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