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         BUG_ON(!PageLocked(page));
 127
 128         write_lock_irq(&mapping->tree_lock);
 129         __remove_from_page_cache(page);
 130         write_unlock_irq(&mapping->tree_lock);
 131 }
 132
 133 static int sync_page(void *word)
 134 {
 135         struct address_space *mapping;
 136         struct page *page;
 137
 138         page = container_of((page_flags_t *)word, struct page, flags);
 139
 140         /*
 141          * page_mapping() is being called without PG_locked held.
 142          * Some knowledge of the state and use of the page is used to
 143          * reduce the requirements down to a memory barrier.
 144          * The danger here is of a stale page_mapping() return value
 145          * indicating a struct address_space different from the one it's
 146          * associated with when it is associated with one.
 147          * After smp_mb(), it's either the correct page_mapping() for
 148          * the page, or an old page_mapping() and the page's own
 149          * page_mapping() has gone NULL.
 150          * The ->sync_page() address_space operation must tolerate
 151          * page_mapping() going NULL. By an amazing coincidence,
 152          * this comes about because none of the users of the page
 153          * in the ->sync_page() methods make essential use of the
 154          * page_mapping(), merely passing the page down to the backing
 155          * device's unplug functions when it's non-NULL, which in turn
 156          * ignore it for all cases but swap, where only page->private is
 157          * of interest. When page_mapping() does go NULL, the entire
 158          * call stack gracefully ignores the page and returns.
 159          * -- wli
 160          */
 161         smp_mb();
 162         mapping = page_mapping(page);
 163         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 164                 mapping->a_ops->sync_page(page);
 165         io_schedule();
 166         return 0;
 167 }
 168
 169 /**
 170  * filemap_fdatawrite_range - start writeback against all of a mapping's
 171  * dirty pages that lie within the byte offsets <start, end>
 172  * @mapping: address space structure to write
 173  * @start: offset in bytes where the range starts
 174  * @end : offset in bytes where the range ends
 175  *
 176  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 177  * opposed to a regular memory * cleansing writeback.  The difference between
 178  * these two operations is that if a dirty page/buffer is encountered, it must
 179  * be waited upon, and not just skipped over.
 180  */
 181 static int __filemap_fdatawrite_range(struct address_space *mapping,
 182         loff_t start, loff_t end, int sync_mode)
 183 {
 184         int ret;
 185         struct writeback_control wbc = {
 186                 .sync_mode = sync_mode,
 187                 .nr_to_write = mapping->nrpages * 2,
 188                 .start = start,
 189                 .end = end,
 190         };
 191
 192         if (!mapping_cap_writeback_dirty(mapping))
 193                 return 0;
 194
 195         ret = do_writepages(mapping, &wbc);
 196         return ret;
 197 }
 198
 199 static inline int __filemap_fdatawrite(struct address_space *mapping,
 200         int sync_mode)
 201 {
 202         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
 203 }
 204
 205 int filemap_fdatawrite(struct address_space *mapping)
 206 {
 207         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 208 }
 209 EXPORT_SYMBOL(filemap_fdatawrite);
 210
 211 static int filemap_fdatawrite_range(struct address_space *mapping,
 212         loff_t start, loff_t end)
 213 {
 214         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
 215 }
 216
 217 /*
 218  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 219  * purposes - I/O may not be started against all dirty pages.
 220  */
 221 int filemap_flush(struct address_space *mapping)
 222 {
 223         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 224 }
 225 EXPORT_SYMBOL(filemap_flush);
 226
 227 /*
 228  * Wait for writeback to complete against pages indexed by start->end
 229  * inclusive
 230  */
 231 static int wait_on_page_writeback_range(struct address_space *mapping,
 232                                 pgoff_t start, pgoff_t end)
 233 {
 234         struct pagevec pvec;
 235         int nr_pages;
 236         int ret = 0;
 237         pgoff_t index;
 238
 239         if (end < start)
 240                 return 0;
 241
 242         pagevec_init(&pvec, 0);
 243         index = start;
 244         while ((index <= end) &&
 245                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 246                         PAGECACHE_TAG_WRITEBACK,
 247                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 248                 unsigned i;
 249
 250                 for (i = 0; i < nr_pages; i++) {
 251                         struct page *page = pvec.pages[i];
 252
 253                         /* until radix tree lookup accepts end_index */
 254                         if (page->index > end)
 255                                 continue;
 256
 257                         wait_on_page_writeback(page);
 258                         if (PageError(page))
 259                                 ret = -EIO;
 260                 }
 261                 pagevec_release(&pvec);
 262                 cond_resched();
 263         }
 264
 265         /* Check for outstanding write errors */
 266         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 267                 ret = -ENOSPC;
 268         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 269                 ret = -EIO;
 270
 271         return ret;
 272 }
 273
 274 /*
 275  * Write and wait upon all the pages in the passed range.  This is a "data
 276  * integrity" operation.  It waits upon in-flight writeout before starting and
 277  * waiting upon new writeout.  If there was an IO error, return it.
 278  *
 279  * We need to re-take i_sem during the generic_osync_inode list walk because
 280  * it is otherwise livelockable.
 281  */
 282 int sync_page_range(struct inode *inode, struct address_space *mapping,
 283                         loff_t pos, size_t count)
 284 {
 285         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 286         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 287         int ret;
 288
 289         if (!mapping_cap_writeback_dirty(mapping) || !count)
 290                 return 0;
 291         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 292         if (ret == 0) {
 293                 down(&inode->i_sem);
 294                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 295                 up(&inode->i_sem);
 296         }
 297         if (ret == 0)
 298                 ret = wait_on_page_writeback_range(mapping, start, end);
 299         return ret;
 300 }
 301 EXPORT_SYMBOL(sync_page_range);
 302
 303 /*
 304  * Note: Holding i_sem across sync_page_range_nolock is not a good idea
 305  * as it forces O_SYNC writers to different parts of the same file
 306  * to be serialised right until io completion.
 307  */
 308 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
 309                         loff_t pos, size_t count)
 310 {
 311         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 312         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 313         int ret;
 314
 315         if (!mapping_cap_writeback_dirty(mapping) || !count)
 316                 return 0;
 317         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 318         if (ret == 0)
 319                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 320         if (ret == 0)
 321                 ret = wait_on_page_writeback_range(mapping, start, end);
 322         return ret;
 323 }
 324 EXPORT_SYMBOL(sync_page_range_nolock);
 325
 326 /**
 327  * filemap_fdatawait - walk the list of under-writeback pages of the given
 328  *     address space and wait for all of them.
 329  *
 330  * @mapping: address space structure to wait for
 331  */
 332 int filemap_fdatawait(struct address_space *mapping)
 333 {
 334         loff_t i_size = i_size_read(mapping->host);
 335
 336         if (i_size == 0)
 337                 return 0;
 338
 339         return wait_on_page_writeback_range(mapping, 0,
 340                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 341 }
 342 EXPORT_SYMBOL(filemap_fdatawait);
 343
 344 int filemap_write_and_wait(struct address_space *mapping)
 345 {
 346         int retval = 0;
 347
 348         if (mapping->nrpages) {
 349                 retval = filemap_fdatawrite(mapping);
 350                 if (retval == 0)
 351                         retval = filemap_fdatawait(mapping);
 352         }
 353         return retval;
 354 }
 355
 356 int filemap_write_and_wait_range(struct address_space *mapping,
 357                                  loff_t lstart, loff_t lend)
 358 {
 359         int retval = 0;
 360
 361         if (mapping->nrpages) {
 362                 retval = __filemap_fdatawrite_range(mapping, lstart, lend,
 363                                                     WB_SYNC_ALL);
 364                 if (retval == 0)
 365                         retval = wait_on_page_writeback_range(mapping,
 366                                                     lstart >> PAGE_CACHE_SHIFT,
 367                                                     lend >> PAGE_CACHE_SHIFT);
 368         }
 369         return retval;
 370 }
 371
 372 /*
 373  * This function is used to add newly allocated pagecache pages:
 374  * the page is new, so we can just run SetPageLocked() against it.
 375  * The other page state flags were set by rmqueue().
 376  *
 377  * This function does not add the page to the LRU.  The caller must do that.
