mm/filemap.c

   1 /*
   2  *      linux/mm/filemap.c
   3  *
   4  * Copyright (C) 1994-1999  Linus Torvalds
   5  */
   6
   7 /*
   8  * This file handles the generic file mmap semantics used by
   9  * most "normal" filesystems (but you don't /have/ to use this:
  10  * the NFS filesystem used to do this differently, for example)
  11  */
  12 #include <linux/config.h>
  13 #include <linux/module.h>
  14 #include <linux/slab.h>
  15 #include <linux/compiler.h>
  16 #include <linux/fs.h>
  17 #include <linux/aio.h>
  18 #include <linux/kernel_stat.h>
  19 #include <linux/mm.h>
  20 #include <linux/swap.h>
  21 #include <linux/mman.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/file.h>
  24 #include <linux/uio.h>
  25 #include <linux/hash.h>
  26 #include <linux/writeback.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/security.h>
  30 #include <linux/syscalls.h>
  31 /*
  32  * This is needed for the following functions:
  33  *  - try_to_release_page
  34  *  - block_invalidatepage
  35  *  - generic_osync_inode
  36  *
  37  * FIXME: remove all knowledge of the buffer layer from the core VM
  38  */
  39 #include <linux/buffer_head.h> /* for generic_osync_inode */
  40
  41 #include <asm/uaccess.h>
  42 #include <asm/mman.h>
  43
  44 /*
  45  * Shared mappings implemented 30.11.1994. It's not fully working yet,
  46  * though.
  47  *
  48  * Shared mappings now work. 15.8.1995  Bruno.
  49  *
  50  * finished 'unifying' the page and buffer cache and SMP-threaded the
  51  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
  52  *
  53  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
  54  */
  55
  56 /*
  57  * Lock ordering:
  58  *
  59  *  ->i_mmap_lock               (vmtruncate)
  60  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
  61  *      ->swap_list_lock
  62  *        ->swap_device_lock    (exclusive_swap_page, others)
  63  *          ->mapping->tree_lock
  64  *
  65  *  ->i_sem
  66  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
  67  *
  68  *  ->mmap_sem
  69  *    ->i_mmap_lock
  70  *      ->page_table_lock       (various places, mainly in mmap.c)
  71  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
  72  *
  73  *  ->mmap_sem
  74  *    ->lock_page               (access_process_vm)
  75  *
  76  *  ->mmap_sem
  77  *    ->i_sem                   (msync)
  78  *
  79  *  ->i_sem
  80  *    ->i_alloc_sem             (various)
  81  *
  82  *  ->inode_lock
  83  *    ->sb_lock                 (fs/fs-writeback.c)
  84  *    ->mapping->tree_lock      (__sync_single_inode)
  85  *
  86  *  ->i_mmap_lock
  87  *    ->anon_vma.lock           (vma_adjust)
  88  *
  89  *  ->anon_vma.lock
  90  *    ->page_table_lock         (anon_vma_prepare and various)
  91  *
  92  *  ->page_table_lock
  93  *    ->swap_device_lock        (try_to_unmap_one)
  94  *    ->private_lock            (try_to_unmap_one)
  95  *    ->tree_lock               (try_to_unmap_one)
  96  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
  97  *    ->private_lock            (page_remove_rmap->set_page_dirty)
  98  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
  99  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
 100  *    ->inode_lock              (zap_pte_range->set_page_dirty)
 101  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
 102  *
 103  *  ->task->proc_lock
 104  *    ->dcache_lock             (proc_pid_lookup)
 105  */
 106
 107 /*
 108  * Remove a page from the page cache and free it. Caller has to make
 109  * sure the page is locked and that nobody else uses it - or that usage
 110  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
 111  */
 112 void __remove_from_page_cache(struct page *page)
 113 {
 114         struct address_space *mapping = page->mapping;
 115
 116         radix_tree_delete(&mapping->page_tree, page->index);
 117         page->mapping = NULL;
 118         mapping->nrpages--;
 119         pagecache_acct(-1);
 120 }
 121
 122 void remove_from_page_cache(struct page *page)
 123 {
 124         struct address_space *mapping = page->mapping;
 125
 126         if (unlikely(!PageLocked(page)))
 127                 PAGE_BUG(page);
 128
 129         write_lock_irq(&mapping->tree_lock);
 130         __remove_from_page_cache(page);
 131         write_unlock_irq(&mapping->tree_lock);
 132 }
 133
 134 static int sync_page(void *word)
 135 {
 136         struct address_space *mapping;
 137         struct page *page;
 138
 139         page = container_of((page_flags_t *)word, struct page, flags);
 140
 141         /*
 142          * FIXME, fercrissake.  What is this barrier here for?
 143          */
 144         smp_mb();
 145         mapping = page_mapping(page);
 146         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 147                 mapping->a_ops->sync_page(page);
 148         io_schedule();
 149         return 0;
 150 }
 151
 152 /**
 153  * filemap_fdatawrite_range - start writeback against all of a mapping's
 154  * dirty pages that lie within the byte offsets <start, end>
 155  * @mapping: address space structure to write
 156  * @start: offset in bytes where the range starts
 157  * @end : offset in bytes where the range ends
 158  *
 159  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 160  * opposed to a regular memory * cleansing writeback.  The difference between
 161  * these two operations is that if a dirty page/buffer is encountered, it must
 162  * be waited upon, and not just skipped over.
 163  */
 164 static int __filemap_fdatawrite_range(struct address_space *mapping,
 165         loff_t start, loff_t end, int sync_mode)
 166 {
 167         int ret;
 168         struct writeback_control wbc = {
 169                 .sync_mode = sync_mode,
 170                 .nr_to_write = mapping->nrpages * 2,
 171                 .start = start,
 172                 .end = end,
 173         };
 174
 175         if (!mapping_cap_writeback_dirty(mapping))
 176                 return 0;
 177
 178         ret = do_writepages(mapping, &wbc);
 179         return ret;
 180 }
 181
 182 static inline int __filemap_fdatawrite(struct address_space *mapping,
 183         int sync_mode)
 184 {
 185         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
 186 }
 187
 188 int filemap_fdatawrite(struct address_space *mapping)
 189 {
 190         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 191 }
 192 EXPORT_SYMBOL(filemap_fdatawrite);
 193
 194 static int filemap_fdatawrite_range(struct address_space *mapping,
 195         loff_t start, loff_t end)
 196 {
 197         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
 198 }
 199
 200 /*
 201  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 202  * purposes - I/O may not be started against all dirty pages.
 203  */
 204 int filemap_flush(struct address_space *mapping)
 205 {
 206         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 207 }
 208 EXPORT_SYMBOL(filemap_flush);
 209
 210 /*
 211  * Wait for writeback to complete against pages indexed by start->end
 212  * inclusive
 213  */
 214 static int wait_on_page_writeback_range(struct address_space *mapping,
 215                                 pgoff_t start, pgoff_t end)
 216 {
 217         struct pagevec pvec;
 218         int nr_pages;
 219         int ret = 0;
 220         pgoff_t index;
 221
 222         if (end < start)
 223                 return 0;
 224
 225         pagevec_init(&pvec, 0);
 226         index = start;
 227         while ((index <= end) &&
 228                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 229                         PAGECACHE_TAG_WRITEBACK,
 230                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 231                 unsigned i;
 232
 233                 for (i = 0; i < nr_pages; i++) {
 234                         struct page *page = pvec.pages[i];
 235
 236                         /* until radix tree lookup accepts end_index */
 237                         if (page->index > end)
 238                                 continue;
 239
 240                         wait_on_page_writeback(page);
 241                         if (PageError(page))
 242                                 ret = -EIO;
 243                 }
 244                 pagevec_release(&pvec);
 245                 cond_resched();
 246         }
 247
 248         /* Check for outstanding write errors */
 249         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 250                 ret = -ENOSPC;
 251         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 252                 ret = -EIO;
 253
 254         return ret;
 255 }
 256
 257 /*
 258  * Write and wait upon all the pages in the passed range.  This is a "data
 259  * integrity" operation.  It waits upon in-flight writeout before starting and
 260  * waiting upon new writeout.  If there was an IO error, return it.
