163
1164         /*
1165          * We arrive here in the unlikely event that someone
1166          * raced with us and added our page to the cache first
1167          * or we are out of memory for radix-tree nodes.
1168          */
1169         page_cache_release(page);
1170         return error == -EEXIST ? 0 : error;
1171 }
1172
1173 #define MMAP_LOTSAMISS  (100)
1174
1175 /*
1176  * filemap_nopage() is invoked via the vma operations vector for a
1177  * mapped memory region to read in file data during a page fault.
1178  *
1179  * The goto's are kind of ugly, but this streamlines the normal case of having
1180  * it in the page cache, and handles the special cases reasonably without
1181  * having a lot of duplicated code.
1182  */
1183 struct page *filemap_nopage(struct vm_area_struct *area,
1184                                 unsigned long address, int *type)
1185 {
1186         int error;
1187         struct file *file = area->vm_file;
1188         struct address_space *mapping = file->f_mapping;
1189         struct file_ra_state *ra = &file->f_ra;
1190         struct inode *inode = mapping->host;
1191         struct page *page;
1192         unsigned long size, pgoff;
1193         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1194
1195         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1196
1197 retry_all:
1198         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1199         if (pgoff >= size)
1200                 goto outside_data_content;
1201
1202         /* If we don't want any read-ahead, don't bother */
1203         if (VM_RandomReadHint(area))
1204                 goto no_cached_page;
1205
1206         /*
1207          * The readahead code wants to be told about each and every page
1208          * so it can build and shrink its windows appropriately
1209          *
1210          * For sequential accesses, we use the generic readahead logic.
1211          */
1212         if (VM_SequentialReadHint(area))
1213                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1214
1215         /*
1216          * Do we have something in the page cache already?
1217          */
1218 retry_find:
1219         page = find_get_page(mapping, pgoff);
1220         if (!page) {
1221                 unsigned long ra_pages;
1222
1223                 if (VM_SequentialReadHint(area)) {
1224                         handle_ra_miss(mapping, ra, pgoff);
1225                         goto no_cached_page;
1226                 }
1227                 ra->mmap_miss++;
1228
1229                 /*
1230                  * Do we miss much more than hit in this file? If so,
1231                  * stop bothering with read-ahead. It will only hurt.
1232                  */
1233                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1234                         goto no_cached_page;
1235
1236                 /*
1237                  * To keep the pgmajfault counter straight, we need to
1238                  * check did_readaround, as this is an inner loop.
1239                  */
1240                 if (!did_readaround) {
1241                         majmin = VM_FAULT_MAJOR;
1242                         inc_page_state(pgmajfault);
1243                 }
1244                 did_readaround = 1;
1245                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1246                 if (ra_pages) {
1247                         pgoff_t start = 0;
1248
1249                         if (pgoff > ra_pages / 2)
1250                                 start = pgoff - ra_pages / 2;
1251                         do_page_cache_readahead(mapping, file, start, ra_pages);
1252                 }
1253                 page = find_get_page(mapping, pgoff);
1254                 if (!page)
1255                         goto no_cached_page;
1256         }
1257
1258         if (!did_readaround)
1259                 ra->mmap_hit++;
1260
1261         /*
1262          * Ok, found a page in the page cache, now we need to check
1263          * that it's up-to-date.
1264          */
1265         if (!PageUptodate(page))
1266                 goto page_not_uptodate;
1267
1268 success:
1269         /*
1270          * Found the page and have a reference on it.
1271          */
1272         mark_page_accessed(page);
1273         if (type)
1274                 *type = majmin;
1275         return page;
1276
1277 outside_data_content:
1278         /*
1279          * An external ptracer can access pages that normally aren't
1280          * accessible..
1281          */
1282         if (area->vm_mm == current->mm)
1283                 return NULL;
1284         /* Fall through to the non-read-ahead case */
1285 no_cached_page:
1286         /*
1287          * We're only likely to ever get here if MADV_RANDOM is in
1288          * effect.
1289          */
1290         error = page_cache_read(file, pgoff);
1291         grab_swap_token();
1292
1293         /*
1294          * The page we want has now been added to the page cache.
1295          * In the unlikely event that someone removed it in the
1296          * meantime, we'll just come back here and read it again.
1297          */
1298         if (error >= 0)
1299                 goto retry_find;
1300
1301         /*
1302          * An error return from page_cache_read can result if the
1303          * system is low on memory, or a problem occurs while trying
1304          * to schedule I/O.