 378  */
 379 int add_to_page_cache(struct page *page, struct address_space *mapping,
 380                 pgoff_t offset, int gfp_mask)
 381 {
 382         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 383
 384         if (error == 0) {
 385                 write_lock_irq(&mapping->tree_lock);
 386                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 387                 if (!error) {
 388                         page_cache_get(page);
 389                         SetPageLocked(page);
 390                         page->mapping = mapping;
 391                         page->index = offset;
 392                         mapping->nrpages++;
 393                         pagecache_acct(1);
 394                 }
 395                 write_unlock_irq(&mapping->tree_lock);
 396                 radix_tree_preload_end();
 397         }
 398         return error;
 399 }
 400
 401 EXPORT_SYMBOL(add_to_page_cache);
 402
 403 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 404                                 pgoff_t offset, int gfp_mask)
 405 {
 406         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 407         if (ret == 0)
 408                 lru_cache_add(page);
 409         return ret;
 410 }
 411
 412 /*
 413  * In order to wait for pages to become available there must be
 414  * waitqueues associated with pages. By using a hash table of
 415  * waitqueues where the bucket discipline is to maintain all
 416  * waiters on the same queue and wake all when any of the pages
 417  * become available, and for the woken contexts to check to be
 418  * sure the appropriate page became available, this saves space
 419  * at a cost of "thundering herd" phenomena during rare hash
 420  * collisions.
 421  */
 422 static wait_queue_head_t *page_waitqueue(struct page *page)
 423 {
 424         const struct zone *zone = page_zone(page);
 425
 426         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 427 }
 428
 429 static inline void wake_up_page(struct page *page, int bit)
 430 {
 431         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 432 }
 433
 434 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 435 {
 436         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 437
 438         if (test_bit(bit_nr, &page->flags))
 439                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 440                                                         TASK_UNINTERRUPTIBLE);
 441 }
 442 EXPORT_SYMBOL(wait_on_page_bit);
 443
 444 /**
 445  * unlock_page() - unlock a locked page
 446  *
 447  * @page: the page
 448  *
 449  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 450  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 451  * mechananism between PageLocked pages and PageWriteback pages is shared.
 452  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 453  *
 454  * The first mb is necessary to safely close the critical section opened by the
 455  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 456  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 457  * parallel wait_on_page_locked()).
 458  */
 459 void fastcall unlock_page(struct page *page)
 460 {
 461         smp_mb__before_clear_bit();
 462         if (!TestClearPageLocked(page))
 463                 BUG();
 464         smp_mb__after_clear_bit();
 465         wake_up_page(page, PG_locked);
 466 }
 467 EXPORT_SYMBOL(unlock_page);
 468
 469 /*
 470  * End writeback against a page.
 471  */
 472 void end_page_writeback(struct page *page)
 473 {
 474         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 475                 if (!test_clear_page_writeback(page))
 476                         BUG();
 477         }
 478         smp_mb__after_clear_bit();
 479         wake_up_page(page, PG_writeback);
 480 }
 481 EXPORT_SYMBOL(end_page_writeback);
 482
 483 /*
 484  * Get a lock on the page, assuming we need to sleep to get it.
 485  *
 486  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 487  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 488  * chances are that on the second loop, the block layer's plug list is empty,
 489  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 490  */
 491 void fastcall __lock_page(struct page *page)
 492 {
 493         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 494
 495         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 496                                                         TASK_UNINTERRUPTIBLE);
 497 }
 498 EXPORT_SYMBOL(__lock_page);
 499
 500 /*
 501  * a rather lightweight function, finding and getting a reference to a
 502  * hashed page atomically.
 503  */
 504 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 505 {
 506         struct page *page;
 507
 508         read_lock_irq(&mapping->tree_lock);
 509         page = radix_tree_lookup(&mapping->page_tree, offset);
 510         if (page)
 511                 page_cache_get(page);
 512         read_unlock_irq(&mapping->tree_lock);
 513         return page;
 514 }
 515
 516 EXPORT_SYMBOL(find_get_page);
 517
 518 /*
 519  * Same as above, but trylock it instead of incrementing the count.
 520  */
 521 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 522 {
 523         struct page *page;
 524
 525         read_lock_irq(&mapping->tree_lock);
 526         page = radix_tree_lookup(&mapping->page_tree, offset);
 527         if (page && TestSetPageLocked(page))
 528                 page = NULL;
 529         read_unlock_irq(&mapping->tree_lock);
 530         return page;
 531 }
 532
 533 EXPORT_SYMBOL(find_trylock_page);
 534
 535 /**
 536  * find_lock_page - locate, pin and lock a pagecache page
 537  *
 538  * @mapping - the address_space to search
 539  * @offset - the page index
 540  *
 541  * Locates the desired pagecache page, locks it, increments its reference
 542  * count and returns its address.
 543  *
 544  * Returns zero if the page was not present. find_lock_page() may sleep.
 545  */
 546 struct page *find_lock_page(struct address_space *mapping,
 547                                 unsigned long offset)
 548 {
 549         struct page *page;
 550
 551         read_lock_irq(&mapping->tree_lock);
 552 repeat:
 553         page = radix_tree_lookup(&mapping->page_tree, offset);
 554         if (page) {
 555                 page_cache_get(page);
 556                 if (TestSetPageLocked(page)) {
 557                         read_unlock_irq(&mapping->tree_lock);
 558                         lock_page(page);
 559                         read_lock_irq(&mapping->tree_lock);
 560
 561                         /* Has the page been truncated while we slept? */
 562                         if (page->mapping != mapping || page->index != offset) {
 563                                 unlock_page(page);
 564                                 page_cache_release(page);
 565                                 goto repeat;
 566                         }
 567                 }
 568         }
 569         read_unlock_irq(&mapping->tree_lock);
 570         return page;
 571 }
 572
 573 EXPORT_SYMBOL(find_lock_page);
 574
 575 /**
 576  * find_or_create_page - locate or add a pagecache page
 577  *
 578  * @mapping - the page's address_space
 579  * @index - the page's index into the mapping
 580  * @gfp_mask - page allocation mode
 581  *
 582  * Locates a page in the pagecache.  If the page is not present, a new page
 583  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 584  * LRU list.  The returned page is locked and has its reference count
 585  * incremented.
 586  *
 587  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 588  * allocation!
 589  *
 590  * find_or_create_page() returns the desired page's address, or zero on
 591  * memory exhaustion.
 592  */
 593 struct page *find_or_create_page(struct address_space *mapping,
 594                 unsigned long index, unsigned int gfp_mask)
 595 {
 596         struct page *page, *cached_page = NULL;
 597         int err;
 598 repeat:
 599         page = find_lock_page(mapping, index);
 600         if (!page) {
 601                 if (!cached_page) {
 602                         cached_page = alloc_page(gfp_mask);
 603                         if (!cached_page)
 604                                 return NULL;
 605                 }
 606                 err = add_to_page_cache_lru(cached_page, mapping,
 607                                         index, gfp_mask);
 608                 if (!err) {
 609                         page = cached_page;
 610                         cached_page = NULL;
 611                 } else if (err == -EEXIST)
 612                         goto repeat;
 613         }
 614         if (cached_page)
 615                 page_cache_release(cached_page);
 616         return page;
 617 }
 618
 619 EXPORT_SYMBOL(find_or_create_page);
 620
 621 /**
 622  * find_get_pages - gang pagecache lookup
 623  * @mapping:    The address_space to search
 624  * @start:      The starting page index
 625  * @nr_pages:   The maximum number of pages
 626  * @pages:      Where the resulting pages are placed
 627  *
 628  * find_get_pages() will search for and return a group of up to
 629  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 630  * find_get_pages() takes a reference against the returned pages.
 631  *
 632  * The search returns a group of mapping-contiguous pages with ascending
 633  * indexes.  There may be holes in the indices due to not-present pages.
 634  *
 635  * find_get_pages() returns the number of pages which were found.
 636  */
 637 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 638                             unsigned int nr_pages, struct page **pages)
 639 {
 640         unsigned int i;
 641         unsigned int ret;
 642
 643         read_lock_irq(&mapping->tree_lock);
 644         ret = radix_tree_gang_lookup(&mapping->page_tree,
 645                                 (void **)pages, start, nr_pages);
 646         for (i = 0; i < ret; i++)
 647                 page_cache_get(pages[i]);
 648         read_unlock_irq(&mapping->tree_lock);
 649         return ret;
 650 }
 651
 652 /*
 653  * Like find_get_pages, except we only return pages which are tagged with
 654  * `tag'.   We update *index to index the next page for the traversal.