 261  *
 262  * We need to re-take i_sem during the generic_osync_inode list walk because
 263  * it is otherwise livelockable.
 264  */
 265 int sync_page_range(struct inode *inode, struct address_space *mapping,
 266                         loff_t pos, size_t count)
 267 {
 268         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 269         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 270         int ret;
 271
 272         if (!mapping_cap_writeback_dirty(mapping) || !count)
 273                 return 0;
 274         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 275         if (ret == 0) {
 276                 down(&inode->i_sem);
 277                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 278                 up(&inode->i_sem);
 279         }
 280         if (ret == 0)
 281                 ret = wait_on_page_writeback_range(mapping, start, end);
 282         return ret;
 283 }
 284 EXPORT_SYMBOL(sync_page_range);
 285
 286 /*
 287  * Note: Holding i_sem across sync_page_range_nolock is not a good idea
 288  * as it forces O_SYNC writers to different parts of the same file
 289  * to be serialised right until io completion.
 290  */
 291 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
 292                         loff_t pos, size_t count)
 293 {
 294         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 295         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 296         int ret;
 297
 298         if (!mapping_cap_writeback_dirty(mapping) || !count)
 299                 return 0;
 300         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 301         if (ret == 0)
 302                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 303         if (ret == 0)
 304                 ret = wait_on_page_writeback_range(mapping, start, end);
 305         return ret;
 306 }
 307 EXPORT_SYMBOL(sync_page_range_nolock);
 308
 309 /**
 310  * filemap_fdatawait - walk the list of under-writeback pages of the given
 311  *     address space and wait for all of them.
 312  *
 313  * @mapping: address space structure to wait for
 314  */
 315 int filemap_fdatawait(struct address_space *mapping)
 316 {
 317         loff_t i_size = i_size_read(mapping->host);
 318
 319         if (i_size == 0)
 320                 return 0;
 321
 322         return wait_on_page_writeback_range(mapping, 0,
 323                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 324 }
 325 EXPORT_SYMBOL(filemap_fdatawait);
 326
 327 int filemap_write_and_wait(struct address_space *mapping)
 328 {
 329         int retval = 0;
 330
 331         if (mapping->nrpages) {
 332                 retval = filemap_fdatawrite(mapping);
 333                 if (retval == 0)
 334                         retval = filemap_fdatawait(mapping);
 335         }
 336         return retval;
 337 }
 338
 339 int filemap_write_and_wait_range(struct address_space *mapping,
 340                                  loff_t lstart, loff_t lend)
 341 {
 342         int retval = 0;
 343
 344         if (mapping->nrpages) {
 345                 retval = __filemap_fdatawrite_range(mapping, lstart, lend,
 346                                                     WB_SYNC_ALL);
 347                 if (retval == 0)
 348                         retval = wait_on_page_writeback_range(mapping,
 349                                                     lstart >> PAGE_CACHE_SHIFT,
 350                                                     lend >> PAGE_CACHE_SHIFT);
 351         }
 352         return retval;
 353 }
 354
 355 /*
 356  * This function is used to add newly allocated pagecache pages:
 357  * the page is new, so we can just run SetPageLocked() against it.
 358  * The other page state flags were set by rmqueue().
 359  *
 360  * This function does not add the page to the LRU.  The caller must do that.
 361  */
 362 int add_to_page_cache(struct page *page, struct address_space *mapping,
 363                 pgoff_t offset, int gfp_mask)
 364 {
 365         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 366
 367         if (error == 0) {
 368                 write_lock_irq(&mapping->tree_lock);
 369                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 370                 if (!error) {
 371                         page_cache_get(page);
 372                         SetPageLocked(page);
 373                         page->mapping = mapping;
 374                         page->index = offset;
 375                         mapping->nrpages++;
 376                         pagecache_acct(1);
 377                 }
 378                 write_unlock_irq(&mapping->tree_lock);
 379                 radix_tree_preload_end();
 380         }
 381         return error;
 382 }
 383
 384 EXPORT_SYMBOL(add_to_page_cache);
 385
 386 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 387                                 pgoff_t offset, int gfp_mask)
 388 {
 389         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 390         if (ret == 0)
 391                 lru_cache_add(page);
 392         return ret;
 393 }
 394
 395 /*
 396  * In order to wait for pages to become available there must be
 397  * waitqueues associated with pages. By using a hash table of
 398  * waitqueues where the bucket discipline is to maintain all
 399  * waiters on the same queue and wake all when any of the pages
 400  * become available, and for the woken contexts to check to be
 401  * sure the appropriate page became available, this saves space
 402  * at a cost of "thundering herd" phenomena during rare hash
 403  * collisions.
 404  */
 405 static wait_queue_head_t *page_waitqueue(struct page *page)
 406 {
 407         const struct zone *zone = page_zone(page);
 408
 409         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 410 }
 411
 412 static inline void wake_up_page(struct page *page, int bit)
 413 {
 414         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 415 }
 416
 417 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 418 {
 419         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 420
 421         if (test_bit(bit_nr, &page->flags))
 422                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 423                                                         TASK_UNINTERRUPTIBLE);
 424 }
 425 EXPORT_SYMBOL(wait_on_page_bit);
 426
 427 /**
 428  * unlock_page() - unlock a locked page
 429  *
 430  * @page: the page
 431  *
 432  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 433  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 434  * mechananism between PageLocked pages and PageWriteback pages is shared.
 435  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 436  *
 437  * The first mb is necessary to safely close the critical section opened by the
 438  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 439  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 440  * parallel wait_on_page_locked()).
 441  */
 442 void fastcall unlock_page(struct page *page)
 443 {
 444         smp_mb__before_clear_bit();
 445         if (!TestClearPageLocked(page))
 446                 BUG();
 447         smp_mb__after_clear_bit();
 448         wake_up_page(page, PG_locked);
 449 }
 450 EXPORT_SYMBOL(unlock_page);
 451
 452 /*
 453  * End writeback against a page.
 454  */
 455 void end_page_writeback(struct page *page)
 456 {
 457         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 458                 if (!test_clear_page_writeback(page))
 459                         BUG();
 460         }
 461         smp_mb__after_clear_bit();
 462         wake_up_page(page, PG_writeback);
 463 }
 464 EXPORT_SYMBOL(end_page_writeback);
 465
 466 /*
 467  * Get a lock on the page, assuming we need to sleep to get it.
 468  *
 469  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 470  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 471  * chances are that on the second loop, the block layer's plug list is empty,
 472  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 473  */
 474 void fastcall __lock_page(struct page *page)
 475 {
 476         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 477
 478         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 479                                                         TASK_UNINTERRUPTIBLE);
 480 }
 481 EXPORT_SYMBOL(__lock_page);
 482
 483 /*
 484  * a rather lightweight function, finding and getting a reference to a
 485  * hashed page atomically.
