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