 655  */
 656 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 657                         int tag, unsigned int nr_pages, struct page **pages)
 658 {
 659         unsigned int i;
 660         unsigned int ret;
 661
 662         read_lock_irq(&mapping->tree_lock);
 663         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 664                                 (void **)pages, *index, nr_pages, tag);
 665         for (i = 0; i < ret; i++)
 666                 page_cache_get(pages[i]);
 667         if (ret)
 668                 *index = pages[ret - 1]->index + 1;
 669         read_unlock_irq(&mapping->tree_lock);
 670         return ret;
 671 }
 672
 673 /*
 674  * Same as grab_cache_page, but do not wait if the page is unavailable.
 675  * This is intended for speculative data generators, where the data can
 676  * be regenerated if the page couldn't be grabbed.  This routine should
 677  * be safe to call while holding the lock for another page.
 678  *
 679  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 680  * and deadlock against the caller's locked page.
 681  */
 682 struct page *
 683 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 684 {
 685         struct page *page = find_get_page(mapping, index);
 686         unsigned int gfp_mask;
 687
 688         if (page) {
 689                 if (!TestSetPageLocked(page))
 690                         return page;
 691                 page_cache_release(page);
 692                 return NULL;
 693         }
 694         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 695         page = alloc_pages(gfp_mask, 0);
 696         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 697                 page_cache_release(page);
 698                 page = NULL;
 699         }
 700         return page;
 701 }
 702
 703 EXPORT_SYMBOL(grab_cache_page_nowait);
 704
 705 /*
 706  * This is a generic file read routine, and uses the
 707  * mapping->a_ops->readpage() function for the actual low-level
 708  * stuff.
 709  *
 710  * This is really ugly. But the goto's actually try to clarify some
 711  * of the logic when it comes to error handling etc.
 712  *
 713  * Note the struct file* is only passed for the use of readpage.  It may be
 714  * NULL.
 715  */
 716 void do_generic_mapping_read(struct address_space *mapping,
 717                              struct file_ra_state *_ra,
 718                              struct file *filp,
 719                              loff_t *ppos,
 720                              read_descriptor_t *desc,
 721                              read_actor_t actor)
 722 {
 723         struct inode *inode = mapping->host;
 724         unsigned long index;
 725         unsigned long end_index;
 726         unsigned long offset;
 727         unsigned long last_index;
 728         unsigned long next_index;
 729         unsigned long prev_index;
 730         loff_t isize;
 731         struct page *cached_page;
 732         int error;
 733         struct file_ra_state ra = *_ra;
 734
 735         cached_page = NULL;
 736         index = *ppos >> PAGE_CACHE_SHIFT;
 737         next_index = index;
 738         prev_index = ra.prev_page;
 739         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 740         offset = *ppos & ~PAGE_CACHE_MASK;
 741
 742         isize = i_size_read(inode);
 743         if (!isize)
 744                 goto out;
 745
 746         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 747         for (;;) {
 748                 struct page *page;
 749                 unsigned long nr, ret;
 750
 751                 /* nr is the maximum number of bytes to copy from this page */
 752                 nr = PAGE_CACHE_SIZE;
 753                 if (index >= end_index) {
 754                         if (index > end_index)
 755                                 goto out;
 756                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 757                         if (nr <= offset) {
 758                                 goto out;
 759                         }
 760                 }
 761                 nr = nr - offset;
 762
 763                 cond_resched();
 764                 if (index == next_index)
 765                         next_index = page_cache_readahead(mapping, &ra, filp,
 766                                         index, last_index - index);
 767
 768 find_page:
 769                 page = find_get_page(mapping, index);
 770                 if (unlikely(page == NULL)) {
 771                         handle_ra_miss(mapping, &ra, index);
 772                         goto no_cached_page;
 773                 }
 774                 if (!PageUptodate(page))
 775                         goto page_not_up_to_date;
 776 page_ok:
 777
 778                 /* If users can be writing to this page using arbitrary
 779                  * virtual addresses, take care about potential aliasing
 780                  * before reading the page on the kernel side.
 781                  */
 782                 if (mapping_writably_mapped(mapping))
 783                         flush_dcache_page(page);
 784
 785                 /*
 786                  * When (part of) the same page is read multiple times
 787                  * in succession, only mark it as accessed the first time.
 788                  */
 789                 if (prev_index != index)
 790                         mark_page_accessed(page);
 791                 prev_index = index;
 792
 793                 /*
 794                  * Ok, we have the page, and it's up-to-date, so
 795                  * now we can copy it to user space...
 796                  *
 797                  * The actor routine returns how many bytes were actually used..
 798                  * NOTE! This may not be the same as how much of a user buffer
 799                  * we filled up (we may be padding etc), so we can only update
 800                  * "pos" here (the actor routine has to update the user buffer
 801                  * pointers and the remaining count).
 802                  */
 803                 ret = actor(desc, page, offset, nr);
 804                 offset += ret;
 805                 index += offset >> PAGE_CACHE_SHIFT;
 806                 offset &= ~PAGE_CACHE_MASK;
 807
 808                 page_cache_release(page);
 809                 if (ret == nr && desc->count)
 810                         continue;
 811                 goto out;
 812
 813 page_not_up_to_date:
 814                 /* Get exclusive access to the page ... */
 815                 lock_page(page);
 816
 817                 /* Did it get unhashed before we got the lock? */
 818                 if (!page->mapping) {
 819                         unlock_page(page);
 820                         page_cache_release(page);
 821                         continue;
 822                 }
 823
 824                 /* Did somebody else fill it already? */
 825                 if (PageUptodate(page)) {
 826                         unlock_page(page);
 827                         goto page_ok;
 828                 }
 829
 830 readpage:
 831                 /* Start the actual read. The read will unlock the page. */
 832                 error = mapping->a_ops->readpage(filp, page);
 833
 834                 if (unlikely(error))
 835                         goto readpage_error;
 836
 837                 if (!PageUptodate(page)) {
 838                         lock_page(page);
 839                         if (!PageUptodate(page)) {
 840                                 if (page->mapping == NULL) {
 841                                         /*
 842                                          * invalidate_inode_pages got it
 843                                          */
 844                                         unlock_page(page);
 845                                         page_cache_release(page);
 846                                         goto find_page;
 847                                 }
 848                                 unlock_page(page);
 849                                 error = -EIO;
 850                                 goto readpage_error;
 851                         }
 852                         unlock_page(page);
 853                 }
 854
 855                 /*
 856                  * i_size must be checked after we have done ->readpage.
 857                  *
 858                  * Checking i_size after the readpage allows us to calculate
 859                  * the correct value for "nr", which means the zero-filled
 860                  * part of the page is not copied back to userspace (unless
 861                  * another truncate extends the file - this is desired though).
 862                  */
 863                 isize = i_size_read(inode);
 864                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 865                 if (unlikely(!isize || index > end_index)) {
 866                         page_cache_release(page);
 867                         goto out;
 868                 }
 869
 870                 /* nr is the maximum number of bytes to copy from this page */
 871                 nr = PAGE_CACHE_SIZE;
 872                 if (index == end_index) {
 873                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 874                         if (nr <= offset) {
 875                                 page_cache_release(page);
 876                                 goto out;
 877                         }
 878                 }
 879                 nr = nr - offset;
 880                 goto page_ok;
 881
 882 readpage_error:
 883                 /* UHHUH! A synchronous read error occurred. Report it */
 884                 desc->error = error;
 885                 page_cache_release(page);
 886                 goto out;
 887
 888 no_cached_page:
 889                 /*
 890                  * Ok, it wasn't cached, so we need to create a new
 891                  * page..