 486  */
 487 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 488 {
 489         struct page *page;
 490
 491         read_lock_irq(&mapping->tree_lock);
 492         page = radix_tree_lookup(&mapping->page_tree, offset);
 493         if (page)
 494                 page_cache_get(page);
 495         read_unlock_irq(&mapping->tree_lock);
 496         return page;
 497 }
 498
 499 EXPORT_SYMBOL(find_get_page);
 500
 501 /*
 502  * Same as above, but trylock it instead of incrementing the count.
 503  */
 504 struct page *find_trylock_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 && TestSetPageLocked(page))
 511                 page = NULL;
 512         read_unlock_irq(&mapping->tree_lock);
 513         return page;
 514 }
 515
 516 EXPORT_SYMBOL(find_trylock_page);
 517
 518 /**
 519  * find_lock_page - locate, pin and lock a pagecache page
 520  *
 521  * @mapping - the address_space to search
 522  * @offset - the page index
 523  *
 524  * Locates the desired pagecache page, locks it, increments its reference
 525  * count and returns its address.
 526  *
 527  * Returns zero if the page was not present. find_lock_page() may sleep.
 528  */
 529 struct page *find_lock_page(struct address_space *mapping,
 530                                 unsigned long offset)
 531 {
 532         struct page *page;
 533
 534         read_lock_irq(&mapping->tree_lock);
 535 repeat:
 536         page = radix_tree_lookup(&mapping->page_tree, offset);
 537         if (page) {
 538                 page_cache_get(page);
 539                 if (TestSetPageLocked(page)) {
 540                         read_unlock_irq(&mapping->tree_lock);
 541                         lock_page(page);
 542                         read_lock_irq(&mapping->tree_lock);
 543
 544                         /* Has the page been truncated while we slept? */
 545                         if (page->mapping != mapping || page->index != offset) {
 546                                 unlock_page(page);
 547                                 page_cache_release(page);
 548                                 goto repeat;
 549                         }
 550                 }
 551         }
 552         read_unlock_irq(&mapping->tree_lock);
 553         return page;
 554 }
 555
 556 EXPORT_SYMBOL(find_lock_page);
 557
 558 /**
 559  * find_or_create_page - locate or add a pagecache page
 560  *
 561  * @mapping - the page's address_space
 562  * @index - the page's index into the mapping
 563  * @gfp_mask - page allocation mode
 564  *
 565  * Locates a page in the pagecache.  If the page is not present, a new page
 566  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 567  * LRU list.  The returned page is locked and has its reference count
 568  * incremented.
 569  *
 570  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 571  * allocation!
 572  *
 573  * find_or_create_page() returns the desired page's address, or zero on
 574  * memory exhaustion.
 575  */
 576 struct page *find_or_create_page(struct address_space *mapping,
 577                 unsigned long index, unsigned int gfp_mask)
 578 {
 579         struct page *page, *cached_page = NULL;
 580         int err;
 581 repeat:
 582         page = find_lock_page(mapping, index);
 583         if (!page) {
 584                 if (!cached_page) {
 585                         cached_page = alloc_page(gfp_mask);
 586                         if (!cached_page)
 587                                 return NULL;
 588                 }
 589                 err = add_to_page_cache_lru(cached_page, mapping,
 590                                         index, gfp_mask);
 591                 if (!err) {
 592                         page = cached_page;
 593                         cached_page = NULL;
 594                 } else if (err == -EEXIST)
 595                         goto repeat;
 596         }
 597         if (cached_page)
 598                 page_cache_release(cached_page);
 599         return page;
 600 }
 601
 602 EXPORT_SYMBOL(find_or_create_page);
 603
 604 /**
 605  * find_get_pages - gang pagecache lookup
 606  * @mapping:    The address_space to search
 607  * @start:      The starting page index
 608  * @nr_pages:   The maximum number of pages
 609  * @pages:      Where the resulting pages are placed
 610  *
 611  * find_get_pages() will search for and return a group of up to
 612  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 613  * find_get_pages() takes a reference against the returned pages.
 614  *
 615  * The search returns a group of mapping-contiguous pages with ascending
 616  * indexes.  There may be holes in the indices due to not-present pages.
 617  *
 618  * find_get_pages() returns the number of pages which were found.
 619  */
 620 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 621                             unsigned int nr_pages, struct page **pages)
 622 {
 623         unsigned int i;
 624         unsigned int ret;
 625
 626         read_lock_irq(&mapping->tree_lock);
 627         ret = radix_tree_gang_lookup(&mapping->page_tree,
 628                                 (void **)pages, start, nr_pages);
 629         for (i = 0; i < ret; i++)
 630                 page_cache_get(pages[i]);
 631         read_unlock_irq(&mapping->tree_lock);
 632         return ret;
 633 }
 634
 635 /*
 636  * Like find_get_pages, except we only return pages which are tagged with
 637  * `tag'.   We update *index to index the next page for the traversal.
 638  */
 639 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 640                         int tag, unsigned int nr_pages, struct page **pages)
 641 {
 642         unsigned int i;
 643         unsigned int ret;
 644
 645         read_lock_irq(&mapping->tree_lock);
 646         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 647                                 (void **)pages, *index, nr_pages, tag);
 648         for (i = 0; i < ret; i++)
 649                 page_cache_get(pages[i]);
 650         if (ret)
 651                 *index = pages[ret - 1]->index + 1;
 652         read_unlock_irq(&mapping->tree_lock);
 653         return ret;
 654 }
 655
 656 /*
 657  * Same as grab_cache_page, but do not wait if the page is unavailable.
 658  * This is intended for speculative data generators, where the data can
 659  * be regenerated if the page couldn't be grabbed.  This routine should
 660  * be safe to call while holding the lock for another page.
 661  *
 662  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 663  * and deadlock against the caller's locked page.
 664  */
 665 struct page *
 666 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 667 {
 668         struct page *page = find_get_page(mapping, index);
 669         unsigned int gfp_mask;
 670
 671         if (page) {
 672                 if (!TestSetPageLocked(page))
 673                         return page;
 674                 page_cache_release(page);
 675                 return NULL;
 676         }
 677         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 678         page = alloc_pages(gfp_mask, 0);
 679         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 680                 page_cache_release(page);
 681                 page = NULL;
 682         }
 683         return page;
 684 }
 685
 686 EXPORT_SYMBOL(grab_cache_page_nowait);
 687
 688 /*
 689  * This is a generic file read routine, and uses the
 690  * mapping->a_ops->readpage() function for the actual low-level
 691  * stuff.
 692  *
 693  * This is really ugly. But the goto's actually try to clarify some
 694  * of the logic when it comes to error handling etc.
 695  *
 696  * Note the struct file* is only passed for the use of readpage.  It may be
 697  * NULL.
