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 <linux/cpuset.h>
  33 #include "filemap.h"
  34 #include "internal.h"
  35
  36 /*
  37  * FIXME: remove all knowledge of the buffer layer from the core VM
  38  */
  39 #include <linux/buffer_head.h> /* for generic_osync_inode */
  40
  41 #include <asm/uaccess.h>
  42 #include <asm/mman.h>
  43
  44 static ssize_t
  45 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
  46         loff_t offset, unsigned long nr_segs);
  47
  48 /*
  49  * Shared mappings implemented 30.11.1994. It's not fully working yet,
  50  * though.
  51  *
  52  * Shared mappings now work. 15.8.1995  Bruno.
  53  *
  54  * finished 'unifying' the page and buffer cache and SMP-threaded the
  55  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
  56  *
  57  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
  58  */
  59
  60 /*
  61  * Lock ordering:
  62  *
  63  *  ->i_mmap_lock               (vmtruncate)
  64  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
  65  *      ->swap_lock             (exclusive_swap_page, others)
  66  *        ->mapping->tree_lock
  67  *
  68  *  ->i_mutex
  69  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
  70  *
  71  *  ->mmap_sem
  72  *    ->i_mmap_lock
  73  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
  74  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
  75  *
  76  *  ->mmap_sem
  77  *    ->lock_page               (access_process_vm)
  78  *
  79  *  ->mmap_sem
  80  *    ->i_mutex                 (msync)
  81  *
  82  *  ->i_mutex
  83  *    ->i_alloc_sem             (various)
  84  *
  85  *  ->inode_lock
  86  *    ->sb_lock                 (fs/fs-writeback.c)
  87  *    ->mapping->tree_lock      (__sync_single_inode)
  88  *
  89  *  ->i_mmap_lock
  90  *    ->anon_vma.lock           (vma_adjust)
  91  *
  92  *  ->anon_vma.lock
  93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
  94  *
  95  *  ->page_table_lock or pte_lock
  96  *    ->swap_lock               (try_to_unmap_one)
  97  *    ->private_lock            (try_to_unmap_one)
  98  *    ->tree_lock               (try_to_unmap_one)
  99  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
 100  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
 101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
 102  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
 103  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
 104  *    ->inode_lock              (zap_pte_range->set_page_dirty)
 105  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
 106  *
 107  *  ->task->proc_lock
 108  *    ->dcache_lock             (proc_pid_lookup)
 109  */
 110
 111 /*
 112  * Remove a page from the page cache and free it. Caller has to make
 113  * sure the page is locked and that nobody else uses it - or that usage
 114  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
 115  */
 116 void __remove_from_page_cache(struct page *page)
 117 {
 118         struct address_space *mapping = page->mapping;
 119
 120         radix_tree_delete(&mapping->page_tree, page->index);
 121         page->mapping = NULL;
 122         mapping->nrpages--;
 123         pagecache_acct(-1);
 124 }
 125
 126 void remove_from_page_cache(struct page *page)
 127 {
 128         struct address_space *mapping = page->mapping;
 129
 130         BUG_ON(!PageLocked(page));
 131
 132         write_lock_irq(&mapping->tree_lock);
 133         __remove_from_page_cache(page);
 134         write_unlock_irq(&mapping->tree_lock);
 135 }
 136
 137 static int sync_page(void *word)
 138 {
 139         struct address_space *mapping;
 140         struct page *page;
 141
 142         page = container_of((unsigned long *)word, struct page, flags);
 143
 144         /*
 145          * page_mapping() is being called without PG_locked held.
 146          * Some knowledge of the state and use of the page is used to
 147          * reduce the requirements down to a memory barrier.
 148          * The danger here is of a stale page_mapping() return value
 149          * indicating a struct address_space different from the one it's
 150          * associated with when it is associated with one.
 151          * After smp_mb(), it's either the correct page_mapping() for
 152          * the page, or an old page_mapping() and the page's own
 153          * page_mapping() has gone NULL.
 154          * The ->sync_page() address_space operation must tolerate
 155          * page_mapping() going NULL. By an amazing coincidence,
 156          * this comes about because none of the users of the page
 157          * in the ->sync_page() methods make essential use of the
 158          * page_mapping(), merely passing the page down to the backing
 159          * device's unplug functions when it's non-NULL, which in turn
 160          * ignore it for all cases but swap, where only page_private(page) is
 161          * of interest. When page_mapping() does go NULL, the entire
 162          * call stack gracefully ignores the page and returns.