1305          */
1306         if (error == -ENOMEM)
1307                 return NOPAGE_OOM;
1308         return NULL;
1309
1310 page_not_uptodate:
1311         if (!did_readaround) {
1312                 majmin = VM_FAULT_MAJOR;
1313                 inc_page_state(pgmajfault);
1314         }
1315         lock_page(page);
1316
1317         /* Did it get unhashed while we waited for it? */
1318         if (!page->mapping) {
1319                 unlock_page(page);
1320                 page_cache_release(page);
1321                 goto retry_all;
1322         }
1323
1324         /* Did somebody else get it up-to-date? */
1325         if (PageUptodate(page)) {
1326                 unlock_page(page);
1327                 goto success;
1328         }
1329
1330         if (!mapping->a_ops->readpage(file, page)) {
1331                 wait_on_page_locked(page);
1332                 if (PageUptodate(page))
1333                         goto success;
1334         }
1335
1336         /*
1337          * Umm, take care of errors if the page isn't up-to-date.
1338          * Try to re-read it _once_. We do this synchronously,
1339          * because there really aren't any performance issues here
1340          * and we need to check for errors.
1341          */
1342         lock_page(page);
1343
1344         /* Somebody truncated the page on us? */
1345         if (!page->mapping) {
1346                 unlock_page(page);
1347                 page_cache_release(page);
1348                 goto retry_all;
1349         }
1350
1351         /* Somebody else successfully read it in? */
1352         if (PageUptodate(page)) {
1353                 unlock_page(page);
1354                 goto success;
1355         }
1356         ClearPageError(page);
1357         if (!mapping->a_ops->readpage(file, page)) {
1358                 wait_on_page_locked(page);
1359                 if (PageUptodate(page))
1360                         goto success;
1361         }
1362
1363         /*
1364          * Things didn't work out. Return zero to tell the
1365          * mm layer so, possibly freeing the page cache page first.
1366          */
1367         page_cache_release(page);
1368         return NULL;
1369 }
1370
1371 EXPORT_SYMBOL(filemap_nopage);
1372
1373 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1374                                         int nonblock)
1375 {
1376         struct address_space *mapping = file->f_mapping;
1377         struct page *page;
1378         int error;
1379
1380         /*
1381          * Do we have something in the page cache already?
1382          */
1383 retry_find:
1384         page = find_get_page(mapping, pgoff);
1385         if (!page) {
1386                 if (nonblock)
1387                         return NULL;
1388                 goto no_cached_page;
1389         }
1390
1391         /*
1392          * Ok, found a page in the page cache, now we need to check
1393          * that it's up-to-date.
1394          */
1395         if (!PageUptodate(page)) {
1396                 if (nonblock) {
1397                         page_cache_release(page);
1398                         return NULL;
1399                 }
1400                 goto page_not_uptodate;
1401         }
1402
1403 success:
1404         /*
1405          * Found the page and have a reference on it.
1406          */
1407         mark_page_accessed(page);
1408         return page;
1409
1410 no_cached_page:
1411         error = page_cache_read(file, pgoff);
1412
1413         /*
1414          * The page we want has now been added to the page cache.
1415          * In the unlikely event that someone removed it in the
1416          * meantime, we'll just come back here and read it again.
1417          */
1418         if (error >= 0)
1419                 goto retry_find;
1420
1421         /*
1422          * An error return from page_cache_read can result if the
1423          * system is low on memory, or a problem occurs while trying
1424          * to schedule I/O.
1425          */
1426         return NULL;
1427
1428 page_not_uptodate:
1429         lock_page(page);
1430
1431         /* Did it get unhashed while we waited for it? */
1432         if (!page->mapping) {
1433                 unlock_page(page);
1434                 goto err;
1435         }
1436
1437         /* Did somebody else get it up-to-date? */
1438         if (PageUptodate(page)) {
1439                 unlock_page(page);
1440                 goto success;
1441         }
1442
1443         if (!mapping->a_ops->readpage(file, page)) {
1444                 wait_on_page_locked(page);
1445                 if (PageUptodate(page))
1446                         goto success;
1447         }
1448
1449         /*
1450          * Umm, take care of errors if the page isn't up-to-date.
1451          * Try to re-read it _once_. We do this synchronously,
1452          * because there really aren't any performance issues here
1453          * and we need to check for errors.