 698  */
 699 void do_generic_mapping_read(struct address_space *mapping,
 700                              struct file_ra_state *_ra,
 701                              struct file *filp,
 702                              loff_t *ppos,
 703                              read_descriptor_t *desc,
 704                              read_actor_t actor)
 705 {
 706         struct inode *inode = mapping->host;
 707         unsigned long index;
 708         unsigned long end_index;
 709         unsigned long offset;
 710         unsigned long last_index;
 711         unsigned long next_index;
 712         unsigned long prev_index;
 713         loff_t isize;
 714         struct page *cached_page;
 715         int error;
 716         struct file_ra_state ra = *_ra;
 717
 718         cached_page = NULL;
 719         index = *ppos >> PAGE_CACHE_SHIFT;
 720         next_index = index;
 721         prev_index = ra.prev_page;
 722         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 723         offset = *ppos & ~PAGE_CACHE_MASK;
 724
 725         isize = i_size_read(inode);
 726         if (!isize)
 727                 goto out;
 728
 729         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 730         for (;;) {
 731                 struct page *page;
 732                 unsigned long nr, ret;
 733
 734                 /* nr is the maximum number of bytes to copy from this page */
 735                 nr = PAGE_CACHE_SIZE;
 736                 if (index >= end_index) {
 737                         if (index > end_index)
 738                                 goto out;
 739                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 740                         if (nr <= offset) {
 741                                 goto out;
 742                         }
 743                 }
 744                 nr = nr - offset;
 745
 746                 cond_resched();
 747                 if (index == next_index)
 748                         next_index = page_cache_readahead(mapping, &ra, filp,
 749                                         index, last_index - index);
 750
 751 find_page:
 752                 page = find_get_page(mapping, index);
 753                 if (unlikely(page == NULL)) {
 754                         handle_ra_miss(mapping, &ra, index);
 755                         goto no_cached_page;
 756                 }
 757                 if (!PageUptodate(page))
 758                         goto page_not_up_to_date;
 759 page_ok:
 760
 761                 /* If users can be writing to this page using arbitrary
 762                  * virtual addresses, take care about potential aliasing
 763                  * before reading the page on the kernel side.
 764                  */
 765                 if (mapping_writably_mapped(mapping))
 766                         flush_dcache_page(page);
 767
 768                 /*
 769                  * When (part of) the same page is read multiple times
 770                  * in succession, only mark it as accessed the first time.
 771                  */
 772                 if (prev_index != index)
 773                         mark_page_accessed(page);
 774                 prev_index = index;
 775
 776                 /*
 777                  * Ok, we have the page, and it's up-to-date, so
 778                  * now we can copy it to user space...
 779                  *
 780                  * The actor routine returns how many bytes were actually used..
 781                  * NOTE! This may not be the same as how much of a user buffer
 782                  * we filled up (we may be padding etc), so we can only update
 783                  * "pos" here (the actor routine has to update the user buffer
 784                  * pointers and the remaining count).
 785                  */
 786                 ret = actor(desc, page, offset, nr);
 787                 offset += ret;
 788                 index += offset >> PAGE_CACHE_SHIFT;
 789                 offset &= ~PAGE_CACHE_MASK;
 790
 791                 page_cache_release(page);
 792                 if (ret == nr && desc->count)
 793                         continue;
 794                 goto out;
 795
 796 page_not_up_to_date:
 797                 /* Get exclusive access to the page ... */
 798                 lock_page(page);
 799
 800                 /* Did it get unhashed before we got the lock? */
 801                 if (!page->mapping) {
 802                         unlock_page(page);
 803                         page_cache_release(page);
 804                         continue;
 805                 }
 806
 807                 /* Did somebody else fill it already? */
 808                 if (PageUptodate(page)) {
 809                         unlock_page(page);
 810                         goto page_ok;
 811                 }
 812
 813 readpage:
 814                 /* Start the actual read. The read will unlock the page. */
 815                 error = mapping->a_ops->readpage(filp, page);
 816
 817                 if (unlikely(error))
 818                         goto readpage_error;
 819
 820                 if (!PageUptodate(page)) {
 821                         lock_page(page);
 822                         if (!PageUptodate(page)) {
 823                                 if (page->mapping == NULL) {
 824                                         /*
 825                                          * invalidate_inode_pages got it
 826                                          */
 827                                         unlock_page(page);
 828                                         page_cache_release(page);
 829                                         goto find_page;
 830                                 }
 831                                 unlock_page(page);
 832                                 error = -EIO;
 833                                 goto readpage_error;
 834                         }
 835                         unlock_page(page);
 836                 }
 837
 838                 /*
 839                  * i_size must be checked after we have done ->readpage.
 840                  *
 841                  * Checking i_size after the readpage allows us to calculate
 842                  * the correct value for "nr", which means the zero-filled
 843                  * part of the page is not copied back to userspace (unless
 844                  * another truncate extends the file - this is desired though).
 845                  */
 846                 isize = i_size_read(inode);
 847                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 848                 if (unlikely(!isize || index > end_index)) {
 849                         page_cache_release(page);
 850                         goto out;
 851                 }
 852
 853                 /* nr is the maximum number of bytes to copy from this page */
 854                 nr = PAGE_CACHE_SIZE;
 855                 if (index == end_index) {
 856                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 857                         if (nr <= offset) {
 858                                 page_cache_release(page);
 859                                 goto out;
 860                         }
 861                 }
 862                 nr = nr - offset;
 863                 goto page_ok;
 864
 865 readpage_error:
 866                 /* UHHUH! A synchronous read error occurred. Report it */
 867                 desc->error = error;
 868                 page_cache_release(page);
 869                 goto out;
 870
 871 no_cached_page:
 872                 /*
 873                  * Ok, it wasn't cached, so we need to create a new
 874                  * page..
 875                  */
 876                 if (!cached_page) {
 877                         cached_page = page_cache_alloc_cold(mapping);
 878                         if (!cached_page) {
 879                                 desc->error = -ENOMEM;
 880                                 goto out;
 881                         }
 882                 }
 883                 error = add_to_page_cache_lru(cached_page, mapping,
 884                                                 index, GFP_KERNEL);
 885                 if (error) {
 886                         if (error == -EEXIST)
 887                                 goto find_page;
 888                         desc->error = error;
 889                         goto out;
 890                 }
 891                 page = cached_page;
 892                 cached_page = NULL;
 893                 goto readpage;
 894         }
 895
 896 out:
 897         *_ra = ra;
 898
 899         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 900         if (cached_page)
 901                 page_cache_release(cached_page);
 902         if (filp)
 903                 file_accessed(filp);
 904 }
 905
 906 EXPORT_SYMBOL(do_generic_mapping_read);
 907
 908 int file_read_actor(read_descriptor_t *desc, struct page *page,
 909                         unsigned long offset, unsigned long size)
 910 {
 911         char *kaddr;
 912         unsigned long left, count = desc->count;
 913
 914         if (size > count)
 915                 size = count;
 916
 917         /*
 918          * Faults on the destination of a read are common, so do it before
 919          * taking the kmap.
 920          */
 921         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 922                 kaddr = kmap_atomic(page, KM_USER0);
 923                 left = __copy_to_user_inatomic(desc->arg.buf,
 924                                                 kaddr + offset, size);
 925                 kunmap_atomic(kaddr, KM_USER0);
 926                 if (left == 0)
 927                         goto success;
 928         }
 929
 930         /* Do it the slow way */
 931         kaddr = kmap(page);
 932         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 933         kunmap(page);
 934
 935         if (left) {
 936                 size -= left;
 937                 desc->error = -EFAULT;
 938         }
 939 success:
 940         desc->count = count - size;
 941         desc->written += size;
 942         desc->arg.buf += size;
 943         return size;
 944 }
 945
 946 /*
 947  * This is the "read()" routine for all filesystems
 948  * that can use the page cache directly.