 892                  */
 893                 if (!cached_page) {
 894                         cached_page = page_cache_alloc_cold(mapping);
 895                         if (!cached_page) {
 896                                 desc->error = -ENOMEM;
 897                                 goto out;
 898                         }
 899                 }
 900                 error = add_to_page_cache_lru(cached_page, mapping,
 901                                                 index, GFP_KERNEL);
 902                 if (error) {
 903                         if (error == -EEXIST)
 904                                 goto find_page;
 905                         desc->error = error;
 906                         goto out;
 907                 }
 908                 page = cached_page;
 909                 cached_page = NULL;
 910                 goto readpage;
 911         }
 912
 913 out:
 914         *_ra = ra;
 915
 916         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 917         if (cached_page)
 918                 page_cache_release(cached_page);
 919         if (filp)
 920                 file_accessed(filp);
 921 }
 922
 923 EXPORT_SYMBOL(do_generic_mapping_read);
 924
 925 int file_read_actor(read_descriptor_t *desc, struct page *page,
 926                         unsigned long offset, unsigned long size)
 927 {
 928         char *kaddr;
 929         unsigned long left, count = desc->count;
 930
 931         if (size > count)
 932                 size = count;
 933
 934         /*
 935          * Faults on the destination of a read are common, so do it before
 936          * taking the kmap.
 937          */
 938         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 939                 kaddr = kmap_atomic(page, KM_USER0);
 940                 left = __copy_to_user_inatomic(desc->arg.buf,
 941                                                 kaddr + offset, size);
 942                 kunmap_atomic(kaddr, KM_USER0);
 943                 if (left == 0)
 944                         goto success;
 945         }
 946
 947         /* Do it the slow way */
 948         kaddr = kmap(page);
 949         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 950         kunmap(page);
 951
 952         if (left) {
 953                 size -= left;
 954                 desc->error = -EFAULT;
 955         }
 956 success:
 957         desc->count = count - size;
 958         desc->written += size;
 959         desc->arg.buf += size;
 960         return size;
 961 }
 962
 963 /*
 964  * This is the "read()" routine for all filesystems
 965  * that can use the page cache directly.
 966  */
 967 ssize_t
 968 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 969                 unsigned long nr_segs, loff_t *ppos)
 970 {
 971         struct file *filp = iocb->ki_filp;
 972         ssize_t retval;
 973         unsigned long seg;
 974         size_t count;
 975
 976         count = 0;
 977         for (seg = 0; seg < nr_segs; seg++) {
 978                 const struct iovec *iv = &iov[seg];
 979
 980                 /*
 981                  * If any segment has a negative length, or the cumulative
 982                  * length ever wraps negative then return -EINVAL.
 983                  */
 984                 count += iv->iov_len;
 985                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
 986                         return -EINVAL;
 987                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
 988                         continue;
 989                 if (seg == 0)
 990                         return -EFAULT;
 991                 nr_segs = seg;
 992                 count -= iv->iov_len;   /* This segment is no good */
 993                 break;
 994         }
 995
 996         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
 997         if (filp->f_flags & O_DIRECT) {
 998                 loff_t pos = *ppos, size;
 999                 struct address_space *mapping;
1000                 struct inode *inode;
1001
1002                 mapping = filp->f_mapping;
1003                 inode = mapping->host;
1004                 retval = 0;
1005                 if (!count)
1006                         goto out; /* skip atime */
1007                 size = i_size_read(inode);
1008                 if (pos < size) {
1009                         retval = generic_file_direct_IO(READ, iocb,
1010                                                 iov, pos, nr_segs);
1011                         if (retval >= 0 && !is_sync_kiocb(iocb))
1012                                 retval = -EIOCBQUEUED;
1013                         if (retval > 0)
1014                                 *ppos = pos + retval;
1015                 }
1016                 file_accessed(filp);
1017                 goto out;
1018         }
1019
1020         retval = 0;
1021         if (count) {
1022                 for (seg = 0; seg < nr_segs; seg++) {
1023                         read_descriptor_t desc;
1024
1025                         desc.written = 0;
1026                         desc.arg.buf = iov[seg].iov_base;
1027                         desc.count = iov[seg].iov_len;
1028                         if (desc.count == 0)
1029                                 continue;
1030                         desc.error = 0;
1031                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1032                         retval += desc.written;
1033                         if (!retval) {
1034                                 retval = desc.error;
1035                                 break;
1036                         }
1037                 }
1038         }
1039 out:
1040         return retval;
1041 }
1042
1043 EXPORT_SYMBOL(__generic_file_aio_read);
1044
1045 ssize_t
1046 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1047 {
1048         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1049
1050         BUG_ON(iocb->ki_pos != pos);
1051         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1052 }
1053
1054 EXPORT_SYMBOL(generic_file_aio_read);
1055
1056 ssize_t
1057 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1058 {
1059         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1060         struct kiocb kiocb;
1061         ssize_t ret;
1062
1063         init_sync_kiocb(&kiocb, filp);
1064         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1065         if (-EIOCBQUEUED == ret)
1066                 ret = wait_on_sync_kiocb(&kiocb);
1067         return ret;
1068 }
1069
1070 EXPORT_SYMBOL(generic_file_read);
1071
1072 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1073 {
1074         ssize_t written;
1075         unsigned long count = desc->count;
1076         struct file *file = desc->arg.data;
1077
1078         if (size > count)
1079                 size = count;
1080
1081         written = file->f_op->sendpage(file, page, offset,
1082                                        size, &file->f_pos, size<count);
1083         if (written < 0) {
1084                 desc->error = written;
1085                 written = 0;
1086         }
1087         desc->count = count - written;
1088         desc->written += written;
1089         return written;
1090 }
1091
1092 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1093                          size_t count, read_actor_t actor, void *target)
1094 {
1095         read_descriptor_t desc;
1096
1097         if (!count)
1098                 return 0;
1099
1100         desc.written = 0;
1101         desc.count = count;
1102         desc.arg.data = target;
1103         desc.error = 0;
1104
1105         do_generic_file_read(in_file, ppos, &desc, actor);
1106         if (desc.written)
1107                 return desc.written;
1108         return desc.error;
1109 }
1110
1111 EXPORT_SYMBOL(generic_file_sendfile);
1112
1113 static ssize_t
1114 do_readahead(struct address_space *mapping, struct file *filp,
1115              unsigned long index, unsigned long nr)
1116 {
1117         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1118                 return -EINVAL;
1119
1120         force_page_cache_readahead(mapping, filp, index,
1121                                         max_sane_readahead(nr));
1122         return 0;
1123 }
1124
1125 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1126 {
1127         ssize_t ret;
1128         struct file *file;
1129
1130         ret = -EBADF;
1131         file = fget(fd);
1132         if (file) {
1133                 if (file->f_mode & FMODE_READ) {
1134                         struct address_space *mapping = file->f_mapping;
1135                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1136                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1137                         unsigned long len = end - start + 1;
1138                         ret = do_readahead(mapping, file, start, len);
1139                 }
1140                 fput(file);
1141         }
1142         return ret;
1143 }
1144
1145 #ifdef CONFIG_MMU
1146 /*
1147  * This adds the requested page to the page cache if it isn't already there,
1148  * and schedules an I/O to read in its contents from disk.
1149  */
1150 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1151 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1152 {
1153         struct address_space *mapping = file->f_mapping;
1154         struct page *page;
1155         int error;
1156
1157         page = page_cache_alloc_cold(mapping);
1158         if (!page)
1159                 return -ENOMEM;
1160
1161         error = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1162         if (!error) {
1163                 error = mapping->a_ops->readpage(file, page);
1164                 page_cache_release(page);
1165                 return error;
1166         }
1167
1168         /*
1169          * We arrive here in the unlikely event that someone
1170          * raced with us and added our page to the cache first
1171          * or we are out of memory for radix-tree nodes.
1172          */
1173         page_cache_release(page);
1174         return error == -EEXIST ? 0 : error;
1175 }
1176
1177 #define MMAP_LOTSAMISS  (100)
1178
1179 /*
1180  * filemap_nopage() is invoked via the vma operations vector for a
1181  * mapped memory region to read in file data during a page fault.
1182  *
1183  * The goto's are kind of ugly, but this streamlines the normal case of having
1184  * it in the page cache, and handles the special cases reasonably without
1185  * having a lot of duplicated code.