1454          */
1455         lock_page(page);
1456
1457         /* Somebody truncated the page on us? */
1458         if (!page->mapping) {
1459                 unlock_page(page);
1460                 goto err;
1461         }
1462         /* Somebody else successfully read it in? */
1463         if (PageUptodate(page)) {
1464                 unlock_page(page);
1465                 goto success;
1466         }
1467
1468         ClearPageError(page);
1469         if (!mapping->a_ops->readpage(file, page)) {
1470                 wait_on_page_locked(page);
1471                 if (PageUptodate(page))
1472                         goto success;
1473         }
1474
1475         /*
1476          * Things didn't work out. Return zero to tell the
1477          * mm layer so, possibly freeing the page cache page first.
1478          */
1479 err:
1480         page_cache_release(page);
1481
1482         return NULL;
1483 }
1484
1485 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1486                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1487                 int nonblock)
1488 {
1489         struct file *file = vma->vm_file;
1490         struct address_space *mapping = file->f_mapping;
1491         struct inode *inode = mapping->host;
1492         unsigned long size;
1493         struct mm_struct *mm = vma->vm_mm;
1494         struct page *page;
1495         int err;
1496
1497         if (!nonblock)
1498                 force_page_cache_readahead(mapping, vma->vm_file,
1499                                         pgoff, len >> PAGE_CACHE_SHIFT);
1500
1501 repeat:
1502         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1503         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1504                 return -EINVAL;
1505
1506         page = filemap_getpage(file, pgoff, nonblock);
1507         if (!page && !nonblock)
1508                 return -ENOMEM;
1509         if (page) {
1510                 err = install_page(mm, vma, addr, page, prot);
1511                 if (err) {
1512                         page_cache_release(page);
1513                         return err;
1514                 }
1515         } else {
1516                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1517                 if (err)
1518                         return err;
1519         }
1520
1521         len -= PAGE_SIZE;
1522         addr += PAGE_SIZE;
1523         pgoff++;
1524         if (len)
1525                 goto repeat;
1526
1527         return 0;
1528 }
1529
1530 struct vm_operations_struct generic_file_vm_ops = {
1531         .nopage         = filemap_nopage,
1532         .populate       = filemap_populate,
1533 };
1534
1535 /* This is used for a general mmap of a disk file */
1536
1537 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1538 {
1539         struct address_space *mapping = file->f_mapping;
1540
1541         if (!mapping->a_ops->readpage)
1542                 return -ENOEXEC;
1543         file_accessed(file);
1544         vma->vm_ops = &generic_file_vm_ops;
1545         return 0;
1546 }
1547 EXPORT_SYMBOL(filemap_populate);
1548
1549 /*
1550  * This is for filesystems which do not implement ->writepage.
1551  */
1552 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1553 {
1554         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1555                 return -EINVAL;
1556         return generic_file_mmap(file, vma);
1557 }
1558 #else
1559 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1560 {
1561         return -ENOSYS;
1562 }
1563 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1564 {
1565         return -ENOSYS;
1566 }
1567 #endif /* CONFIG_MMU */
1568
1569 EXPORT_SYMBOL(generic_file_mmap);
1570 EXPORT_SYMBOL(generic_file_readonly_mmap);
1571
1572 static inline struct page *__read_cache_page(struct address_space *mapping,
1573                                 unsigned long index,
1574                                 int (*filler)(void *,struct page*),
1575                                 void *data)
1576 {
1577         struct page *page, *cached_page = NULL;
1578         int err;
1579 repeat:
1580         page = find_get_page(mapping, index);
1581         if (!page) {
1582                 if (!cached_page) {
1583                         cached_page = page_cache_alloc_cold(mapping);
1584                         if (!cached_page)
1585                                 return ERR_PTR(-ENOMEM);
1586                 }
1587                 err = add_to_page_cache_lru(cached_page, mapping,
1588                                         index, GFP_KERNEL);
1589                 if (err == -EEXIST)
1590                         goto repeat;
1591                 if (err < 0) {
1592                         /* Presumably ENOMEM for radix tree node */
1593                         page_cache_release(cached_page);
1594                         return ERR_PTR(err);
1595                 }
1596                 page = cached_page;
1597                 cached_page = NULL;
1598                 err = filler(data, page);
1599                 if (err < 0) {
1600                         page_cache_release(page);
1601                         page = ERR_PTR(err);
1602                 }
1603         }
1604         if (cached_page)
1605                 page_cache_release(cached_page);
1606         return page;
1607 }
1608
1609 /*
1610  * Read into the page cache. If a page already exists,
1611  * and PageUptodate() is not set, try to fill the page.