 949  */
 950 ssize_t
 951 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 952                 unsigned long nr_segs, loff_t *ppos)
 953 {
 954         struct file *filp = iocb->ki_filp;
 955         ssize_t retval;
 956         unsigned long seg;
 957         size_t count;
 958
 959         count = 0;
 960         for (seg = 0; seg < nr_segs; seg++) {
 961                 const struct iovec *iv = &iov[seg];
 962
 963                 /*
 964                  * If any segment has a negative length, or the cumulative
 965                  * length ever wraps negative then return -EINVAL.
 966                  */
 967                 count += iv->iov_len;
 968                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
 969                         return -EINVAL;
 970                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
 971                         continue;
 972                 if (seg == 0)
 973                         return -EFAULT;
 974                 nr_segs = seg;
 975                 count -= iv->iov_len;   /* This segment is no good */
 976                 break;
 977         }
 978
 979         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
 980         if (filp->f_flags & O_DIRECT) {
 981                 loff_t pos = *ppos, size;
 982                 struct address_space *mapping;
 983                 struct inode *inode;
 984
 985                 mapping = filp->f_mapping;
 986                 inode = mapping->host;
 987                 retval = 0;
 988                 if (!count)
 989                         goto out; /* skip atime */
 990                 size = i_size_read(inode);
 991                 if (pos < size) {
 992                         retval = generic_file_direct_IO(READ, iocb,
 993                                                 iov, pos, nr_segs);
 994                         if (retval >= 0 && !is_sync_kiocb(iocb))
 995                                 retval = -EIOCBQUEUED;
 996                         if (retval > 0)
 997                                 *ppos = pos + retval;
 998                 }
 999                 file_accessed(filp);
1000                 goto out;
1001         }
1002
1003         retval = 0;
1004         if (count) {
1005                 for (seg = 0; seg < nr_segs; seg++) {
1006                         read_descriptor_t desc;
1007
1008                         desc.written = 0;
1009                         desc.arg.buf = iov[seg].iov_base;
1010                         desc.count = iov[seg].iov_len;
1011                         if (desc.count == 0)
1012                                 continue;
1013                         desc.error = 0;
1014                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1015                         retval += desc.written;
1016                         if (!retval) {
1017                                 retval = desc.error;
1018                                 break;
1019                         }
1020                 }
1021         }
1022 out:
1023         return retval;
1024 }
1025
1026 EXPORT_SYMBOL(__generic_file_aio_read);
1027
1028 ssize_t
1029 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1030 {
1031         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1032
1033         BUG_ON(iocb->ki_pos != pos);
1034         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1035 }
1036
1037 EXPORT_SYMBOL(generic_file_aio_read);
1038
1039 ssize_t
1040 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1041 {
1042         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1043         struct kiocb kiocb;
1044         ssize_t ret;
1045
1046         init_sync_kiocb(&kiocb, filp);
1047         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1048         if (-EIOCBQUEUED == ret)
1049                 ret = wait_on_sync_kiocb(&kiocb);
1050         return ret;
1051 }
1052
1053 EXPORT_SYMBOL(generic_file_read);
1054
1055 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1056 {
1057         ssize_t written;
1058         unsigned long count = desc->count;
1059         struct file *file = desc->arg.data;
1060
1061         if (size > count)
1062                 size = count;
1063
1064         written = file->f_op->sendpage(file, page, offset,
1065                                        size, &file->f_pos, size<count);
1066         if (written < 0) {
1067                 desc->error = written;
1068                 written = 0;
1069         }
1070         desc->count = count - written;
1071         desc->written += written;
1072         return written;
1073 }
1074
1075 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1076                          size_t count, read_actor_t actor, void *target)
1077 {
1078         read_descriptor_t desc;
1079
1080         if (!count)
1081                 return 0;
1082
1083         desc.written = 0;
1084         desc.count = count;
1085         desc.arg.data = target;
1086         desc.error = 0;
1087
1088         do_generic_file_read(in_file, ppos, &desc, actor);
1089         if (desc.written)
1090                 return desc.written;
1091         return desc.error;
1092 }
1093
1094 EXPORT_SYMBOL(generic_file_sendfile);
1095
1096 static ssize_t
1097 do_readahead(struct address_space *mapping, struct file *filp,
1098              unsigned long index, unsigned long nr)
1099 {
1100         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1101                 return -EINVAL;
1102
1103         force_page_cache_readahead(mapping, filp, index,
1104                                         max_sane_readahead(nr));
1105         return 0;
1106 }
1107
1108 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1109 {
1110         ssize_t ret;
1111         struct file *file;
1112
1113         ret = -EBADF;
1114         file = fget(fd);
1115         if (file) {
1116                 if (file->f_mode & FMODE_READ) {
1117                         struct address_space *mapping = file->f_mapping;
1118                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1119                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1120                         unsigned long len = end - start + 1;
1121                         ret = do_readahead(mapping, file, start, len);
1122                 }
1123                 fput(file);
1124         }
1125         return ret;
1126 }
1127
1128 #ifdef CONFIG_MMU
1129 /*
1130  * This adds the requested page to the page cache if it isn't already there,
1131  * and schedules an I/O to read in its contents from disk.
1132  */
1133 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1134 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1135 {
1136         struct address_space *mapping = file->f_mapping;
1137         struct page *page;
1138         int error;
1139
1140         page = page_cache_alloc_cold(mapping);
1141         if (!page)
1142                 return -ENOMEM;
1143
1144         error = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1145         if (!error) {
1146                 error = mapping->a_ops->readpage(file, page);
1147                 page_cache_release(page);
1148                 return error;
1149         }
1150
1151         /*
1152          * We arrive here in the unlikely event that someone
1153          * raced with us and added our page to the cache first
1154          * or we are out of memory for radix-tree nodes.
1155          */
1156         page_cache_release(page);
1157         return error == -EEXIST ? 0 : error;
1158 }
1159
1160 #define MMAP_LOTSAMISS  (100)
1161
1162 /*
1163  * filemap_nopage() is invoked via the vma operations vector for a
1164  * mapped memory region to read in file data during a page fault.
1165  *
1166  * The goto's are kind of ugly, but this streamlines the normal case of having
1167  * it in the page cache, and handles the special cases reasonably without
1168  * having a lot of duplicated code.
1169  */
1170 struct page *filemap_nopage(struct vm_area_struct *area,
1171                                 unsigned long address, int *type)
1172 {
1173         int error;
1174         struct file *file = area->vm_file;
1175         struct address_space *mapping = file->f_mapping;
1176         struct file_ra_state *ra = &file->f_ra;
1177         struct inode *inode = mapping->host;
1178         struct page *page;
1179         unsigned long size, pgoff;
1180         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1181
1182         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1183
1184 retry_all:
1185         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1186         if (pgoff >= size)
1187                 goto outside_data_content;
1188
1189         /* If we don't want any read-ahead, don't bother */
1190         if (VM_RandomReadHint(area))
1191                 goto no_cached_page;
1192
1193         /*
1194          * The readahead code wants to be told about each and every page
1195          * so it can build and shrink its windows appropriately
1196          *
1197          * For sequential accesses, we use the generic readahead logic.
1198          */
1199         if (VM_SequentialReadHint(area))
1200                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1201
1202         /*
1203          * Do we have something in the page cache already?
1204          */
1205 retry_find:
1206         page = find_get_page(mapping, pgoff);
1207         if (!page) {
1208                 unsigned long ra_pages;
1209
1210                 if (VM_SequentialReadHint(area)) {
1211                         handle_ra_miss(mapping, ra, pgoff);
1212                         goto no_cached_page;
1213                 }
1214                 ra->mmap_miss++;
1215
1216                 /*
1217                  * Do we miss much more than hit in this file? If so,
1218                  * stop bothering with read-ahead. It will only hurt.
