git.oblomov.eu Git - linux-2.6/blob - 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_mutex
  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_mutex                 (msync)
  77  *
  78  *  ->i_mutex
  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_mutex 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, loff_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                 mutex_lock(&inode->i_mutex);
 294                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 295                 mutex_unlock(&inode->i_mutex);
 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_mutex across sync_page_range_nolock is not a good idea
 305  * as it forces O_SYNC writers to different parts of the same file
 306  * to be serialised right until io completion.
 307  */
 308 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
 309                            loff_t pos, loff_t count)
 310 {
 311         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 312         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 313         int ret;
 314
 315         if (!mapping_cap_writeback_dirty(mapping) || !count)
 316                 return 0;
 317         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 318         if (ret == 0)
 319                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 320         if (ret == 0)
 321                 ret = wait_on_page_writeback_range(mapping, start, end);
 322         return ret;
 323 }
 324 EXPORT_SYMBOL(sync_page_range_nolock);
 325
 326 /**
 327  * filemap_fdatawait - walk the list of under-writeback pages of the given
 328  *     address space and wait for all of them.
 329  *
 330  * @mapping: address space structure to wait for
 331  */
 332 int filemap_fdatawait(struct address_space *mapping)
 333 {
 334         loff_t i_size = i_size_read(mapping->host);
 335
 336         if (i_size == 0)
 337                 return 0;
 338
 339         return wait_on_page_writeback_range(mapping, 0,
 340                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 341 }
 342 EXPORT_SYMBOL(filemap_fdatawait);
 343
 344 int filemap_write_and_wait(struct address_space *mapping)
 345 {
 346         int err = 0;
 347
 348         if (mapping->nrpages) {
 349                 err = filemap_fdatawrite(mapping);
 350                 /*
 351                  * Even if the above returned error, the pages may be
 352                  * written partially (e.g. -ENOSPC), so we wait for it.
 353                  * But the -EIO is special case, it may indicate the worst
 354                  * thing (e.g. bug) happened, so we avoid waiting for it.
 355                  */
 356                 if (err != -EIO) {
 357                         int err2 = filemap_fdatawait(mapping);
 358                         if (!err)
 359                                 err = err2;
 360                 }
 361         }
 362         return err;
 363 }
 364 EXPORT_SYMBOL(filemap_write_and_wait);
 365
 366 int filemap_write_and_wait_range(struct address_space *mapping,
 367                                  loff_t lstart, loff_t lend)
 368 {
 369         int err = 0;
 370
 371         if (mapping->nrpages) {
 372                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
 373                                                  WB_SYNC_ALL);
 374                 /* See comment of filemap_write_and_wait() */
 375                 if (err != -EIO) {
 376                         int err2 = wait_on_page_writeback_range(mapping,
 377                                                 lstart >> PAGE_CACHE_SHIFT,
 378                                                 lend >> PAGE_CACHE_SHIFT);
 379                         if (!err)
 380                                 err = err2;
 381                 }
 382         }
 383         return err;
 384 }
 385
 386 /*
 387  * This function is used to add newly allocated pagecache pages:
 388  * the page is new, so we can just run SetPageLocked() against it.
 389  * The other page state flags were set by rmqueue().
 390  *
 391  * This function does not add the page to the LRU.  The caller must do that.
 392  */
 393 int add_to_page_cache(struct page *page, struct address_space *mapping,
 394                 pgoff_t offset, gfp_t gfp_mask)
 395 {
 396         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 397
 398         if (error == 0) {
 399                 write_lock_irq(&mapping->tree_lock);
 400                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 401                 if (!error) {
 402                         page_cache_get(page);
 403                         SetPageLocked(page);
 404                         page->mapping = mapping;
 405                         page->index = offset;
 406                         mapping->nrpages++;
 407                         pagecache_acct(1);
 408                 }
 409                 write_unlock_irq(&mapping->tree_lock);
 410                 radix_tree_preload_end();
 411         }
 412         return error;
 413 }
 414
 415 EXPORT_SYMBOL(add_to_page_cache);
 416
 417 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 418                                 pgoff_t offset, gfp_t gfp_mask)
 419 {
 420         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 421         if (ret == 0)
 422                 lru_cache_add(page);
 423         return ret;
 424 }
 425
 426 /*
 427  * In order to wait for pages to become available there must be
 428  * waitqueues associated with pages. By using a hash table of
 429  * waitqueues where the bucket discipline is to maintain all
 430  * waiters on the same queue and wake all when any of the pages
 431  * become available, and for the woken contexts to check to be
 432  * sure the appropriate page became available, this saves space
 433  * at a cost of "thundering herd" phenomena during rare hash
 434  * collisions.
 435  */
 436 static wait_queue_head_t *page_waitqueue(struct page *page)
 437 {
 438         const struct zone *zone = page_zone(page);
 439
 440         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 441 }
 442
 443 static inline void wake_up_page(struct page *page, int bit)
 444 {
 445         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 446 }
 447
 448 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 449 {
 450         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 451
 452         if (test_bit(bit_nr, &page->flags))
 453                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 454                                                         TASK_UNINTERRUPTIBLE);
 455 }
 456 EXPORT_SYMBOL(wait_on_page_bit);
 457
 458 /**
 459  * unlock_page() - unlock a locked page
 460  *
 461  * @page: the page
 462  *
 463  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 464  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 465  * mechananism between PageLocked pages and PageWriteback pages is shared.
 466  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 467  *
 468  * The first mb is necessary to safely close the critical section opened by the
 469  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 470  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 471  * parallel wait_on_page_locked()).
 472  */
 473 void fastcall unlock_page(struct page *page)
 474 {
 475         smp_mb__before_clear_bit();
 476         if (!TestClearPageLocked(page))
 477                 BUG();
 478         smp_mb__after_clear_bit();
 479         wake_up_page(page, PG_locked);
 480 }
 481 EXPORT_SYMBOL(unlock_page);
 482
 483 /*
 484  * End writeback against a page.
 485  */
 486 void end_page_writeback(struct page *page)
 487 {
 488         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 489                 if (!test_clear_page_writeback(page))
 490                         BUG();
 491         }
 492         smp_mb__after_clear_bit();
 493         wake_up_page(page, PG_writeback);
 494 }
 495 EXPORT_SYMBOL(end_page_writeback);
 496
 497 /*
 498  * Get a lock on the page, assuming we need to sleep to get it.
 499  *
 500  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 501  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 502  * chances are that on the second loop, the block layer's plug list is empty,
 503  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 504  */
 505 void fastcall __lock_page(struct page *page)
 506 {
 507         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 508
 509         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 510                                                         TASK_UNINTERRUPTIBLE);
 511 }
 512 EXPORT_SYMBOL(__lock_page);
 513
 514 /*
 515  * a rather lightweight function, finding and getting a reference to a
 516  * hashed page atomically.
 517  */
 518 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 519 {
 520         struct page *page;
 521
 522         read_lock_irq(&mapping->tree_lock);
 523         page = radix_tree_lookup(&mapping->page_tree, offset);
 524         if (page)
 525                 page_cache_get(page);
 526         read_unlock_irq(&mapping->tree_lock);
 527         return page;
 528 }
 529
 530 EXPORT_SYMBOL(find_get_page);
 531
 532 /*
 533  * Same as above, but trylock it instead of incrementing the count.
 534  */
 535 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 536 {
 537         struct page *page;
 538
 539         read_lock_irq(&mapping->tree_lock);
 540         page = radix_tree_lookup(&mapping->page_tree, offset);
 541         if (page && TestSetPageLocked(page))
 542                 page = NULL;
 543         read_unlock_irq(&mapping->tree_lock);
 544         return page;
 545 }
 546
 547 EXPORT_SYMBOL(find_trylock_page);
 548
 549 /**
 550  * find_lock_page - locate, pin and lock a pagecache page
 551  *
 552  * @mapping: the address_space to search
 553  * @offset: the page index
 554  *
 555  * Locates the desired pagecache page, locks it, increments its reference
 556  * count and returns its address.