 163          * -- wli
 164          */
 165         smp_mb();
 166         mapping = page_mapping(page);
 167         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 168                 mapping->a_ops->sync_page(page);
 169         io_schedule();
 170         return 0;
 171 }
 172
 173 /**
 174  * filemap_fdatawrite_range - start writeback against all of a mapping's
 175  * dirty pages that lie within the byte offsets <start, end>
 176  * @mapping:    address space structure to write
 177  * @start:      offset in bytes where the range starts
 178  * @end:        offset in bytes where the range ends (inclusive)
 179  * @sync_mode:  enable synchronous operation
 180  *
 181  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 182  * opposed to a regular memory * cleansing writeback.  The difference between
 183  * these two operations is that if a dirty page/buffer is encountered, it must
 184  * be waited upon, and not just skipped over.
 185  */
 186 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
 187                                 loff_t end, int sync_mode)
 188 {
 189         int ret;
 190         struct writeback_control wbc = {
 191                 .sync_mode = sync_mode,
 192                 .nr_to_write = mapping->nrpages * 2,
 193                 .start = start,
 194                 .end = end,
 195         };
 196
 197         if (!mapping_cap_writeback_dirty(mapping))
 198                 return 0;
 199
 200         ret = do_writepages(mapping, &wbc);
 201         return ret;
 202 }
 203
 204 static inline int __filemap_fdatawrite(struct address_space *mapping,
 205         int sync_mode)
 206 {
 207         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
 208 }
 209
 210 int filemap_fdatawrite(struct address_space *mapping)
 211 {
 212         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 213 }
 214 EXPORT_SYMBOL(filemap_fdatawrite);
 215
 216 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
 217                                 loff_t end)
 218 {
 219         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
 220 }
 221
 222 /*
 223  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 224  * purposes - I/O may not be started against all dirty pages.
 225  */
 226 int filemap_flush(struct address_space *mapping)
 227 {
 228         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 229 }
 230 EXPORT_SYMBOL(filemap_flush);
 231
 232 /*
 233  * Wait for writeback to complete against pages indexed by start->end
 234  * inclusive
 235  */
 236 int wait_on_page_writeback_range(struct address_space *mapping,
 237                                 pgoff_t start, pgoff_t end)
 238 {
 239         struct pagevec pvec;
 240         int nr_pages;
 241         int ret = 0;
 242         pgoff_t index;
 243
 244         if (end < start)
 245                 return 0;
 246
 247         pagevec_init(&pvec, 0);
 248         index = start;
 249         while ((index <= end) &&
 250                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 251                         PAGECACHE_TAG_WRITEBACK,
 252                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 253                 unsigned i;
 254
 255                 for (i = 0; i < nr_pages; i++) {
 256                         struct page *page = pvec.pages[i];
 257
 258                         /* until radix tree lookup accepts end_index */
 259                         if (page->index > end)
 260                                 continue;
 261
 262                         wait_on_page_writeback(page);
 263                         if (PageError(page))
 264                                 ret = -EIO;
 265                 }
 266                 pagevec_release(&pvec);
 267                 cond_resched();
 268         }
 269
 270         /* Check for outstanding write errors */
 271         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 272                 ret = -ENOSPC;
 273         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 274                 ret = -EIO;
 275
 276         return ret;
 277 }
 278
 279 /*
 280  * Write and wait upon all the pages in the passed range.  This is a "data
 281  * integrity" operation.  It waits upon in-flight writeout before starting and
 282  * waiting upon new writeout.  If there was an IO error, return it.
 283  *
 284  * We need to re-take i_mutex during the generic_osync_inode list walk because
 285  * it is otherwise livelockable.
 286  */
 287 int sync_page_range(struct inode *inode, struct address_space *mapping,
 288                         loff_t pos, loff_t count)
 289 {
 290         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 291         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 292         int ret;
 293
 294         if (!mapping_cap_writeback_dirty(mapping) || !count)
 295                 return 0;
 296         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 297         if (ret == 0) {
 298                 mutex_lock(&inode->i_mutex);
 299                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 300                 mutex_unlock(&inode->i_mutex);
 301         }
 302         if (ret == 0)
 303                 ret = wait_on_page_writeback_range(mapping, start, end);
 304         return ret;
 305 }
 306 EXPORT_SYMBOL(sync_page_range);
 307
 308 /*
 309  * Note: Holding i_mutex across sync_page_range_nolock is not a good idea
 310  * as it forces O_SYNC writers to different parts of the same file
 311  * to be serialised right until io completion.
 312  */
 313 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
 314                            loff_t pos, loff_t count)
 315 {
 316         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 317         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 318         int ret;
 319
 320         if (!mapping_cap_writeback_dirty(mapping) || !count)
 321                 return 0;
 322         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 323         if (ret == 0)
 324                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 325         if (ret == 0)
 326                 ret = wait_on_page_writeback_range(mapping, start, end);
 327         return ret;
 328 }
 329 EXPORT_SYMBOL(sync_page_range_nolock);
 330
 331 /**
 332  * filemap_fdatawait - walk the list of under-writeback pages of the given
 333  *     address space and wait for all of them.