1219                  */
1220                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1221                         goto no_cached_page;
1222
1223                 /*
1224                  * To keep the pgmajfault counter straight, we need to
1225                  * check did_readaround, as this is an inner loop.
1226                  */
1227                 if (!did_readaround) {
1228                         majmin = VM_FAULT_MAJOR;
1229                         inc_page_state(pgmajfault);
1230                 }
1231                 did_readaround = 1;
1232                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1233                 if (ra_pages) {
1234                         pgoff_t start = 0;
1235
1236                         if (pgoff > ra_pages / 2)
1237                                 start = pgoff - ra_pages / 2;
1238                         do_page_cache_readahead(mapping, file, start, ra_pages);
1239                 }
1240                 page = find_get_page(mapping, pgoff);
1241                 if (!page)
1242                         goto no_cached_page;
1243         }
1244
1245         if (!did_readaround)
1246                 ra->mmap_hit++;
1247
1248         /*
1249          * Ok, found a page in the page cache, now we need to check
1250          * that it's up-to-date.
1251          */
1252         if (!PageUptodate(page))
1253                 goto page_not_uptodate;
1254
1255 success:
1256         /*
1257          * Found the page and have a reference on it.
1258          */
1259         mark_page_accessed(page);
1260         if (type)
1261                 *type = majmin;
1262         return page;
1263
1264 outside_data_content:
1265         /*
1266          * An external ptracer can access pages that normally aren't
1267          * accessible..
1268          */
1269         if (area->vm_mm == current->mm)
1270                 return NULL;
1271         /* Fall through to the non-read-ahead case */
1272 no_cached_page:
1273         /*
1274          * We're only likely to ever get here if MADV_RANDOM is in
1275          * effect.
1276          */
1277         error = page_cache_read(file, pgoff);
1278         grab_swap_token();
1279
1280         /*
1281          * The page we want has now been added to the page cache.
1282          * In the unlikely event that someone removed it in the
1283          * meantime, we'll just come back here and read it again.
1284          */
1285         if (error >= 0)
1286                 goto retry_find;
1287
1288         /*
1289          * An error return from page_cache_read can result if the
1290          * system is low on memory, or a problem occurs while trying
1291          * to schedule I/O.
1292          */
1293         if (error == -ENOMEM)
1294                 return NOPAGE_OOM;
1295         return NULL;
1296
1297 page_not_uptodate:
1298         if (!did_readaround) {
1299                 majmin = VM_FAULT_MAJOR;
1300                 inc_page_state(pgmajfault);
1301         }
1302         lock_page(page);
1303
1304         /* Did it get unhashed while we waited for it? */
1305         if (!page->mapping) {
1306                 unlock_page(page);
1307                 page_cache_release(page);
1308                 goto retry_all;
1309         }
1310
1311         /* Did somebody else get it up-to-date? */
1312         if (PageUptodate(page)) {
1313                 unlock_page(page);
1314                 goto success;
1315         }
1316
1317         if (!mapping->a_ops->readpage(file, page)) {
1318                 wait_on_page_locked(page);
1319                 if (PageUptodate(page))
1320                         goto success;
1321         }
1322
1323         /*
1324          * Umm, take care of errors if the page isn't up-to-date.
1325          * Try to re-read it _once_. We do this synchronously,
1326          * because there really aren't any performance issues here
1327          * and we need to check for errors.
1328          */
1329         lock_page(page);
1330
1331         /* Somebody truncated the page on us? */
1332         if (!page->mapping) {
1333                 unlock_page(page);
1334                 page_cache_release(page);
1335                 goto retry_all;
1336         }
1337
1338         /* Somebody else successfully read it in? */
1339         if (PageUptodate(page)) {
1340                 unlock_page(page);
1341                 goto success;
1342         }
1343         ClearPageError(page);
1344         if (!mapping->a_ops->readpage(file, page)) {
1345                 wait_on_page_locked(page);
1346                 if (PageUptodate(page))
1347                         goto success;
1348         }
1349
1350         /*
1351          * Things didn't work out. Return zero to tell the
1352          * mm layer so, possibly freeing the page cache page first.
1353          */
1354         page_cache_release(page);
1355         return NULL;
1356 }
1357
1358 EXPORT_SYMBOL(filemap_nopage);
1359
1360 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1361                                         int nonblock)
1362 {
1363         struct address_space *mapping = file->f_mapping;
1364         struct page *page;
1365         int error;
1366
1367         /*
1368          * Do we have something in the page cache already?
1369          */
1370 retry_find:
1371         page = find_get_page(mapping, pgoff);
1372         if (!page) {
1373                 if (nonblock)
1374                         return NULL;
1375                 goto no_cached_page;
1376         }
1377
1378         /*
1379          * Ok, found a page in the page cache, now we need to check
1380          * that it's up-to-date.
1381          */
1382         if (!PageUptodate(page))
1383                 goto page_not_uptodate;
1384
1385 success:
1386         /*
1387          * Found the page and have a reference on it.
1388          */
1389         mark_page_accessed(page);
1390         return page;
1391
1392 no_cached_page:
1393         error = page_cache_read(file, pgoff);
1394
1395         /*
1396          * The page we want has now been added to the page cache.
1397          * In the unlikely event that someone removed it in the
1398          * meantime, we'll just come back here and read it again.
1399          */
1400         if (error >= 0)
1401                 goto retry_find;
1402
1403         /*
1404          * An error return from page_cache_read can result if the
1405          * system is low on memory, or a problem occurs while trying
1406          * to schedule I/O.
1407          */
1408         return NULL;
1409
1410 page_not_uptodate:
1411         lock_page(page);
1412
1413         /* Did it get unhashed while we waited for it? */
1414         if (!page->mapping) {
1415                 unlock_page(page);
1416                 goto err;
1417         }
1418
1419         /* Did somebody else get it up-to-date? */
1420         if (PageUptodate(page)) {
1421                 unlock_page(page);
1422                 goto success;
1423         }
1424
1425         if (!mapping->a_ops->readpage(file, page)) {
1426                 wait_on_page_locked(page);
1427                 if (PageUptodate(page))
1428                         goto success;
1429         }
1430
1431         /*
1432          * Umm, take care of errors if the page isn't up-to-date.
1433          * Try to re-read it _once_. We do this synchronously,
1434          * because there really aren't any performance issues here
1435          * and we need to check for errors.
1436          */
1437         lock_page(page);
1438
1439         /* Somebody truncated the page on us? */
1440         if (!page->mapping) {
1441                 unlock_page(page);
1442                 goto err;
1443         }
1444         /* Somebody else successfully read it in? */
1445         if (PageUptodate(page)) {
1446                 unlock_page(page);
1447                 goto success;
1448         }
1449
1450         ClearPageError(page);
1451         if (!mapping->a_ops->readpage(file, page)) {
1452                 wait_on_page_locked(page);
1453                 if (PageUptodate(page))
1454                         goto success;
1455         }
1456
1457         /*
1458          * Things didn't work out. Return zero to tell the
1459          * mm layer so, possibly freeing the page cache page first.