 557  *
 558  * Returns zero if the page was not present. find_lock_page() may sleep.
 559  */
 560 struct page *find_lock_page(struct address_space *mapping,
 561                                 unsigned long offset)
 562 {
 563         struct page *page;
 564
 565         read_lock_irq(&mapping->tree_lock);
 566 repeat:
 567         page = radix_tree_lookup(&mapping->page_tree, offset);
 568         if (page) {
 569                 page_cache_get(page);
 570                 if (TestSetPageLocked(page)) {
 571                         read_unlock_irq(&mapping->tree_lock);
 572                         __lock_page(page);
 573                         read_lock_irq(&mapping->tree_lock);
 574
 575                         /* Has the page been truncated while we slept? */
 576                         if (unlikely(page->mapping != mapping ||
 577                                      page->index != offset)) {
 578                                 unlock_page(page);
 579                                 page_cache_release(page);
 580                                 goto repeat;
 581                         }
 582                 }
 583         }
 584         read_unlock_irq(&mapping->tree_lock);
 585         return page;
 586 }
 587
 588 EXPORT_SYMBOL(find_lock_page);
 589
 590 /**
 591  * find_or_create_page - locate or add a pagecache page
 592  *
 593  * @mapping: the page's address_space
 594  * @index: the page's index into the mapping
 595  * @gfp_mask: page allocation mode
 596  *
 597  * Locates a page in the pagecache.  If the page is not present, a new page
 598  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 599  * LRU list.  The returned page is locked and has its reference count
 600  * incremented.
 601  *
 602  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 603  * allocation!
 604  *
 605  * find_or_create_page() returns the desired page's address, or zero on
 606  * memory exhaustion.
 607  */
 608 struct page *find_or_create_page(struct address_space *mapping,
 609                 unsigned long index, gfp_t gfp_mask)
 610 {
 611         struct page *page, *cached_page = NULL;
 612         int err;
 613 repeat:
 614         page = find_lock_page(mapping, index);
 615         if (!page) {
 616                 if (!cached_page) {
 617                         cached_page = alloc_page(gfp_mask);
 618                         if (!cached_page)
 619                                 return NULL;
 620                 }
 621                 err = add_to_page_cache_lru(cached_page, mapping,
 622                                         index, gfp_mask);
 623                 if (!err) {
 624                         page = cached_page;
 625                         cached_page = NULL;
 626                 } else if (err == -EEXIST)
 627                         goto repeat;
 628         }
 629         if (cached_page)
 630                 page_cache_release(cached_page);
 631         return page;
 632 }
 633
 634 EXPORT_SYMBOL(find_or_create_page);
 635
 636 /**
 637  * find_get_pages - gang pagecache lookup
 638  * @mapping:    The address_space to search
 639  * @start:      The starting page index
 640  * @nr_pages:   The maximum number of pages
 641  * @pages:      Where the resulting pages are placed
 642  *
 643  * find_get_pages() will search for and return a group of up to
 644  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 645  * find_get_pages() takes a reference against the returned pages.
 646  *
 647  * The search returns a group of mapping-contiguous pages with ascending
 648  * indexes.  There may be holes in the indices due to not-present pages.
 649  *
 650  * find_get_pages() returns the number of pages which were found.
 651  */
 652 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 653                             unsigned int nr_pages, struct page **pages)
 654 {
 655         unsigned int i;
 656         unsigned int ret;
 657
 658         read_lock_irq(&mapping->tree_lock);
 659         ret = radix_tree_gang_lookup(&mapping->page_tree,
 660                                 (void **)pages, start, nr_pages);
 661         for (i = 0; i < ret; i++)
 662                 page_cache_get(pages[i]);
 663         read_unlock_irq(&mapping->tree_lock);
 664         return ret;
 665 }
 666
 667 /*
 668  * Like find_get_pages, except we only return pages which are tagged with
 669  * `tag'.   We update *index to index the next page for the traversal.
 670  */
 671 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 672                         int tag, unsigned int nr_pages, struct page **pages)
 673 {
 674         unsigned int i;
 675         unsigned int ret;
 676
 677         read_lock_irq(&mapping->tree_lock);
 678         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 679                                 (void **)pages, *index, nr_pages, tag);
 680         for (i = 0; i < ret; i++)
 681                 page_cache_get(pages[i]);
 682         if (ret)
 683                 *index = pages[ret - 1]->index + 1;
 684         read_unlock_irq(&mapping->tree_lock);
 685         return ret;
 686 }
 687
 688 /*
 689  * Same as grab_cache_page, but do not wait if the page is unavailable.
 690  * This is intended for speculative data generators, where the data can
 691  * be regenerated if the page couldn't be grabbed.  This routine should
 692  * be safe to call while holding the lock for another page.
 693  *
 694  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 695  * and deadlock against the caller's locked page.
 696  */
 697 struct page *
 698 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 699 {
 700         struct page *page = find_get_page(mapping, index);
 701         gfp_t gfp_mask;
 702
 703         if (page) {
 704                 if (!TestSetPageLocked(page))
 705                         return page;
 706                 page_cache_release(page);
 707                 return NULL;
 708         }
 709         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 710         page = alloc_pages(gfp_mask, 0);
 711         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 712                 page_cache_release(page);
 713                 page = NULL;
 714         }
 715         return page;
 716 }
 717
 718 EXPORT_SYMBOL(grab_cache_page_nowait);
 719
 720 /*
 721  * This is a generic file read routine, and uses the
 722  * mapping->a_ops->readpage() function for the actual low-level
 723  * stuff.
 724  *
 725  * This is really ugly. But the goto's actually try to clarify some
 726  * of the logic when it comes to error handling etc.
 727  *
 728  * Note the struct file* is only passed for the use of readpage.  It may be
 729  * NULL.
 730  */
 731 void do_generic_mapping_read(struct address_space *mapping,
 732                              struct file_ra_state *_ra,
 733                              struct file *filp,
 734                              loff_t *ppos,
 735                              read_descriptor_t *desc,
 736                              read_actor_t actor)
 737 {
 738         struct inode *inode = mapping->host;
 739         unsigned long index;
 740         unsigned long end_index;
 741         unsigned long offset;
 742         unsigned long last_index;
 743         unsigned long next_index;
 744         unsigned long prev_index;
 745         loff_t isize;
 746         struct page *cached_page;
 747         int error;
 748         struct file_ra_state ra = *_ra;
 749
 750         cached_page = NULL;
 751         index = *ppos >> PAGE_CACHE_SHIFT;
 752         next_index = index;
 753         prev_index = ra.prev_page;
 754         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 755         offset = *ppos & ~PAGE_CACHE_MASK;
 756
 757         isize = i_size_read(inode);
 758         if (!isize)
 759                 goto out;
 760
 761         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 762         for (;;) {
 763                 struct page *page;
 764                 unsigned long nr, ret;
 765
 766                 /* nr is the maximum number of bytes to copy from this page */
 767                 nr = PAGE_CACHE_SIZE;
 768                 if (index >= end_index) {
 769                         if (index > end_index)
 770                                 goto out;
 771                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 772                         if (nr <= offset) {
 773                                 goto out;
 774                         }
 775                 }
 776                 nr = nr - offset;
 777
 778                 cond_resched();
 779                 if (index == next_index)
 780                         next_index = page_cache_readahead(mapping, &ra, filp,
 781                                         index, last_index - index);
 782
 783 find_page:
 784                 page = find_get_page(mapping, index);
 785                 if (unlikely(page == NULL)) {
 786                         handle_ra_miss(mapping, &ra, index);
 787                         goto no_cached_page;
 788                 }
 789                 if (!PageUptodate(page))
 790                         goto page_not_up_to_date;
 791 page_ok:
 792
 793                 /* If users can be writing to this page using arbitrary
 794                  * virtual addresses, take care about potential aliasing
 795                  * before reading the page on the kernel side.
