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 #include "filemap.h"
  32 /*
  33  * FIXME: remove all knowledge of the buffer layer from the core VM
  34  */
  35 #include <linux/buffer_head.h> /* for generic_osync_inode */
  36
  37 #include <asm/uaccess.h>
  38 #include <asm/mman.h>
  39
  40 static ssize_t
  41 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
  42         loff_t offset, unsigned long nr_segs);
  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_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 or pte_lock   (various, mainly in memory.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 or pte_lock     (anon_vma_prepare and various)
  90  *
  91  *  ->page_table_lock or pte_lock
  92  *    ->swap_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         BUG_ON(!PageLocked(page));
 126
 127         write_lock_irq(&mapping->tree_lock);
 128         __remove_from_page_cache(page);
 129         write_unlock_irq(&mapping->tree_lock);
 130 }
 131
 132 static int sync_page(void *word)
 133 {
 134         struct address_space *mapping;
 135         struct page *page;
 136
 137         page = container_of((unsigned long *)word, struct page, flags);
 138
 139         /*
 140          * page_mapping() is being called without PG_locked held.
 141          * Some knowledge of the state and use of the page is used to
 142          * reduce the requirements down to a memory barrier.
 143          * The danger here is of a stale page_mapping() return value
 144          * indicating a struct address_space different from the one it's
 145          * associated with when it is associated with one.
 146          * After smp_mb(), it's either the correct page_mapping() for
 147          * the page, or an old page_mapping() and the page's own
 148          * page_mapping() has gone NULL.
 149          * The ->sync_page() address_space operation must tolerate
 150          * page_mapping() going NULL. By an amazing coincidence,
 151          * this comes about because none of the users of the page
 152          * in the ->sync_page() methods make essential use of the
 153          * page_mapping(), merely passing the page down to the backing
 154          * device's unplug functions when it's non-NULL, which in turn
 155          * ignore it for all cases but swap, where only page_private(page) is
 156          * of interest. When page_mapping() does go NULL, the entire
 157          * call stack gracefully ignores the page and returns.
 158          * -- wli
 159          */
 160         smp_mb();
 161         mapping = page_mapping(page);
 162         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 163                 mapping->a_ops->sync_page(page);
 164         io_schedule();
 165         return 0;
 166 }
 167
 168 /**
 169  * filemap_fdatawrite_range - start writeback against all of a mapping's
 170  * dirty pages that lie within the byte offsets <start, end>
 171  * @mapping:    address space structure to write
 172  * @start:      offset in bytes where the range starts
 173  * @end:        offset in bytes where the range ends
 174  * @sync_mode:  enable synchronous operation
 175  *
 176  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 177  * opposed to a regular memory * cleansing writeback.  The difference between
 178  * these two operations is that if a dirty page/buffer is encountered, it must
 179  * be waited upon, and not just skipped over.
 180  */
 181 static int __filemap_fdatawrite_range(struct address_space *mapping,
 182         loff_t start, loff_t end, int sync_mode)
 183 {
 184         int ret;
 185         struct writeback_control wbc = {
 186                 .sync_mode = sync_mode,
 187                 .nr_to_write = mapping->nrpages * 2,
 188                 .start = start,
 189                 .end = end,
 190         };
 191
 192         if (!mapping_cap_writeback_dirty(mapping))
 193                 return 0;
 194
 195         ret = do_writepages(mapping, &wbc);
 196         return ret;
 197 }
 198
 199 static inline int __filemap_fdatawrite(struct address_space *mapping,
 200         int sync_mode)
 201 {
 202         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
 203 }
 204
 205 int filemap_fdatawrite(struct address_space *mapping)
 206 {
 207         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 208 }
 209 EXPORT_SYMBOL(filemap_fdatawrite);
 210
 211 static int filemap_fdatawrite_range(struct address_space *mapping,
 212         loff_t start, loff_t end)
 213 {
 214         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
 215 }
 216
 217 /*
 218  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 219  * purposes - I/O may not be started against all dirty pages.
 220  */
 221 int filemap_flush(struct address_space *mapping)
 222 {
 223         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 224 }
 225 EXPORT_SYMBOL(filemap_flush);
 226
 227 /*
 228  * Wait for writeback to complete against pages indexed by start->end
 229  * inclusive
 230  */
 231 static int wait_on_page_writeback_range(struct address_space *mapping,
 232                                 pgoff_t start, pgoff_t end)
 233 {
 234         struct pagevec pvec;
 235         int nr_pages;
 236         int ret = 0;
 237         pgoff_t index;
 238
 239         if (end < start)
 240                 return 0;
 241
 242         pagevec_init(&pvec, 0);
 243         index = start;
 244         while ((index <= end) &&
 245                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 246                         PAGECACHE_TAG_WRITEBACK,
 247                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 248                 unsigned i;
 249
 250                 for (i = 0; i < nr_pages; i++) {
 251                         struct page *page = pvec.pages[i];
 252
 253                         /* until radix tree lookup accepts end_index */
 254                         if (page->index > end)
 255                                 continue;
 256
 257                         wait_on_page_writeback(page);
 258                         if (PageError(page))
 259                                 ret = -EIO;
 260                 }
 261                 pagevec_release(&pvec);
 262                 cond_resched();
 263         }
 264
 265         /* Check for outstanding write errors */
 266         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 267                 ret = -ENOSPC;
 268         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 269                 ret = -EIO;
 270
 271         return ret;
 272 }
 273
 274 /*
 275  * Write and wait upon all the pages in the passed range.  This is a "data
 276  * integrity" operation.  It waits upon in-flight writeout before starting and
 277  * waiting upon new writeout.  If there was an IO error, return it.
 278  *
 279  * We need to re-take i_sem during the generic_osync_inode list walk because
 280  * it is otherwise livelockable.
 281  */
 282 int sync_page_range(struct inode *inode, struct address_space *mapping,
 283                         loff_t pos, size_t count)
 284 {
 285         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 286         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 287         int ret;
 288
 289         if (!mapping_cap_writeback_dirty(mapping) || !count)
 290                 return 0;
 291         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 292         if (ret == 0) {
 293                 down(&inode->i_sem);
 294                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 295                 up(&inode->i_sem);
 296         }
 297         if (ret == 0)
 298                 ret = wait_on_page_writeback_range(mapping, start, end);
 299         return ret;
 300 }
 301 EXPORT_SYMBOL(sync_page_range);
 302
 303 /*
 304  * Note: Holding i_sem across sync_page_range_nolock is not a good idea
 305  * as it forces O_SYNC writers to different parts of the same file
 306  * to be serialised right until io completion.
 307  */
 308 static int sync_page_range_nolock(struct inode *inode,
 309                                   struct address_space *mapping,
 310                                   loff_t pos, size_t count)
 311 {
 312         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 313         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 314         int ret;
 315
 316         if (!mapping_cap_writeback_dirty(mapping) || !count)
 317                 return 0;
 318         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 319         if (ret == 0)
 320                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 321         if (ret == 0)
 322                 ret = wait_on_page_writeback_range(mapping, start, end);
 323         return ret;
 324 }
 325
 326 /**
 327  * filemap_fdatawait - walk the list of under-writeback pages of the given
 328  *     address space and wait for all of them.
 329  *
 330  * @mapping: address space structure to wait for
 331  */
 332 int filemap_fdatawait(struct address_space *mapping)
 333 {
 334         loff_t i_size = i_size_read(mapping->host);
 335
 336         if (i_size == 0)
 337                 return 0;
 338
 339         return wait_on_page_writeback_range(mapping, 0,
 340                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 341 }
 342 EXPORT_SYMBOL(filemap_fdatawait);
 343
 344 int filemap_write_and_wait(struct address_space *mapping)
 345 {
 346         int retval = 0;
 347
 348         if (mapping->nrpages) {
 349                 retval = filemap_fdatawrite(mapping);
 350                 if (retval == 0)
 351                         retval = filemap_fdatawait(mapping);
 352         }
 353         return retval;
 354 }
 355
 356 int filemap_write_and_wait_range(struct address_space *mapping,
 357                                  loff_t lstart, loff_t lend)
 358 {
 359         int retval = 0;
 360
 361         if (mapping->nrpages) {
 362                 retval = __filemap_fdatawrite_range(mapping, lstart, lend,
 363                                                     WB_SYNC_ALL);
 364                 if (retval == 0)
 365                         retval = wait_on_page_writeback_range(mapping,
 366                                                     lstart >> PAGE_CACHE_SHIFT,
 367                                                     lend >> PAGE_CACHE_SHIFT);
 368         }
 369         return retval;
 370 }
 371
 372 /*
 373  * This function is used to add newly allocated pagecache pages:
 374  * the page is new, so we can just run SetPageLocked() against it.
