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/capability.h>
  19 #include <linux/kernel_stat.h>
  20 #include <linux/mm.h>
  21 #include <linux/swap.h>
  22 #include <linux/mman.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/file.h>
  25 #include <linux/uio.h>
  26 #include <linux/hash.h>
  27 #include <linux/writeback.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/blkdev.h>
  30 #include <linux/security.h>
  31 #include <linux/syscalls.h>
  32 #include "filemap.h"
  33 #include "internal.h"
  34
  35 /*
  36  * FIXME: remove all knowledge of the buffer layer from the core VM
  37  */
  38 #include <linux/buffer_head.h> /* for generic_osync_inode */
  39
  40 #include <asm/uaccess.h>
  41 #include <asm/mman.h>
  42
  43 static ssize_t
  44 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
  45         loff_t offset, unsigned long nr_segs);
  46
  47 /*
  48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
  49  * though.
  50  *
  51  * Shared mappings now work. 15.8.1995  Bruno.
  52  *
  53  * finished 'unifying' the page and buffer cache and SMP-threaded the
  54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
  55  *
  56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
  57  */
  58
  59 /*
  60  * Lock ordering:
  61  *
  62  *  ->i_mmap_lock               (vmtruncate)
  63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
  64  *      ->swap_lock             (exclusive_swap_page, others)
  65  *        ->mapping->tree_lock
  66  *
  67  *  ->i_mutex
  68  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
  69  *
  70  *  ->mmap_sem
  71  *    ->i_mmap_lock
  72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
  73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
  74  *
  75  *  ->mmap_sem
  76  *    ->lock_page               (access_process_vm)
  77  *
  78  *  ->mmap_sem
  79  *    ->i_mutex                 (msync)
  80  *
  81  *  ->i_mutex
  82  *    ->i_alloc_sem             (various)
  83  *
  84  *  ->inode_lock
  85  *    ->sb_lock                 (fs/fs-writeback.c)
  86  *    ->mapping->tree_lock      (__sync_single_inode)
  87  *
  88  *  ->i_mmap_lock
  89  *    ->anon_vma.lock           (vma_adjust)
  90  *
  91  *  ->anon_vma.lock
  92  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
  93  *
  94  *  ->page_table_lock or pte_lock
  95  *    ->swap_lock               (try_to_unmap_one)
  96  *    ->private_lock            (try_to_unmap_one)
  97  *    ->tree_lock               (try_to_unmap_one)
  98  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
  99  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
 100  *    ->private_lock            (page_remove_rmap->set_page_dirty)
 101  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
 102  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
 103  *    ->inode_lock              (zap_pte_range->set_page_dirty)
 104  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
 105  *
 106  *  ->task->proc_lock
 107  *    ->dcache_lock             (proc_pid_lookup)
 108  */
 109
 110 /*
 111  * Remove a page from the page cache and free it. Caller has to make
 112  * sure the page is locked and that nobody else uses it - or that usage
 113  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
 114  */
 115 void __remove_from_page_cache(struct page *page)
 116 {
 117         struct address_space *mapping = page->mapping;
 118
 119         radix_tree_delete(&mapping->page_tree, page->index);
 120         page->mapping = NULL;
 121         mapping->nrpages--;
 122         pagecache_acct(-1);
 123 }
 124
 125 void remove_from_page_cache(struct page *page)
 126 {
 127         struct address_space *mapping = page->mapping;
 128
 129         BUG_ON(!PageLocked(page));
 130
 131         write_lock_irq(&mapping->tree_lock);
 132         __remove_from_page_cache(page);
 133         write_unlock_irq(&mapping->tree_lock);
 134 }
 135
 136 static int sync_page(void *word)
 137 {
 138         struct address_space *mapping;
 139         struct page *page;
 140
 141         page = container_of((unsigned long *)word, struct page, flags);
 142
 143         /*
 144          * page_mapping() is being called without PG_locked held.
 145          * Some knowledge of the state and use of the page is used to
 146          * reduce the requirements down to a memory barrier.
 147          * The danger here is of a stale page_mapping() return value
 148          * indicating a struct address_space different from the one it's
 149          * associated with when it is associated with one.
 150          * After smp_mb(), it's either the correct page_mapping() for
 151          * the page, or an old page_mapping() and the page's own
 152          * page_mapping() has gone NULL.
 153          * The ->sync_page() address_space operation must tolerate
 154          * page_mapping() going NULL. By an amazing coincidence,
 155          * this comes about because none of the users of the page
 156          * in the ->sync_page() methods make essential use of the
 157          * page_mapping(), merely passing the page down to the backing
 158          * device's unplug functions when it's non-NULL, which in turn
 159          * ignore it for all cases but swap, where only page_private(page) is
 160          * of interest. When page_mapping() does go NULL, the entire
 161          * call stack gracefully ignores the page and returns.
 162          * -- wli
 163          */
 164         smp_mb();
 165         mapping = page_mapping(page);
 166         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 167                 mapping->a_ops->sync_page(page);
 168         io_schedule();
 169         return 0;
 170 }
 171
 172 /**
 173  * filemap_fdatawrite_range - start writeback against all of a mapping's
 174  * dirty pages that lie within the byte offsets <start, end>
 175  * @mapping:    address space structure to write
 176  * @start:      offset in bytes where the range starts
 177  * @end:        offset in bytes where the range ends
 178  * @sync_mode:  enable synchronous operation
 179  *
 180  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 181  * opposed to a regular memory * cleansing writeback.  The difference between
 182  * these two operations is that if a dirty page/buffer is encountered, it must
 183  * be waited upon, and not just skipped over.
 184  */
 185 static int __filemap_fdatawrite_range(struct address_space *mapping,
 186         loff_t start, loff_t end, int sync_mode)
 187 {
 188         int ret;
 189         struct writeback_control wbc = {
 190                 .sync_mode = sync_mode,
 191                 .nr_to_write = mapping->nrpages * 2,
 192                 .start = start,
 193                 .end = end,
 194         };
 195
 196         if (!mapping_cap_writeback_dirty(mapping))
 197                 return 0;
 198
 199         ret = do_writepages(mapping, &wbc);
 200         return ret;
 201 }
 202
 203 static inline int __filemap_fdatawrite(struct address_space *mapping,
 204         int sync_mode)
 205 {
 206         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
 207 }
 208
 209 int filemap_fdatawrite(struct address_space *mapping)
 210 {
 211         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 212 }
 213 EXPORT_SYMBOL(filemap_fdatawrite);
 214
 215 static int filemap_fdatawrite_range(struct address_space *mapping,
 216         loff_t start, loff_t end)
 217 {
 218         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
 219 }
 220
 221 /*
 222  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 223  * purposes - I/O may not be started against all dirty pages.
 224  */
 225 int filemap_flush(struct address_space *mapping)
 226 {
 227         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 228 }
 229 EXPORT_SYMBOL(filemap_flush);
 230
 231 /*
 232  * Wait for writeback to complete against pages indexed by start->end
 233  * inclusive
 234  */
 235 static int wait_on_page_writeback_range(struct address_space *mapping,
 236                                 pgoff_t start, pgoff_t end)
 237 {
 238         struct pagevec pvec;
 239         int nr_pages;
 240         int ret = 0;
 241         pgoff_t index;
 242
 243         if (end < start)
 244                 return 0;
 245
 246         pagevec_init(&pvec, 0);
 247         index = start;
 248         while ((index <= end) &&
 249                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 250                         PAGECACHE_TAG_WRITEBACK,
 251                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 252                 unsigned i;
 253
 254                 for (i = 0; i < nr_pages; i++) {
 255                         struct page *page = pvec.pages[i];
 256
 257                         /* until radix tree lookup accepts end_index */
 258                         if (page->index > end)
 259                                 continue;
 260
 261                         wait_on_page_writeback(page);
 262                         if (PageError(page))
 263                                 ret = -EIO;
 264                 }
 265                 pagevec_release(&pvec);
 266                 cond_resched();
 267         }
 268
 269         /* Check for outstanding write errors */
 270         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 271                 ret = -ENOSPC;
 272         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 273                 ret = -EIO;
 274
 275         return ret;
 276 }
 277
 278 /*
 279  * Write and wait upon all the pages in the passed range.  This is a "data
 280  * integrity" operation.  It waits upon in-flight writeout before starting and
 281  * waiting upon new writeout.  If there was an IO error, return it.
 282  *
 283  * We need to re-take i_mutex during the generic_osync_inode list walk because
 284  * it is otherwise livelockable.