1460          */
1461 err:
1462         page_cache_release(page);
1463
1464         return NULL;
1465 }
1466
1467 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1468                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1469                 int nonblock)
1470 {
1471         struct file *file = vma->vm_file;
1472         struct address_space *mapping = file->f_mapping;
1473         struct inode *inode = mapping->host;
1474         unsigned long size;
1475         struct mm_struct *mm = vma->vm_mm;
1476         struct page *page;
1477         int err;
1478
1479         if (!nonblock)
1480                 force_page_cache_readahead(mapping, vma->vm_file,
1481                                         pgoff, len >> PAGE_CACHE_SHIFT);
1482
1483 repeat:
1484         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1485         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1486                 return -EINVAL;
1487
1488         page = filemap_getpage(file, pgoff, nonblock);
1489         if (!page && !nonblock)
1490                 return -ENOMEM;
1491         if (page) {
1492                 err = install_page(mm, vma, addr, page, prot);
1493                 if (err) {
1494                         page_cache_release(page);
1495                         return err;
1496                 }
1497         } else {
1498                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1499                 if (err)
1500                         return err;
1501         }
1502
1503         len -= PAGE_SIZE;
1504         addr += PAGE_SIZE;
1505         pgoff++;
1506         if (len)
1507                 goto repeat;
1508
1509         return 0;
1510 }
1511
1512 struct vm_operations_struct generic_file_vm_ops = {
1513         .nopage         = filemap_nopage,
1514         .populate       = filemap_populate,
1515 };
1516
1517 /* This is used for a general mmap of a disk file */
1518
1519 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1520 {
1521         struct address_space *mapping = file->f_mapping;
1522
1523         if (!mapping->a_ops->readpage)
1524                 return -ENOEXEC;
1525         file_accessed(file);
1526         vma->vm_ops = &generic_file_vm_ops;
1527         return 0;
1528 }
1529 EXPORT_SYMBOL(filemap_populate);
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->signal->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 EXPORT_SYMBOL(generic_write_checks);
1882
1883 ssize_t
1884 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1885                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1886                 size_t count, size_t ocount)
1887 {
1888         struct file     *file = iocb->ki_filp;
1889         struct address_space *mapping = file->f_mapping;
1890         struct inode    *inode = mapping->host;
1891         ssize_t         written;
1892
1893         if (count != ocount)
1894                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1895
1896         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1897         if (written > 0) {
1898                 loff_t end = pos + written;
1899                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1900                         i_size_write(inode,  end);
1901                         mark_inode_dirty(inode);
1902                 }
1903                 *ppos = end;
1904         }
1905
1906         /*
1907          * Sync the fs metadata but not the minor inode changes and
1908          * of course not the data as we did direct DMA for the IO.
1909          * i_sem is held, which protects generic_osync_inode() from
1910          * livelocking.
1911          */
1912         if (written >= 0 && file->f_flags & O_SYNC)
1913                 generic_osync_inode(inode, mapping, OSYNC_METADATA);
1914         if (written == count && !is_sync_kiocb(iocb))
1915                 written = -EIOCBQUEUED;
1916         return written;
1917 }
1918 EXPORT_SYMBOL(generic_file_direct_write);
1919
1920 ssize_t
1921 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1922                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1923                 size_t count, ssize_t written)
1924 {
1925         struct file *file = iocb->ki_filp;
1926         struct address_space * mapping = file->f_mapping;
1927         struct address_space_operations *a_ops = mapping->a_ops;
1928         struct inode    *inode = mapping->host;
1929         long            status = 0;
1930         struct page     *page;
1931         struct page     *cached_page = NULL;
1932         size_t          bytes;
1933         struct pagevec  lru_pvec;
1934         const struct iovec *cur_iov = iov; /* current iovec */
1935         size_t          iov_base = 0;      /* offset in the current iovec */
1936         char __user     *buf;
1937
1938         pagevec_init(&lru_pvec, 0);
1939
1940         /*
1941          * handle partial DIO write.  Adjust cur_iov if needed.
1942          */
1943         if (likely(nr_segs == 1))
1944                 buf = iov->iov_base + written;
1945         else {
1946                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1947                 buf = iov->iov_base + iov_base;
1948         }
1949
1950         do {
1951                 unsigned long index;
1952                 unsigned long offset;
1953                 size_t copied;
1954
1955                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1956                 index = pos >> PAGE_CACHE_SHIFT;
1957                 bytes = PAGE_CACHE_SIZE - offset;
1958                 if (bytes > count)
1959                         bytes = count;
1960
1961                 /*
1962                  * Bring in the user page that we will copy from _first_.
1963                  * Otherwise there's a nasty deadlock on copying from the
1964                  * same page as we're writing to, without it being marked
1965                  * up-to-date.
1966                  */
1967                 fault_in_pages_readable(buf, bytes);
1968
1969                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1970                 if (!page) {
1971                         status = -ENOMEM;
1972                         break;
1973                 }
1974
1975                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1976                 if (unlikely(status)) {
1977                         loff_t isize = i_size_read(inode);
1978                         /*
1979                          * prepare_write() may have instantiated a few blocks
1980                          * outside i_size.  Trim these off again.
1981                          */
1982                         unlock_page(page);
1983                         page_cache_release(page);
1984                         if (pos + bytes > isize)
1985                                 vmtruncate(inode, isize);
1986                         break;
1987                 }
1988                 if (likely(nr_segs == 1))
1989                         copied = filemap_copy_from_user(page, offset,
1990                                                         buf, bytes);
1991                 else
1992                         copied = filemap_copy_from_user_iovec(page, offset,
1993                                                 cur_iov, iov_base, bytes);
1994                 flush_dcache_page(page);
1995                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1996                 if (likely(copied > 0)) {
1997                         if (!status)
1998                                 status = copied;
1999
2000                         if (status >= 0) {
2001                                 written += status;
2002                                 count -= status;
2003                                 pos += status;
2004                                 buf += status;
2005                                 if (unlikely(nr_segs > 1))
2006                                         filemap_set_next_iovec(&cur_iov,
2007                                                         &iov_base, status);
2008                         }
2009                 }
2010                 if (unlikely(copied != bytes))
2011                         if (status >= 0)
2012                                 status = -EFAULT;
2013                 unlock_page(page);
2014                 mark_page_accessed(page);
2015                 page_cache_release(page);
2016                 if (status < 0)
2017                         break;
2018                 balance_dirty_pages_ratelimited(mapping);
2019                 cond_resched();
2020         } while (count);
2021         *ppos = pos;
2022
2023         if (cached_page)
2024                 page_cache_release(cached_page);
2025
2026         /*
2027          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2028          */
2029         if (likely(status >= 0)) {
2030                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2031                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2032                                 status = generic_osync_inode(inode, mapping,
2033                                                 OSYNC_METADATA|OSYNC_DATA);
2034                 }
2035         }
2036
2037         /*
2038          * If we get here for O_DIRECT writes then we must have fallen through
2039          * to buffered writes (block instantiation inside i_size).  So we sync
2040          * the file data here, to try to honour O_DIRECT expectations.
2041          */
2042         if (unlikely(file->f_flags & O_DIRECT) && written)
2043                 status = filemap_write_and_wait(mapping);
2044
2045         pagevec_lru_add(&lru_pvec);
2046         return written ? written : status;
2047 }
2048 EXPORT_SYMBOL(generic_file_buffered_write);
2049
2050 ssize_t
2051 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2052                                 unsigned long nr_segs, loff_t *ppos)
2053 {
2054         struct file *file = iocb->ki_filp;
2055         struct address_space * mapping = file->f_mapping;
2056         size_t ocount;          /* original count */
2057         size_t count;           /* after file limit checks */
2058         struct inode    *inode = mapping->host;
2059         unsigned long   seg;
2060         loff_t          pos;
2061         ssize_t         written;
2062         ssize_t         err;
2063
2064         ocount = 0;
2065         for (seg = 0; seg < nr_segs; seg++) {
2066                 const struct iovec *iv = &iov[seg];
2067
2068                 /*
2069                  * If any segment has a negative length, or the cumulative
2070                  * length ever wraps negative then return -EINVAL.