 334  *
 335  * @mapping: address space structure to wait for
 336  */
 337 int filemap_fdatawait(struct address_space *mapping)
 338 {
 339         loff_t i_size = i_size_read(mapping->host);
 340
 341         if (i_size == 0)
 342                 return 0;
 343
 344         return wait_on_page_writeback_range(mapping, 0,
 345                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 346 }
 347 EXPORT_SYMBOL(filemap_fdatawait);
 348
 349 int filemap_write_and_wait(struct address_space *mapping)
 350 {
 351         int err = 0;
 352
 353         if (mapping->nrpages) {
 354                 err = filemap_fdatawrite(mapping);
 355                 /*
 356                  * Even if the above returned error, the pages may be
 357                  * written partially (e.g. -ENOSPC), so we wait for it.
 358                  * But the -EIO is special case, it may indicate the worst
 359                  * thing (e.g. bug) happened, so we avoid waiting for it.
 360                  */
 361                 if (err != -EIO) {
 362                         int err2 = filemap_fdatawait(mapping);
 363                         if (!err)
 364                                 err = err2;
 365                 }
 366         }
 367         return err;
 368 }
 369 EXPORT_SYMBOL(filemap_write_and_wait);
 370
 371 /*
 372  * Write out and wait upon file offsets lstart->lend, inclusive.
 373  *
 374  * Note that `lend' is inclusive (describes the last byte to be written) so
 375  * that this function can be used to write to the very end-of-file (end = -1).
 376  */
 377 int filemap_write_and_wait_range(struct address_space *mapping,
 378                                  loff_t lstart, loff_t lend)
 379 {
 380         int err = 0;
 381
 382         if (mapping->nrpages) {
 383                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
 384                                                  WB_SYNC_ALL);
 385                 /* See comment of filemap_write_and_wait() */
 386                 if (err != -EIO) {
 387                         int err2 = wait_on_page_writeback_range(mapping,
 388                                                 lstart >> PAGE_CACHE_SHIFT,
 389                                                 lend >> PAGE_CACHE_SHIFT);
 390                         if (!err)
 391                                 err = err2;
 392                 }
 393         }
 394         return err;
 395 }
 396
 397 /*
 398  * This function is used to add newly allocated pagecache pages:
 399  * the page is new, so we can just run SetPageLocked() against it.
 400  * The other page state flags were set by rmqueue().
 401  *
 402  * This function does not add the page to the LRU.  The caller must do that.
 403  */
 404 int add_to_page_cache(struct page *page, struct address_space *mapping,
 405                 pgoff_t offset, gfp_t gfp_mask)
 406 {
 407         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 408
 409         if (error == 0) {
 410                 write_lock_irq(&mapping->tree_lock);
 411                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 412                 if (!error) {
 413                         page_cache_get(page);
 414                         SetPageLocked(page);
 415                         page->mapping = mapping;
 416                         page->index = offset;
 417                         mapping->nrpages++;
 418                         pagecache_acct(1);
 419                 }
 420                 write_unlock_irq(&mapping->tree_lock);
 421                 radix_tree_preload_end();
 422         }
 423         return error;
 424 }
 425
 426 EXPORT_SYMBOL(add_to_page_cache);
 427
 428 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 429                                 pgoff_t offset, gfp_t gfp_mask)
 430 {
 431         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 432         if (ret == 0)
 433                 lru_cache_add(page);
 434         return ret;
 435 }
 436
 437 #ifdef CONFIG_NUMA
 438 struct page *page_cache_alloc(struct address_space *x)
 439 {
 440         if (cpuset_do_page_mem_spread()) {
 441                 int n = cpuset_mem_spread_node();
 442                 return alloc_pages_node(n, mapping_gfp_mask(x), 0);
 443         }
 444         return alloc_pages(mapping_gfp_mask(x), 0);
 445 }
 446 EXPORT_SYMBOL(page_cache_alloc);
 447
 448 struct page *page_cache_alloc_cold(struct address_space *x)
 449 {
 450         if (cpuset_do_page_mem_spread()) {
 451                 int n = cpuset_mem_spread_node();
 452                 return alloc_pages_node(n, mapping_gfp_mask(x)|__GFP_COLD, 0);
 453         }
 454         return alloc_pages(mapping_gfp_mask(x)|__GFP_COLD, 0);
 455 }
 456 EXPORT_SYMBOL(page_cache_alloc_cold);
 457 #endif
 458
 459 /*
 460  * In order to wait for pages to become available there must be
 461  * waitqueues associated with pages. By using a hash table of
 462  * waitqueues where the bucket discipline is to maintain all
 463  * waiters on the same queue and wake all when any of the pages
 464  * become available, and for the woken contexts to check to be
 465  * sure the appropriate page became available, this saves space
 466  * at a cost of "thundering herd" phenomena during rare hash
 467  * collisions.