 285  */
 286 int sync_page_range(struct inode *inode, struct address_space *mapping,
 287                         loff_t pos, loff_t count)
 288 {
 289         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 290         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 291         int ret;
 292
 293         if (!mapping_cap_writeback_dirty(mapping) || !count)
 294                 return 0;
 295         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 296         if (ret == 0) {
 297                 mutex_lock(&inode->i_mutex);
 298                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 299                 mutex_unlock(&inode->i_mutex);
 300         }
 301         if (ret == 0)
 302                 ret = wait_on_page_writeback_range(mapping, start, end);
 303         return ret;
 304 }
 305 EXPORT_SYMBOL(sync_page_range);
 306
 307 /*
 308  * Note: Holding i_mutex across sync_page_range_nolock is not a good idea
 309  * as it forces O_SYNC writers to different parts of the same file
 310  * to be serialised right until io completion.
 311  */
 312 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
 313                            loff_t pos, loff_t count)
 314 {
 315         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 316         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 317         int ret;
 318
 319         if (!mapping_cap_writeback_dirty(mapping) || !count)
 320                 return 0;
 321         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 322         if (ret == 0)
 323                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 324         if (ret == 0)
 325                 ret = wait_on_page_writeback_range(mapping, start, end);
 326         return ret;
 327 }
 328 EXPORT_SYMBOL(sync_page_range_nolock);
 329
 330 /**
 331  * filemap_fdatawait - walk the list of under-writeback pages of the given
 332  *     address space and wait for all of them.
 333  *
 334  * @mapping: address space structure to wait for
 335  */
 336 int filemap_fdatawait(struct address_space *mapping)
 337 {
 338         loff_t i_size = i_size_read(mapping->host);
 339
 340         if (i_size == 0)
 341                 return 0;
 342
 343         return wait_on_page_writeback_range(mapping, 0,
 344                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 345 }
 346 EXPORT_SYMBOL(filemap_fdatawait);
 347
 348 int filemap_write_and_wait(struct address_space *mapping)
 349 {
 350         int err = 0;
 351
 352         if (mapping->nrpages) {
 353                 err = filemap_fdatawrite(mapping);
 354                 /*
 355                  * Even if the above returned error, the pages may be
 356                  * written partially (e.g. -ENOSPC), so we wait for it.
 357                  * But the -EIO is special case, it may indicate the worst
 358                  * thing (e.g. bug) happened, so we avoid waiting for it.
 359                  */
 360                 if (err != -EIO) {
 361                         int err2 = filemap_fdatawait(mapping);
 362                         if (!err)
 363                                 err = err2;
 364                 }
 365         }
 366         return err;
 367 }
 368 EXPORT_SYMBOL(filemap_write_and_wait);
 369
 370 int filemap_write_and_wait_range(struct address_space *mapping,
 371                                  loff_t lstart, loff_t lend)
 372 {
 373         int err = 0;
 374
 375         if (mapping->nrpages) {
 376                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
 377                                                  WB_SYNC_ALL);
 378                 /* See comment of filemap_write_and_wait() */
 379                 if (err != -EIO) {
 380                         int err2 = wait_on_page_writeback_range(mapping,
 381                                                 lstart >> PAGE_CACHE_SHIFT,
 382                                                 lend >> PAGE_CACHE_SHIFT);
 383                         if (!err)
 384                                 err = err2;
 385                 }
 386         }
 387         return err;
 388 }
 389
 390 /*
 391  * This function is used to add newly allocated pagecache pages:
 392  * the page is new, so we can just run SetPageLocked() against it.
 393  * The other page state flags were set by rmqueue().
 394  *
 395  * This function does not add the page to the LRU.  The caller must do that.
 396  */
 397 int add_to_page_cache(struct page *page, struct address_space *mapping,
 398                 pgoff_t offset, gfp_t gfp_mask)
 399 {
 400         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 401
 402         if (error == 0) {
 403                 write_lock_irq(&mapping->tree_lock);
 404                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 405                 if (!error) {
 406                         page_cache_get(page);
 407                         SetPageLocked(page);
 408                         page->mapping = mapping;
 409                         page->index = offset;
 410                         mapping->nrpages++;
 411                         pagecache_acct(1);
 412                 }
 413                 write_unlock_irq(&mapping->tree_lock);
 414                 radix_tree_preload_end();
 415         }
 416         return error;
 417 }
 418
 419 EXPORT_SYMBOL(add_to_page_cache);
 420
 421 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 422                                 pgoff_t offset, gfp_t gfp_mask)
 423 {
 424         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 425         if (ret == 0)
 426                 lru_cache_add(page);
 427         return ret;
 428 }
 429
 430 /*
 431  * In order to wait for pages to become available there must be
 432  * waitqueues associated with pages. By using a hash table of
 433  * waitqueues where the bucket discipline is to maintain all
 434  * waiters on the same queue and wake all when any of the pages
 435  * become available, and for the woken contexts to check to be
 436  * sure the appropriate page became available, this saves space
 437  * at a cost of "thundering herd" phenomena during rare hash
 438  * collisions.
 439  */
 440 static wait_queue_head_t *page_waitqueue(struct page *page)
 441 {
 442         const struct zone *zone = page_zone(page);
 443
 444         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 445 }
 446
 447 static inline void wake_up_page(struct page *page, int bit)
 448 {
 449         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 450 }
 451
 452 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 453 {
 454         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 455
 456         if (test_bit(bit_nr, &page->flags))
 457                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 458                                                         TASK_UNINTERRUPTIBLE);
 459 }
 460 EXPORT_SYMBOL(wait_on_page_bit);
 461
 462 /**
 463  * unlock_page() - unlock a locked page
 464  *
 465  * @page: the page
 466  *
 467  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 468  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 469  * mechananism between PageLocked pages and PageWriteback pages is shared.
 470  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 471  *
 472  * The first mb is necessary to safely close the critical section opened by the
 473  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 474  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 475  * parallel wait_on_page_locked()).
 476  */
 477 void fastcall unlock_page(struct page *page)
 478 {
 479         smp_mb__before_clear_bit();
 480         if (!TestClearPageLocked(page))
 481                 BUG();
 482         smp_mb__after_clear_bit();
 483         wake_up_page(page, PG_locked);
 484 }
 485 EXPORT_SYMBOL(unlock_page);
 486
 487 /*
 488  * End writeback against a page.
 489  */
 490 void end_page_writeback(struct page *page)
 491 {
 492         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 493                 if (!test_clear_page_writeback(page))
 494                         BUG();
 495         }
 496         smp_mb__after_clear_bit();
 497         wake_up_page(page, PG_writeback);
 498 }
 499 EXPORT_SYMBOL(end_page_writeback);
 500
 501 /*
 502  * Get a lock on the page, assuming we need to sleep to get it.
 503  *
 504  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 505  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 506  * chances are that on the second loop, the block layer's plug list is empty,
 507  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 508  */
 509 void fastcall __lock_page(struct page *page)
 510 {
 511         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 512
 513         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 514                                                         TASK_UNINTERRUPTIBLE);
 515 }
 516 EXPORT_SYMBOL(__lock_page);
 517
 518 /*
 519  * a rather lightweight function, finding and getting a reference to a
 520  * hashed page atomically.
 521  */
 522 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 523 {
 524         struct page *page;
 525
 526         read_lock_irq(&mapping->tree_lock);
 527         page = radix_tree_lookup(&mapping->page_tree, offset);
 528         if (page)
 529                 page_cache_get(page);
 530         read_unlock_irq(&mapping->tree_lock);
 531         return page;
 532 }
 533
 534 EXPORT_SYMBOL(find_get_page);
 535
 536 /*
 537  * Same as above, but trylock it instead of incrementing the count.
 538  */
 539 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 540 {
 541         struct page *page;
 542
 543         read_lock_irq(&mapping->tree_lock);
 544         page = radix_tree_lookup(&mapping->page_tree, offset);
 545         if (page && TestSetPageLocked(page))
 546                 page = NULL;
 547         read_unlock_irq(&mapping->tree_lock);
 548         return page;
 549 }
 550
 551 EXPORT_SYMBOL(find_trylock_page);
 552
 553 /**
 554  * find_lock_page - locate, pin and lock a pagecache page
 555  *
 556  * @mapping: the address_space to search
 557  * @offset: the page index
 558  *
 559  * Locates the desired pagecache page, locks it, increments its reference
 560  * count and returns its address.
 561  *
 562  * Returns zero if the page was not present. find_lock_page() may sleep.