1186  */
1187 struct page *filemap_nopage(struct vm_area_struct *area,
1188                                 unsigned long address, int *type)
1189 {
1190         int error;
1191         struct file *file = area->vm_file;
1192         struct address_space *mapping = file->f_mapping;
1193         struct file_ra_state *ra = &file->f_ra;
1194         struct inode *inode = mapping->host;
1195         struct page *page;
1196         unsigned long size, pgoff;
1197         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1198
1199         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1200
1201 retry_all:
1202         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1203         if (pgoff >= size)
1204                 goto outside_data_content;
1205
1206         /* If we don't want any read-ahead, don't bother */
1207         if (VM_RandomReadHint(area))
1208                 goto no_cached_page;
1209
1210         /*
1211          * The readahead code wants to be told about each and every page
1212          * so it can build and shrink its windows appropriately
1213          *
1214          * For sequential accesses, we use the generic readahead logic.
1215          */
1216         if (VM_SequentialReadHint(area))
1217                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1218
1219         /*
1220          * Do we have something in the page cache already?
1221          */
1222 retry_find:
1223         page = find_get_page(mapping, pgoff);
1224         if (!page) {
1225                 unsigned long ra_pages;
1226
1227                 if (VM_SequentialReadHint(area)) {
1228                         handle_ra_miss(mapping, ra, pgoff);
1229                         goto no_cached_page;
1230                 }
1231                 ra->mmap_miss++;
1232
1233                 /*
1234                  * Do we miss much more than hit in this file? If so,
1235                  * stop bothering with read-ahead. It will only hurt.
1236                  */
1237                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1238                         goto no_cached_page;
1239
1240                 /*
1241                  * To keep the pgmajfault counter straight, we need to
1242                  * check did_readaround, as this is an inner loop.
1243                  */
1244                 if (!did_readaround) {
1245                         majmin = VM_FAULT_MAJOR;
1246                         inc_page_state(pgmajfault);
1247                 }
1248                 did_readaround = 1;
1249                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1250                 if (ra_pages) {
1251                         pgoff_t start = 0;
1252
1253                         if (pgoff > ra_pages / 2)
1254                                 start = pgoff - ra_pages / 2;
1255                         do_page_cache_readahead(mapping, file, start, ra_pages);
1256                 }
1257                 page = find_get_page(mapping, pgoff);
1258                 if (!page)
1259                         goto no_cached_page;
1260         }
1261
1262         if (!did_readaround)
1263                 ra->mmap_hit++;
1264
1265         /*
1266          * Ok, found a page in the page cache, now we need to check
1267          * that it's up-to-date.
1268          */
1269         if (!PageUptodate(page))
1270                 goto page_not_uptodate;
1271
1272 success:
1273         /*
1274          * Found the page and have a reference on it.
1275          */
1276         mark_page_accessed(page);
1277         if (type)
1278                 *type = majmin;
1279         return page;
1280
1281 outside_data_content:
1282         /*
1283          * An external ptracer can access pages that normally aren't
1284          * accessible..
1285          */
1286         if (area->vm_mm == current->mm)
1287                 return NULL;
1288         /* Fall through to the non-read-ahead case */
1289 no_cached_page:
1290         /*
1291          * We're only likely to ever get here if MADV_RANDOM is in
1292          * effect.
1293          */
1294         error = page_cache_read(file, pgoff);
1295         grab_swap_token();
1296
1297         /*
1298          * The page we want has now been added to the page cache.
1299          * In the unlikely event that someone removed it in the
1300          * meantime, we'll just come back here and read it again.
1301          */
1302         if (error >= 0)
1303                 goto retry_find;
1304
1305         /*
1306          * An error return from page_cache_read can result if the
1307          * system is low on memory, or a problem occurs while trying
1308          * to schedule I/O.
1309          */
1310         if (error == -ENOMEM)
1311                 return NOPAGE_OOM;
1312         return NULL;
1313
1314 page_not_uptodate:
1315         if (!did_readaround) {
1316                 majmin = VM_FAULT_MAJOR;
1317                 inc_page_state(pgmajfault);
1318         }
1319         lock_page(page);
1320
1321         /* Did it get unhashed while we waited for it? */
1322         if (!page->mapping) {
1323                 unlock_page(page);
1324                 page_cache_release(page);
1325                 goto retry_all;
1326         }
1327
1328         /* Did somebody else get it up-to-date? */
1329         if (PageUptodate(page)) {
1330                 unlock_page(page);
1331                 goto success;
1332         }
1333
1334         if (!mapping->a_ops->readpage(file, page)) {
1335                 wait_on_page_locked(page);
1336                 if (PageUptodate(page))
1337                         goto success;
1338         }
1339
1340         /*
1341          * Umm, take care of errors if the page isn't up-to-date.
1342          * Try to re-read it _once_. We do this synchronously,
1343          * because there really aren't any performance issues here
1344          * and we need to check for errors.
1345          */
1346         lock_page(page);
1347
1348         /* Somebody truncated the page on us? */
1349         if (!page->mapping) {
1350                 unlock_page(page);
1351                 page_cache_release(page);
1352                 goto retry_all;
1353         }
1354
1355         /* Somebody else successfully read it in? */
1356         if (PageUptodate(page)) {
1357                 unlock_page(page);
1358                 goto success;
1359         }
1360         ClearPageError(page);
1361         if (!mapping->a_ops->readpage(file, page)) {
1362                 wait_on_page_locked(page);
1363                 if (PageUptodate(page))
1364                         goto success;
1365         }
1366
1367         /*
1368          * Things didn't work out. Return zero to tell the
1369          * mm layer so, possibly freeing the page cache page first.
1370          */
1371         page_cache_release(page);
1372         return NULL;
1373 }
1374
1375 EXPORT_SYMBOL(filemap_nopage);
1376
1377 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1378                                         int nonblock)
1379 {
1380         struct address_space *mapping = file->f_mapping;
1381         struct page *page;
1382         int error;
1383
1384         /*
1385          * Do we have something in the page cache already?
1386          */
1387 retry_find:
1388         page = find_get_page(mapping, pgoff);
1389         if (!page) {
1390                 if (nonblock)
1391                         return NULL;
1392                 goto no_cached_page;
1393         }
1394
1395         /*
1396          * Ok, found a page in the page cache, now we need to check
1397          * that it's up-to-date.
1398          */
1399         if (!PageUptodate(page)) {
1400                 if (nonblock) {
1401                         page_cache_release(page);
1402                         return NULL;
1403                 }
1404                 goto page_not_uptodate;
1405         }
1406
1407 success:
1408         /*
1409          * Found the page and have a reference on it.
1410          */
1411         mark_page_accessed(page);
1412         return page;
1413
1414 no_cached_page:
1415         error = page_cache_read(file, pgoff);
1416
1417         /*
1418          * The page we want has now been added to the page cache.
1419          * In the unlikely event that someone removed it in the
1420          * meantime, we'll just come back here and read it again.
1421          */
1422         if (error >= 0)
1423                 goto retry_find;
1424
1425         /*
1426          * An error return from page_cache_read can result if the
1427          * system is low on memory, or a problem occurs while trying
1428          * to schedule I/O.
1429          */
1430         return NULL;
1431
1432 page_not_uptodate:
1433         lock_page(page);
1434
1435         /* Did it get unhashed while we waited for it? */
1436         if (!page->mapping) {
1437                 unlock_page(page);
1438                 goto err;
1439         }
1440
1441         /* Did somebody else get it up-to-date? */
1442         if (PageUptodate(page)) {
1443                 unlock_page(page);
1444                 goto success;
1445         }
1446
1447         if (!mapping->a_ops->readpage(file, page)) {
1448                 wait_on_page_locked(page);
1449                 if (PageUptodate(page))
1450                         goto success;
1451         }
1452
1453         /*
1454          * Umm, take care of errors if the page isn't up-to-date.
1455          * Try to re-read it _once_. We do this synchronously,
1456          * because there really aren't any performance issues here
1457          * and we need to check for errors.
1458          */
1459         lock_page(page);
1460
1461         /* Somebody truncated the page on us? */
1462         if (!page->mapping) {
1463                 unlock_page(page);
1464                 goto err;
1465         }
1466         /* Somebody else successfully read it in? */
1467         if (PageUptodate(page)) {
1468                 unlock_page(page);
1469                 goto success;
1470         }
1471
1472         ClearPageError(page);
1473         if (!mapping->a_ops->readpage(file, page)) {
1474                 wait_on_page_locked(page);
1475                 if (PageUptodate(page))
1476                         goto success;
1477         }
1478
1479         /*
1480          * Things didn't work out. Return zero to tell the
1481          * mm layer so, possibly freeing the page cache page first.