 796                  */
 797                 if (mapping_writably_mapped(mapping))
 798                         flush_dcache_page(page);
 799
 800                 /*
 801                  * When (part of) the same page is read multiple times
 802                  * in succession, only mark it as accessed the first time.
 803                  */
 804                 if (prev_index != index)
 805                         mark_page_accessed(page);
 806                 prev_index = index;
 807
 808                 /*
 809                  * Ok, we have the page, and it's up-to-date, so
 810                  * now we can copy it to user space...
 811                  *
 812                  * The actor routine returns how many bytes were actually used..
 813                  * NOTE! This may not be the same as how much of a user buffer
 814                  * we filled up (we may be padding etc), so we can only update
 815                  * "pos" here (the actor routine has to update the user buffer
 816                  * pointers and the remaining count).
 817                  */
 818                 ret = actor(desc, page, offset, nr);
 819                 offset += ret;
 820                 index += offset >> PAGE_CACHE_SHIFT;
 821                 offset &= ~PAGE_CACHE_MASK;
 822
 823                 page_cache_release(page);
 824                 if (ret == nr && desc->count)
 825                         continue;
 826                 goto out;
 827
 828 page_not_up_to_date:
 829                 /* Get exclusive access to the page ... */
 830                 lock_page(page);
 831
 832                 /* Did it get unhashed before we got the lock? */
 833                 if (!page->mapping) {
 834                         unlock_page(page);
 835                         page_cache_release(page);
 836                         continue;
 837                 }
 838
 839                 /* Did somebody else fill it already? */
 840                 if (PageUptodate(page)) {
 841                         unlock_page(page);
 842                         goto page_ok;
 843                 }
 844
 845 readpage:
 846                 /* Start the actual read. The read will unlock the page. */
 847                 error = mapping->a_ops->readpage(filp, page);
 848
 849                 if (unlikely(error)) {
 850                         if (error == AOP_TRUNCATED_PAGE) {
 851                                 page_cache_release(page);
 852                                 goto find_page;
 853                         }
 854                         goto readpage_error;
 855                 }
 856
 857                 if (!PageUptodate(page)) {
 858                         lock_page(page);
 859                         if (!PageUptodate(page)) {
 860                                 if (page->mapping == NULL) {
 861                                         /*
 862                                          * invalidate_inode_pages got it
 863                                          */
 864                                         unlock_page(page);
 865                                         page_cache_release(page);
 866                                         goto find_page;
 867                                 }
 868                                 unlock_page(page);
 869                                 error = -EIO;
 870                                 goto readpage_error;
 871                         }
 872                         unlock_page(page);
 873                 }
 874
 875                 /*
 876                  * i_size must be checked after we have done ->readpage.
 877                  *
 878                  * Checking i_size after the readpage allows us to calculate
 879                  * the correct value for "nr", which means the zero-filled
 880                  * part of the page is not copied back to userspace (unless
 881                  * another truncate extends the file - this is desired though).
 882                  */
 883                 isize = i_size_read(inode);
 884                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 885                 if (unlikely(!isize || index > end_index)) {
 886                         page_cache_release(page);
 887                         goto out;
 888                 }
 889
 890                 /* nr is the maximum number of bytes to copy from this page */
 891                 nr = PAGE_CACHE_SIZE;
 892                 if (index == end_index) {
 893                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 894                         if (nr <= offset) {
 895                                 page_cache_release(page);
 896                                 goto out;
 897                         }
 898                 }
 899                 nr = nr - offset;
 900                 goto page_ok;
 901
 902 readpage_error:
 903                 /* UHHUH! A synchronous read error occurred. Report it */
 904                 desc->error = error;
 905                 page_cache_release(page);
 906                 goto out;
 907
 908 no_cached_page:
 909                 /*
 910                  * Ok, it wasn't cached, so we need to create a new
 911                  * page..
 912                  */
 913                 if (!cached_page) {
 914                         cached_page = page_cache_alloc_cold(mapping);
 915                         if (!cached_page) {
 916                                 desc->error = -ENOMEM;
 917                                 goto out;
 918                         }
 919                 }
 920                 error = add_to_page_cache_lru(cached_page, mapping,
 921                                                 index, GFP_KERNEL);
 922                 if (error) {
 923                         if (error == -EEXIST)
 924                                 goto find_page;
 925                         desc->error = error;
 926                         goto out;
 927                 }
 928                 page = cached_page;
 929                 cached_page = NULL;
 930                 goto readpage;
 931         }
 932
 933 out:
 934         *_ra = ra;
 935
 936         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 937         if (cached_page)
 938                 page_cache_release(cached_page);
 939         if (filp)
 940                 file_accessed(filp);
 941 }
 942
 943 EXPORT_SYMBOL(do_generic_mapping_read);
 944
 945 int file_read_actor(read_descriptor_t *desc, struct page *page,
 946                         unsigned long offset, unsigned long size)
 947 {
 948         char *kaddr;
 949         unsigned long left, count = desc->count;
 950
 951         if (size > count)
 952                 size = count;
 953
 954         /*
 955          * Faults on the destination of a read are common, so do it before
 956          * taking the kmap.
 957          */
 958         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 959                 kaddr = kmap_atomic(page, KM_USER0);
 960                 left = __copy_to_user_inatomic(desc->arg.buf,
 961                                                 kaddr + offset, size);
 962                 kunmap_atomic(kaddr, KM_USER0);
 963                 if (left == 0)
 964                         goto success;
 965         }
 966
 967         /* Do it the slow way */
 968         kaddr = kmap(page);
 969         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 970         kunmap(page);
 971
 972         if (left) {
 973                 size -= left;
 974                 desc->error = -EFAULT;
 975         }
 976 success:
 977         desc->count = count - size;
 978         desc->written += size;
 979         desc->arg.buf += size;
 980         return size;
 981 }
 982
 983 /*
 984  * This is the "read()" routine for all filesystems
 985  * that can use the page cache directly.
 986  */
 987 ssize_t
 988 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 989                 unsigned long nr_segs, loff_t *ppos)
 990 {
 991         struct file *filp = iocb->ki_filp;
 992         ssize_t retval;
 993         unsigned long seg;
 994         size_t count;
 995
 996         count = 0;
 997         for (seg = 0; seg < nr_segs; seg++) {
 998                 const struct iovec *iv = &iov[seg];
 999
1000                 /*
1001                  * If any segment has a negative length, or the cumulative
1002                  * length ever wraps negative then return -EINVAL.