 563  */
 564 struct page *find_lock_page(struct address_space *mapping,
 565                                 unsigned long offset)
 566 {
 567         struct page *page;
 568
 569         read_lock_irq(&mapping->tree_lock);
 570 repeat:
 571         page = radix_tree_lookup(&mapping->page_tree, offset);
 572         if (page) {
 573                 page_cache_get(page);
 574                 if (TestSetPageLocked(page)) {
 575                         read_unlock_irq(&mapping->tree_lock);
 576                         __lock_page(page);
 577                         read_lock_irq(&mapping->tree_lock);
 578
 579                         /* Has the page been truncated while we slept? */
 580                         if (unlikely(page->mapping != mapping ||
 581                                      page->index != offset)) {
 582                                 unlock_page(page);
 583                                 page_cache_release(page);
 584                                 goto repeat;
 585                         }
 586                 }
 587         }
 588         read_unlock_irq(&mapping->tree_lock);
 589         return page;
 590 }
 591
 592 EXPORT_SYMBOL(find_lock_page);
 593
 594 /**
 595  * find_or_create_page - locate or add a pagecache page
 596  *
 597  * @mapping: the page's address_space
 598  * @index: the page's index into the mapping
 599  * @gfp_mask: page allocation mode
 600  *
 601  * Locates a page in the pagecache.  If the page is not present, a new page
 602  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 603  * LRU list.  The returned page is locked and has its reference count
 604  * incremented.
 605  *
 606  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 607  * allocation!
 608  *
 609  * find_or_create_page() returns the desired page's address, or zero on
 610  * memory exhaustion.
 611  */
 612 struct page *find_or_create_page(struct address_space *mapping,
 613                 unsigned long index, gfp_t gfp_mask)
 614 {
 615         struct page *page, *cached_page = NULL;
 616         int err;
 617 repeat:
 618         page = find_lock_page(mapping, index);
 619         if (!page) {
 620                 if (!cached_page) {
 621                         cached_page = alloc_page(gfp_mask);
 622                         if (!cached_page)
 623                                 return NULL;
 624                 }
 625                 err = add_to_page_cache_lru(cached_page, mapping,
 626                                         index, gfp_mask);
 627                 if (!err) {
 628                         page = cached_page;
 629                         cached_page = NULL;
 630                 } else if (err == -EEXIST)
 631                         goto repeat;
 632         }
 633         if (cached_page)
 634                 page_cache_release(cached_page);
 635         return page;
 636 }
 637
 638 EXPORT_SYMBOL(find_or_create_page);
 639
 640 /**
 641  * find_get_pages - gang pagecache lookup
 642  * @mapping:    The address_space to search
 643  * @start:      The starting page index
 644  * @nr_pages:   The maximum number of pages
 645  * @pages:      Where the resulting pages are placed
 646  *
 647  * find_get_pages() will search for and return a group of up to
 648  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 649  * find_get_pages() takes a reference against the returned pages.
 650  *
 651  * The search returns a group of mapping-contiguous pages with ascending
 652  * indexes.  There may be holes in the indices due to not-present pages.
 653  *
 654  * find_get_pages() returns the number of pages which were found.
 655  */
 656 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 657                             unsigned int nr_pages, struct page **pages)
 658 {
 659         unsigned int i;
 660         unsigned int ret;
 661
 662         read_lock_irq(&mapping->tree_lock);
 663         ret = radix_tree_gang_lookup(&mapping->page_tree,
 664                                 (void **)pages, start, nr_pages);
 665         for (i = 0; i < ret; i++)
 666                 page_cache_get(pages[i]);
 667         read_unlock_irq(&mapping->tree_lock);
 668         return ret;
 669 }
 670
 671 /*
 672  * Like find_get_pages, except we only return pages which are tagged with
 673  * `tag'.   We update *index to index the next page for the traversal.
 674  */
 675 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 676                         int tag, unsigned int nr_pages, struct page **pages)
 677 {
 678         unsigned int i;
 679         unsigned int ret;
 680
 681         read_lock_irq(&mapping->tree_lock);
 682         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 683                                 (void **)pages, *index, nr_pages, tag);
 684         for (i = 0; i < ret; i++)
 685                 page_cache_get(pages[i]);
 686         if (ret)
 687                 *index = pages[ret - 1]->index + 1;
 688         read_unlock_irq(&mapping->tree_lock);
 689         return ret;
 690 }
 691
 692 /*
 693  * Same as grab_cache_page, but do not wait if the page is unavailable.
 694  * This is intended for speculative data generators, where the data can
 695  * be regenerated if the page couldn't be grabbed.  This routine should
 696  * be safe to call while holding the lock for another page.
 697  *
 698  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 699  * and deadlock against the caller's locked page.
 700  */
 701 struct page *
 702 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 703 {
 704         struct page *page = find_get_page(mapping, index);
 705         gfp_t gfp_mask;
 706
 707         if (page) {
 708                 if (!TestSetPageLocked(page))
 709                         return page;
 710                 page_cache_release(page);
 711                 return NULL;
 712         }
 713         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 714         page = alloc_pages(gfp_mask, 0);
 715         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 716                 page_cache_release(page);
 717                 page = NULL;
 718         }
 719         return page;
 720 }
 721
 722 EXPORT_SYMBOL(grab_cache_page_nowait);
 723
 724 /*
 725  * This is a generic file read routine, and uses the
 726  * mapping->a_ops->readpage() function for the actual low-level
 727  * stuff.
 728  *
 729  * This is really ugly. But the goto's actually try to clarify some
 730  * of the logic when it comes to error handling etc.
 731  *
 732  * Note the struct file* is only passed for the use of readpage.  It may be
 733  * NULL.
 734  */
 735 void do_generic_mapping_read(struct address_space *mapping,
 736                              struct file_ra_state *_ra,
 737                              struct file *filp,
 738                              loff_t *ppos,
 739                              read_descriptor_t *desc,
 740                              read_actor_t actor)
 741 {
 742         struct inode *inode = mapping->host;
 743         unsigned long index;
 744         unsigned long end_index;
 745         unsigned long offset;
 746         unsigned long last_index;
 747         unsigned long next_index;
 748         unsigned long prev_index;
 749         loff_t isize;
 750         struct page *cached_page;
 751         int error;
 752         struct file_ra_state ra = *_ra;
 753
 754         cached_page = NULL;
 755         index = *ppos >> PAGE_CACHE_SHIFT;
 756         next_index = index;
 757         prev_index = ra.prev_page;
 758         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 759         offset = *ppos & ~PAGE_CACHE_MASK;
 760
 761         isize = i_size_read(inode);
 762         if (!isize)
 763                 goto out;
 764
 765         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 766         for (;;) {
 767                 struct page *page;
 768                 unsigned long nr, ret;
 769
 770                 /* nr is the maximum number of bytes to copy from this page */
 771                 nr = PAGE_CACHE_SIZE;
 772                 if (index >= end_index) {
 773                         if (index > end_index)
 774                                 goto out;
 775                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 776                         if (nr <= offset) {
 777                                 goto out;
 778                         }
 779                 }
 780                 nr = nr - offset;
 781
 782                 cond_resched();
 783                 if (index == next_index)
 784                         next_index = page_cache_readahead(mapping, &ra, filp,
 785                                         index, last_index - index);
 786
 787 find_page:
 788                 page = find_get_page(mapping, index);
 789                 if (unlikely(page == NULL)) {
 790                         handle_ra_miss(mapping, &ra, index);
 791                         goto no_cached_page;
 792                 }
 793                 if (!PageUptodate(page))
 794                         goto page_not_up_to_date;
 795 page_ok:
 796
 797                 /* If users can be writing to this page using arbitrary
 798                  * virtual addresses, take care about potential aliasing
 799                  * before reading the page on the kernel side.
 800                  */
 801                 if (mapping_writably_mapped(mapping))
 802                         flush_dcache_page(page);
 803
 804                 /*
 805                  * When (part of) the same page is read multiple times
 806                  * in succession, only mark it as accessed the first time.
 807                  */
 808                 if (prev_index != index)
 809                         mark_page_accessed(page);
 810                 prev_index = index;
 811
 812                 /*
 813                  * Ok, we have the page, and it's up-to-date, so
 814                  * now we can copy it to user space...
 815                  *
 816                  * The actor routine returns how many bytes were actually used..
 817                  * NOTE! This may not be the same as how much of a user buffer
 818                  * we filled up (we may be padding etc), so we can only update
 819                  * "pos" here (the actor routine has to update the user buffer
 820                  * pointers and the remaining count).