1612  */
1613 struct page *read_cache_page(struct address_space *mapping,
1614                                 unsigned long index,
1615                                 int (*filler)(void *,struct page*),
1616                                 void *data)
1617 {
1618         struct page *page;
1619         int err;
1620
1621 retry:
1622         page = __read_cache_page(mapping, index, filler, data);
1623         if (IS_ERR(page))
1624                 goto out;
1625         mark_page_accessed(page);
1626         if (PageUptodate(page))
1627                 goto out;
1628
1629         lock_page(page);
1630         if (!page->mapping) {
1631                 unlock_page(page);
1632                 page_cache_release(page);
1633                 goto retry;
1634         }
1635         if (PageUptodate(page)) {
1636                 unlock_page(page);
1637                 goto out;
1638         }
1639         err = filler(data, page);
1640         if (err < 0) {
1641                 page_cache_release(page);
1642                 page = ERR_PTR(err);
1643         }
1644  out:
1645         return page;
1646 }
1647
1648 EXPORT_SYMBOL(read_cache_page);
1649
1650 /*
1651  * If the page was newly created, increment its refcount and add it to the
1652  * caller's lru-buffering pagevec.  This function is specifically for
1653  * generic_file_write().
1654  */
1655 static inline struct page *
1656 __grab_cache_page(struct address_space *mapping, unsigned long index,
1657                         struct page **cached_page, struct pagevec *lru_pvec)
1658 {
1659         int err;
1660         struct page *page;
1661 repeat:
1662         page = find_lock_page(mapping, index);
1663         if (!page) {
1664                 if (!*cached_page) {
1665                         *cached_page = page_cache_alloc(mapping);
1666                         if (!*cached_page)
1667                                 return NULL;
1668                 }
1669                 err = add_to_page_cache(*cached_page, mapping,
1670                                         index, GFP_KERNEL);
1671                 if (err == -EEXIST)
1672                         goto repeat;
1673                 if (err == 0) {
1674                         page = *cached_page;
1675                         page_cache_get(page);
1676                         if (!pagevec_add(lru_pvec, page))
1677                                 __pagevec_lru_add(lru_pvec);
1678                         *cached_page = NULL;
1679                 }
1680         }
1681         return page;
1682 }
1683
1684 /*
1685  * The logic we want is
1686  *
1687  *      if suid or (sgid and xgrp)
1688  *              remove privs
1689  */
1690 int remove_suid(struct dentry *dentry)
1691 {
1692         mode_t mode = dentry->d_inode->i_mode;
1693         int kill = 0;
1694         int result = 0;
1695
1696         /* suid always must be killed */
1697         if (unlikely(mode & S_ISUID))
1698                 kill = ATTR_KILL_SUID;
1699
1700         /*
1701          * sgid without any exec bits is just a mandatory locking mark; leave
1702          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1703          */
1704         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1705                 kill |= ATTR_KILL_SGID;
1706
1707         if (unlikely(kill && !capable(CAP_FSETID))) {
1708                 struct iattr newattrs;
1709
1710                 newattrs.ia_valid = ATTR_FORCE | kill;
1711                 result = notify_change(dentry, &newattrs);
1712         }
1713         return result;
1714 }
1715 EXPORT_SYMBOL(remove_suid);
1716
1717 /*
1718  * Copy as much as we can into the page and return the number of bytes which
1719  * were sucessfully copied.  If a fault is encountered then clear the page
1720  * out to (offset+bytes) and return the number of bytes which were copied.
1721  */
1722 static inline size_t
1723 filemap_copy_from_user(struct page *page, unsigned long offset,
1724                         const char __user *buf, unsigned bytes)
1725 {
1726         char *kaddr;
1727         int left;
1728
1729         kaddr = kmap_atomic(page, KM_USER0);
1730         left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1731         kunmap_atomic(kaddr, KM_USER0);
1732
1733         if (left != 0) {
1734                 /* Do it the slow way */
1735                 kaddr = kmap(page);
1736                 left = __copy_from_user(kaddr + offset, buf, bytes);
1737                 kunmap(page);
1738         }
1739         return bytes - left;
1740 }
1741
1742 static size_t
1743 __filemap_copy_from_user_iovec(char *vaddr,
1744                         const struct iovec *iov, size_t base, size_t bytes)
1745 {
1746         size_t copied = 0, left = 0;
1747
1748         while (bytes) {
1749                 char __user *buf = iov->iov_base + base;
1750                 int copy = min(bytes, iov->iov_len - base);
1751
1752                 base = 0;
1753                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1754                 copied += copy;
1755                 bytes -= copy;
1756                 vaddr += copy;
1757                 iov++;
1758
1759                 if (unlikely(left)) {
1760                         /* zero the rest of the target like __copy_from_user */
1761                         if (bytes)
1762                                 memset(vaddr, 0, bytes);
1763                         break;
1764                 }
1765         }
1766         return copied - left;
1767 }
1768
1769 /*
1770  * This has the same sideeffects and return value as filemap_copy_from_user().