 375  * The other page state flags were set by rmqueue().
 376  *
 377  * This function does not add the page to the LRU.  The caller must do that.
 378  */
 379 int add_to_page_cache(struct page *page, struct address_space *mapping,
 380                 pgoff_t offset, gfp_t gfp_mask)
 381 {
 382         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 383
 384         if (error == 0) {
 385                 write_lock_irq(&mapping->tree_lock);
 386                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 387                 if (!error) {
 388                         page_cache_get(page);
 389                         SetPageLocked(page);
 390                         page->mapping = mapping;
 391                         page->index = offset;
 392                         mapping->nrpages++;
 393                         pagecache_acct(1);
 394                 }
 395                 write_unlock_irq(&mapping->tree_lock);
 396                 radix_tree_preload_end();
 397         }
 398         return error;
 399 }
 400
 401 EXPORT_SYMBOL(add_to_page_cache);
 402
 403 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 404                                 pgoff_t offset, gfp_t gfp_mask)
 405 {
 406         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 407         if (ret == 0)
 408                 lru_cache_add(page);
 409         return ret;
 410 }
 411
 412 /*
 413  * In order to wait for pages to become available there must be
 414  * waitqueues associated with pages. By using a hash table of
 415  * waitqueues where the bucket discipline is to maintain all
 416  * waiters on the same queue and wake all when any of the pages
 417  * become available, and for the woken contexts to check to be
 418  * sure the appropriate page became available, this saves space
 419  * at a cost of "thundering herd" phenomena during rare hash
 420  * collisions.
 421  */
 422 static wait_queue_head_t *page_waitqueue(struct page *page)
 423 {
 424         const struct zone *zone = page_zone(page);
 425
 426         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 427 }
 428
 429 static inline void wake_up_page(struct page *page, int bit)
 430 {
 431         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 432 }
 433
 434 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 435 {
 436         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 437
 438         if (test_bit(bit_nr, &page->flags))
 439                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 440                                                         TASK_UNINTERRUPTIBLE);
 441 }
 442 EXPORT_SYMBOL(wait_on_page_bit);
 443
 444 /**
 445  * unlock_page() - unlock a locked page
 446  *
 447  * @page: the page
 448  *
 449  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 450  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 451  * mechananism between PageLocked pages and PageWriteback pages is shared.
 452  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 453  *
 454  * The first mb is necessary to safely close the critical section opened by the
 455  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 456  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 457  * parallel wait_on_page_locked()).
 458  */
 459 void fastcall unlock_page(struct page *page)
 460 {
 461         smp_mb__before_clear_bit();
 462         if (!TestClearPageLocked(page))
 463                 BUG();
 464         smp_mb__after_clear_bit();
 465         wake_up_page(page, PG_locked);
 466 }
 467 EXPORT_SYMBOL(unlock_page);
 468
 469 /*
 470  * End writeback against a page.
 471  */
 472 void end_page_writeback(struct page *page)
 473 {
 474         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 475                 if (!test_clear_page_writeback(page))
 476                         BUG();
 477         }
 478         smp_mb__after_clear_bit();
 479         wake_up_page(page, PG_writeback);
 480 }
 481 EXPORT_SYMBOL(end_page_writeback);
 482
 483 /*
 484  * Get a lock on the page, assuming we need to sleep to get it.
 485  *
 486  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 487  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 488  * chances are that on the second loop, the block layer's plug list is empty,
 489  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 490  */
 491 void fastcall __lock_page(struct page *page)
 492 {
 493         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 494
 495         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 496                                                         TASK_UNINTERRUPTIBLE);
 497 }
 498 EXPORT_SYMBOL(__lock_page);
 499
 500 /*
 501  * a rather lightweight function, finding and getting a reference to a
 502  * hashed page atomically.
 503  */
 504 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 505 {
 506         struct page *page;
 507
 508         read_lock_irq(&mapping->tree_lock);
 509         page = radix_tree_lookup(&mapping->page_tree, offset);
 510         if (page)
 511                 page_cache_get(page);
 512         read_unlock_irq(&mapping->tree_lock);
 513         return page;
 514 }
 515
 516 EXPORT_SYMBOL(find_get_page);
 517
 518 /*
 519  * Same as above, but trylock it instead of incrementing the count.
 520  */
 521 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 522 {
 523         struct page *page;
 524
 525         read_lock_irq(&mapping->tree_lock);
 526         page = radix_tree_lookup(&mapping->page_tree, offset);
 527         if (page && TestSetPageLocked(page))
 528                 page = NULL;
 529         read_unlock_irq(&mapping->tree_lock);
 530         return page;
 531 }
 532
 533 EXPORT_SYMBOL(find_trylock_page);
 534
 535 /**
 536  * find_lock_page - locate, pin and lock a pagecache page
 537  *
 538  * @mapping: the address_space to search
 539  * @offset: the page index
 540  *
 541  * Locates the desired pagecache page, locks it, increments its reference
 542  * count and returns its address.
 543  *
 544  * Returns zero if the page was not present. find_lock_page() may sleep.
 545  */
 546 struct page *find_lock_page(struct address_space *mapping,
 547                                 unsigned long offset)
 548 {
 549         struct page *page;
 550
 551         read_lock_irq(&mapping->tree_lock);
 552 repeat:
 553         page = radix_tree_lookup(&mapping->page_tree, offset);
 554         if (page) {
 555                 page_cache_get(page);
 556                 if (TestSetPageLocked(page)) {
 557                         read_unlock_irq(&mapping->tree_lock);
 558                         __lock_page(page);
 559                         read_lock_irq(&mapping->tree_lock);
 560
 561                         /* Has the page been truncated while we slept? */
 562                         if (unlikely(page->mapping != mapping ||
 563                                      page->index != offset)) {
 564                                 unlock_page(page);
 565                                 page_cache_release(page);
 566                                 goto repeat;
 567                         }
 568                 }
 569         }
 570         read_unlock_irq(&mapping->tree_lock);
 571         return page;
 572 }
 573
 574 EXPORT_SYMBOL(find_lock_page);
 575
 576 /**
 577  * find_or_create_page - locate or add a pagecache page
 578  *
 579  * @mapping: the page's address_space
 580  * @index: the page's index into the mapping
 581  * @gfp_mask: page allocation mode
 582  *
 583  * Locates a page in the pagecache.  If the page is not present, a new page
 584  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 585  * LRU list.  The returned page is locked and has its reference count
 586  * incremented.
 587  *
 588  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 589  * allocation!
 590  *
 591  * find_or_create_page() returns the desired page's address, or zero on
 592  * memory exhaustion.
 593  */
 594 struct page *find_or_create_page(struct address_space *mapping,
 595                 unsigned long index, gfp_t gfp_mask)
 596 {
 597         struct page *page, *cached_page = NULL;
 598         int err;
 599 repeat:
 600         page = find_lock_page(mapping, index);
 601         if (!page) {
 602                 if (!cached_page) {
 603                         cached_page = alloc_page(gfp_mask);
 604                         if (!cached_page)
 605                                 return NULL;
 606                 }
 607                 err = add_to_page_cache_lru(cached_page, mapping,
 608                                         index, gfp_mask);
 609                 if (!err) {
 610                         page = cached_page;
 611                         cached_page = NULL;
 612                 } else if (err == -EEXIST)
 613                         goto repeat;
 614         }
 615         if (cached_page)
 616                 page_cache_release(cached_page);
 617         return page;
 618 }
 619
 620 EXPORT_SYMBOL(find_or_create_page);
 621
 622 /**
 623  * find_get_pages - gang pagecache lookup
 624  * @mapping:    The address_space to search
 625  * @start:      The starting page index
 626  * @nr_pages:   The maximum number of pages
 627  * @pages:      Where the resulting pages are placed
 628  *
 629  * find_get_pages() will search for and return a group of up to
 630  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 631  * find_get_pages() takes a reference against the returned pages.