 821                  */
 822                 ret = actor(desc, page, offset, nr);
 823                 offset += ret;
 824                 index += offset >> PAGE_CACHE_SHIFT;
 825                 offset &= ~PAGE_CACHE_MASK;
 826
 827                 page_cache_release(page);
 828                 if (ret == nr && desc->count)
 829                         continue;
 830                 goto out;
 831
 832 page_not_up_to_date:
 833                 /* Get exclusive access to the page ... */
 834                 lock_page(page);
 835
 836                 /* Did it get unhashed before we got the lock? */
 837                 if (!page->mapping) {
 838                         unlock_page(page);
 839                         page_cache_release(page);
 840                         continue;
 841                 }
 842
 843                 /* Did somebody else fill it already? */
 844                 if (PageUptodate(page)) {
 845                         unlock_page(page);
 846                         goto page_ok;
 847                 }
 848
 849 readpage:
 850                 /* Start the actual read. The read will unlock the page. */
 851                 error = mapping->a_ops->readpage(filp, page);
 852
 853                 if (unlikely(error)) {
 854                         if (error == AOP_TRUNCATED_PAGE) {
 855                                 page_cache_release(page);
 856                                 goto find_page;
 857                         }
 858                         goto readpage_error;
 859                 }
 860
 861                 if (!PageUptodate(page)) {
 862                         lock_page(page);
 863                         if (!PageUptodate(page)) {
 864                                 if (page->mapping == NULL) {
 865                                         /*
 866                                          * invalidate_inode_pages got it
 867                                          */
 868                                         unlock_page(page);
 869                                         page_cache_release(page);
 870                                         goto find_page;
 871                                 }
 872                                 unlock_page(page);
 873                                 error = -EIO;
 874                                 goto readpage_error;
 875                         }
 876                         unlock_page(page);
 877                 }
 878
 879                 /*
 880                  * i_size must be checked after we have done ->readpage.
 881                  *
 882                  * Checking i_size after the readpage allows us to calculate
 883                  * the correct value for "nr", which means the zero-filled
 884                  * part of the page is not copied back to userspace (unless
 885                  * another truncate extends the file - this is desired though).
 886                  */
 887                 isize = i_size_read(inode);
 888                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 889                 if (unlikely(!isize || index > end_index)) {
 890                         page_cache_release(page);
 891                         goto out;
 892                 }
 893
 894                 /* nr is the maximum number of bytes to copy from this page */
 895                 nr = PAGE_CACHE_SIZE;
 896                 if (index == end_index) {
 897                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 898                         if (nr <= offset) {
 899                                 page_cache_release(page);
 900                                 goto out;
 901                         }
 902                 }
 903                 nr = nr - offset;
 904                 goto page_ok;
 905
 906 readpage_error:
 907                 /* UHHUH! A synchronous read error occurred. Report it */
 908                 desc->error = error;
 909                 page_cache_release(page);
 910                 goto out;
 911
 912 no_cached_page:
 913                 /*
 914                  * Ok, it wasn't cached, so we need to create a new
 915                  * page..
 916                  */
 917                 if (!cached_page) {
 918                         cached_page = page_cache_alloc_cold(mapping);
 919                         if (!cached_page) {
 920                                 desc->error = -ENOMEM;
 921                                 goto out;
 922                         }
 923                 }
 924                 error = add_to_page_cache_lru(cached_page, mapping,
 925                                                 index, GFP_KERNEL);
 926                 if (error) {
 927                         if (error == -EEXIST)
 928                                 goto find_page;
 929                         desc->error = error;
 930                         goto out;
 931                 }
 932                 page = cached_page;
 933                 cached_page = NULL;
 934                 goto readpage;
 935         }
 936
 937 out:
 938         *_ra = ra;
 939
 940         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 941         if (cached_page)
 942                 page_cache_release(cached_page);
 943         if (filp)
 944                 file_accessed(filp);
 945 }
 946
 947 EXPORT_SYMBOL(do_generic_mapping_read);
 948
 949 int file_read_actor(read_descriptor_t *desc, struct page *page,
 950                         unsigned long offset, unsigned long size)
 951 {
 952         char *kaddr;
 953         unsigned long left, count = desc->count;
 954
 955         if (size > count)
 956                 size = count;
 957
 958         /*
 959          * Faults on the destination of a read are common, so do it before
 960          * taking the kmap.
 961          */
 962         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 963                 kaddr = kmap_atomic(page, KM_USER0);
 964                 left = __copy_to_user_inatomic(desc->arg.buf,
 965                                                 kaddr + offset, size);
 966                 kunmap_atomic(kaddr, KM_USER0);
 967                 if (left == 0)
 968                         goto success;
 969         }
 970
 971         /* Do it the slow way */
 972         kaddr = kmap(page);
 973         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 974         kunmap(page);
 975
 976         if (left) {
 977                 size -= left;
 978                 desc->error = -EFAULT;
 979         }
 980 success:
 981         desc->count = count - size;
 982         desc->written += size;
 983         desc->arg.buf += size;
 984         return size;
 985 }
 986
 987 /*
 988  * This is the "read()" routine for all filesystems
 989  * that can use the page cache directly.
 990  */
 991 ssize_t
 992 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 993                 unsigned long nr_segs, loff_t *ppos)
 994 {
 995         struct file *filp = iocb->ki_filp;
 996         ssize_t retval;
 997         unsigned long seg;
 998         size_t count;
 999
1000         count = 0;
1001         for (seg = 0; seg < nr_segs; seg++) {
1002                 const struct iovec *iv = &iov[seg];
1003
1004                 /*
1005                  * If any segment has a negative length, or the cumulative
1006                  * length ever wraps negative then return -EINVAL.
1007                  */
1008                 count += iv->iov_len;
1009                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1010                         return -EINVAL;
1011                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1012                         continue;
1013                 if (seg == 0)
1014                         return -EFAULT;
1015                 nr_segs = seg;
1016                 count -= iv->iov_len;   /* This segment is no good */
1017                 break;
1018         }
1019
1020         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1021         if (filp->f_flags & O_DIRECT) {
1022                 loff_t pos = *ppos, size;
1023                 struct address_space *mapping;
1024                 struct inode *inode;
1025
1026                 mapping = filp->f_mapping;
1027                 inode = mapping->host;
1028                 retval = 0;
1029                 if (!count)
1030                         goto out; /* skip atime */
1031                 size = i_size_read(inode);
1032                 if (pos < size) {
1033                         retval = generic_file_direct_IO(READ, iocb,
1034                                                 iov, pos, nr_segs);
1035                         if (retval > 0 && !is_sync_kiocb(iocb))
1036                                 retval = -EIOCBQUEUED;
1037                         if (retval > 0)
1038                                 *ppos = pos + retval;
1039                 }
1040                 file_accessed(filp);
1041                 goto out;
1042         }
1043
1044         retval = 0;
1045         if (count) {
1046                 for (seg = 0; seg < nr_segs; seg++) {
1047                         read_descriptor_t desc;
1048
1049                         desc.written = 0;
1050                         desc.arg.buf = iov[seg].iov_base;
1051                         desc.count = iov[seg].iov_len;
1052                         if (desc.count == 0)
1053                                 continue;
1054                         desc.error = 0;
1055                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1056                         retval += desc.written;
1057                         if (desc.error) {
1058                                 retval = retval ?: desc.error;
1059                                 break;
1060                         }
1061                 }
1062         }
1063 out:
1064         return retval;
1065 }
1066
1067 EXPORT_SYMBOL(__generic_file_aio_read);
1068
1069 ssize_t
1070 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1071 {
1072         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1073
1074         BUG_ON(iocb->ki_pos != pos);
1075         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1076 }
1077
1078 EXPORT_SYMBOL(generic_file_aio_read);
1079
1080 ssize_t
1081 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1082 {
1083         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1084         struct kiocb kiocb;
1085         ssize_t ret;
1086
1087         init_sync_kiocb(&kiocb, filp);
1088         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1089         if (-EIOCBQUEUED == ret)
1090                 ret = wait_on_sync_kiocb(&kiocb);
1091         return ret;
1092 }
1093
1094 EXPORT_SYMBOL(generic_file_read);
1095
1096 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1097 {
1098         ssize_t written;
1099         unsigned long count = desc->count;
1100         struct file *file = desc->arg.data;
1101
1102         if (size > count)
1103                 size = count;
1104
1105         written = file->f_op->sendpage(file, page, offset,
1106                                        size, &file->f_pos, size<count);
1107         if (written < 0) {
1108                 desc->error = written;
1109                 written = 0;
1110         }
1111         desc->count = count - written;
1112         desc->written += written;
1113         return written;
1114 }
1115
1116 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1117                          size_t count, read_actor_t actor, void *target)
1118 {
1119         read_descriptor_t desc;
1120
1121         if (!count)
1122                 return 0;
1123
1124         desc.written = 0;
1125         desc.count = count;
1126         desc.arg.data = target;
1127         desc.error = 0;
1128
1129         do_generic_file_read(in_file, ppos, &desc, actor);
1130         if (desc.written)
1131                 return desc.written;
1132         return desc.error;
1133 }
1134
1135 EXPORT_SYMBOL(generic_file_sendfile);
1136
1137 static ssize_t
1138 do_readahead(struct address_space *mapping, struct file *filp,
1139              unsigned long index, unsigned long nr)
1140 {
1141         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1142                 return -EINVAL;
1143
1144         force_page_cache_readahead(mapping, filp, index,
1145                                         max_sane_readahead(nr));
1146         return 0;
1147 }
1148
1149 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1150 {
1151         ssize_t ret;
1152         struct file *file;
1153
1154         ret = -EBADF;
1155         file = fget(fd);
1156         if (file) {
1157                 if (file->f_mode & FMODE_READ) {
1158                         struct address_space *mapping = file->f_mapping;
1159                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1160                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1161                         unsigned long len = end - start + 1;
1162                         ret = do_readahead(mapping, file, start, len);
1163                 }
1164                 fput(file);
1165         }
1166         return ret;
1167 }
1168
1169 #ifdef CONFIG_MMU
1170 /*
1171  * This adds the requested page to the page cache if it isn't already there,
1172  * and schedules an I/O to read in its contents from disk.