1003                  */
1004                 count += iv->iov_len;
1005                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1006                         return -EINVAL;
1007                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1008                         continue;
1009                 if (seg == 0)
1010                         return -EFAULT;
1011                 nr_segs = seg;
1012                 count -= iv->iov_len;   /* This segment is no good */
1013                 break;
1014         }
1015
1016         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1017         if (filp->f_flags & O_DIRECT) {
1018                 loff_t pos = *ppos, size;
1019                 struct address_space *mapping;
1020                 struct inode *inode;
1021
1022                 mapping = filp->f_mapping;
1023                 inode = mapping->host;
1024                 retval = 0;
1025                 if (!count)
1026                         goto out; /* skip atime */
1027                 size = i_size_read(inode);
1028                 if (pos < size) {
1029                         retval = generic_file_direct_IO(READ, iocb,
1030                                                 iov, pos, nr_segs);
1031                         if (retval > 0 && !is_sync_kiocb(iocb))
1032                                 retval = -EIOCBQUEUED;
1033                         if (retval > 0)
1034                                 *ppos = pos + retval;
1035                 }
1036                 file_accessed(filp);
1037                 goto out;
1038         }
1039
1040         retval = 0;
1041         if (count) {
1042                 for (seg = 0; seg < nr_segs; seg++) {
1043                         read_descriptor_t desc;
1044
1045                         desc.written = 0;
1046                         desc.arg.buf = iov[seg].iov_base;
1047                         desc.count = iov[seg].iov_len;
1048                         if (desc.count == 0)
1049                                 continue;
1050                         desc.error = 0;
1051                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1052                         retval += desc.written;
1053                         if (desc.error) {
1054                                 retval = retval ?: desc.error;
1055                                 break;
1056                         }
1057                 }
1058         }
1059 out:
1060         return retval;
1061 }
1062
1063 EXPORT_SYMBOL(__generic_file_aio_read);
1064
1065 ssize_t
1066 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1067 {
1068         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1069
1070         BUG_ON(iocb->ki_pos != pos);
1071         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1072 }
1073
1074 EXPORT_SYMBOL(generic_file_aio_read);
1075
1076 ssize_t
1077 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1078 {
1079         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1080         struct kiocb kiocb;
1081         ssize_t ret;
1082
1083         init_sync_kiocb(&kiocb, filp);
1084         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1085         if (-EIOCBQUEUED == ret)
1086                 ret = wait_on_sync_kiocb(&kiocb);
1087         return ret;
1088 }
1089
1090 EXPORT_SYMBOL(generic_file_read);
1091
1092 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1093 {
1094         ssize_t written;
1095         unsigned long count = desc->count;
1096         struct file *file = desc->arg.data;
1097
1098         if (size > count)
1099                 size = count;
1100
1101         written = file->f_op->sendpage(file, page, offset,
1102                                        size, &file->f_pos, size<count);
1103         if (written < 0) {
1104                 desc->error = written;
1105                 written = 0;
1106         }
1107         desc->count = count - written;
1108         desc->written += written;
1109         return written;
1110 }
1111
1112 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1113                          size_t count, read_actor_t actor, void *target)
1114 {
1115         read_descriptor_t desc;
1116
1117         if (!count)
1118                 return 0;
1119
1120         desc.written = 0;
1121         desc.count = count;
1122         desc.arg.data = target;
1123         desc.error = 0;
1124
1125         do_generic_file_read(in_file, ppos, &desc, actor);
1126         if (desc.written)
1127                 return desc.written;
1128         return desc.error;
1129 }
1130
1131 EXPORT_SYMBOL(generic_file_sendfile);
1132
1133 static ssize_t
1134 do_readahead(struct address_space *mapping, struct file *filp,
1135              unsigned long index, unsigned long nr)
1136 {
1137         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1138                 return -EINVAL;
1139
1140         force_page_cache_readahead(mapping, filp, index,
1141                                         max_sane_readahead(nr));
1142         return 0;
1143 }
1144
1145 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1146 {
1147         ssize_t ret;
1148         struct file *file;
1149
1150         ret = -EBADF;
1151         file = fget(fd);
1152         if (file) {
1153                 if (file->f_mode & FMODE_READ) {
1154                         struct address_space *mapping = file->f_mapping;
1155                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1156                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1157                         unsigned long len = end - start + 1;
1158                         ret = do_readahead(mapping, file, start, len);
1159                 }
1160                 fput(file);
1161         }
1162         return ret;
1163 }
1164
1165 #ifdef CONFIG_MMU
1166 /*
1167  * This adds the requested page to the page cache if it isn't already there,
1168  * and schedules an I/O to read in its contents from disk.
1169  */
1170 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1171 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1172 {
1173         struct address_space *mapping = file->f_mapping;
1174         struct page *page;
1175         int ret;
1176
1177         do {
1178                 page = page_cache_alloc_cold(mapping);
1179                 if (!page)
1180                         return -ENOMEM;
1181
1182                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1183                 if (ret == 0)
1184                         ret = mapping->a_ops->readpage(file, page);
1185                 else if (ret == -EEXIST)
1186                         ret = 0; /* losing race to add is OK */
1187
1188                 page_cache_release(page);
1189
1190         } while (ret == AOP_TRUNCATED_PAGE);
1191
1192         return ret;
1193 }
1194
1195 #define MMAP_LOTSAMISS  (100)
1196
1197 /*
1198  * filemap_nopage() is invoked via the vma operations vector for a
1199  * mapped memory region to read in file data during a page fault.
1200  *
1201  * The goto's are kind of ugly, but this streamlines the normal case of having
1202  * it in the page cache, and handles the special cases reasonably without
1203  * having a lot of duplicated code.
1204  */
1205 struct page *filemap_nopage(struct vm_area_struct *area,
1206                                 unsigned long address, int *type)
1207 {
1208         int error;
1209         struct file *file = area->vm_file;
1210         struct address_space *mapping = file->f_mapping;
1211         struct file_ra_state *ra = &file->f_ra;
1212         struct inode *inode = mapping->host;
1213         struct page *page;
1214         unsigned long size, pgoff;
1215         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1216
1217         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1218
1219 retry_all:
1220         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1221         if (pgoff >= size)
1222                 goto outside_data_content;
1223
1224         /* If we don't want any read-ahead, don't bother */
1225         if (VM_RandomReadHint(area))
1226                 goto no_cached_page;
1227
1228         /*
1229          * The readahead code wants to be told about each and every page
1230          * so it can build and shrink its windows appropriately
1231          *
1232          * For sequential accesses, we use the generic readahead logic.
1233          */
1234         if (VM_SequentialReadHint(area))
1235                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1236
1237         /*
1238          * Do we have something in the page cache already?
1239          */
1240 retry_find:
1241         page = find_get_page(mapping, pgoff);
1242         if (!page) {
1243                 unsigned long ra_pages;
1244
1245                 if (VM_SequentialReadHint(area)) {
1246                         handle_ra_miss(mapping, ra, pgoff);
1247                         goto no_cached_page;
1248                 }
1249                 ra->mmap_miss++;
1250
1251                 /*
1252                  * Do we miss much more than hit in this file? If so,
1253                  * stop bothering with read-ahead. It will only hurt.
1254                  */
1255                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1256                         goto no_cached_page;
1257
1258                 /*
1259                  * To keep the pgmajfault counter straight, we need to
1260                  * check did_readaround, as this is an inner loop.
1261                  */
1262                 if (!did_readaround) {
1263                         majmin = VM_FAULT_MAJOR;
1264                         inc_page_state(pgmajfault);
1265                 }
1266                 did_readaround = 1;
1267                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1268                 if (ra_pages) {
1269                         pgoff_t start = 0;
1270
1271                         if (pgoff > ra_pages / 2)
1272                                 start = pgoff - ra_pages / 2;
1273                         do_page_cache_readahead(mapping, file, start, ra_pages);
1274                 }
1275                 page = find_get_page(mapping, pgoff);
1276                 if (!page)
1277                         goto no_cached_page;
1278         }
1279
1280         if (!did_readaround)
1281                 ra->mmap_hit++;
1282
1283         /*
1284          * Ok, found a page in the page cache, now we need to check
1285          * that it's up-to-date.
1286          */
1287         if (!PageUptodate(page))
1288                 goto page_not_uptodate;
1289
1290 success:
1291         /*
1292          * Found the page and have a reference on it.