 632  *
 633  * The search returns a group of mapping-contiguous pages with ascending
 634  * indexes.  There may be holes in the indices due to not-present pages.
 635  *
 636  * find_get_pages() returns the number of pages which were found.
 637  */
 638 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 639                             unsigned int nr_pages, struct page **pages)
 640 {
 641         unsigned int i;
 642         unsigned int ret;
 643
 644         read_lock_irq(&mapping->tree_lock);
 645         ret = radix_tree_gang_lookup(&mapping->page_tree,
 646                                 (void **)pages, start, nr_pages);
 647         for (i = 0; i < ret; i++)
 648                 page_cache_get(pages[i]);
 649         read_unlock_irq(&mapping->tree_lock);
 650         return ret;
 651 }
 652
 653 /*
 654  * Like find_get_pages, except we only return pages which are tagged with
 655  * `tag'.   We update *index to index the next page for the traversal.
 656  */
 657 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 658                         int tag, unsigned int nr_pages, struct page **pages)
 659 {
 660         unsigned int i;
 661         unsigned int ret;
 662
 663         read_lock_irq(&mapping->tree_lock);
 664         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 665                                 (void **)pages, *index, nr_pages, tag);
 666         for (i = 0; i < ret; i++)
 667                 page_cache_get(pages[i]);
 668         if (ret)
 669                 *index = pages[ret - 1]->index + 1;
 670         read_unlock_irq(&mapping->tree_lock);
 671         return ret;
 672 }
 673
 674 /*
 675  * Same as grab_cache_page, but do not wait if the page is unavailable.
 676  * This is intended for speculative data generators, where the data can
 677  * be regenerated if the page couldn't be grabbed.  This routine should
 678  * be safe to call while holding the lock for another page.
 679  *
 680  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 681  * and deadlock against the caller's locked page.
 682  */
 683 struct page *
 684 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 685 {
 686         struct page *page = find_get_page(mapping, index);
 687         gfp_t gfp_mask;
 688
 689         if (page) {
 690                 if (!TestSetPageLocked(page))
 691                         return page;
 692                 page_cache_release(page);
 693                 return NULL;
 694         }
 695         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 696         page = alloc_pages(gfp_mask, 0);
 697         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 698                 page_cache_release(page);
 699                 page = NULL;
 700         }
 701         return page;
 702 }
 703
 704 EXPORT_SYMBOL(grab_cache_page_nowait);
 705
 706 /*
 707  * This is a generic file read routine, and uses the
 708  * mapping->a_ops->readpage() function for the actual low-level
 709  * stuff.
 710  *
 711  * This is really ugly. But the goto's actually try to clarify some
 712  * of the logic when it comes to error handling etc.
 713  *
 714  * Note the struct file* is only passed for the use of readpage.  It may be
 715  * NULL.
 716  */
 717 void do_generic_mapping_read(struct address_space *mapping,
 718                              struct file_ra_state *_ra,
 719                              struct file *filp,
 720                              loff_t *ppos,
 721                              read_descriptor_t *desc,
 722                              read_actor_t actor)
 723 {
 724         struct inode *inode = mapping->host;
 725         unsigned long index;
 726         unsigned long end_index;
 727         unsigned long offset;
 728         unsigned long last_index;
 729         unsigned long next_index;
 730         unsigned long prev_index;
 731         loff_t isize;
 732         struct page *cached_page;
 733         int error;
 734         struct file_ra_state ra = *_ra;
 735
 736         cached_page = NULL;
 737         index = *ppos >> PAGE_CACHE_SHIFT;
 738         next_index = index;
 739         prev_index = ra.prev_page;
 740         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 741         offset = *ppos & ~PAGE_CACHE_MASK;
 742
 743         isize = i_size_read(inode);
 744         if (!isize)
 745                 goto out;
 746
 747         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 748         for (;;) {
 749                 struct page *page;
 750                 unsigned long nr, ret;
 751
 752                 /* nr is the maximum number of bytes to copy from this page */
 753                 nr = PAGE_CACHE_SIZE;
 754                 if (index >= end_index) {
 755                         if (index > end_index)
 756                                 goto out;
 757                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 758                         if (nr <= offset) {
 759                                 goto out;
 760                         }
 761                 }
 762                 nr = nr - offset;
 763
 764                 cond_resched();
 765                 if (index == next_index)
 766                         next_index = page_cache_readahead(mapping, &ra, filp,
 767                                         index, last_index - index);
 768
 769 find_page:
 770                 page = find_get_page(mapping, index);
 771                 if (unlikely(page == NULL)) {
 772                         handle_ra_miss(mapping, &ra, index);
 773                         goto no_cached_page;
 774                 }
 775                 if (!PageUptodate(page))
 776                         goto page_not_up_to_date;
 777 page_ok:
 778
 779                 /* If users can be writing to this page using arbitrary
 780                  * virtual addresses, take care about potential aliasing
 781                  * before reading the page on the kernel side.
 782                  */
 783                 if (mapping_writably_mapped(mapping))
 784                         flush_dcache_page(page);
 785
 786                 /*
 787                  * When (part of) the same page is read multiple times
 788                  * in succession, only mark it as accessed the first time.
 789                  */
 790                 if (prev_index != index)
 791                         mark_page_accessed(page);
 792                 prev_index = index;
 793
 794                 /*
 795                  * Ok, we have the page, and it's up-to-date, so
 796                  * now we can copy it to user space...
 797                  *
 798                  * The actor routine returns how many bytes were actually used..
 799                  * NOTE! This may not be the same as how much of a user buffer
 800                  * we filled up (we may be padding etc), so we can only update
 801                  * "pos" here (the actor routine has to update the user buffer
 802                  * pointers and the remaining count).
 803                  */
 804                 ret = actor(desc, page, offset, nr);
 805                 offset += ret;
 806                 index += offset >> PAGE_CACHE_SHIFT;
 807                 offset &= ~PAGE_CACHE_MASK;
 808
 809                 page_cache_release(page);
 810                 if (ret == nr && desc->count)
 811                         continue;
 812                 goto out;
 813
 814 page_not_up_to_date:
 815                 /* Get exclusive access to the page ... */
 816                 lock_page(page);
 817
 818                 /* Did it get unhashed before we got the lock? */
 819                 if (!page->mapping) {
 820                         unlock_page(page);
 821                         page_cache_release(page);
 822                         continue;
 823                 }
 824
 825                 /* Did somebody else fill it already? */
 826                 if (PageUptodate(page)) {
 827                         unlock_page(page);
 828                         goto page_ok;
 829                 }
 830
 831 readpage:
 832                 /* Start the actual read. The read will unlock the page. */
 833                 error = mapping->a_ops->readpage(filp, page);
 834
 835                 if (unlikely(error)) {
 836                         if (error == AOP_TRUNCATED_PAGE) {
 837                                 page_cache_release(page);
 838                                 goto find_page;
 839                         }
 840                         goto readpage_error;
 841                 }
 842
 843                 if (!PageUptodate(page)) {
 844                         lock_page(page);
 845                         if (!PageUptodate(page)) {
 846                                 if (page->mapping == NULL) {
 847                                         /*
 848                                          * invalidate_inode_pages got it
 849                                          */
 850                                         unlock_page(page);
 851                                         page_cache_release(page);
 852                                         goto find_page;
 853                                 }
 854                                 unlock_page(page);
 855                                 error = -EIO;
 856                                 goto readpage_error;
 857                         }
 858                         unlock_page(page);
 859                 }
 860
 861                 /*
 862                  * i_size must be checked after we have done ->readpage.
 863                  *
 864                  * Checking i_size after the readpage allows us to calculate
 865                  * the correct value for "nr", which means the zero-filled
 866                  * part of the page is not copied back to userspace (unless
 867                  * another truncate extends the file - this is desired though).