1173  */
1174 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1175 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1176 {
1177         struct address_space *mapping = file->f_mapping;
1178         struct page *page;
1179         int ret;
1180
1181         do {
1182                 page = page_cache_alloc_cold(mapping);
1183                 if (!page)
1184                         return -ENOMEM;
1185
1186                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1187                 if (ret == 0)
1188                         ret = mapping->a_ops->readpage(file, page);
1189                 else if (ret == -EEXIST)
1190                         ret = 0; /* losing race to add is OK */
1191
1192                 page_cache_release(page);
1193
1194         } while (ret == AOP_TRUNCATED_PAGE);
1195
1196         return ret;
1197 }
1198
1199 #define MMAP_LOTSAMISS  (100)
1200
1201 /*
1202  * filemap_nopage() is invoked via the vma operations vector for a
1203  * mapped memory region to read in file data during a page fault.
1204  *
1205  * The goto's are kind of ugly, but this streamlines the normal case of having
1206  * it in the page cache, and handles the special cases reasonably without
1207  * having a lot of duplicated code.
1208  */
1209 struct page *filemap_nopage(struct vm_area_struct *area,
1210                                 unsigned long address, int *type)
1211 {
1212         int error;
1213         struct file *file = area->vm_file;
1214         struct address_space *mapping = file->f_mapping;
1215         struct file_ra_state *ra = &file->f_ra;
1216         struct inode *inode = mapping->host;
1217         struct page *page;
1218         unsigned long size, pgoff;
1219         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1220
1221         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1222
1223 retry_all:
1224         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1225         if (pgoff >= size)
1226                 goto outside_data_content;
1227
1228         /* If we don't want any read-ahead, don't bother */
1229         if (VM_RandomReadHint(area))
1230                 goto no_cached_page;
1231
1232         /*
1233          * The readahead code wants to be told about each and every page
1234          * so it can build and shrink its windows appropriately
1235          *
1236          * For sequential accesses, we use the generic readahead logic.
1237          */
1238         if (VM_SequentialReadHint(area))
1239                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1240
1241         /*
1242          * Do we have something in the page cache already?
1243          */
1244 retry_find:
1245         page = find_get_page(mapping, pgoff);
1246         if (!page) {
1247                 unsigned long ra_pages;
1248
1249                 if (VM_SequentialReadHint(area)) {
1250                         handle_ra_miss(mapping, ra, pgoff);
1251                         goto no_cached_page;
1252                 }
1253                 ra->mmap_miss++;
1254
1255                 /*
1256                  * Do we miss much more than hit in this file? If so,
1257                  * stop bothering with read-ahead. It will only hurt.
1258                  */
1259                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1260                         goto no_cached_page;
1261
1262                 /*
1263                  * To keep the pgmajfault counter straight, we need to
1264                  * check did_readaround, as this is an inner loop.
1265                  */
1266                 if (!did_readaround) {
1267                         majmin = VM_FAULT_MAJOR;
1268                         inc_page_state(pgmajfault);
1269                 }
1270                 did_readaround = 1;
1271                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1272                 if (ra_pages) {
1273                         pgoff_t start = 0;
1274
1275                         if (pgoff > ra_pages / 2)
1276                                 start = pgoff - ra_pages / 2;
1277                         do_page_cache_readahead(mapping, file, start, ra_pages);
1278                 }
1279                 page = find_get_page(mapping, pgoff);
1280                 if (!page)
1281                         goto no_cached_page;
1282         }
1283
1284         if (!did_readaround)
1285                 ra->mmap_hit++;
1286
1287         /*
1288          * Ok, found a page in the page cache, now we need to check
1289          * that it's up-to-date.
1290          */
1291         if (!PageUptodate(page))
1292                 goto page_not_uptodate;
1293
1294 success:
1295         /*
1296          * Found the page and have a reference on it.
1297          */
1298         mark_page_accessed(page);
1299         if (type)
1300                 *type = majmin;
1301         return page;
1302
1303 outside_data_content:
1304         /*
1305          * An external ptracer can access pages that normally aren't
1306          * accessible..
1307          */
1308         if (area->vm_mm == current->mm)
1309                 return NULL;
1310         /* Fall through to the non-read-ahead case */
1311 no_cached_page:
1312         /*
1313          * We're only likely to ever get here if MADV_RANDOM is in
1314          * effect.
1315          */
1316         error = page_cache_read(file, pgoff);
1317         grab_swap_token();
1318
1319         /*
1320          * The page we want has now been added to the page cache.
1321          * In the unlikely event that someone removed it in the
1322          * meantime, we'll just come back here and read it again.
1323          */
1324         if (error >= 0)
1325                 goto retry_find;
1326
1327         /*
1328          * An error return from page_cache_read can result if the
1329          * system is low on memory, or a problem occurs while trying
1330          * to schedule I/O.
1331          */
1332         if (error == -ENOMEM)
1333                 return NOPAGE_OOM;
1334         return NULL;
1335
1336 page_not_uptodate:
1337         if (!did_readaround) {
1338                 majmin = VM_FAULT_MAJOR;
1339                 inc_page_state(pgmajfault);
1340         }
1341         lock_page(page);
1342
1343         /* Did it get unhashed while we waited for it? */
1344         if (!page->mapping) {
1345                 unlock_page(page);
1346                 page_cache_release(page);
1347                 goto retry_all;
1348         }
1349
1350         /* Did somebody else get it up-to-date? */
1351         if (PageUptodate(page)) {
1352                 unlock_page(page);
1353                 goto success;
1354         }
1355
1356         error = mapping->a_ops->readpage(file, page);
1357         if (!error) {
1358                 wait_on_page_locked(page);
1359                 if (PageUptodate(page))
1360                         goto success;
1361         } else if (error == AOP_TRUNCATED_PAGE) {
1362                 page_cache_release(page);
1363                 goto retry_find;
1364         }
1365
1366         /*
1367          * Umm, take care of errors if the page isn't up-to-date.
1368          * Try to re-read it _once_. We do this synchronously,
1369          * because there really aren't any performance issues here
1370          * and we need to check for errors.
1371          */
1372         lock_page(page);
1373
1374         /* Somebody truncated the page on us? */
1375         if (!page->mapping) {
1376                 unlock_page(page);
1377                 page_cache_release(page);
1378                 goto retry_all;
1379         }
1380
1381         /* Somebody else successfully read it in? */
1382         if (PageUptodate(page)) {
1383                 unlock_page(page);
1384                 goto success;
1385         }
1386         ClearPageError(page);
1387         error = mapping->a_ops->readpage(file, page);
1388         if (!error) {
1389                 wait_on_page_locked(page);
1390                 if (PageUptodate(page))
1391                         goto success;
1392         } else if (error == AOP_TRUNCATED_PAGE) {
1393                 page_cache_release(page);
1394                 goto retry_find;
1395         }
1396
1397         /*
1398          * Things didn't work out. Return zero to tell the
1399          * mm layer so, possibly freeing the page cache page first.
1400          */
1401         page_cache_release(page);
1402         return NULL;
1403 }
1404
1405 EXPORT_SYMBOL(filemap_nopage);
1406
1407 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1408                                         int nonblock)
1409 {
1410         struct address_space *mapping = file->f_mapping;
1411         struct page *page;
1412         int error;
1413
1414         /*
1415          * Do we have something in the page cache already?
1416          */
1417 retry_find:
1418         page = find_get_page(mapping, pgoff);
1419         if (!page) {
1420                 if (nonblock)
1421                         return NULL;
1422                 goto no_cached_page;
1423         }
1424
1425         /*
1426          * Ok, found a page in the page cache, now we need to check
1427          * that it's up-to-date.
1428          */
1429         if (!PageUptodate(page)) {
1430                 if (nonblock) {
1431                         page_cache_release(page);
1432                         return NULL;
1433                 }
1434                 goto page_not_uptodate;
1435         }
1436
1437 success:
1438         /*
1439          * Found the page and have a reference on it.
1440          */
1441         mark_page_accessed(page);
1442         return page;
1443
1444 no_cached_page:
1445         error = page_cache_read(file, pgoff);
1446
1447         /*
1448          * The page we want has now been added to the page cache.
1449          * In the unlikely event that someone removed it in the
1450          * meantime, we'll just come back here and read it again.