1293          */
1294         mark_page_accessed(page);
1295         if (type)
1296                 *type = majmin;
1297         return page;
1298
1299 outside_data_content:
1300         /*
1301          * An external ptracer can access pages that normally aren't
1302          * accessible..
1303          */
1304         if (area->vm_mm == current->mm)
1305                 return NULL;
1306         /* Fall through to the non-read-ahead case */
1307 no_cached_page:
1308         /*
1309          * We're only likely to ever get here if MADV_RANDOM is in
1310          * effect.
1311          */
1312         error = page_cache_read(file, pgoff);
1313         grab_swap_token();
1314
1315         /*
1316          * The page we want has now been added to the page cache.
1317          * In the unlikely event that someone removed it in the
1318          * meantime, we'll just come back here and read it again.
1319          */
1320         if (error >= 0)
1321                 goto retry_find;
1322
1323         /*
1324          * An error return from page_cache_read can result if the
1325          * system is low on memory, or a problem occurs while trying
1326          * to schedule I/O.
1327          */
1328         if (error == -ENOMEM)
1329                 return NOPAGE_OOM;
1330         return NULL;
1331
1332 page_not_uptodate:
1333         if (!did_readaround) {
1334                 majmin = VM_FAULT_MAJOR;
1335                 inc_page_state(pgmajfault);
1336         }
1337         lock_page(page);
1338
1339         /* Did it get unhashed while we waited for it? */
1340         if (!page->mapping) {
1341                 unlock_page(page);
1342                 page_cache_release(page);
1343                 goto retry_all;
1344         }
1345
1346         /* Did somebody else get it up-to-date? */
1347         if (PageUptodate(page)) {
1348                 unlock_page(page);
1349                 goto success;
1350         }
1351
1352         error = mapping->a_ops->readpage(file, page);
1353         if (!error) {
1354                 wait_on_page_locked(page);
1355                 if (PageUptodate(page))
1356                         goto success;
1357         } else if (error == AOP_TRUNCATED_PAGE) {
1358                 page_cache_release(page);
1359                 goto retry_find;
1360         }
1361
1362         /*
1363          * Umm, take care of errors if the page isn't up-to-date.
1364          * Try to re-read it _once_. We do this synchronously,
1365          * because there really aren't any performance issues here
1366          * and we need to check for errors.
1367          */
1368         lock_page(page);
1369
1370         /* Somebody truncated the page on us? */
1371         if (!page->mapping) {
1372                 unlock_page(page);
1373                 page_cache_release(page);
1374                 goto retry_all;
1375         }
1376
1377         /* Somebody else successfully read it in? */
1378         if (PageUptodate(page)) {
1379                 unlock_page(page);
1380                 goto success;
1381         }
1382         ClearPageError(page);
1383         error = mapping->a_ops->readpage(file, page);
1384         if (!error) {
1385                 wait_on_page_locked(page);
1386                 if (PageUptodate(page))
1387                         goto success;
1388         } else if (error == AOP_TRUNCATED_PAGE) {
1389                 page_cache_release(page);
1390                 goto retry_find;
1391         }
1392
1393         /*
1394          * Things didn't work out. Return zero to tell the
1395          * mm layer so, possibly freeing the page cache page first.
1396          */
1397         page_cache_release(page);
1398         return NULL;
1399 }
1400
1401 EXPORT_SYMBOL(filemap_nopage);
1402
1403 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1404                                         int nonblock)
1405 {
1406         struct address_space *mapping = file->f_mapping;
1407         struct page *page;
1408         int error;
1409
1410         /*
1411          * Do we have something in the page cache already?
1412          */
1413 retry_find:
1414         page = find_get_page(mapping, pgoff);
1415         if (!page) {
1416                 if (nonblock)
1417                         return NULL;
1418                 goto no_cached_page;
1419         }
1420
1421         /*
1422          * Ok, found a page in the page cache, now we need to check
1423          * that it's up-to-date.
1424          */
1425         if (!PageUptodate(page)) {
1426                 if (nonblock) {
1427                         page_cache_release(page);
1428                         return NULL;
1429                 }
1430                 goto page_not_uptodate;
1431         }
1432
1433 success:
1434         /*
1435          * Found the page and have a reference on it.
1436          */
1437         mark_page_accessed(page);
1438         return page;
1439
1440 no_cached_page:
1441         error = page_cache_read(file, pgoff);
1442
1443         /*
1444          * The page we want has now been added to the page cache.
1445          * In the unlikely event that someone removed it in the
1446          * meantime, we'll just come back here and read it again.
1447          */
1448         if (error >= 0)
1449                 goto retry_find;
1450
1451         /*
1452          * An error return from page_cache_read can result if the
1453          * system is low on memory, or a problem occurs while trying
1454          * to schedule I/O.
1455          */
1456         return NULL;
1457
1458 page_not_uptodate:
1459         lock_page(page);
1460
1461         /* Did it get unhashed while we waited for it? */
1462         if (!page->mapping) {
1463                 unlock_page(page);
1464                 goto err;
1465         }
1466
1467         /* Did somebody else get it up-to-date? */
1468         if (PageUptodate(page)) {
1469                 unlock_page(page);
1470                 goto success;
1471         }
1472
1473         error = mapping->a_ops->readpage(file, page);
1474         if (!error) {
1475                 wait_on_page_locked(page);
1476                 if (PageUptodate(page))
1477                         goto success;
1478         } else if (error == AOP_TRUNCATED_PAGE) {
1479                 page_cache_release(page);
1480                 goto retry_find;
1481         }
1482
1483         /*
1484          * Umm, take care of errors if the page isn't up-to-date.
1485          * Try to re-read it _once_. We do this synchronously,
1486          * because there really aren't any performance issues here
1487          * and we need to check for errors.
1488          */
1489         lock_page(page);
1490
1491         /* Somebody truncated the page on us? */
1492         if (!page->mapping) {
1493                 unlock_page(page);
1494                 goto err;
1495         }
1496         /* Somebody else successfully read it in? */
1497         if (PageUptodate(page)) {
1498                 unlock_page(page);
1499                 goto success;
1500         }
1501
1502         ClearPageError(page);
1503         error = mapping->a_ops->readpage(file, page);
1504         if (!error) {
1505                 wait_on_page_locked(page);
1506                 if (PageUptodate(page))
1507                         goto success;
1508         } else if (error == AOP_TRUNCATED_PAGE) {
1509                 page_cache_release(page);
1510                 goto retry_find;
1511         }
1512
1513         /*
1514          * Things didn't work out. Return zero to tell the
1515          * mm layer so, possibly freeing the page cache page first.
1516          */
1517 err:
1518         page_cache_release(page);
1519
1520         return NULL;
1521 }
1522
1523 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1524                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1525                 int nonblock)
1526 {
1527         struct file *file = vma->vm_file;
1528         struct address_space *mapping = file->f_mapping;
1529         struct inode *inode = mapping->host;
1530         unsigned long size;
1531         struct mm_struct *mm = vma->vm_mm;
1532         struct page *page;
1533         int err;
1534
1535         if (!nonblock)
1536                 force_page_cache_readahead(mapping, vma->vm_file,
1537                                         pgoff, len >> PAGE_CACHE_SHIFT);
1538
1539 repeat:
1540         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1541         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1542                 return -EINVAL;
1543
1544         page = filemap_getpage(file, pgoff, nonblock);
1545
1546         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1547          * done in shmem_populate calling shmem_getpage */
1548         if (!page && !nonblock)
1549                 return -ENOMEM;
1550
1551         if (page) {
1552                 err = install_page(mm, vma, addr, page, prot);
1553                 if (err) {
1554                         page_cache_release(page);
1555                         return err;
1556                 }
1557         } else if (vma->vm_flags & VM_NONLINEAR) {
1558                 /* No page was found just because we can't read it in now (being
1559                  * here implies nonblock != 0), but the page may exist, so set
1560                  * the PTE to fault it in later. */
1561                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1562                 if (err)
1563                         return err;
1564         }
1565
1566         len -= PAGE_SIZE;
1567         addr += PAGE_SIZE;
1568         pgoff++;
1569         if (len)
1570                 goto repeat;
1571
1572         return 0;
1573 }
1574 EXPORT_SYMBOL(filemap_populate);
1575
1576 struct vm_operations_struct generic_file_vm_ops = {
1577         .nopage         = filemap_nopage,
1578         .populate       = filemap_populate,
1579 };
1580
1581 /* This is used for a general mmap of a disk file */
1582
1583 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1584 {
1585         struct address_space *mapping = file->f_mapping;
1586
1587         if (!mapping->a_ops->readpage)
1588                 return -ENOEXEC;
1589         file_accessed(file);
1590         vma->vm_ops = &generic_file_vm_ops;
1591         return 0;
1592 }
1593
1594 /*
1595  * This is for filesystems which do not implement ->writepage.