1482          */
1483 err:
1484         page_cache_release(page);
1485
1486         return NULL;
1487 }
1488
1489 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1490                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1491                 int nonblock)
1492 {
1493         struct file *file = vma->vm_file;
1494         struct address_space *mapping = file->f_mapping;
1495         struct inode *inode = mapping->host;
1496         unsigned long size;
1497         struct mm_struct *mm = vma->vm_mm;
1498         struct page *page;
1499         int err;
1500
1501         if (!nonblock)
1502                 force_page_cache_readahead(mapping, vma->vm_file,
1503                                         pgoff, len >> PAGE_CACHE_SHIFT);
1504
1505 repeat:
1506         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1507         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1508                 return -EINVAL;
1509
1510         page = filemap_getpage(file, pgoff, nonblock);
1511         if (!page && !nonblock)
1512                 return -ENOMEM;
1513         if (page) {
1514                 err = install_page(mm, vma, addr, page, prot);
1515                 if (err) {
1516                         page_cache_release(page);
1517                         return err;
1518                 }
1519         } else {
1520                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1521                 if (err)
1522                         return err;
1523         }
1524
1525         len -= PAGE_SIZE;
1526         addr += PAGE_SIZE;
1527         pgoff++;
1528         if (len)
1529                 goto repeat;
1530
1531         return 0;
1532 }
1533
1534 struct vm_operations_struct generic_file_vm_ops = {
1535         .nopage         = filemap_nopage,
1536         .populate       = filemap_populate,
1537 };
1538
1539 /* This is used for a general mmap of a disk file */
1540
1541 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1542 {
1543         struct address_space *mapping = file->f_mapping;
1544
1545         if (!mapping->a_ops->readpage)
1546                 return -ENOEXEC;
1547         file_accessed(file);
1548         vma->vm_ops = &generic_file_vm_ops;
1549         return 0;
1550 }
1551 EXPORT_SYMBOL(filemap_populate);
1552
1553 /*
1554  * This is for filesystems which do not implement ->writepage.
1555  */
1556 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1557 {
1558         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1559                 return -EINVAL;
1560         return generic_file_mmap(file, vma);
1561 }
1562 #else
1563 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1564 {
1565         return -ENOSYS;
1566 }
1567 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1568 {
1569         return -ENOSYS;
1570 }
1571 #endif /* CONFIG_MMU */
1572
1573 EXPORT_SYMBOL(generic_file_mmap);
1574 EXPORT_SYMBOL(generic_file_readonly_mmap);
1575
1576 static inline struct page *__read_cache_page(struct address_space *mapping,
1577                                 unsigned long index,
1578                                 int (*filler)(void *,struct page*),
1579                                 void *data)
1580 {
1581         struct page *page, *cached_page = NULL;
1582         int err;
1583 repeat:
1584         page = find_get_page(mapping, index);
1585         if (!page) {
1586                 if (!cached_page) {
1587                         cached_page = page_cache_alloc_cold(mapping);
1588                         if (!cached_page)
1589                                 return ERR_PTR(-ENOMEM);
1590                 }
1591                 err = add_to_page_cache_lru(cached_page, mapping,
1592                                         index, GFP_KERNEL);
1593                 if (err == -EEXIST)
1594                         goto repeat;
1595                 if (err < 0) {
1596                         /* Presumably ENOMEM for radix tree node */
1597                         page_cache_release(cached_page);
1598                         return ERR_PTR(err);
1599                 }
1600                 page = cached_page;
1601                 cached_page = NULL;
1602                 err = filler(data, page);
1603                 if (err < 0) {
1604                         page_cache_release(page);
1605                         page = ERR_PTR(err);
1606                 }
1607         }
1608         if (cached_page)
1609                 page_cache_release(cached_page);
1610         return page;
1611 }
1612
1613 /*
1614  * Read into the page cache. If a page already exists,
1615  * and PageUptodate() is not set, try to fill the page.
1616  */
1617 struct page *read_cache_page(struct address_space *mapping,
1618                                 unsigned long index,
1619                                 int (*filler)(void *,struct page*),
1620                                 void *data)
1621 {
1622         struct page *page;
1623         int err;
1624
1625 retry:
1626         page = __read_cache_page(mapping, index, filler, data);
1627         if (IS_ERR(page))
1628                 goto out;
1629         mark_page_accessed(page);
1630         if (PageUptodate(page))
1631                 goto out;
1632
1633         lock_page(page);
1634         if (!page->mapping) {
1635                 unlock_page(page);
1636                 page_cache_release(page);
1637                 goto retry;
1638         }
1639         if (PageUptodate(page)) {
1640                 unlock_page(page);
1641                 goto out;
1642         }
1643         err = filler(data, page);
1644         if (err < 0) {
1645                 page_cache_release(page);
1646                 page = ERR_PTR(err);
1647         }
1648  out:
1649         return page;
1650 }
1651
1652 EXPORT_SYMBOL(read_cache_page);
1653
1654 /*
1655  * If the page was newly created, increment its refcount and add it to the
1656  * caller's lru-buffering pagevec.  This function is specifically for
1657  * generic_file_write().
1658  */
1659 static inline struct page *
1660 __grab_cache_page(struct address_space *mapping, unsigned long index,
1661                         struct page **cached_page, struct pagevec *lru_pvec)
1662 {
1663         int err;
1664         struct page *page;
1665 repeat:
1666         page = find_lock_page(mapping, index);
1667         if (!page) {
1668                 if (!*cached_page) {
1669                         *cached_page = page_cache_alloc(mapping);
1670                         if (!*cached_page)
1671                                 return NULL;
1672                 }
1673                 err = add_to_page_cache(*cached_page, mapping,
1674                                         index, GFP_KERNEL);
1675                 if (err == -EEXIST)
1676                         goto repeat;
1677                 if (err == 0) {
1678                         page = *cached_page;
1679                         page_cache_get(page);
1680                         if (!pagevec_add(lru_pvec, page))
1681                                 __pagevec_lru_add(lru_pvec);
1682                         *cached_page = NULL;
1683                 }
1684         }
1685         return page;
1686 }
1687
1688 /*
1689  * The logic we want is
1690  *
1691  *      if suid or (sgid and xgrp)
1692  *              remove privs
1693  */
1694 int remove_suid(struct dentry *dentry)
1695 {
1696         mode_t mode = dentry->d_inode->i_mode;
1697         int kill = 0;
1698         int result = 0;
1699
1700         /* suid always must be killed */
1701         if (unlikely(mode & S_ISUID))
1702                 kill = ATTR_KILL_SUID;
1703
1704         /*
1705          * sgid without any exec bits is just a mandatory locking mark; leave
1706          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1707          */
1708         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1709                 kill |= ATTR_KILL_SGID;
1710
1711         if (unlikely(kill && !capable(CAP_FSETID))) {
1712                 struct iattr newattrs;
1713
1714                 newattrs.ia_valid = ATTR_FORCE | kill;
1715                 result = notify_change(dentry, &newattrs);
1716         }
1717         return result;
1718 }
1719 EXPORT_SYMBOL(remove_suid);
1720
1721 /*
1722  * Copy as much as we can into the page and return the number of bytes which
1723  * were sucessfully copied.  If a fault is encountered then clear the page
1724  * out to (offset+bytes) and return the number of bytes which were copied.
1725  */
1726 static inline size_t
1727 filemap_copy_from_user(struct page *page, unsigned long offset,
1728                         const char __user *buf, unsigned bytes)
1729 {
1730         char *kaddr;
1731         int left;
1732
1733         kaddr = kmap_atomic(page, KM_USER0);
1734         left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1735         kunmap_atomic(kaddr, KM_USER0);
1736
1737         if (left != 0) {
1738                 /* Do it the slow way */
1739                 kaddr = kmap(page);
1740                 left = __copy_from_user(kaddr + offset, buf, bytes);
1741                 kunmap(page);
1742         }
1743         return bytes - left;
1744 }
1745
1746 static size_t
1747 __filemap_copy_from_user_iovec(char *vaddr,
1748                         const struct iovec *iov, size_t base, size_t bytes)
1749 {
1750         size_t copied = 0, left = 0;
1751
1752         while (bytes) {
1753                 char __user *buf = iov->iov_base + base;
1754                 int copy = min(bytes, iov->iov_len - base);
1755
1756                 base = 0;
1757                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1758                 copied += copy;
1759                 bytes -= copy;
1760                 vaddr += copy;
1761                 iov++;
1762
1763                 if (unlikely(left)) {
1764                         /* zero the rest of the target like __copy_from_user */
1765                         if (bytes)
1766                                 memset(vaddr, 0, bytes);
1767                         break;
1768                 }
1769         }
1770         return copied - left;
1771 }
1772
1773 /*
1774  * This has the same sideeffects and return value as filemap_copy_from_user().