1451          */
1452         if (error >= 0)
1453                 goto retry_find;
1454
1455         /*
1456          * An error return from page_cache_read can result if the
1457          * system is low on memory, or a problem occurs while trying
1458          * to schedule I/O.
1459          */
1460         return NULL;
1461
1462 page_not_uptodate:
1463         lock_page(page);
1464
1465         /* Did it get unhashed while we waited for it? */
1466         if (!page->mapping) {
1467                 unlock_page(page);
1468                 goto err;
1469         }
1470
1471         /* Did somebody else get it up-to-date? */
1472         if (PageUptodate(page)) {
1473                 unlock_page(page);
1474                 goto success;
1475         }
1476
1477         error = mapping->a_ops->readpage(file, page);
1478         if (!error) {
1479                 wait_on_page_locked(page);
1480                 if (PageUptodate(page))
1481                         goto success;
1482         } else if (error == AOP_TRUNCATED_PAGE) {
1483                 page_cache_release(page);
1484                 goto retry_find;
1485         }
1486
1487         /*
1488          * Umm, take care of errors if the page isn't up-to-date.
1489          * Try to re-read it _once_. We do this synchronously,
1490          * because there really aren't any performance issues here
1491          * and we need to check for errors.
1492          */
1493         lock_page(page);
1494
1495         /* Somebody truncated the page on us? */
1496         if (!page->mapping) {
1497                 unlock_page(page);
1498                 goto err;
1499         }
1500         /* Somebody else successfully read it in? */
1501         if (PageUptodate(page)) {
1502                 unlock_page(page);
1503                 goto success;
1504         }
1505
1506         ClearPageError(page);
1507         error = mapping->a_ops->readpage(file, page);
1508         if (!error) {
1509                 wait_on_page_locked(page);
1510                 if (PageUptodate(page))
1511                         goto success;
1512         } else if (error == AOP_TRUNCATED_PAGE) {
1513                 page_cache_release(page);
1514                 goto retry_find;
1515         }
1516
1517         /*
1518          * Things didn't work out. Return zero to tell the
1519          * mm layer so, possibly freeing the page cache page first.
1520          */
1521 err:
1522         page_cache_release(page);
1523
1524         return NULL;
1525 }
1526
1527 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1528                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1529                 int nonblock)
1530 {
1531         struct file *file = vma->vm_file;
1532         struct address_space *mapping = file->f_mapping;
1533         struct inode *inode = mapping->host;
1534         unsigned long size;
1535         struct mm_struct *mm = vma->vm_mm;
1536         struct page *page;
1537         int err;
1538
1539         if (!nonblock)
1540                 force_page_cache_readahead(mapping, vma->vm_file,
1541                                         pgoff, len >> PAGE_CACHE_SHIFT);
1542
1543 repeat:
1544         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1545         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1546                 return -EINVAL;
1547
1548         page = filemap_getpage(file, pgoff, nonblock);
1549
1550         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1551          * done in shmem_populate calling shmem_getpage */
1552         if (!page && !nonblock)
1553                 return -ENOMEM;
1554
1555         if (page) {
1556                 err = install_page(mm, vma, addr, page, prot);
1557                 if (err) {
1558                         page_cache_release(page);
1559                         return err;
1560                 }
1561         } else if (vma->vm_flags & VM_NONLINEAR) {
1562                 /* No page was found just because we can't read it in now (being
1563                  * here implies nonblock != 0), but the page may exist, so set
1564                  * the PTE to fault it in later. */
1565                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1566                 if (err)
1567                         return err;
1568         }
1569
1570         len -= PAGE_SIZE;
1571         addr += PAGE_SIZE;
1572         pgoff++;
1573         if (len)
1574                 goto repeat;
1575
1576         return 0;
1577 }
1578 EXPORT_SYMBOL(filemap_populate);
1579
1580 struct vm_operations_struct generic_file_vm_ops = {
1581         .nopage         = filemap_nopage,
1582         .populate       = filemap_populate,
1583 };
1584
1585 /* This is used for a general mmap of a disk file */
1586
1587 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1588 {
1589         struct address_space *mapping = file->f_mapping;
1590
1591         if (!mapping->a_ops->readpage)
1592                 return -ENOEXEC;
1593         file_accessed(file);
1594         vma->vm_ops = &generic_file_vm_ops;
1595         return 0;
1596 }
1597
1598 /*
1599  * This is for filesystems which do not implement ->writepage.
1600  */
1601 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1602 {
1603         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1604                 return -EINVAL;
1605         return generic_file_mmap(file, vma);
1606 }
1607 #else
1608 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1609 {
1610         return -ENOSYS;
1611 }
1612 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1613 {
1614         return -ENOSYS;
1615 }
1616 #endif /* CONFIG_MMU */
1617
1618 EXPORT_SYMBOL(generic_file_mmap);
1619 EXPORT_SYMBOL(generic_file_readonly_mmap);
1620
1621 static inline struct page *__read_cache_page(struct address_space *mapping,
1622                                 unsigned long index,
1623                                 int (*filler)(void *,struct page*),
1624                                 void *data)
1625 {
1626         struct page *page, *cached_page = NULL;
1627         int err;
1628 repeat:
1629         page = find_get_page(mapping, index);
1630         if (!page) {
1631                 if (!cached_page) {
1632                         cached_page = page_cache_alloc_cold(mapping);
1633                         if (!cached_page)
1634                                 return ERR_PTR(-ENOMEM);
1635                 }
1636                 err = add_to_page_cache_lru(cached_page, mapping,
1637                                         index, GFP_KERNEL);
1638                 if (err == -EEXIST)
1639                         goto repeat;
1640                 if (err < 0) {
1641                         /* Presumably ENOMEM for radix tree node */
1642                         page_cache_release(cached_page);
1643                         return ERR_PTR(err);
1644                 }
1645                 page = cached_page;
1646                 cached_page = NULL;
1647                 err = filler(data, page);
1648                 if (err < 0) {
1649                         page_cache_release(page);
1650                         page = ERR_PTR(err);
1651                 }
1652         }
1653         if (cached_page)
1654                 page_cache_release(cached_page);
1655         return page;
1656 }
1657
1658 /*
1659  * Read into the page cache. If a page already exists,
1660  * and PageUptodate() is not set, try to fill the page.
1661  */
1662 struct page *read_cache_page(struct address_space *mapping,
1663                                 unsigned long index,
1664                                 int (*filler)(void *,struct page*),
1665                                 void *data)
1666 {
1667         struct page *page;
1668         int err;
1669
1670 retry:
1671         page = __read_cache_page(mapping, index, filler, data);
1672         if (IS_ERR(page))
1673                 goto out;
1674         mark_page_accessed(page);
1675         if (PageUptodate(page))
1676                 goto out;
1677
1678         lock_page(page);
1679         if (!page->mapping) {
1680                 unlock_page(page);
1681                 page_cache_release(page);
1682                 goto retry;
1683         }
1684         if (PageUptodate(page)) {
1685                 unlock_page(page);
1686                 goto out;
1687         }
1688         err = filler(data, page);
1689         if (err < 0) {
1690                 page_cache_release(page);
1691                 page = ERR_PTR(err);
1692         }
1693  out:
1694         return page;
1695 }
1696
1697 EXPORT_SYMBOL(read_cache_page);
1698
1699 /*
1700  * If the page was newly created, increment its refcount and add it to the
1701  * caller's lru-buffering pagevec.  This function is specifically for
1702  * generic_file_write().