1596  */
1597 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1598 {
1599         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1600                 return -EINVAL;
1601         return generic_file_mmap(file, vma);
1602 }
1603 #else
1604 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1605 {
1606         return -ENOSYS;
1607 }
1608 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1609 {
1610         return -ENOSYS;
1611 }
1612 #endif /* CONFIG_MMU */
1613
1614 EXPORT_SYMBOL(generic_file_mmap);
1615 EXPORT_SYMBOL(generic_file_readonly_mmap);
1616
1617 static inline struct page *__read_cache_page(struct address_space *mapping,
1618                                 unsigned long index,
1619                                 int (*filler)(void *,struct page*),
1620                                 void *data)
1621 {
1622         struct page *page, *cached_page = NULL;
1623         int err;
1624 repeat:
1625         page = find_get_page(mapping, index);
1626         if (!page) {
1627                 if (!cached_page) {
1628                         cached_page = page_cache_alloc_cold(mapping);
1629                         if (!cached_page)
1630                                 return ERR_PTR(-ENOMEM);
1631                 }
1632                 err = add_to_page_cache_lru(cached_page, mapping,
1633                                         index, GFP_KERNEL);
1634                 if (err == -EEXIST)
1635                         goto repeat;
1636                 if (err < 0) {
1637                         /* Presumably ENOMEM for radix tree node */
1638                         page_cache_release(cached_page);
1639                         return ERR_PTR(err);
1640                 }
1641                 page = cached_page;
1642                 cached_page = NULL;
1643                 err = filler(data, page);
1644                 if (err < 0) {
1645                         page_cache_release(page);
1646                         page = ERR_PTR(err);
1647                 }
1648         }
1649         if (cached_page)
1650                 page_cache_release(cached_page);
1651         return page;
1652 }
1653
1654 /*
1655  * Read into the page cache. If a page already exists,
1656  * and PageUptodate() is not set, try to fill the page.
1657  */
1658 struct page *read_cache_page(struct address_space *mapping,
1659                                 unsigned long index,
1660                                 int (*filler)(void *,struct page*),
1661                                 void *data)
1662 {
1663         struct page *page;
1664         int err;
1665
1666 retry:
1667         page = __read_cache_page(mapping, index, filler, data);
1668         if (IS_ERR(page))
1669                 goto out;
1670         mark_page_accessed(page);
1671         if (PageUptodate(page))
1672                 goto out;
1673
1674         lock_page(page);
1675         if (!page->mapping) {
1676                 unlock_page(page);
1677                 page_cache_release(page);
1678                 goto retry;
1679         }
1680         if (PageUptodate(page)) {
1681                 unlock_page(page);
1682                 goto out;
1683         }
1684         err = filler(data, page);
1685         if (err < 0) {
1686                 page_cache_release(page);
1687                 page = ERR_PTR(err);
1688         }
1689  out:
1690         return page;
1691 }
1692
1693 EXPORT_SYMBOL(read_cache_page);
1694
1695 /*
1696  * If the page was newly created, increment its refcount and add it to the
1697  * caller's lru-buffering pagevec.  This function is specifically for
1698  * generic_file_write().
1699  */
1700 static inline struct page *
1701 __grab_cache_page(struct address_space *mapping, unsigned long index,
1702                         struct page **cached_page, struct pagevec *lru_pvec)
1703 {
1704         int err;
1705         struct page *page;
1706 repeat:
1707         page = find_lock_page(mapping, index);
1708         if (!page) {
1709                 if (!*cached_page) {
1710                         *cached_page = page_cache_alloc(mapping);
1711                         if (!*cached_page)
1712                                 return NULL;
1713                 }
1714                 err = add_to_page_cache(*cached_page, mapping,
1715                                         index, GFP_KERNEL);
1716                 if (err == -EEXIST)
1717                         goto repeat;
1718                 if (err == 0) {
1719                         page = *cached_page;
1720                         page_cache_get(page);
1721                         if (!pagevec_add(lru_pvec, page))
1722                                 __pagevec_lru_add(lru_pvec);
1723                         *cached_page = NULL;
1724                 }
1725         }
1726         return page;
1727 }
1728
1729 /*
1730  * The logic we want is
1731  *
1732  *      if suid or (sgid and xgrp)
1733  *              remove privs
1734  */
1735 int remove_suid(struct dentry *dentry)
1736 {
1737         mode_t mode = dentry->d_inode->i_mode;
1738         int kill = 0;
1739         int result = 0;
1740
1741         /* suid always must be killed */
1742         if (unlikely(mode & S_ISUID))
1743                 kill = ATTR_KILL_SUID;
1744
1745         /*
1746          * sgid without any exec bits is just a mandatory locking mark; leave
1747          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1748          */
1749         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1750                 kill |= ATTR_KILL_SGID;
1751
1752         if (unlikely(kill && !capable(CAP_FSETID))) {
1753                 struct iattr newattrs;
1754
1755                 newattrs.ia_valid = ATTR_FORCE | kill;
1756                 result = notify_change(dentry, &newattrs);
1757         }
1758         return result;
1759 }
1760 EXPORT_SYMBOL(remove_suid);
1761
1762 size_t
1763 __filemap_copy_from_user_iovec(char *vaddr,
1764                         const struct iovec *iov, size_t base, size_t bytes)
1765 {
1766         size_t copied = 0, left = 0;
1767
1768         while (bytes) {
1769                 char __user *buf = iov->iov_base + base;
1770                 int copy = min(bytes, iov->iov_len - base);
1771
1772                 base = 0;
1773                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1774                 copied += copy;
1775                 bytes -= copy;
1776                 vaddr += copy;
1777                 iov++;
1778
1779                 if (unlikely(left)) {
1780                         /* zero the rest of the target like __copy_from_user */
1781                         if (bytes)
1782                                 memset(vaddr, 0, bytes);
1783                         break;
1784                 }
1785         }
1786         return copied - left;
1787 }
1788
1789 /*
1790  * Performs necessary checks before doing a write
1791  *
1792  * Can adjust writing position aor amount of bytes to write.
1793  * Returns appropriate error code that caller should return or
1794  * zero in case that write should be allowed.