1775  * The difference is that on a fault we need to memset the remainder of the
1776  * page (out to offset+bytes), to emulate filemap_copy_from_user()'s
1777  * single-segment behaviour.
1778  */
1779 static inline size_t
1780 filemap_copy_from_user_iovec(struct page *page, unsigned long offset,
1781                         const struct iovec *iov, size_t base, size_t bytes)
1782 {
1783         char *kaddr;
1784         size_t copied;
1785
1786         kaddr = kmap_atomic(page, KM_USER0);
1787         copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1788                                                 base, bytes);
1789         kunmap_atomic(kaddr, KM_USER0);
1790         if (copied != bytes) {
1791                 kaddr = kmap(page);
1792                 copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1793                                                         base, bytes);
1794                 kunmap(page);
1795         }
1796         return copied;
1797 }
1798
1799 static inline void
1800 filemap_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1801 {
1802         const struct iovec *iov = *iovp;
1803         size_t base = *basep;
1804
1805         while (bytes) {
1806                 int copy = min(bytes, iov->iov_len - base);
1807
1808                 bytes -= copy;
1809                 base += copy;
1810                 if (iov->iov_len == base) {
1811                         iov++;
1812                         base = 0;
1813                 }
1814         }
1815         *iovp = iov;
1816         *basep = base;
1817 }
1818
1819 /*
1820  * Performs necessary checks before doing a write
1821  *
1822  * Can adjust writing position aor amount of bytes to write.
1823  * Returns appropriate error code that caller should return or
1824  * zero in case that write should be allowed.
1825  */
1826 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1827 {
1828         struct inode *inode = file->f_mapping->host;
1829         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1830
1831         if (unlikely(*pos < 0))
1832                 return -EINVAL;
1833
1834         if (unlikely(file->f_error)) {
1835                 int err = file->f_error;
1836                 file->f_error = 0;
1837                 return err;
1838         }
1839
1840         if (!isblk) {
1841                 /* FIXME: this is for backwards compatibility with 2.4 */
1842                 if (file->f_flags & O_APPEND)
1843                         *pos = i_size_read(inode);
1844
1845                 if (limit != RLIM_INFINITY) {
1846                         if (*pos >= limit) {
1847                                 send_sig(SIGXFSZ, current, 0);
1848                                 return -EFBIG;
1849                         }
1850                         if (*count > limit - (typeof(limit))*pos) {
1851                                 *count = limit - (typeof(limit))*pos;
1852                         }
1853                 }
1854         }
1855
1856         /*
1857          * LFS rule
1858          */
1859         if (unlikely(*pos + *count > MAX_NON_LFS &&
1860                                 !(file->f_flags & O_LARGEFILE))) {
1861                 if (*pos >= MAX_NON_LFS) {
1862                         send_sig(SIGXFSZ, current, 0);
1863                         return -EFBIG;
1864                 }
1865                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1866                         *count = MAX_NON_LFS - (unsigned long)*pos;
1867                 }
1868         }
1869
1870         /*
1871          * Are we about to exceed the fs block limit ?
1872          *
1873          * If we have written data it becomes a short write.  If we have
1874          * exceeded without writing data we send a signal and return EFBIG.
1875          * Linus frestrict idea will clean these up nicely..
1876          */
1877         if (likely(!isblk)) {
1878                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1879                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1880                                 send_sig(SIGXFSZ, current, 0);
1881                                 return -EFBIG;
1882                         }
1883                         /* zero-length writes at ->s_maxbytes are OK */
1884                 }
1885
1886                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1887                         *count = inode->i_sb->s_maxbytes - *pos;
1888         } else {
1889                 loff_t isize;
1890                 if (bdev_read_only(I_BDEV(inode)))
1891                         return -EPERM;
1892                 isize = i_size_read(inode);
1893                 if (*pos >= isize) {
1894                         if (*count || *pos > isize)
1895                                 return -ENOSPC;
1896                 }
1897
1898                 if (*pos + *count > isize)
1899                         *count = isize - *pos;
1900         }
1901         return 0;
1902 }
1903 EXPORT_SYMBOL(generic_write_checks);
1904
1905 ssize_t
1906 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1907                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1908                 size_t count, size_t ocount)
1909 {
1910         struct file     *file = iocb->ki_filp;
1911         struct address_space *mapping = file->f_mapping;
1912         struct inode    *inode = mapping->host;
1913         ssize_t         written;
1914
1915         if (count != ocount)
1916                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1917
1918         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1919         if (written > 0) {
1920                 loff_t end = pos + written;
1921                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1922                         i_size_write(inode,  end);
1923                         mark_inode_dirty(inode);
1924                 }
1925                 *ppos = end;
1926         }
1927
1928         /*
1929          * Sync the fs metadata but not the minor inode changes and
1930          * of course not the data as we did direct DMA for the IO.
1931          * i_sem is held, which protects generic_osync_inode() from
1932          * livelocking.
1933          */
1934         if (written >= 0 && file->f_flags & O_SYNC)
1935                 generic_osync_inode(inode, mapping, OSYNC_METADATA);
1936         if (written == count && !is_sync_kiocb(iocb))
1937                 written = -EIOCBQUEUED;
1938         return written;
1939 }
1940 EXPORT_SYMBOL(generic_file_direct_write);
1941
1942 ssize_t
1943 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1944                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1945                 size_t count, ssize_t written)
1946 {
1947         struct file *file = iocb->ki_filp;
1948         struct address_space * mapping = file->f_mapping;
1949         struct address_space_operations *a_ops = mapping->a_ops;
1950         struct inode    *inode = mapping->host;
1951         long            status = 0;
1952         struct page     *page;
1953         struct page     *cached_page = NULL;
1954         size_t          bytes;
1955         struct pagevec  lru_pvec;
1956         const struct iovec *cur_iov = iov; /* current iovec */
1957         size_t          iov_base = 0;      /* offset in the current iovec */
1958         char __user     *buf;
1959
1960         pagevec_init(&lru_pvec, 0);
1961
1962         /*
1963          * handle partial DIO write.  Adjust cur_iov if needed.
1964          */
1965         if (likely(nr_segs == 1))
1966                 buf = iov->iov_base + written;
1967         else {
1968                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1969                 buf = cur_iov->iov_base + iov_base;
1970         }
1971
1972         do {
1973                 unsigned long index;
1974                 unsigned long offset;
1975                 size_t copied;
1976
1977                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1978                 index = pos >> PAGE_CACHE_SHIFT;
1979                 bytes = PAGE_CACHE_SIZE - offset;
1980                 if (bytes > count)
1981                         bytes = count;
1982
1983                 /*
1984                  * Bring in the user page that we will copy from _first_.
1985                  * Otherwise there's a nasty deadlock on copying from the
1986                  * same page as we're writing to, without it being marked
1987                  * up-to-date.
1988                  */
1989                 fault_in_pages_readable(buf, bytes);
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                                 }
2032                         }
2033                 }
2034                 if (unlikely(copied != bytes))
2035                         if (status >= 0)
2036                                 status = -EFAULT;
2037                 unlock_page(page);
2038                 mark_page_accessed(page);
2039                 page_cache_release(page);
2040                 if (status < 0)
2041                         break;
2042                 balance_dirty_pages_ratelimited(mapping);
2043                 cond_resched();
2044         } while (count);
2045         *ppos = pos;
2046
2047         if (cached_page)
2048                 page_cache_release(cached_page);
2049
2050         /*
2051          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2052          */
2053         if (likely(status >= 0)) {
2054                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2055                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2056                                 status = generic_osync_inode(inode, mapping,
2057                                                 OSYNC_METADATA|OSYNC_DATA);
2058                 }
2059         }
2060
2061         /*
2062          * If we get here for O_DIRECT writes then we must have fallen through
2063          * to buffered writes (block instantiation inside i_size).  So we sync
2064          * the file data here, to try to honour O_DIRECT expectations.