 868                  */
 869                 isize = i_size_read(inode);
 870                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 871                 if (unlikely(!isize || index > end_index)) {
 872                         page_cache_release(page);
 873                         goto out;
 874                 }
 875
 876                 /* nr is the maximum number of bytes to copy from this page */
 877                 nr = PAGE_CACHE_SIZE;
 878                 if (index == end_index) {
 879                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 880                         if (nr <= offset) {
 881                                 page_cache_release(page);
 882                                 goto out;
 883                         }
 884                 }
 885                 nr = nr - offset;
 886                 goto page_ok;
 887
 888 readpage_error:
 889                 /* UHHUH! A synchronous read error occurred. Report it */
 890                 desc->error = error;
 891                 page_cache_release(page);
 892                 goto out;
 893
 894 no_cached_page:
 895                 /*
 896                  * Ok, it wasn't cached, so we need to create a new
 897                  * page..
 898                  */
 899                 if (!cached_page) {
 900                         cached_page = page_cache_alloc_cold(mapping);
 901                         if (!cached_page) {
 902                                 desc->error = -ENOMEM;
 903                                 goto out;
 904                         }
 905                 }
 906                 error = add_to_page_cache_lru(cached_page, mapping,
 907                                                 index, GFP_KERNEL);
 908                 if (error) {
 909                         if (error == -EEXIST)
 910                                 goto find_page;
 911                         desc->error = error;
 912                         goto out;
 913                 }
 914                 page = cached_page;
 915                 cached_page = NULL;
 916                 goto readpage;
 917         }
 918
 919 out:
 920         *_ra = ra;
 921
 922         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 923         if (cached_page)
 924                 page_cache_release(cached_page);
 925         if (filp)
 926                 file_accessed(filp);
 927 }
 928
 929 EXPORT_SYMBOL(do_generic_mapping_read);
 930
 931 int file_read_actor(read_descriptor_t *desc, struct page *page,
 932                         unsigned long offset, unsigned long size)
 933 {
 934         char *kaddr;
 935         unsigned long left, count = desc->count;
 936
 937         if (size > count)
 938                 size = count;
 939
 940         /*
 941          * Faults on the destination of a read are common, so do it before
 942          * taking the kmap.
 943          */
 944         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 945                 kaddr = kmap_atomic(page, KM_USER0);
 946                 left = __copy_to_user_inatomic(desc->arg.buf,
 947                                                 kaddr + offset, size);
 948                 kunmap_atomic(kaddr, KM_USER0);
 949                 if (left == 0)
 950                         goto success;
 951         }
 952
 953         /* Do it the slow way */
 954         kaddr = kmap(page);
 955         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 956         kunmap(page);
 957
 958         if (left) {
 959                 size -= left;
 960                 desc->error = -EFAULT;
 961         }
 962 success:
 963         desc->count = count - size;
 964         desc->written += size;
 965         desc->arg.buf += size;
 966         return size;
 967 }
 968
 969 /*
 970  * This is the "read()" routine for all filesystems
 971  * that can use the page cache directly.
 972  */
 973 ssize_t
 974 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 975                 unsigned long nr_segs, loff_t *ppos)
 976 {
 977         struct file *filp = iocb->ki_filp;
 978         ssize_t retval;
 979         unsigned long seg;
 980         size_t count;
 981
 982         count = 0;
 983         for (seg = 0; seg < nr_segs; seg++) {
 984                 const struct iovec *iv = &iov[seg];
 985
 986                 /*
 987                  * If any segment has a negative length, or the cumulative
 988                  * length ever wraps negative then return -EINVAL.
 989                  */
 990                 count += iv->iov_len;
 991                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
 992                         return -EINVAL;
 993                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
 994                         continue;
 995                 if (seg == 0)
 996                         return -EFAULT;
 997                 nr_segs = seg;
 998                 count -= iv->iov_len;   /* This segment is no good */
 999                 break;
1000         }
1001
1002         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1003         if (filp->f_flags & O_DIRECT) {
1004                 loff_t pos = *ppos, size;
1005                 struct address_space *mapping;
1006                 struct inode *inode;
1007
1008                 mapping = filp->f_mapping;
1009                 inode = mapping->host;
1010                 retval = 0;
1011                 if (!count)
1012                         goto out; /* skip atime */
1013                 size = i_size_read(inode);
1014                 if (pos < size) {
1015                         retval = generic_file_direct_IO(READ, iocb,
1016                                                 iov, pos, nr_segs);
1017                         if (retval > 0 && !is_sync_kiocb(iocb))
1018                                 retval = -EIOCBQUEUED;
1019                         if (retval > 0)
1020                                 *ppos = pos + retval;
1021                 }
1022                 file_accessed(filp);
1023                 goto out;
1024         }
1025
1026         retval = 0;
1027         if (count) {
1028                 for (seg = 0; seg < nr_segs; seg++) {
1029                         read_descriptor_t desc;
1030
1031                         desc.written = 0;
1032                         desc.arg.buf = iov[seg].iov_base;
1033                         desc.count = iov[seg].iov_len;
1034                         if (desc.count == 0)
1035                                 continue;
1036                         desc.error = 0;
1037                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1038                         retval += desc.written;
1039                         if (desc.error) {
1040                                 retval = retval ?: desc.error;
1041                                 break;
1042                         }
1043                 }
1044         }
1045 out:
1046         return retval;
1047 }
1048
1049 EXPORT_SYMBOL(__generic_file_aio_read);
1050
1051 ssize_t
1052 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1053 {
1054         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1055
1056         BUG_ON(iocb->ki_pos != pos);
1057         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1058 }
1059
1060 EXPORT_SYMBOL(generic_file_aio_read);
1061
1062 ssize_t
1063 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1064 {
1065         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1066         struct kiocb kiocb;
1067         ssize_t ret;
1068
1069         init_sync_kiocb(&kiocb, filp);
1070         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1071         if (-EIOCBQUEUED == ret)
1072                 ret = wait_on_sync_kiocb(&kiocb);
1073         return ret;
1074 }
1075
1076 EXPORT_SYMBOL(generic_file_read);
1077
1078 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1079 {
1080         ssize_t written;
1081         unsigned long count = desc->count;
1082         struct file *file = desc->arg.data;
1083
1084         if (size > count)
1085                 size = count;
1086
1087         written = file->f_op->sendpage(file, page, offset,
1088                                        size, &file->f_pos, size<count);
1089         if (written < 0) {
1090                 desc->error = written;
1091                 written = 0;
1092         }
1093         desc->count = count - written;
1094         desc->written += written;
1095         return written;
1096 }
1097
1098 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1099                          size_t count, read_actor_t actor, void *target)
1100 {
1101         read_descriptor_t desc;
1102
1103         if (!count)
1104                 return 0;
1105
1106         desc.written = 0;
1107         desc.count = count;
1108         desc.arg.data = target;
1109         desc.error = 0;
1110
1111         do_generic_file_read(in_file, ppos, &desc, actor);
1112         if (desc.written)
1113                 return desc.written;
1114         return desc.error;
1115 }
1116
1117 EXPORT_SYMBOL(generic_file_sendfile);
1118
1119 static ssize_t
1120 do_readahead(struct address_space *mapping, struct file *filp,
1121              unsigned long index, unsigned long nr)
1122 {
1123         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1124                 return -EINVAL;
1125
1126         force_page_cache_readahead(mapping, filp, index,
1127                                         max_sane_readahead(nr));
1128         return 0;
1129 }
1130
1131 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1132 {
1133         ssize_t ret;
1134         struct file *file;
1135
1136         ret = -EBADF;
1137         file = fget(fd);
1138         if (file) {
1139                 if (file->f_mode & FMODE_READ) {
1140                         struct address_space *mapping = file->f_mapping;
1141                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1142                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1143                         unsigned long len = end - start + 1;
1144                         ret = do_readahead(mapping, file, start, len);
1145                 }
1146                 fput(file);
1147         }
1148         return ret;
1149 }
1150
1151 #ifdef CONFIG_MMU
1152 /*
1153  * This adds the requested page to the page cache if it isn't already there,
1154  * and schedules an I/O to read in its contents from disk.
1155  */
1156 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1157 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1158 {
1159         struct address_space *mapping = file->f_mapping;
1160         struct page *page;
1161         int ret;
1162
1163         do {
1164                 page = page_cache_alloc_cold(mapping);
1165                 if (!page)
1166                         return -ENOMEM;
1167
1168                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1169                 if (ret == 0)
1170                         ret = mapping->a_ops->readpage(file, page);
1171                 else if (ret == -EEXIST)
1172                         ret = 0; /* losing race to add is OK */
1173
1174                 page_cache_release(page);
1175
1176         } while (ret == AOP_TRUNCATED_PAGE);
1177
1178         return ret;
1179 }
1180
1181 #define MMAP_LOTSAMISS  (100)
1182
1183 /*
1184  * filemap_nopage() is invoked via the vma operations vector for a
1185  * mapped memory region to read in file data during a page fault.