1795  */
1796 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1797 {
1798         struct inode *inode = file->f_mapping->host;
1799         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1800
1801         if (unlikely(*pos < 0))
1802                 return -EINVAL;
1803
1804         if (!isblk) {
1805                 /* FIXME: this is for backwards compatibility with 2.4 */
1806                 if (file->f_flags & O_APPEND)
1807                         *pos = i_size_read(inode);
1808
1809                 if (limit != RLIM_INFINITY) {
1810                         if (*pos >= limit) {
1811                                 send_sig(SIGXFSZ, current, 0);
1812                                 return -EFBIG;
1813                         }
1814                         if (*count > limit - (typeof(limit))*pos) {
1815                                 *count = limit - (typeof(limit))*pos;
1816                         }
1817                 }
1818         }
1819
1820         /*
1821          * LFS rule
1822          */
1823         if (unlikely(*pos + *count > MAX_NON_LFS &&
1824                                 !(file->f_flags & O_LARGEFILE))) {
1825                 if (*pos >= MAX_NON_LFS) {
1826                         send_sig(SIGXFSZ, current, 0);
1827                         return -EFBIG;
1828                 }
1829                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1830                         *count = MAX_NON_LFS - (unsigned long)*pos;
1831                 }
1832         }
1833
1834         /*
1835          * Are we about to exceed the fs block limit ?
1836          *
1837          * If we have written data it becomes a short write.  If we have
1838          * exceeded without writing data we send a signal and return EFBIG.
1839          * Linus frestrict idea will clean these up nicely..
1840          */
1841         if (likely(!isblk)) {
1842                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1843                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1844                                 send_sig(SIGXFSZ, current, 0);
1845                                 return -EFBIG;
1846                         }
1847                         /* zero-length writes at ->s_maxbytes are OK */
1848                 }
1849
1850                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1851                         *count = inode->i_sb->s_maxbytes - *pos;
1852         } else {
1853                 loff_t isize;
1854                 if (bdev_read_only(I_BDEV(inode)))
1855                         return -EPERM;
1856                 isize = i_size_read(inode);
1857                 if (*pos >= isize) {
1858                         if (*count || *pos > isize)
1859                                 return -ENOSPC;
1860                 }
1861
1862                 if (*pos + *count > isize)
1863                         *count = isize - *pos;
1864         }
1865         return 0;
1866 }
1867 EXPORT_SYMBOL(generic_write_checks);
1868
1869 ssize_t
1870 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1871                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1872                 size_t count, size_t ocount)
1873 {
1874         struct file     *file = iocb->ki_filp;
1875         struct address_space *mapping = file->f_mapping;
1876         struct inode    *inode = mapping->host;
1877         ssize_t         written;
1878
1879         if (count != ocount)
1880                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1881
1882         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1883         if (written > 0) {
1884                 loff_t end = pos + written;
1885                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1886                         i_size_write(inode,  end);
1887                         mark_inode_dirty(inode);
1888                 }
1889                 *ppos = end;
1890         }
1891
1892         /*
1893          * Sync the fs metadata but not the minor inode changes and
1894          * of course not the data as we did direct DMA for the IO.
1895          * i_mutex is held, which protects generic_osync_inode() from
1896          * livelocking.
1897          */
1898         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1899                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1900                 if (err < 0)
1901                         written = err;
1902         }
1903         if (written == count && !is_sync_kiocb(iocb))
1904                 written = -EIOCBQUEUED;
1905         return written;
1906 }
1907 EXPORT_SYMBOL(generic_file_direct_write);
1908
1909 ssize_t
1910 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1911                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1912                 size_t count, ssize_t written)
1913 {
1914         struct file *file = iocb->ki_filp;
1915         struct address_space * mapping = file->f_mapping;
1916         struct address_space_operations *a_ops = mapping->a_ops;
1917         struct inode    *inode = mapping->host;
1918         long            status = 0;
1919         struct page     *page;
1920         struct page     *cached_page = NULL;
1921         size_t          bytes;
1922         struct pagevec  lru_pvec;
1923         const struct iovec *cur_iov = iov; /* current iovec */
1924         size_t          iov_base = 0;      /* offset in the current iovec */
1925         char __user     *buf;
1926
1927         pagevec_init(&lru_pvec, 0);
1928
1929         /*
1930          * handle partial DIO write.  Adjust cur_iov if needed.
1931          */
1932         if (likely(nr_segs == 1))
1933                 buf = iov->iov_base + written;
1934         else {
1935                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1936                 buf = cur_iov->iov_base + iov_base;
1937         }
1938
1939         do {
1940                 unsigned long index;
1941                 unsigned long offset;
1942                 unsigned long maxlen;
1943                 size_t copied;
1944
1945                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1946                 index = pos >> PAGE_CACHE_SHIFT;
1947                 bytes = PAGE_CACHE_SIZE - offset;
1948                 if (bytes > count)
1949                         bytes = count;
1950
1951                 /*
1952                  * Bring in the user page that we will copy from _first_.
1953                  * Otherwise there's a nasty deadlock on copying from the
1954                  * same page as we're writing to, without it being marked
1955                  * up-to-date.
1956                  */
1957                 maxlen = cur_iov->iov_len - iov_base;
1958                 if (maxlen > bytes)
1959                         maxlen = bytes;
1960                 fault_in_pages_readable(buf, maxlen);
1961
1962                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1963                 if (!page) {
1964                         status = -ENOMEM;
1965                         break;
1966                 }
1967
1968                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1969                 if (unlikely(status)) {
1970                         loff_t isize = i_size_read(inode);
1971
1972                         if (status != AOP_TRUNCATED_PAGE)
1973                                 unlock_page(page);
1974                         page_cache_release(page);
1975                         if (status == AOP_TRUNCATED_PAGE)
1976                                 continue;
1977                         /*
1978                          * prepare_write() may have instantiated a few blocks
1979                          * outside i_size.  Trim these off again.
1980                          */
1981                         if (pos + bytes > isize)
1982                                 vmtruncate(inode, isize);
1983                         break;
1984                 }
1985                 if (likely(nr_segs == 1))
1986                         copied = filemap_copy_from_user(page, offset,
1987                                                         buf, bytes);
1988                 else
1989                         copied = filemap_copy_from_user_iovec(page, offset,
1990                                                 cur_iov, iov_base, bytes);
1991                 flush_dcache_page(page);
1992                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1993                 if (status == AOP_TRUNCATED_PAGE) {
1994                         page_cache_release(page);
1995                         continue;
1996                 }
1997                 if (likely(copied > 0)) {
1998                         if (!status)
1999                                 status = copied;
2000
2001                         if (status >= 0) {
2002                                 written += status;
2003                                 count -= status;
2004                                 pos += status;
2005                                 buf += status;
2006                                 if (unlikely(nr_segs > 1)) {
2007                                         filemap_set_next_iovec(&cur_iov,
2008                                                         &iov_base, status);
2009                                         if (count)
2010                                                 buf = cur_iov->iov_base +
2011                                                         iov_base;
2012                                 } else {
2013                                         iov_base += status;
2014                                 }
2015                         }
2016                 }
2017                 if (unlikely(copied != bytes))
2018                         if (status >= 0)
2019                                 status = -EFAULT;
2020                 unlock_page(page);
2021                 mark_page_accessed(page);
2022                 page_cache_release(page);
2023                 if (status < 0)
2024                         break;
2025                 balance_dirty_pages_ratelimited(mapping);
2026                 cond_resched();
2027         } while (count);
2028         *ppos = pos;
2029
2030         if (cached_page)
2031                 page_cache_release(cached_page);
2032
2033         /*
2034          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2035          */
2036         if (likely(status >= 0)) {
2037                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2038                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2039                                 status = generic_osync_inode(inode, mapping,
2040                                                 OSYNC_METADATA|OSYNC_DATA);
2041                 }
2042         }
2043
2044         /*
2045          * If we get here for O_DIRECT writes then we must have fallen through
2046          * to buffered writes (block instantiation inside i_size).  So we sync
2047          * the file data here, to try to honour O_DIRECT expectations.