2065          */
2066         if (unlikely(file->f_flags & O_DIRECT) && written)
2067                 status = filemap_write_and_wait(mapping);
2068
2069         pagevec_lru_add(&lru_pvec);
2070         return written ? written : status;
2071 }
2072 EXPORT_SYMBOL(generic_file_buffered_write);
2073
2074 ssize_t
2075 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2076                                 unsigned long nr_segs, loff_t *ppos)
2077 {
2078         struct file *file = iocb->ki_filp;
2079         struct address_space * mapping = file->f_mapping;
2080         size_t ocount;          /* original count */
2081         size_t count;           /* after file limit checks */
2082         struct inode    *inode = mapping->host;
2083         unsigned long   seg;
2084         loff_t          pos;
2085         ssize_t         written;
2086         ssize_t         err;
2087
2088         ocount = 0;
2089         for (seg = 0; seg < nr_segs; seg++) {
2090                 const struct iovec *iv = &iov[seg];
2091
2092                 /*
2093                  * If any segment has a negative length, or the cumulative
2094                  * length ever wraps negative then return -EINVAL.
2095                  */
2096                 ocount += iv->iov_len;
2097                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2098                         return -EINVAL;
2099                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2100                         continue;
2101                 if (seg == 0)
2102                         return -EFAULT;
2103                 nr_segs = seg;
2104                 ocount -= iv->iov_len;  /* This segment is no good */
2105                 break;
2106         }
2107
2108         count = ocount;
2109         pos = *ppos;
2110
2111         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2112
2113         /* We can write back this queue in page reclaim */
2114         current->backing_dev_info = mapping->backing_dev_info;
2115         written = 0;
2116
2117         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2118         if (err)
2119                 goto out;
2120
2121         if (count == 0)
2122                 goto out;
2123
2124         err = remove_suid(file->f_dentry);
2125         if (err)
2126                 goto out;
2127
2128         inode_update_time(inode, 1);
2129
2130         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2131         if (unlikely(file->f_flags & O_DIRECT)) {
2132                 written = generic_file_direct_write(iocb, iov,
2133                                 &nr_segs, pos, ppos, count, ocount);
2134                 if (written < 0 || written == count)
2135                         goto out;
2136                 /*
2137                  * direct-io write to a hole: fall through to buffered I/O
2138                  * for completing the rest of the request.
2139                  */
2140                 pos += written;
2141                 count -= written;
2142         }
2143
2144         written = generic_file_buffered_write(iocb, iov, nr_segs,
2145                         pos, ppos, count, written);
2146 out:
2147         current->backing_dev_info = NULL;
2148         return written ? written : err;
2149 }
2150 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2151
2152 ssize_t
2153 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2154                                 unsigned long nr_segs, loff_t *ppos)
2155 {
2156         struct file *file = iocb->ki_filp;
2157         struct address_space *mapping = file->f_mapping;
2158         struct inode *inode = mapping->host;
2159         ssize_t ret;
2160         loff_t pos = *ppos;
2161
2162         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2163
2164         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2165                 int err;
2166
2167                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2168                 if (err < 0)
2169                         ret = err;
2170         }
2171         return ret;
2172 }
2173
2174 ssize_t
2175 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2176                                 unsigned long nr_segs, loff_t *ppos)
2177 {
2178         struct kiocb kiocb;
2179         ssize_t ret;
2180
2181         init_sync_kiocb(&kiocb, file);
2182         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2183         if (ret == -EIOCBQUEUED)
2184                 ret = wait_on_sync_kiocb(&kiocb);
2185         return ret;
2186 }
2187
2188 ssize_t
2189 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2190                                 unsigned long nr_segs, loff_t *ppos)
2191 {
2192         struct kiocb kiocb;
2193         ssize_t ret;
2194
2195         init_sync_kiocb(&kiocb, file);
2196         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2197         if (-EIOCBQUEUED == ret)
2198                 ret = wait_on_sync_kiocb(&kiocb);
2199         return ret;
2200 }
2201 EXPORT_SYMBOL(generic_file_write_nolock);
2202
2203 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2204                                size_t count, loff_t pos)
2205 {
2206         struct file *file = iocb->ki_filp;
2207         struct address_space *mapping = file->f_mapping;
2208         struct inode *inode = mapping->host;
2209         ssize_t ret;
2210         struct iovec local_iov = { .iov_base = (void __user *)buf,
2211                                         .iov_len = count };
2212
2213         BUG_ON(iocb->ki_pos != pos);
2214
2215         down(&inode->i_sem);
2216         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2217                                                 &iocb->ki_pos);
2218         up(&inode->i_sem);
2219
2220         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2221                 ssize_t err;
2222
2223                 err = sync_page_range(inode, mapping, pos, ret);
2224                 if (err < 0)
2225                         ret = err;
2226         }
2227         return ret;
2228 }
2229 EXPORT_SYMBOL(generic_file_aio_write);
2230
2231 ssize_t generic_file_write(struct file *file, const char __user *buf,
2232                            size_t count, loff_t *ppos)
2233 {
2234         struct address_space *mapping = file->f_mapping;
2235         struct inode *inode = mapping->host;
2236         ssize_t ret;
2237         struct iovec local_iov = { .iov_base = (void __user *)buf,
2238                                         .iov_len = count };
2239
2240         down(&inode->i_sem);
2241         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2242         up(&inode->i_sem);
2243
2244         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2245                 ssize_t err;
2246
2247                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2248                 if (err < 0)
2249                         ret = err;
2250         }
2251         return ret;
2252 }
2253 EXPORT_SYMBOL(generic_file_write);
2254
2255 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2256                         unsigned long nr_segs, loff_t *ppos)
2257 {
2258         struct kiocb kiocb;
2259         ssize_t ret;
2260
2261         init_sync_kiocb(&kiocb, filp);
2262         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2263         if (-EIOCBQUEUED == ret)
2264                 ret = wait_on_sync_kiocb(&kiocb);
2265         return ret;
2266 }
2267 EXPORT_SYMBOL(generic_file_readv);
2268
2269 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2270                         unsigned long nr_segs, loff_t *ppos)
2271 {
2272         struct address_space *mapping = file->f_mapping;
2273         struct inode *inode = mapping->host;
2274         ssize_t ret;
2275
2276         down(&inode->i_sem);
2277         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2278         up(&inode->i_sem);
2279
2280         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2281                 int err;
2282
2283                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2284                 if (err < 0)
2285                         ret = err;
2286         }
2287         return ret;
2288 }
2289 EXPORT_SYMBOL(generic_file_writev);
2290
2291 /*
2292  * Called under i_sem for writes to S_ISREG files.   Returns -EIO if something
2293  * went wrong during pagecache shootdown.
2294  */
2295 ssize_t
2296 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2297         loff_t offset, unsigned long nr_segs)
2298 {
2299         struct file *file = iocb->ki_filp;
2300         struct address_space *mapping = file->f_mapping;
2301         ssize_t retval;
2302         size_t write_len = 0;
2303
2304         /*
2305          * If it's a write, unmap all mmappings of the file up-front.  This
2306          * will cause any pte dirty bits to be propagated into the pageframes
2307          * for the subsequent filemap_write_and_wait().
2308          */
2309         if (rw == WRITE) {
2310                 write_len = iov_length(iov, nr_segs);
2311                 if (mapping_mapped(mapping))
2312                         unmap_mapping_range(mapping, offset, write_len, 0);
2313         }
2314
2315         retval = filemap_write_and_wait(mapping);
2316         if (retval == 0) {
2317                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2318                                                 offset, nr_segs);
2319                 if (rw == WRITE && mapping->nrpages) {
2320                         pgoff_t end = (offset + write_len - 1)
2321                                                 >> PAGE_CACHE_SHIFT;
2322                         int err = invalidate_inode_pages2_range(mapping,
2323                                         offset >> PAGE_CACHE_SHIFT, end);
2324                         if (err)
2325                                 retval = err;
2326                 }
2327         }
2328         return retval;
2329 }
2330 EXPORT_SYMBOL_GPL(generic_file_direct_IO);