1186  *
1187  * The goto's are kind of ugly, but this streamlines the normal case of having
1188  * it in the page cache, and handles the special cases reasonably without
1189  * having a lot of duplicated code.
1190  */
1191 struct page *filemap_nopage(struct vm_area_struct *area,
1192                                 unsigned long address, int *type)
1193 {
1194         int error;
1195         struct file *file = area->vm_file;
1196         struct address_space *mapping = file->f_mapping;
1197         struct file_ra_state *ra = &file->f_ra;
1198         struct inode *inode = mapping->host;
1199         struct page *page;
1200         unsigned long size, pgoff;
1201         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1202
1203         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1204
1205 retry_all:
1206         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1207         if (pgoff >= size)
1208                 goto outside_data_content;
1209
1210         /* If we don't want any read-ahead, don't bother */
1211         if (VM_RandomReadHint(area))
1212                 goto no_cached_page;
1213
1214         /*
1215          * The readahead code wants to be told about each and every page
1216          * so it can build and shrink its windows appropriately
1217          *
1218          * For sequential accesses, we use the generic readahead logic.
1219          */
1220         if (VM_SequentialReadHint(area))
1221                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1222
1223         /*
1224          * Do we have something in the page cache already?
1225          */
1226 retry_find:
1227         page = find_get_page(mapping, pgoff);
1228         if (!page) {
1229                 unsigned long ra_pages;
1230
1231                 if (VM_SequentialReadHint(area)) {
1232                         handle_ra_miss(mapping, ra, pgoff);
1233                         goto no_cached_page;
1234                 }
1235                 ra->mmap_miss++;
1236
1237                 /*
1238                  * Do we miss much more than hit in this file? If so,
1239                  * stop bothering with read-ahead. It will only hurt.
1240                  */
1241                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1242                         goto no_cached_page;
1243
1244                 /*
1245                  * To keep the pgmajfault counter straight, we need to
1246                  * check did_readaround, as this is an inner loop.
1247                  */
1248                 if (!did_readaround) {
1249                         majmin = VM_FAULT_MAJOR;
1250                         inc_page_state(pgmajfault);
1251                 }
1252                 did_readaround = 1;
1253                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1254                 if (ra_pages) {
1255                         pgoff_t start = 0;
1256
1257                         if (pgoff > ra_pages / 2)
1258                                 start = pgoff - ra_pages / 2;
1259                         do_page_cache_readahead(mapping, file, start, ra_pages);
1260                 }
1261                 page = find_get_page(mapping, pgoff);
1262                 if (!page)
1263                         goto no_cached_page;
1264         }
1265
1266         if (!did_readaround)
1267                 ra->mmap_hit++;
1268
1269         /*
1270          * Ok, found a page in the page cache, now we need to check
1271          * that it's up-to-date.
1272          */
1273         if (!PageUptodate(page))
1274                 goto page_not_uptodate;
1275
1276 success:
1277         /*
1278          * Found the page and have a reference on it.
1279          */
1280         mark_page_accessed(page);
1281         if (type)
1282                 *type = majmin;
1283         return page;
1284
1285 outside_data_content:
1286         /*
1287          * An external ptracer can access pages that normally aren't
1288          * accessible..
1289          */
1290         if (area->vm_mm == current->mm)
1291                 return NULL;
1292         /* Fall through to the non-read-ahead case */
1293 no_cached_page:
1294         /*
1295          * We're only likely to ever get here if MADV_RANDOM is in
1296          * effect.
1297          */
1298         error = page_cache_read(file, pgoff);
1299         grab_swap_token();
1300
1301         /*
1302          * The page we want has now been added to the page cache.
1303          * In the unlikely event that someone removed it in the
1304          * meantime, we'll just come back here and read it again.
1305          */
1306         if (error >= 0)
1307                 goto retry_find;
1308
1309         /*
1310          * An error return from page_cache_read can result if the
1311          * system is low on memory, or a problem occurs while trying
1312          * to schedule I/O.
1313          */
1314         if (error == -ENOMEM)
1315                 return NOPAGE_OOM;
1316         return NULL;
1317
1318 page_not_uptodate:
1319         if (!did_readaround) {
1320                 majmin = VM_FAULT_MAJOR;
1321                 inc_page_state(pgmajfault);
1322         }
1323         lock_page(page);
1324
1325         /* Did it get unhashed while we waited for it? */
1326         if (!page->mapping) {
1327                 unlock_page(page);
1328                 page_cache_release(page);
1329                 goto retry_all;
1330         }
1331
1332         /* Did somebody else get it up-to-date? */
1333         if (PageUptodate(page)) {
1334                 unlock_page(page);
1335                 goto success;
1336         }
1337
1338         error = mapping->a_ops->readpage(file, page);
1339         if (!error) {
1340                 wait_on_page_locked(page);
1341                 if (PageUptodate(page))
1342                         goto success;
1343         } else if (error == AOP_TRUNCATED_PAGE) {
1344                 page_cache_release(page);
1345                 goto retry_find;
1346         }
1347
1348         /*
1349          * Umm, take care of errors if the page isn't up-to-date.
1350          * Try to re-read it _once_. We do this synchronously,
1351          * because there really aren't any performance issues here
1352          * and we need to check for errors.
1353          */
1354         lock_page(page);
1355
1356         /* Somebody truncated the page on us? */
1357         if (!page->mapping) {
1358                 unlock_page(page);
1359                 page_cache_release(page);
1360                 goto retry_all;
1361         }
1362
1363         /* Somebody else successfully read it in? */
1364         if (PageUptodate(page)) {
1365                 unlock_page(page);
1366                 goto success;
1367         }
1368         ClearPageError(page);
1369         error = mapping->a_ops->readpage(file, page);
1370         if (!error) {
1371                 wait_on_page_locked(page);
1372                 if (PageUptodate(page))
1373                         goto success;
1374         } else if (error == AOP_TRUNCATED_PAGE) {
1375                 page_cache_release(page);
1376                 goto retry_find;
1377         }
1378
1379         /*
1380          * Things didn't work out. Return zero to tell the
1381          * mm layer so, possibly freeing the page cache page first.
1382          */
1383         page_cache_release(page);
1384         return NULL;
1385 }
1386
1387 EXPORT_SYMBOL(filemap_nopage);
1388
1389 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1390                                         int nonblock)
1391 {
1392         struct address_space *mapping = file->f_mapping;
1393         struct page *page;
1394         int error;
1395
1396         /*
1397          * Do we have something in the page cache already?
1398          */
1399 retry_find:
1400         page = find_get_page(mapping, pgoff);
1401         if (!page) {
1402                 if (nonblock)
1403                         return NULL;
1404                 goto no_cached_page;
1405         }
1406
1407         /*
1408          * Ok, found a page in the page cache, now we need to check
1409          * that it's up-to-date.
1410          */
1411         if (!PageUptodate(page)) {
1412                 if (nonblock) {
1413                         page_cache_release(page);
1414                         return NULL;
1415                 }
1416                 goto page_not_uptodate;
1417         }
1418
1419 success:
1420         /*
1421          * Found the page and have a reference on it.
1422          */
1423         mark_page_accessed(page);
1424         return page;
1425
1426 no_cached_page:
1427         error = page_cache_read(file, pgoff);
1428
1429         /*
1430          * The page we want has now been added to the page cache.
1431          * In the unlikely event that someone removed it in the
1432          * meantime, we'll just come back here and read it again.
1433          */
1434         if (error >= 0)
1435                 goto retry_find;
1436
1437         /*
1438          * An error return from page_cache_read can result if the
1439          * system is low on memory, or a problem occurs while trying
1440          * to schedule I/O.