1703  */
1704 static inline struct page *
1705 __grab_cache_page(struct address_space *mapping, unsigned long index,
1706                         struct page **cached_page, struct pagevec *lru_pvec)
1707 {
1708         int err;
1709         struct page *page;
1710 repeat:
1711         page = find_lock_page(mapping, index);
1712         if (!page) {
1713                 if (!*cached_page) {
1714                         *cached_page = page_cache_alloc(mapping);
1715                         if (!*cached_page)
1716                                 return NULL;
1717                 }
1718                 err = add_to_page_cache(*cached_page, mapping,
1719                                         index, GFP_KERNEL);
1720                 if (err == -EEXIST)
1721                         goto repeat;
1722                 if (err == 0) {
1723                         page = *cached_page;
1724                         page_cache_get(page);
1725                         if (!pagevec_add(lru_pvec, page))
1726                                 __pagevec_lru_add(lru_pvec);
1727                         *cached_page = NULL;
1728                 }
1729         }
1730         return page;
1731 }
1732
1733 /*
1734  * The logic we want is
1735  *
1736  *      if suid or (sgid and xgrp)
1737  *              remove privs
1738  */
1739 int remove_suid(struct dentry *dentry)
1740 {
1741         mode_t mode = dentry->d_inode->i_mode;
1742         int kill = 0;
1743         int result = 0;
1744
1745         /* suid always must be killed */
1746         if (unlikely(mode & S_ISUID))
1747                 kill = ATTR_KILL_SUID;
1748
1749         /*
1750          * sgid without any exec bits is just a mandatory locking mark; leave
1751          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1752          */
1753         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1754                 kill |= ATTR_KILL_SGID;
1755
1756         if (unlikely(kill && !capable(CAP_FSETID))) {
1757                 struct iattr newattrs;
1758
1759                 newattrs.ia_valid = ATTR_FORCE | kill;
1760                 result = notify_change(dentry, &newattrs);
1761         }
1762         return result;
1763 }
1764 EXPORT_SYMBOL(remove_suid);
1765
1766 size_t
1767 __filemap_copy_from_user_iovec(char *vaddr,
1768                         const struct iovec *iov, size_t base, size_t bytes)
1769 {
1770         size_t copied = 0, left = 0;
1771
1772         while (bytes) {
1773                 char __user *buf = iov->iov_base + base;
1774                 int copy = min(bytes, iov->iov_len - base);
1775
1776                 base = 0;
1777                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1778                 copied += copy;
1779                 bytes -= copy;
1780                 vaddr += copy;
1781                 iov++;
1782
1783                 if (unlikely(left)) {
1784                         /* zero the rest of the target like __copy_from_user */
1785                         if (bytes)
1786                                 memset(vaddr, 0, bytes);
1787                         break;
1788                 }
1789         }
1790         return copied - left;
1791 }
1792
1793 /*
1794  * Performs necessary checks before doing a write
1795  *
1796  * Can adjust writing position aor amount of bytes to write.
1797  * Returns appropriate error code that caller should return or
1798  * zero in case that write should be allowed.
1799  */
1800 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1801 {
1802         struct inode *inode = file->f_mapping->host;
1803         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1804
1805         if (unlikely(*pos < 0))
1806                 return -EINVAL;
1807
1808         if (!isblk) {
1809                 /* FIXME: this is for backwards compatibility with 2.4 */
1810                 if (file->f_flags & O_APPEND)
1811                         *pos = i_size_read(inode);
1812
1813                 if (limit != RLIM_INFINITY) {
1814                         if (*pos >= limit) {
1815                                 send_sig(SIGXFSZ, current, 0);
1816                                 return -EFBIG;
1817                         }
1818                         if (*count > limit - (typeof(limit))*pos) {
1819                                 *count = limit - (typeof(limit))*pos;
1820                         }
1821                 }
1822         }
1823
1824         /*
1825          * LFS rule
1826          */
1827         if (unlikely(*pos + *count > MAX_NON_LFS &&
1828                                 !(file->f_flags & O_LARGEFILE))) {
1829                 if (*pos >= MAX_NON_LFS) {
1830                         send_sig(SIGXFSZ, current, 0);
1831                         return -EFBIG;
1832                 }
1833                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1834                         *count = MAX_NON_LFS - (unsigned long)*pos;
1835                 }
1836         }
1837
1838         /*
1839          * Are we about to exceed the fs block limit ?
1840          *
1841          * If we have written data it becomes a short write.  If we have
1842          * exceeded without writing data we send a signal and return EFBIG.
1843          * Linus frestrict idea will clean these up nicely..
1844          */
1845         if (likely(!isblk)) {
1846                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1847                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1848                                 send_sig(SIGXFSZ, current, 0);
1849                                 return -EFBIG;
1850                         }
1851                         /* zero-length writes at ->s_maxbytes are OK */
1852                 }
1853
1854                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1855                         *count = inode->i_sb->s_maxbytes - *pos;
1856         } else {
1857                 loff_t isize;
1858                 if (bdev_read_only(I_BDEV(inode)))
1859                         return -EPERM;
1860                 isize = i_size_read(inode);
1861                 if (*pos >= isize) {
1862                         if (*count || *pos > isize)
1863                                 return -ENOSPC;
1864                 }
1865
1866                 if (*pos + *count > isize)
1867                         *count = isize - *pos;
1868         }
1869         return 0;
1870 }
1871 EXPORT_SYMBOL(generic_write_checks);
1872
1873 ssize_t
1874 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1875                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1876                 size_t count, size_t ocount)
1877 {
1878         struct file     *file = iocb->ki_filp;
1879         struct address_space *mapping = file->f_mapping;
1880         struct inode    *inode = mapping->host;
1881         ssize_t         written;
1882
1883         if (count != ocount)
1884                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1885
1886         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1887         if (written > 0) {
1888                 loff_t end = pos + written;
1889                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1890                         i_size_write(inode,  end);
1891                         mark_inode_dirty(inode);
1892                 }
1893                 *ppos = end;
1894         }
1895
1896         /*
1897          * Sync the fs metadata but not the minor inode changes and
1898          * of course not the data as we did direct DMA for the IO.
1899          * i_mutex is held, which protects generic_osync_inode() from
1900          * livelocking.
1901          */
1902         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1903                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1904                 if (err < 0)
1905                         written = err;
1906         }
1907         if (written == count && !is_sync_kiocb(iocb))
1908                 written = -EIOCBQUEUED;
1909         return written;
1910 }
1911 EXPORT_SYMBOL(generic_file_direct_write);
1912
1913 ssize_t
1914 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1915                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1916                 size_t count, ssize_t written)
1917 {
1918         struct file *file = iocb->ki_filp;
1919         struct address_space * mapping = file->f_mapping;
1920         struct address_space_operations *a_ops = mapping->a_ops;
1921         struct inode    *inode = mapping->host;
1922         long            status = 0;
1923         struct page     *page;
1924         struct page     *cached_page = NULL;
1925         size_t          bytes;
1926         struct pagevec  lru_pvec;
1927         const struct iovec *cur_iov = iov; /* current iovec */
1928         size_t          iov_base = 0;      /* offset in the current iovec */
1929         char __user     *buf;
1930
1931         pagevec_init(&lru_pvec, 0);
1932
1933         /*
1934          * handle partial DIO write.  Adjust cur_iov if needed.
1935          */
1936         if (likely(nr_segs == 1))
1937                 buf = iov->iov_base + written;
1938         else {
1939                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1940                 buf = cur_iov->iov_base + iov_base;
1941         }
1942
1943         do {
1944                 unsigned long index;
1945                 unsigned long offset;
1946                 unsigned long maxlen;
1947                 size_t copied;
1948
1949                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1950                 index = pos >> PAGE_CACHE_SHIFT;
1951                 bytes = PAGE_CACHE_SIZE - offset;
1952                 if (bytes > count)
1953                         bytes = count;
1954
1955                 /*
1956                  * Bring in the user page that we will copy from _first_.
1957                  * Otherwise there's a nasty deadlock on copying from the
1958                  * same page as we're writing to, without it being marked
1959                  * up-to-date.
1960                  */
1961                 maxlen = cur_iov->iov_len - iov_base;
1962                 if (maxlen > bytes)
1963                         maxlen = bytes;
1964                 fault_in_pages_readable(buf, maxlen);
1965
1966                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1967                 if (!page) {
1968                         status = -ENOMEM;
1969                         break;
1970                 }
1971
1972                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1973                 if (unlikely(status)) {
1974                         loff_t isize = i_size_read(inode);
1975
1976                         if (status != AOP_TRUNCATED_PAGE)
1977                                 unlock_page(page);
1978                         page_cache_release(page);
1979                         if (status == AOP_TRUNCATED_PAGE)
1980                                 continue;
1981                         /*
1982                          * prepare_write() may have instantiated a few blocks
1983                          * outside i_size.  Trim these off again.
1984                          */
1985                         if (pos + bytes > isize)
1986                                 vmtruncate(inode, isize);
1987                         break;
1988                 }
1989                 if (likely(nr_segs == 1))
1990                         copied = filemap_copy_from_user(page, offset,
1991                                                         buf, bytes);
1992                 else
1993                         copied = filemap_copy_from_user_iovec(page, offset,
1994                                                 cur_iov, iov_base, bytes);
1995                 flush_dcache_page(page);
1996                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1997                 if (status == AOP_TRUNCATED_PAGE) {
1998                         page_cache_release(page);
1999                         continue;
2000                 }
2001                 if (likely(copied > 0)) {
2002                         if (!status)
2003                                 status = copied;
2004
2005                         if (status >= 0) {
2006                                 written += status;
2007                                 count -= status;
2008                                 pos += status;
2009                                 buf += status;
2010                                 if (unlikely(nr_segs > 1)) {
2011                                         filemap_set_next_iovec(&cur_iov,
2012                                                         &iov_base, status);
2013                                         if (count)
2014                                                 buf = cur_iov->iov_base +
2015                                                         iov_base;
2016                                 } else {
2017                                         iov_base += status;
2018                                 }
2019                         }
2020                 }
2021                 if (unlikely(copied != bytes))
2022                         if (status >= 0)
2023                                 status = -EFAULT;
2024                 unlock_page(page);
2025                 mark_page_accessed(page);
2026                 page_cache_release(page);
2027                 if (status < 0)
2028                         break;
2029                 balance_dirty_pages_ratelimited(mapping);
2030                 cond_resched();
2031         } while (count);
2032         *ppos = pos;
2033
2034         if (cached_page)
2035                 page_cache_release(cached_page);
2036
2037         /*
2038          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2039          */
2040         if (likely(status >= 0)) {
2041                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2042                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2043                                 status = generic_osync_inode(inode, mapping,
2044                                                 OSYNC_METADATA|OSYNC_DATA);
2045                 }
2046         }
2047
2048         /*
2049          * If we get here for O_DIRECT writes then we must have fallen through
2050          * to buffered writes (block instantiation inside i_size).  So we sync
2051          * the file data here, to try to honour O_DIRECT expectations.