1771  * The difference is that on a fault we need to memset the remainder of the
1772  * page (out to offset+bytes), to emulate filemap_copy_from_user()'s
1773  * single-segment behaviour.
1774  */
1775 static inline size_t
1776 filemap_copy_from_user_iovec(struct page *page, unsigned long offset,
1777                         const struct iovec *iov, size_t base, size_t bytes)
1778 {
1779         char *kaddr;
1780         size_t copied;
1781
1782         kaddr = kmap_atomic(page, KM_USER0);
1783         copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1784                                                 base, bytes);
1785         kunmap_atomic(kaddr, KM_USER0);
1786         if (copied != bytes) {
1787                 kaddr = kmap(page);
1788                 copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1789                                                         base, bytes);
1790                 kunmap(page);
1791         }
1792         return copied;
1793 }
1794
1795 static inline void
1796 filemap_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1797 {
1798         const struct iovec *iov = *iovp;
1799         size_t base = *basep;
1800
1801         while (bytes) {
1802                 int copy = min(bytes, iov->iov_len - base);
1803
1804                 bytes -= copy;
1805                 base += copy;
1806                 if (iov->iov_len == base) {
1807                         iov++;
1808                         base = 0;
1809                 }
1810         }
1811         *iovp = iov;
1812         *basep = base;
1813 }
1814
1815 /*
1816  * Performs necessary checks before doing a write
1817  *
1818  * Can adjust writing position aor amount of bytes to write.
1819  * Returns appropriate error code that caller should return or
1820  * zero in case that write should be allowed.
1821  */
1822 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1823 {
1824         struct inode *inode = file->f_mapping->host;
1825         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1826
1827         if (unlikely(*pos < 0))
1828                 return -EINVAL;
1829
1830         if (unlikely(file->f_error)) {
1831                 int err = file->f_error;
1832                 file->f_error = 0;
1833                 return err;
1834         }
1835
1836         if (!isblk) {
1837                 /* FIXME: this is for backwards compatibility with 2.4 */
1838                 if (file->f_flags & O_APPEND)
1839                         *pos = i_size_read(inode);
1840
1841                 if (limit != RLIM_INFINITY) {
1842                         if (*pos >= limit) {
1843                                 send_sig(SIGXFSZ, current, 0);
1844                                 return -EFBIG;
1845                         }
1846                         if (*count > limit - (typeof(limit))*pos) {
1847                                 *count = limit - (typeof(limit))*pos;
1848                         }
1849                 }
1850         }
1851
1852         /*
1853          * LFS rule
1854          */
1855         if (unlikely(*pos + *count > MAX_NON_LFS &&
1856                                 !(file->f_flags & O_LARGEFILE))) {
1857                 if (*pos >= MAX_NON_LFS) {
1858                         send_sig(SIGXFSZ, current, 0);
1859                         return -EFBIG;
1860                 }
1861                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1862                         *count = MAX_NON_LFS - (unsigned long)*pos;
1863                 }
1864         }
1865
1866         /*
1867          * Are we about to exceed the fs block limit ?
1868          *
1869          * If we have written data it becomes a short write.  If we have
1870          * exceeded without writing data we send a signal and return EFBIG.
1871          * Linus frestrict idea will clean these up nicely..