1441          */
1442         return NULL;
1443
1444 page_not_uptodate:
1445         lock_page(page);
1446
1447         /* Did it get unhashed while we waited for it? */
1448         if (!page->mapping) {
1449                 unlock_page(page);
1450                 goto err;
1451         }
1452
1453         /* Did somebody else get it up-to-date? */
1454         if (PageUptodate(page)) {
1455                 unlock_page(page);
1456                 goto success;
1457         }
1458
1459         error = mapping->a_ops->readpage(file, page);
1460         if (!error) {
1461                 wait_on_page_locked(page);
1462                 if (PageUptodate(page))
1463                         goto success;
1464         } else if (error == AOP_TRUNCATED_PAGE) {
1465                 page_cache_release(page);
1466                 goto retry_find;
1467         }
1468
1469         /*
1470          * Umm, take care of errors if the page isn't up-to-date.
1471          * Try to re-read it _once_. We do this synchronously,
1472          * because there really aren't any performance issues here
1473          * and we need to check for errors.
1474          */
1475         lock_page(page);
1476
1477         /* Somebody truncated the page on us? */
1478         if (!page->mapping) {
1479                 unlock_page(page);
1480                 goto err;
1481         }
1482         /* Somebody else successfully read it in? */
1483         if (PageUptodate(page)) {
1484                 unlock_page(page);
1485                 goto success;
1486         }
1487
1488         ClearPageError(page);
1489         error = mapping->a_ops->readpage(file, page);
1490         if (!error) {
1491                 wait_on_page_locked(page);
1492                 if (PageUptodate(page))
1493                         goto success;
1494         } else if (error == AOP_TRUNCATED_PAGE) {
1495                 page_cache_release(page);
1496                 goto retry_find;
1497         }
1498
1499         /*
1500          * Things didn't work out. Return zero to tell the
1501          * mm layer so, possibly freeing the page cache page first.
1502          */
1503 err:
1504         page_cache_release(page);
1505
1506         return NULL;
1507 }
1508
1509 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1510                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1511                 int nonblock)
1512 {
1513         struct file *file = vma->vm_file;
1514         struct address_space *mapping = file->f_mapping;
1515         struct inode *inode = mapping->host;
1516         unsigned long size;
1517         struct mm_struct *mm = vma->vm_mm;
1518         struct page *page;
1519         int err;
1520
1521         if (!nonblock)
1522                 force_page_cache_readahead(mapping, vma->vm_file,
1523                                         pgoff, len >> PAGE_CACHE_SHIFT);
1524
1525 repeat:
1526         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1527         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1528                 return -EINVAL;
1529
1530         page = filemap_getpage(file, pgoff, nonblock);
1531
1532         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1533          * done in shmem_populate calling shmem_getpage */
1534         if (!page && !nonblock)
1535                 return -ENOMEM;
1536
1537         if (page) {
1538                 err = install_page(mm, vma, addr, page, prot);
1539                 if (err) {
1540                         page_cache_release(page);
1541                         return err;
1542                 }
1543         } else if (vma->vm_flags & VM_NONLINEAR) {
1544                 /* No page was found just because we can't read it in now (being
1545                  * here implies nonblock != 0), but the page may exist, so set
1546                  * the PTE to fault it in later. */
1547                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1548                 if (err)
1549                         return err;
1550         }
1551
1552         len -= PAGE_SIZE;
1553         addr += PAGE_SIZE;
1554         pgoff++;
1555         if (len)
1556                 goto repeat;
1557
1558         return 0;
1559 }
1560 EXPORT_SYMBOL(filemap_populate);
1561
1562 struct vm_operations_struct generic_file_vm_ops = {
1563         .nopage         = filemap_nopage,
1564         .populate       = filemap_populate,
1565 };
1566
1567 /* This is used for a general mmap of a disk file */
1568
1569 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1570 {
1571         struct address_space *mapping = file->f_mapping;
1572
1573         if (!mapping->a_ops->readpage)
1574                 return -ENOEXEC;
1575         file_accessed(file);
1576         vma->vm_ops = &generic_file_vm_ops;
1577         return 0;
1578 }
1579
1580 /*
1581  * This is for filesystems which do not implement ->writepage.
1582  */
1583 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1584 {
1585         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1586                 return -EINVAL;
1587         return generic_file_mmap(file, vma);
1588 }
1589 #else
1590 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1591 {
1592         return -ENOSYS;
1593 }
1594 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1595 {
1596         return -ENOSYS;
1597 }
1598 #endif /* CONFIG_MMU */
1599
1600 EXPORT_SYMBOL(generic_file_mmap);
1601 EXPORT_SYMBOL(generic_file_readonly_mmap);
1602
1603 static inline struct page *__read_cache_page(struct address_space *mapping,
1604                                 unsigned long index,
1605                                 int (*filler)(void *,struct page*),
1606                                 void *data)
1607 {
1608         struct page *page, *cached_page = NULL;
1609         int err;
1610 repeat:
1611         page = find_get_page(mapping, index);
1612         if (!page) {
1613                 if (!cached_page) {
1614                         cached_page = page_cache_alloc_cold(mapping);
1615                         if (!cached_page)
1616                                 return ERR_PTR(-ENOMEM);
1617                 }
1618                 err = add_to_page_cache_lru(cached_page, mapping,
1619                                         index, GFP_KERNEL);
1620                 if (err == -EEXIST)
1621                         goto repeat;
1622                 if (err < 0) {
1623                         /* Presumably ENOMEM for radix tree node */
1624                         page_cache_release(cached_page);
1625                         return ERR_PTR(err);
1626                 }
1627                 page = cached_page;
1628                 cached_page = NULL;
1629                 err = filler(data, page);
1630                 if (err < 0) {
1631                         page_cache_release(page);
1632                         page = ERR_PTR(err);
1633                 }
1634         }
1635         if (cached_page)
1636                 page_cache_release(cached_page);
1637         return page;
1638 }
1639
1640 /*
1641  * Read into the page cache. If a page already exists,
1642  * and PageUptodate() is not set, try to fill the page.
1643  */
1644 struct page *read_cache_page(struct address_space *mapping,
1645                                 unsigned long index,
1646                                 int (*filler)(void *,struct page*),
1647                                 void *data)
1648 {
1649         struct page *page;
1650         int err;
1651
1652 retry:
1653         page = __read_cache_page(mapping, index, filler, data);
1654         if (IS_ERR(page))
1655                 goto out;
1656         mark_page_accessed(page);
1657         if (PageUptodate(page))
1658                 goto out;
1659
1660         lock_page(page);
1661         if (!page->mapping) {
1662                 unlock_page(page);
1663                 page_cache_release(page);
1664                 goto retry;
1665         }
1666         if (PageUptodate(page)) {
1667                 unlock_page(page);
1668                 goto out;
1669         }
1670         err = filler(data, page);
1671         if (err < 0) {
1672                 page_cache_release(page);
1673                 page = ERR_PTR(err);
1674         }
1675  out:
1676         return page;
1677 }
1678
1679 EXPORT_SYMBOL(read_cache_page);
1680
1681 /*
1682  * If the page was newly created, increment its refcount and add it to the
1683  * caller's lru-buffering pagevec.  This function is specifically for
1684  * generic_file_write().