2048          */
2049         if (unlikely(file->f_flags & O_DIRECT) && written)
2050                 status = filemap_write_and_wait(mapping);
2051
2052         pagevec_lru_add(&lru_pvec);
2053         return written ? written : status;
2054 }
2055 EXPORT_SYMBOL(generic_file_buffered_write);
2056
2057 static ssize_t
2058 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2059                                 unsigned long nr_segs, loff_t *ppos)
2060 {
2061         struct file *file = iocb->ki_filp;
2062         struct address_space * mapping = file->f_mapping;
2063         size_t ocount;          /* original count */
2064         size_t count;           /* after file limit checks */
2065         struct inode    *inode = mapping->host;
2066         unsigned long   seg;
2067         loff_t          pos;
2068         ssize_t         written;
2069         ssize_t         err;
2070
2071         ocount = 0;
2072         for (seg = 0; seg < nr_segs; seg++) {
2073                 const struct iovec *iv = &iov[seg];
2074
2075                 /*
2076                  * If any segment has a negative length, or the cumulative
2077                  * length ever wraps negative then return -EINVAL.
2078                  */
2079                 ocount += iv->iov_len;
2080                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2081                         return -EINVAL;
2082                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2083                         continue;
2084                 if (seg == 0)
2085                         return -EFAULT;
2086                 nr_segs = seg;
2087                 ocount -= iv->iov_len;  /* This segment is no good */
2088                 break;
2089         }
2090
2091         count = ocount;
2092         pos = *ppos;
2093
2094         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2095
2096         /* We can write back this queue in page reclaim */
2097         current->backing_dev_info = mapping->backing_dev_info;
2098         written = 0;
2099
2100         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2101         if (err)
2102                 goto out;
2103
2104         if (count == 0)
2105                 goto out;
2106
2107         err = remove_suid(file->f_dentry);
2108         if (err)
2109                 goto out;
2110
2111         file_update_time(file);
2112
2113         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2114         if (unlikely(file->f_flags & O_DIRECT)) {
2115                 written = generic_file_direct_write(iocb, iov,
2116                                 &nr_segs, pos, ppos, count, ocount);
2117                 if (written < 0 || written == count)
2118                         goto out;
2119                 /*
2120                  * direct-io write to a hole: fall through to buffered I/O
2121                  * for completing the rest of the request.
2122                  */
2123                 pos += written;
2124                 count -= written;
2125         }
2126
2127         written = generic_file_buffered_write(iocb, iov, nr_segs,
2128                         pos, ppos, count, written);
2129 out:
2130         current->backing_dev_info = NULL;
2131         return written ? written : err;
2132 }
2133 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2134
2135 ssize_t
2136 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2137                                 unsigned long nr_segs, loff_t *ppos)
2138 {
2139         struct file *file = iocb->ki_filp;
2140         struct address_space *mapping = file->f_mapping;
2141         struct inode *inode = mapping->host;
2142         ssize_t ret;
2143         loff_t pos = *ppos;
2144
2145         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2146
2147         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2148                 int err;
2149
2150                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2151                 if (err < 0)
2152                         ret = err;
2153         }
2154         return ret;
2155 }
2156
2157 static 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 (ret == -EIOCBQUEUED)
2167                 ret = wait_on_sync_kiocb(&kiocb);
2168         return ret;
2169 }
2170
2171 ssize_t
2172 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2173                                 unsigned long nr_segs, loff_t *ppos)
2174 {
2175         struct kiocb kiocb;
2176         ssize_t ret;
2177
2178         init_sync_kiocb(&kiocb, file);
2179         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2180         if (-EIOCBQUEUED == ret)
2181                 ret = wait_on_sync_kiocb(&kiocb);
2182         return ret;
2183 }
2184 EXPORT_SYMBOL(generic_file_write_nolock);
2185
2186 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2187                                size_t count, loff_t pos)
2188 {
2189         struct file *file = iocb->ki_filp;
2190         struct address_space *mapping = file->f_mapping;
2191         struct inode *inode = mapping->host;
2192         ssize_t ret;
2193         struct iovec local_iov = { .iov_base = (void __user *)buf,
2194                                         .iov_len = count };
2195
2196         BUG_ON(iocb->ki_pos != pos);
2197
2198         mutex_lock(&inode->i_mutex);
2199         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2200                                                 &iocb->ki_pos);
2201         mutex_unlock(&inode->i_mutex);
2202
2203         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2204                 ssize_t err;
2205
2206                 err = sync_page_range(inode, mapping, pos, ret);
2207                 if (err < 0)
2208                         ret = err;
2209         }
2210         return ret;
2211 }
2212 EXPORT_SYMBOL(generic_file_aio_write);
2213
2214 ssize_t generic_file_write(struct file *file, const char __user *buf,
2215                            size_t count, loff_t *ppos)
2216 {
2217         struct address_space *mapping = file->f_mapping;
2218         struct inode *inode = mapping->host;
2219         ssize_t ret;
2220         struct iovec local_iov = { .iov_base = (void __user *)buf,
2221                                         .iov_len = count };
2222
2223         mutex_lock(&inode->i_mutex);
2224         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2225         mutex_unlock(&inode->i_mutex);
2226
2227         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2228                 ssize_t err;
2229
2230                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2231                 if (err < 0)
2232                         ret = err;
2233         }
2234         return ret;
2235 }
2236 EXPORT_SYMBOL(generic_file_write);
2237
2238 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2239                         unsigned long nr_segs, loff_t *ppos)
2240 {
2241         struct kiocb kiocb;
2242         ssize_t ret;
2243
2244         init_sync_kiocb(&kiocb, filp);
2245         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2246         if (-EIOCBQUEUED == ret)
2247                 ret = wait_on_sync_kiocb(&kiocb);
2248         return ret;
2249 }
2250 EXPORT_SYMBOL(generic_file_readv);
2251
2252 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2253                         unsigned long nr_segs, loff_t *ppos)
2254 {
2255         struct address_space *mapping = file->f_mapping;
2256         struct inode *inode = mapping->host;
2257         ssize_t ret;
2258
2259         mutex_lock(&inode->i_mutex);
2260         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2261         mutex_unlock(&inode->i_mutex);
2262
2263         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2264                 int err;
2265
2266                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2267                 if (err < 0)
2268                         ret = err;
2269         }
2270         return ret;
2271 }
2272 EXPORT_SYMBOL(generic_file_writev);
2273
2274 /*
2275  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2276  * went wrong during pagecache shootdown.
2277  */
2278 static ssize_t
2279 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2280         loff_t offset, unsigned long nr_segs)
2281 {
2282         struct file *file = iocb->ki_filp;
2283         struct address_space *mapping = file->f_mapping;
2284         ssize_t retval;
2285         size_t write_len = 0;
2286
2287         /*
2288          * If it's a write, unmap all mmappings of the file up-front.  This
2289          * will cause any pte dirty bits to be propagated into the pageframes
2290          * for the subsequent filemap_write_and_wait().
2291          */
2292         if (rw == WRITE) {
2293                 write_len = iov_length(iov, nr_segs);
2294                 if (mapping_mapped(mapping))
2295                         unmap_mapping_range(mapping, offset, write_len, 0);
2296         }
2297
2298         retval = filemap_write_and_wait(mapping);
2299         if (retval == 0) {
2300                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2301                                                 offset, nr_segs);
2302                 if (rw == WRITE && mapping->nrpages) {
2303                         pgoff_t end = (offset + write_len - 1)
2304                                                 >> PAGE_CACHE_SHIFT;
2305                         int err = invalidate_inode_pages2_range(mapping,
2306                                         offset >> PAGE_CACHE_SHIFT, end);
2307                         if (err)
2308                                 retval = err;
2309                 }
2310         }
2311         return retval;
2312 }