1872          */
1873         if (likely(!isblk)) {
1874                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1875                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1876                                 send_sig(SIGXFSZ, current, 0);
1877                                 return -EFBIG;
1878                         }
1879                         /* zero-length writes at ->s_maxbytes are OK */
1880                 }
1881
1882                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1883                         *count = inode->i_sb->s_maxbytes - *pos;
1884         } else {
1885                 loff_t isize;
1886                 if (bdev_read_only(I_BDEV(inode)))
1887                         return -EPERM;
1888                 isize = i_size_read(inode);
1889                 if (*pos >= isize) {
1890                         if (*count || *pos > isize)
1891                                 return -ENOSPC;
1892                 }
1893
1894                 if (*pos + *count > isize)
1895                         *count = isize - *pos;
1896         }
1897         return 0;
1898 }
1899 EXPORT_SYMBOL(generic_write_checks);
1900
1901 ssize_t
1902 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1903                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1904                 size_t count, size_t ocount)
1905 {
1906         struct file     *file = iocb->ki_filp;
1907         struct address_space *mapping = file->f_mapping;
1908         struct inode    *inode = mapping->host;
1909         ssize_t         written;
1910
1911         if (count != ocount)
1912                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1913
1914         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1915         if (written > 0) {
1916                 loff_t end = pos + written;
1917                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1918                         i_size_write(inode,  end);
1919                         mark_inode_dirty(inode);
1920                 }
1921                 *ppos = end;
1922         }
1923
1924         /*
1925          * Sync the fs metadata but not the minor inode changes and
1926          * of course not the data as we did direct DMA for the IO.
1927          * i_sem is held, which protects generic_osync_inode() from
1928          * livelocking.
1929          */
1930         if (written >= 0 && file->f_flags & O_SYNC)
1931                 generic_osync_inode(inode, mapping, OSYNC_METADATA);
1932         if (written == count && !is_sync_kiocb(iocb))
1933                 written = -EIOCBQUEUED;
1934         return written;
1935 }
1936 EXPORT_SYMBOL(generic_file_direct_write);
1937
1938 ssize_t
1939 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1940                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1941                 size_t count, ssize_t written)
1942 {
1943         struct file *file = iocb->ki_filp;
1944         struct address_space * mapping = file->f_mapping;
1945         struct address_space_operations *a_ops = mapping->a_ops;
1946         struct inode    *inode = mapping->host;
1947         long            status = 0;
1948         struct page     *page;
1949         struct page     *cached_page = NULL;
1950         size_t          bytes;
1951         struct pagevec  lru_pvec;
1952         const struct iovec *cur_iov = iov; /* current iovec */
1953         size_t          iov_base = 0;      /* offset in the current iovec */
1954         char __user     *buf;
1955
1956         pagevec_init(&lru_pvec, 0);
1957
1958         /*
1959          * handle partial DIO write.  Adjust cur_iov if needed.
1960          */
1961         if (likely(nr_segs == 1))
1962                 buf = iov->iov_base + written;
1963         else {
1964                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1965                 buf = cur_iov->iov_base + iov_base;
1966         }
1967
1968         do {
1969                 unsigned long index;
1970                 unsigned long offset;
1971                 unsigned long maxlen;
1972                 size_t copied;
1973
1974                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1975                 index = pos >> PAGE_CACHE_SHIFT;
1976                 bytes = PAGE_CACHE_SIZE - offset;
1977                 if (bytes > count)
1978                         bytes = count;
1979
1980                 /*
1981                  * Bring in the user page that we will copy from _first_.
1982                  * Otherwise there's a nasty deadlock on copying from the
1983                  * same page as we're writing to, without it being marked
1984                  * up-to-date.
1985                  */
1986                 maxlen = cur_iov->iov_len - iov_base;
1987                 if (maxlen > bytes)
1988                         maxlen = bytes;
1989                 fault_in_pages_readable(buf, maxlen);
1990
1991                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1992                 if (!page) {
1993                         status = -ENOMEM;
1994                         break;
1995                 }
1996
1997                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1998                 if (unlikely(status)) {
1999                         loff_t isize = i_size_read(inode);
2000                         /*
2001                          * prepare_write() may have instantiated a few blocks
2002                          * outside i_size.  Trim these off again.