1685  */
1686 static inline struct page *
1687 __grab_cache_page(struct address_space *mapping, unsigned long index,
1688                         struct page **cached_page, struct pagevec *lru_pvec)
1689 {
1690         int err;
1691         struct page *page;
1692 repeat:
1693         page = find_lock_page(mapping, index);
1694         if (!page) {
1695                 if (!*cached_page) {
1696                         *cached_page = page_cache_alloc(mapping);
1697                         if (!*cached_page)
1698                                 return NULL;
1699                 }
1700                 err = add_to_page_cache(*cached_page, mapping,
1701                                         index, GFP_KERNEL);
1702                 if (err == -EEXIST)
1703                         goto repeat;
1704                 if (err == 0) {
1705                         page = *cached_page;
1706                         page_cache_get(page);
1707                         if (!pagevec_add(lru_pvec, page))
1708                                 __pagevec_lru_add(lru_pvec);
1709                         *cached_page = NULL;
1710                 }
1711         }
1712         return page;
1713 }
1714
1715 /*
1716  * The logic we want is
1717  *
1718  *      if suid or (sgid and xgrp)
1719  *              remove privs
1720  */
1721 int remove_suid(struct dentry *dentry)
1722 {
1723         mode_t mode = dentry->d_inode->i_mode;
1724         int kill = 0;
1725         int result = 0;
1726
1727         /* suid always must be killed */
1728         if (unlikely(mode & S_ISUID))
1729                 kill = ATTR_KILL_SUID;
1730
1731         /*
1732          * sgid without any exec bits is just a mandatory locking mark; leave
1733          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1734          */
1735         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1736                 kill |= ATTR_KILL_SGID;
1737
1738         if (unlikely(kill && !capable(CAP_FSETID))) {
1739                 struct iattr newattrs;
1740
1741                 newattrs.ia_valid = ATTR_FORCE | kill;
1742                 result = notify_change(dentry, &newattrs);
1743         }
1744         return result;
1745 }
1746 EXPORT_SYMBOL(remove_suid);
1747
1748 size_t
1749 __filemap_copy_from_user_iovec(char *vaddr,
1750                         const struct iovec *iov, size_t base, size_t bytes)
1751 {
1752         size_t copied = 0, left = 0;
1753
1754         while (bytes) {
1755                 char __user *buf = iov->iov_base + base;
1756                 int copy = min(bytes, iov->iov_len - base);
1757
1758                 base = 0;
1759                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1760                 copied += copy;
1761                 bytes -= copy;
1762                 vaddr += copy;
1763                 iov++;
1764
1765                 if (unlikely(left)) {
1766                         /* zero the rest of the target like __copy_from_user */
1767                         if (bytes)
1768                                 memset(vaddr, 0, bytes);
1769                         break;
1770                 }
1771         }
1772         return copied - left;
1773 }
1774
1775 /*
1776  * Performs necessary checks before doing a write
1777  *
1778  * Can adjust writing position aor amount of bytes to write.
1779  * Returns appropriate error code that caller should return or
1780  * zero in case that write should be allowed.
1781  */
1782 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1783 {
1784         struct inode *inode = file->f_mapping->host;
1785         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1786
1787         if (unlikely(*pos < 0))
1788                 return -EINVAL;
1789
1790         if (!isblk) {
1791                 /* FIXME: this is for backwards compatibility with 2.4 */
1792                 if (file->f_flags & O_APPEND)
1793                         *pos = i_size_read(inode);
1794
1795                 if (limit != RLIM_INFINITY) {
1796                         if (*pos >= limit) {
1797                                 send_sig(SIGXFSZ, current, 0);
1798                                 return -EFBIG;
1799                         }
1800                         if (*count > limit - (typeof(limit))*pos) {
1801                                 *count = limit - (typeof(limit))*pos;
1802                         }
1803                 }
1804         }
1805
1806         /*
1807          * LFS rule
1808          */
1809         if (unlikely(*pos + *count > MAX_NON_LFS &&
1810                                 !(file->f_flags & O_LARGEFILE))) {
1811                 if (*pos >= MAX_NON_LFS) {
1812                         send_sig(SIGXFSZ, current, 0);
1813                         return -EFBIG;
1814                 }
1815                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1816                         *count = MAX_NON_LFS - (unsigned long)*pos;
1817                 }
1818         }
1819
1820         /*
1821          * Are we about to exceed the fs block limit ?
1822          *
1823          * If we have written data it becomes a short write.  If we have
1824          * exceeded without writing data we send a signal and return EFBIG.
1825          * Linus frestrict idea will clean these up nicely..
1826          */
1827         if (likely(!isblk)) {
1828                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1829                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1830                                 send_sig(SIGXFSZ, current, 0);
1831                                 return -EFBIG;
1832                         }
1833                         /* zero-length writes at ->s_maxbytes are OK */
1834                 }
1835
1836                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1837                         *count = inode->i_sb->s_maxbytes - *pos;
1838         } else {
1839                 loff_t isize;
1840                 if (bdev_read_only(I_BDEV(inode)))
1841                         return -EPERM;
1842                 isize = i_size_read(inode);
1843                 if (*pos >= isize) {
1844                         if (*count || *pos > isize)
1845                                 return -ENOSPC;
1846                 }
1847
1848                 if (*pos + *count > isize)
1849                         *count = isize - *pos;
1850         }
1851         return 0;
1852 }
1853 EXPORT_SYMBOL(generic_write_checks);
1854
1855 ssize_t
1856 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1857                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1858                 size_t count, size_t ocount)
1859 {
1860         struct file     *file = iocb->ki_filp;
1861         struct address_space *mapping = file->f_mapping;
1862         struct inode    *inode = mapping->host;
1863         ssize_t         written;
1864
1865         if (count != ocount)
1866                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1867
1868         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1869         if (written > 0) {
1870                 loff_t end = pos + written;
1871                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1872                         i_size_write(inode,  end);
1873                         mark_inode_dirty(inode);
1874                 }
1875                 *ppos = end;
1876         }
1877
1878         /*
1879          * Sync the fs metadata but not the minor inode changes and
1880          * of course not the data as we did direct DMA for the IO.
1881          * i_sem is held, which protects generic_osync_inode() from
1882          * livelocking.
1883          */
1884         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1885                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1886                 if (err < 0)
1887                         written = err;
1888         }
1889         if (written == count && !is_sync_kiocb(iocb))
1890                 written = -EIOCBQUEUED;
1891         return written;
1892 }
1893 EXPORT_SYMBOL(generic_file_direct_write);
1894
1895 ssize_t
1896 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1897                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1898                 size_t count, ssize_t written)
1899 {
1900         struct file *file = iocb->ki_filp;
1901         struct address_space * mapping = file->f_mapping;
1902         struct address_space_operations *a_ops = mapping->a_ops;
1903         struct inode    *inode = mapping->host;
1904         long            status = 0;
1905         struct page     *page;
1906         struct page     *cached_page = NULL;
1907         size_t          bytes;
1908         struct pagevec  lru_pvec;
1909         const struct iovec *cur_iov = iov; /* current iovec */
1910         size_t          iov_base = 0;      /* offset in the current iovec */
1911         char __user     *buf;
1912
1913         pagevec_init(&lru_pvec, 0);
1914
1915         /*
1916          * handle partial DIO write.  Adjust cur_iov if needed.
1917          */
1918         if (likely(nr_segs == 1))
1919                 buf = iov->iov_base + written;
1920         else {
1921                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1922                 buf = cur_iov->iov_base + iov_base;
1923         }
1924
1925         do {
1926                 unsigned long index;
1927                 unsigned long offset;
1928                 unsigned long maxlen;
1929                 size_t copied;
1930
1931                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1932                 index = pos >> PAGE_CACHE_SHIFT;
1933                 bytes = PAGE_CACHE_SIZE - offset;
1934                 if (bytes > count)
1935                         bytes = count;
1936
1937                 /*
1938                  * Bring in the user page that we will copy from _first_.
1939                  * Otherwise there's a nasty deadlock on copying from the
1940                  * same page as we're writing to, without it being marked
1941                  * up-to-date.
1942                  */
1943                 maxlen = cur_iov->iov_len - iov_base;
1944                 if (maxlen > bytes)
1945                         maxlen = bytes;
1946                 fault_in_pages_readable(buf, maxlen);
1947
1948                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1949                 if (!page) {
1950                         status = -ENOMEM;
1951                         break;
1952                 }
1953
1954                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1955                 if (unlikely(status)) {
1956                         loff_t isize = i_size_read(inode);
1957
1958                         if (status != AOP_TRUNCATED_PAGE)
1959                                 unlock_page(page);
1960                         page_cache_release(page);
1961                         if (status == AOP_TRUNCATED_PAGE)
1962                                 continue;
1963                         /*
1964                          * prepare_write() may have instantiated a few blocks
1965                          * outside i_size.  Trim these off again.