2071                  */
2072                 ocount += iv->iov_len;
2073                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2074                         return -EINVAL;
2075                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2076                         continue;
2077                 if (seg == 0)
2078                         return -EFAULT;
2079                 nr_segs = seg;
2080                 ocount -= iv->iov_len;  /* This segment is no good */
2081                 break;
2082         }
2083
2084         count = ocount;
2085         pos = *ppos;
2086
2087         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2088
2089         /* We can write back this queue in page reclaim */
2090         current->backing_dev_info = mapping->backing_dev_info;
2091         written = 0;
2092
2093         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2094         if (err)
2095                 goto out;
2096
2097         if (count == 0)
2098                 goto out;
2099
2100         err = remove_suid(file->f_dentry);
2101         if (err)
2102                 goto out;
2103
2104         inode_update_time(inode, 1);
2105
2106         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2107         if (unlikely(file->f_flags & O_DIRECT)) {
2108                 written = generic_file_direct_write(iocb, iov,
2109                                 &nr_segs, pos, ppos, count, ocount);
2110                 if (written < 0 || written == count)
2111                         goto out;
2112                 /*
2113                  * direct-io write to a hole: fall through to buffered I/O
2114                  * for completing the rest of the request.
2115                  */
2116                 pos += written;
2117                 count -= written;
2118         }
2119
2120         written = generic_file_buffered_write(iocb, iov, nr_segs,
2121                         pos, ppos, count, written);
2122 out:
2123         current->backing_dev_info = NULL;
2124         return written ? written : err;
2125 }
2126 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2127
2128 ssize_t
2129 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2130                                 unsigned long nr_segs, loff_t *ppos)
2131 {
2132         struct file *file = iocb->ki_filp;
2133         struct address_space *mapping = file->f_mapping;
2134         struct inode *inode = mapping->host;
2135         ssize_t ret;
2136         loff_t pos = *ppos;
2137
2138         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2139
2140         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2141                 int err;
2142
2143                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2144                 if (err < 0)
2145                         ret = err;
2146         }
2147         return ret;
2148 }
2149
2150 ssize_t
2151 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2152                                 unsigned long nr_segs, loff_t *ppos)
2153 {
2154         struct kiocb kiocb;
2155         ssize_t ret;
2156
2157         init_sync_kiocb(&kiocb, file);
2158         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2159         if (ret == -EIOCBQUEUED)
2160                 ret = wait_on_sync_kiocb(&kiocb);
2161         return ret;
2162 }
2163
2164 ssize_t
2165 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2166                                 unsigned long nr_segs, loff_t *ppos)
2167 {
2168         struct kiocb kiocb;
2169         ssize_t ret;
2170
2171         init_sync_kiocb(&kiocb, file);
2172         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2173         if (-EIOCBQUEUED == ret)
2174                 ret = wait_on_sync_kiocb(&kiocb);
2175         return ret;
2176 }
2177 EXPORT_SYMBOL(generic_file_write_nolock);
2178
2179 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2180                                size_t count, loff_t pos)
2181 {
2182         struct file *file = iocb->ki_filp;
2183         struct address_space *mapping = file->f_mapping;
2184         struct inode *inode = mapping->host;
2185         ssize_t ret;
2186         struct iovec local_iov = { .iov_base = (void __user *)buf,
2187                                         .iov_len = count };
2188
2189         BUG_ON(iocb->ki_pos != pos);
2190
2191         down(&inode->i_sem);
2192         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2193                                                 &iocb->ki_pos);
2194         up(&inode->i_sem);
2195
2196         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2197                 ssize_t err;
2198
2199                 err = sync_page_range(inode, mapping, pos, ret);
2200                 if (err < 0)
2201                         ret = err;
2202         }
2203         return ret;
2204 }
2205 EXPORT_SYMBOL(generic_file_aio_write);
2206
2207 ssize_t generic_file_write(struct file *file, const char __user *buf,
2208                            size_t count, loff_t *ppos)
2209 {
2210         struct address_space *mapping = file->f_mapping;
2211         struct inode *inode = mapping->host;
2212         ssize_t ret;
2213         struct iovec local_iov = { .iov_base = (void __user *)buf,
2214                                         .iov_len = count };
2215
2216         down(&inode->i_sem);
2217         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
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, *ppos - ret, ret);
2224                 if (err < 0)
2225                         ret = err;
2226         }
2227         return ret;
2228 }
2229 EXPORT_SYMBOL(generic_file_write);
2230
2231 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2232                         unsigned long nr_segs, loff_t *ppos)
2233 {
2234         struct kiocb kiocb;
2235         ssize_t ret;
2236
2237         init_sync_kiocb(&kiocb, filp);
2238         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2239         if (-EIOCBQUEUED == ret)
2240                 ret = wait_on_sync_kiocb(&kiocb);
2241         return ret;
2242 }
2243 EXPORT_SYMBOL(generic_file_readv);
2244
2245 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2246                         unsigned long nr_segs, loff_t *ppos)
2247 {
2248         struct address_space *mapping = file->f_mapping;
2249         struct inode *inode = mapping->host;
2250         ssize_t ret;
2251
2252         down(&inode->i_sem);
2253         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2254         up(&inode->i_sem);
2255
2256         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2257                 int err;
2258
2259                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2260                 if (err < 0)
2261                         ret = err;
2262         }
2263         return ret;
2264 }
2265 EXPORT_SYMBOL(generic_file_writev);
2266
2267 /*
2268  * Called under i_sem for writes to S_ISREG files.   Returns -EIO if something
2269  * went wrong during pagecache shootdown.
2270  */
2271 ssize_t
2272 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2273         loff_t offset, unsigned long nr_segs)
2274 {
2275         struct file *file = iocb->ki_filp;
2276         struct address_space *mapping = file->f_mapping;
2277         ssize_t retval;
2278         size_t write_len = 0;
2279
2280         /*
2281          * If it's a write, unmap all mmappings of the file up-front.  This
2282          * will cause any pte dirty bits to be propagated into the pageframes
2283          * for the subsequent filemap_write_and_wait().
2284          */
2285         if (rw == WRITE) {
2286                 write_len = iov_length(iov, nr_segs);
2287                 if (mapping_mapped(mapping))
2288                         unmap_mapping_range(mapping, offset, write_len, 0);
2289         }
2290
2291         retval = filemap_write_and_wait(mapping);
2292         if (retval == 0) {
2293                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2294                                                 offset, nr_segs);
2295                 if (rw == WRITE && mapping->nrpages) {
2296                         pgoff_t end = (offset + write_len - 1)
2297                                                 >> PAGE_CACHE_SHIFT;
2298                         int err = invalidate_inode_pages2_range(mapping,
2299                                         offset >> PAGE_CACHE_SHIFT, end);
2300                         if (err)
2301                                 retval = err;
2302                 }
2303         }
2304         return retval;
2305 }
2306 EXPORT_SYMBOL_GPL(generic_file_direct_IO);