2003                          */
2004                         unlock_page(page);
2005                         page_cache_release(page);
2006                         if (pos + bytes > isize)
2007                                 vmtruncate(inode, isize);
2008                         break;
2009                 }
2010                 if (likely(nr_segs == 1))
2011                         copied = filemap_copy_from_user(page, offset,
2012                                                         buf, bytes);
2013                 else
2014                         copied = filemap_copy_from_user_iovec(page, offset,
2015                                                 cur_iov, iov_base, bytes);
2016                 flush_dcache_page(page);
2017                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2018                 if (likely(copied > 0)) {
2019                         if (!status)
2020                                 status = copied;
2021
2022                         if (status >= 0) {
2023                                 written += status;
2024                                 count -= status;
2025                                 pos += status;
2026                                 buf += status;
2027                                 if (unlikely(nr_segs > 1)) {
2028                                         filemap_set_next_iovec(&cur_iov,
2029                                                         &iov_base, status);
2030                                         buf = cur_iov->iov_base + iov_base;
2031                                 } else {
2032                                         iov_base += status;
2033                                 }
2034                         }
2035                 }
2036                 if (unlikely(copied != bytes))
2037                         if (status >= 0)
2038                                 status = -EFAULT;
2039                 unlock_page(page);
2040                 mark_page_accessed(page);
2041                 page_cache_release(page);
2042                 if (status < 0)
2043                         break;
2044                 balance_dirty_pages_ratelimited(mapping);
2045                 cond_resched();
2046         } while (count);
2047         *ppos = pos;
2048
2049         if (cached_page)
2050                 page_cache_release(cached_page);
2051
2052         /*
2053          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2054          */
2055         if (likely(status >= 0)) {
2056                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2057                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2058                                 status = generic_osync_inode(inode, mapping,
2059                                                 OSYNC_METADATA|OSYNC_DATA);
2060                 }
2061         }
2062
2063         /*
2064          * If we get here for O_DIRECT writes then we must have fallen through
2065          * to buffered writes (block instantiation inside i_size).  So we sync
2066          * the file data here, to try to honour O_DIRECT expectations.
2067          */
2068         if (unlikely(file->f_flags & O_DIRECT) && written)
2069                 status = filemap_write_and_wait(mapping);
2070
2071         pagevec_lru_add(&lru_pvec);
2072         return written ? written : status;
2073 }
2074 EXPORT_SYMBOL(generic_file_buffered_write);
2075
2076 ssize_t
2077 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2078                                 unsigned long nr_segs, loff_t *ppos)
2079 {
2080         struct file *file = iocb->ki_filp;
2081         struct address_space * mapping = file->f_mapping;
2082         size_t ocount;          /* original count */
2083         size_t count;           /* after file limit checks */
2084         struct inode    *inode = mapping->host;
2085         unsigned long   seg;
2086         loff_t          pos;
2087         ssize_t         written;
2088         ssize_t         err;
2089
2090         ocount = 0;
2091         for (seg = 0; seg < nr_segs; seg++) {
2092                 const struct iovec *iv = &iov[seg];
2093
2094                 /*
2095                  * If any segment has a negative length, or the cumulative
2096                  * length ever wraps negative then return -EINVAL.
2097                  */
2098                 ocount += iv->iov_len;
2099                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2100                         return -EINVAL;
2101                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2102                         continue;
2103                 if (seg == 0)
2104                         return -EFAULT;
2105                 nr_segs = seg;
2106                 ocount -= iv->iov_len;  /* This segment is no good */
2107                 break;
2108         }
2109
2110         count = ocount;
2111         pos = *ppos;
2112
2113         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2114
2115         /* We can write back this queue in page reclaim */
2116         current->backing_dev_info = mapping->backing_dev_info;
2117         written = 0;
2118
2119         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2120         if (err)
2121                 goto out;
2122
2123         if (count == 0)
2124                 goto out;
2125
2126         err = remove_suid(file->f_dentry);
2127         if (err)
2128                 goto out;
2129
2130         inode_update_time(inode, 1);
2131
2132         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2133         if (unlikely(file->f_flags & O_DIRECT)) {
2134                 written = generic_file_direct_write(iocb, iov,
2135                                 &nr_segs, pos, ppos, count, ocount);
2136                 if (written < 0 || written == count)
2137                         goto out;
2138                 /*
2139                  * direct-io write to a hole: fall through to buffered I/O
2140                  * for completing the rest of the request.