1966                          */
1967                         if (pos + bytes > isize)
1968                                 vmtruncate(inode, isize);
1969                         break;
1970                 }
1971                 if (likely(nr_segs == 1))
1972                         copied = filemap_copy_from_user(page, offset,
1973                                                         buf, bytes);
1974                 else
1975                         copied = filemap_copy_from_user_iovec(page, offset,
1976                                                 cur_iov, iov_base, bytes);
1977                 flush_dcache_page(page);
1978                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1979                 if (status == AOP_TRUNCATED_PAGE) {
1980                         page_cache_release(page);
1981                         continue;
1982                 }
1983                 if (likely(copied > 0)) {
1984                         if (!status)
1985                                 status = copied;
1986
1987                         if (status >= 0) {
1988                                 written += status;
1989                                 count -= status;
1990                                 pos += status;
1991                                 buf += status;
1992                                 if (unlikely(nr_segs > 1)) {
1993                                         filemap_set_next_iovec(&cur_iov,
1994                                                         &iov_base, status);
1995                                         if (count)
1996                                                 buf = cur_iov->iov_base +
1997                                                         iov_base;
1998                                 } else {
1999                                         iov_base += status;
2000                                 }
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 EXPORT_SYMBOL(generic_file_buffered_write);
2042
2043 static ssize_t
2044 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2045                                 unsigned long nr_segs, loff_t *ppos)
2046 {
2047         struct file *file = iocb->ki_filp;
2048         struct address_space * mapping = file->f_mapping;
2049         size_t ocount;          /* original count */
2050         size_t count;           /* after file limit checks */
2051         struct inode    *inode = mapping->host;
2052         unsigned long   seg;
2053         loff_t          pos;
2054         ssize_t         written;
2055         ssize_t         err;
2056
2057         ocount = 0;
2058         for (seg = 0; seg < nr_segs; seg++) {
2059                 const struct iovec *iv = &iov[seg];
2060
2061                 /*
2062                  * If any segment has a negative length, or the cumulative
2063                  * length ever wraps negative then return -EINVAL.
2064                  */
2065                 ocount += iv->iov_len;
2066                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2067                         return -EINVAL;
2068                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2069                         continue;
2070                 if (seg == 0)
2071                         return -EFAULT;
2072                 nr_segs = seg;
2073                 ocount -= iv->iov_len;  /* This segment is no good */
2074                 break;
2075         }
2076
2077         count = ocount;
2078         pos = *ppos;
2079
2080         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2081
2082         /* We can write back this queue in page reclaim */
2083         current->backing_dev_info = mapping->backing_dev_info;
2084         written = 0;
2085
2086         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2087         if (err)
2088                 goto out;
2089
2090         if (count == 0)
2091                 goto out;
2092
2093         err = remove_suid(file->f_dentry);
2094         if (err)
2095                 goto out;
2096
2097         inode_update_time(inode, 1);
2098
2099         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2100         if (unlikely(file->f_flags & O_DIRECT)) {
2101                 written = generic_file_direct_write(iocb, iov,
2102                                 &nr_segs, pos, ppos, count, ocount);
2103                 if (written < 0 || written == count)
2104                         goto out;
2105                 /*
2106                  * direct-io write to a hole: fall through to buffered I/O
2107                  * for completing the rest of the request.
2108                  */
2109                 pos += written;
2110                 count -= written;
2111         }
2112
2113         written = generic_file_buffered_write(iocb, iov, nr_segs,
2114                         pos, ppos, count, written);
2115 out:
2116         current->backing_dev_info = NULL;
2117         return written ? written : err;
2118 }
2119 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2120
2121 ssize_t
2122 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2123                                 unsigned long nr_segs, loff_t *ppos)
2124 {
2125         struct file *file = iocb->ki_filp;
2126         struct address_space *mapping = file->f_mapping;
2127         struct inode *inode = mapping->host;
2128         ssize_t ret;
2129         loff_t pos = *ppos;
2130
2131         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2132
2133         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2134                 int err;
2135
2136                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2137                 if (err < 0)
2138                         ret = err;
2139         }
2140         return ret;
2141 }
2142
2143 static ssize_t
2144 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2145                                 unsigned long nr_segs, loff_t *ppos)
2146 {
2147         struct kiocb kiocb;
2148         ssize_t ret;
2149
2150         init_sync_kiocb(&kiocb, file);
2151         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2152         if (ret == -EIOCBQUEUED)
2153                 ret = wait_on_sync_kiocb(&kiocb);
2154         return ret;
2155 }
2156
2157 ssize_t
2158 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2159                                 unsigned long nr_segs, loff_t *ppos)
2160 {
2161         struct kiocb kiocb;
2162         ssize_t ret;
2163
2164         init_sync_kiocb(&kiocb, file);
2165         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2166         if (-EIOCBQUEUED == ret)
2167                 ret = wait_on_sync_kiocb(&kiocb);
2168         return ret;
2169 }
2170 EXPORT_SYMBOL(generic_file_write_nolock);
2171
2172 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2173                                size_t count, loff_t pos)
2174 {
2175         struct file *file = iocb->ki_filp;
2176         struct address_space *mapping = file->f_mapping;
2177         struct inode *inode = mapping->host;
2178         ssize_t ret;
2179         struct iovec local_iov = { .iov_base = (void __user *)buf,
2180                                         .iov_len = count };
2181
2182         BUG_ON(iocb->ki_pos != pos);
2183
2184         down(&inode->i_sem);
2185         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2186                                                 &iocb->ki_pos);
2187         up(&inode->i_sem);
2188
2189         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2190                 ssize_t err;
2191
2192                 err = sync_page_range(inode, mapping, pos, ret);
2193                 if (err < 0)
2194                         ret = err;
2195         }
2196         return ret;
2197 }
2198 EXPORT_SYMBOL(generic_file_aio_write);
2199
2200 ssize_t generic_file_write(struct file *file, const char __user *buf,
2201                            size_t count, loff_t *ppos)
2202 {
2203         struct address_space *mapping = file->f_mapping;
2204         struct inode *inode = mapping->host;
2205         ssize_t ret;
2206         struct iovec local_iov = { .iov_base = (void __user *)buf,
2207                                         .iov_len = count };
2208
2209         down(&inode->i_sem);
2210         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2211         up(&inode->i_sem);
2212
2213         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2214                 ssize_t err;
2215
2216                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2217                 if (err < 0)
2218                         ret = err;
2219         }
2220         return ret;
2221 }
2222 EXPORT_SYMBOL(generic_file_write);
2223
2224 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2225                         unsigned long nr_segs, loff_t *ppos)
2226 {
2227         struct kiocb kiocb;
2228         ssize_t ret;
2229
2230         init_sync_kiocb(&kiocb, filp);
2231         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2232         if (-EIOCBQUEUED == ret)
2233                 ret = wait_on_sync_kiocb(&kiocb);
2234         return ret;
2235 }
2236 EXPORT_SYMBOL(generic_file_readv);
2237
2238 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2239                         unsigned long nr_segs, loff_t *ppos)
2240 {
2241         struct address_space *mapping = file->f_mapping;
2242         struct inode *inode = mapping->host;
2243         ssize_t ret;
2244
2245         down(&inode->i_sem);
2246         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2247         up(&inode->i_sem);
2248
2249         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2250                 int err;
2251
2252                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2253                 if (err < 0)
2254                         ret = err;
2255         }
2256         return ret;
2257 }
2258 EXPORT_SYMBOL(generic_file_writev);
2259
2260 /*
2261  * Called under i_sem for writes to S_ISREG files.   Returns -EIO if something
2262  * went wrong during pagecache shootdown.
2263  */
2264 static ssize_t
2265 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2266         loff_t offset, unsigned long nr_segs)
2267 {
2268         struct file *file = iocb->ki_filp;
2269         struct address_space *mapping = file->f_mapping;
2270         ssize_t retval;
2271         size_t write_len = 0;
2272
2273         /*
2274          * If it's a write, unmap all mmappings of the file up-front.  This
2275          * will cause any pte dirty bits to be propagated into the pageframes
2276          * for the subsequent filemap_write_and_wait().
2277          */
2278         if (rw == WRITE) {
2279                 write_len = iov_length(iov, nr_segs);
2280                 if (mapping_mapped(mapping))
2281                         unmap_mapping_range(mapping, offset, write_len, 0);
2282         }
2283
2284         retval = filemap_write_and_wait(mapping);
2285         if (retval == 0) {
2286                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2287                                                 offset, nr_segs);
2288                 if (rw == WRITE && mapping->nrpages) {
2289                         pgoff_t end = (offset + write_len - 1)
2290                                                 >> PAGE_CACHE_SHIFT;
2291                         int err = invalidate_inode_pages2_range(mapping,
2292                                         offset >> PAGE_CACHE_SHIFT, end);
2293                         if (err)
2294                                 retval = err;
2295                 }
2296         }
2297         return retval;
2298 }