2052          */
2053         if (unlikely(file->f_flags & O_DIRECT) && written)
2054                 status = filemap_write_and_wait(mapping);
2055
2056         pagevec_lru_add(&lru_pvec);
2057         return written ? written : status;
2058 }
2059 EXPORT_SYMBOL(generic_file_buffered_write);
2060
2061 static ssize_t
2062 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2063                                 unsigned long nr_segs, loff_t *ppos)
2064 {
2065         struct file *file = iocb->ki_filp;
2066         struct address_space * mapping = file->f_mapping;
2067         size_t ocount;          /* original count */
2068         size_t count;           /* after file limit checks */
2069         struct inode    *inode = mapping->host;
2070         unsigned long   seg;
2071         loff_t          pos;
2072         ssize_t         written;
2073         ssize_t         err;
2074
2075         ocount = 0;
2076         for (seg = 0; seg < nr_segs; seg++) {
2077                 const struct iovec *iv = &iov[seg];
2078
2079                 /*
2080                  * If any segment has a negative length, or the cumulative
2081                  * length ever wraps negative then return -EINVAL.
2082                  */
2083                 ocount += iv->iov_len;
2084                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2085                         return -EINVAL;
2086                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2087                         continue;
2088                 if (seg == 0)
2089                         return -EFAULT;
2090                 nr_segs = seg;
2091                 ocount -= iv->iov_len;  /* This segment is no good */
2092                 break;
2093         }
2094
2095         count = ocount;
2096         pos = *ppos;
2097
2098         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2099
2100         /* We can write back this queue in page reclaim */
2101         current->backing_dev_info = mapping->backing_dev_info;
2102         written = 0;
2103
2104         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2105         if (err)
2106                 goto out;
2107
2108         if (count == 0)
2109                 goto out;
2110
2111         err = remove_suid(file->f_dentry);
2112         if (err)
2113                 goto out;
2114
2115         file_update_time(file);
2116
2117         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2118         if (unlikely(file->f_flags & O_DIRECT)) {
2119                 written = generic_file_direct_write(iocb, iov,
2120                                 &nr_segs, pos, ppos, count, ocount);
2121                 if (written < 0 || written == count)
2122                         goto out;
2123                 /*
2124                  * direct-io write to a hole: fall through to buffered I/O
2125                  * for completing the rest of the request.
2126                  */
2127                 pos += written;
2128                 count -= written;
2129         }
2130
2131         written = generic_file_buffered_write(iocb, iov, nr_segs,
2132                         pos, ppos, count, written);
2133 out:
2134         current->backing_dev_info = NULL;
2135         return written ? written : err;
2136 }
2137 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2138
2139 ssize_t
2140 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2141                                 unsigned long nr_segs, loff_t *ppos)
2142 {
2143         struct file *file = iocb->ki_filp;
2144         struct address_space *mapping = file->f_mapping;
2145         struct inode *inode = mapping->host;
2146         ssize_t ret;
2147         loff_t pos = *ppos;
2148
2149         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2150
2151         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2152                 int err;
2153
2154                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2155                 if (err < 0)
2156                         ret = err;
2157         }
2158         return ret;
2159 }
2160
2161 static ssize_t
2162 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2163                                 unsigned long nr_segs, loff_t *ppos)
2164 {
2165         struct kiocb kiocb;
2166         ssize_t ret;
2167
2168         init_sync_kiocb(&kiocb, file);
2169         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2170         if (ret == -EIOCBQUEUED)
2171                 ret = wait_on_sync_kiocb(&kiocb);
2172         return ret;
2173 }
2174
2175 ssize_t
2176 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2177                                 unsigned long nr_segs, loff_t *ppos)
2178 {
2179         struct kiocb kiocb;
2180         ssize_t ret;
2181
2182         init_sync_kiocb(&kiocb, file);
2183         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2184         if (-EIOCBQUEUED == ret)
2185                 ret = wait_on_sync_kiocb(&kiocb);
2186         return ret;
2187 }
2188 EXPORT_SYMBOL(generic_file_write_nolock);
2189
2190 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2191                                size_t count, loff_t pos)
2192 {
2193         struct file *file = iocb->ki_filp;
2194         struct address_space *mapping = file->f_mapping;
2195         struct inode *inode = mapping->host;
2196         ssize_t ret;
2197         struct iovec local_iov = { .iov_base = (void __user *)buf,
2198                                         .iov_len = count };
2199
2200         BUG_ON(iocb->ki_pos != pos);
2201
2202         mutex_lock(&inode->i_mutex);
2203         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2204                                                 &iocb->ki_pos);
2205         mutex_unlock(&inode->i_mutex);
2206
2207         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2208                 ssize_t err;
2209
2210                 err = sync_page_range(inode, mapping, pos, ret);
2211                 if (err < 0)
2212                         ret = err;
2213         }
2214         return ret;
2215 }
2216 EXPORT_SYMBOL(generic_file_aio_write);
2217
2218 ssize_t generic_file_write(struct file *file, const char __user *buf,
2219                            size_t count, loff_t *ppos)
2220 {
2221         struct address_space *mapping = file->f_mapping;
2222         struct inode *inode = mapping->host;
2223         ssize_t ret;
2224         struct iovec local_iov = { .iov_base = (void __user *)buf,
2225                                         .iov_len = count };
2226
2227         mutex_lock(&inode->i_mutex);
2228         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2229         mutex_unlock(&inode->i_mutex);
2230
2231         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2232                 ssize_t err;
2233
2234                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2235                 if (err < 0)
2236                         ret = err;
2237         }
2238         return ret;
2239 }
2240 EXPORT_SYMBOL(generic_file_write);
2241
2242 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2243                         unsigned long nr_segs, loff_t *ppos)
2244 {
2245         struct kiocb kiocb;
2246         ssize_t ret;
2247
2248         init_sync_kiocb(&kiocb, filp);
2249         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2250         if (-EIOCBQUEUED == ret)
2251                 ret = wait_on_sync_kiocb(&kiocb);
2252         return ret;
2253 }
2254 EXPORT_SYMBOL(generic_file_readv);
2255
2256 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2257                         unsigned long nr_segs, loff_t *ppos)
2258 {
2259         struct address_space *mapping = file->f_mapping;
2260         struct inode *inode = mapping->host;
2261         ssize_t ret;
2262
2263         mutex_lock(&inode->i_mutex);
2264         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2265         mutex_unlock(&inode->i_mutex);
2266
2267         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2268                 int err;
2269
2270                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2271                 if (err < 0)
2272                         ret = err;
2273         }
2274         return ret;
2275 }
2276 EXPORT_SYMBOL(generic_file_writev);
2277
2278 /*
2279  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2280  * went wrong during pagecache shootdown.
2281  */
2282 static ssize_t
2283 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2284         loff_t offset, unsigned long nr_segs)
2285 {
2286         struct file *file = iocb->ki_filp;
2287         struct address_space *mapping = file->f_mapping;
2288         ssize_t retval;
2289         size_t write_len = 0;
2290
2291         /*
2292          * If it's a write, unmap all mmappings of the file up-front.  This
2293          * will cause any pte dirty bits to be propagated into the pageframes
2294          * for the subsequent filemap_write_and_wait().
2295          */
2296         if (rw == WRITE) {
2297                 write_len = iov_length(iov, nr_segs);
2298                 if (mapping_mapped(mapping))
2299                         unmap_mapping_range(mapping, offset, write_len, 0);
2300         }
2301
2302         retval = filemap_write_and_wait(mapping);
2303         if (retval == 0) {
2304                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2305                                                 offset, nr_segs);
2306                 if (rw == WRITE && mapping->nrpages) {
2307                         pgoff_t end = (offset + write_len - 1)
2308                                                 >> PAGE_CACHE_SHIFT;
2309                         int err = invalidate_inode_pages2_range(mapping,
2310                                         offset >> PAGE_CACHE_SHIFT, end);
2311                         if (err)
2312                                 retval = err;
2313                 }
2314         }
2315         return retval;
2316 }