2141                  */
2142                 pos += written;
2143                 count -= written;
2144         }
2145
2146         written = generic_file_buffered_write(iocb, iov, nr_segs,
2147                         pos, ppos, count, written);
2148 out:
2149         current->backing_dev_info = NULL;
2150         return written ? written : err;
2151 }
2152 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2153
2154 ssize_t
2155 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2156                                 unsigned long nr_segs, loff_t *ppos)
2157 {
2158         struct file *file = iocb->ki_filp;
2159         struct address_space *mapping = file->f_mapping;
2160         struct inode *inode = mapping->host;
2161         ssize_t ret;
2162         loff_t pos = *ppos;
2163
2164         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2165
2166         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2167                 int err;
2168
2169                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2170                 if (err < 0)
2171                         ret = err;
2172         }
2173         return ret;
2174 }
2175
2176 ssize_t
2177 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2178                                 unsigned long nr_segs, loff_t *ppos)
2179 {
2180         struct kiocb kiocb;
2181         ssize_t ret;
2182
2183         init_sync_kiocb(&kiocb, file);
2184         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2185         if (ret == -EIOCBQUEUED)
2186                 ret = wait_on_sync_kiocb(&kiocb);
2187         return ret;
2188 }
2189
2190 ssize_t
2191 generic_file_write_nolock(struct file *file, 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, file);
2198         ret = generic_file_aio_write_nolock(&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_write_nolock);
2204
2205 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2206                                size_t count, loff_t pos)
2207 {
2208         struct file *file = iocb->ki_filp;
2209         struct address_space *mapping = file->f_mapping;
2210         struct inode *inode = mapping->host;
2211         ssize_t ret;
2212         struct iovec local_iov = { .iov_base = (void __user *)buf,
2213                                         .iov_len = count };
2214
2215         BUG_ON(iocb->ki_pos != pos);
2216
2217         down(&inode->i_sem);
2218         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2219                                                 &iocb->ki_pos);
2220         up(&inode->i_sem);
2221
2222         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2223                 ssize_t err;
2224
2225                 err = sync_page_range(inode, mapping, pos, ret);
2226                 if (err < 0)
2227                         ret = err;
2228         }
2229         return ret;
2230 }
2231 EXPORT_SYMBOL(generic_file_aio_write);
2232
2233 ssize_t generic_file_write(struct file *file, const char __user *buf,
2234                            size_t count, loff_t *ppos)
2235 {
2236         struct address_space *mapping = file->f_mapping;
2237         struct inode *inode = mapping->host;
2238         ssize_t ret;
2239         struct iovec local_iov = { .iov_base = (void __user *)buf,
2240                                         .iov_len = count };
2241
2242         down(&inode->i_sem);
2243         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2244         up(&inode->i_sem);
2245
2246         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2247                 ssize_t err;
2248
2249                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2250                 if (err < 0)
2251                         ret = err;
2252         }
2253         return ret;
2254 }
2255 EXPORT_SYMBOL(generic_file_write);
2256
2257 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2258                         unsigned long nr_segs, loff_t *ppos)
2259 {
2260         struct kiocb kiocb;
2261         ssize_t ret;
2262
2263         init_sync_kiocb(&kiocb, filp);
2264         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2265         if (-EIOCBQUEUED == ret)
2266                 ret = wait_on_sync_kiocb(&kiocb);
2267         return ret;
2268 }
2269 EXPORT_SYMBOL(generic_file_readv);
2270
2271 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2272                         unsigned long nr_segs, loff_t *ppos)
2273 {
2274         struct address_space *mapping = file->f_mapping;
2275         struct inode *inode = mapping->host;
2276         ssize_t ret;
2277
2278         down(&inode->i_sem);
2279         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2280         up(&inode->i_sem);
2281
2282         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2283                 int err;
2284
2285                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2286                 if (err < 0)
2287                         ret = err;
2288         }
2289         return ret;
2290 }
2291 EXPORT_SYMBOL(generic_file_writev);
2292
2293 /*
2294  * Called under i_sem for writes to S_ISREG files.   Returns -EIO if something
2295  * went wrong during pagecache shootdown.
2296  */
2297 ssize_t
2298 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2299         loff_t offset, unsigned long nr_segs)
2300 {
2301         struct file *file = iocb->ki_filp;
2302         struct address_space *mapping = file->f_mapping;
2303         ssize_t retval;
2304         size_t write_len = 0;
2305
2306         /*
2307          * If it's a write, unmap all mmappings of the file up-front.  This
2308          * will cause any pte dirty bits to be propagated into the pageframes
2309          * for the subsequent filemap_write_and_wait().
2310          */
2311         if (rw == WRITE) {
2312                 write_len = iov_length(iov, nr_segs);
2313                 if (mapping_mapped(mapping))
2314                         unmap_mapping_range(mapping, offset, write_len, 0);
2315         }
2316
2317         retval = filemap_write_and_wait(mapping);
2318         if (retval == 0) {
2319                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2320                                                 offset, nr_segs);
2321                 if (rw == WRITE && mapping->nrpages) {
2322                         pgoff_t end = (offset + write_len - 1)
2323                                                 >> PAGE_CACHE_SHIFT;
2324                         int err = invalidate_inode_pages2_range(mapping,
2325                                         offset >> PAGE_CACHE_SHIFT, end);
2326                         if (err)
2327                                 retval = err;
2328                 }
2329         }
2330         return retval;
2331 }
2332 EXPORT_SYMBOL_GPL(generic_file_direct_IO);