 468  */
 469 static wait_queue_head_t *page_waitqueue(struct page *page)
 470 {
 471         const struct zone *zone = page_zone(page);
 472
 473         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 474 }
 475
 476 static inline void wake_up_page(struct page *page, int bit)
 477 {
 478         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 479 }
 480
 481 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 482 {
 483         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 484
 485         if (test_bit(bit_nr, &page->flags))
 486                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 487                                                         TASK_UNINTERRUPTIBLE);
 488 }
 489 EXPORT_SYMBOL(wait_on_page_bit);
 490
 491 /**
 492  * unlock_page() - unlock a locked page
 493  *
 494  * @page: the page
 495  *
 496  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 497  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 498  * mechananism between PageLocked pages and PageWriteback pages is shared.
 499  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 500  *
 501  * The first mb is necessary to safely close the critical section opened by the
 502  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 503  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 504  * parallel wait_on_page_locked()).
 505  */
 506 void fastcall unlock_page(struct page *page)
 507 {
 508         smp_mb__before_clear_bit();
 509         if (!TestClearPageLocked(page))
 510                 BUG();
 511         smp_mb__after_clear_bit();
 512         wake_up_page(page, PG_locked);
 513 }
 514 EXPORT_SYMBOL(unlock_page);
 515
 516 /*
 517  * End writeback against a page.
 518  */
 519 void end_page_writeback(struct page *page)
 520 {
 521         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 522                 if (!test_clear_page_writeback(page))
 523                         BUG();
 524         }
 525         smp_mb__after_clear_bit();
 526         wake_up_page(page, PG_writeback);
 527 }
 528 EXPORT_SYMBOL(end_page_writeback);
 529
 530 /*
 531  * Get a lock on the page, assuming we need to sleep to get it.
 532  *
 533  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 534  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 535  * chances are that on the second loop, the block layer's plug list is empty,
 536  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 537  */
 538 void fastcall __lock_page(struct page *page)
 539 {
 540         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 541
 542         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 543                                                         TASK_UNINTERRUPTIBLE);
 544 }
 545 EXPORT_SYMBOL(__lock_page);
 546
 547 /*
 548  * a rather lightweight function, finding and getting a reference to a
 549  * hashed page atomically.
 550  */
 551 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 552 {
 553         struct page *page;
 554
 555         read_lock_irq(&mapping->tree_lock);
 556         page = radix_tree_lookup(&mapping->page_tree, offset);
 557         if (page)
 558                 page_cache_get(page);
 559         read_unlock_irq(&mapping->tree_lock);
 560         return page;
 561 }
 562
 563 EXPORT_SYMBOL(find_get_page);
 564
 565 /*
 566  * Same as above, but trylock it instead of incrementing the count.
 567  */
 568 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 569 {
 570         struct page *page;
 571
 572         read_lock_irq(&mapping->tree_lock);
 573         page = radix_tree_lookup(&mapping->page_tree, offset);
 574         if (page && TestSetPageLocked(page))
 575                 page = NULL;
 576         read_unlock_irq(&mapping->tree_lock);
 577         return page;
 578 }
 579
 580 EXPORT_SYMBOL(find_trylock_page);
 581
 582 /**
 583  * find_lock_page - locate, pin and lock a pagecache page
 584  *
 585  * @mapping: the address_space to search
 586  * @offset: the page index
 587  *
 588  * Locates the desired pagecache page, locks it, increments its reference
 589  * count and returns its address.
 590  *
 591  * Returns zero if the page was not present. find_lock_page() may sleep.
 592  */
 593 struct page *find_lock_page(struct address_space *mapping,
 594                                 unsigned long offset)
 595 {
 596         struct page *page;
 597
 598         read_lock_irq(&mapping->tree_lock);
 599 repeat:
 600         page = radix_tree_lookup(&mapping->page_tree, offset);
 601         if (page) {
 602                 page_cache_get(page);
 603                 if (TestSetPageLocked(page)) {
 604                         read_unlock_irq(&mapping->tree_lock);
 605                         __lock_page(page);
 606                         read_lock_irq(&mapping->tree_lock);
 607
 608                         /* Has the page been truncated while we slept? */
 609                         if (unlikely(page->mapping != mapping ||
 610                                      page->index != offset)) {
 611                                 unlock_page(page);
 612                                 page_cache_release(page);
 613                                 goto repeat;
 614                         }
 615                 }
 616         }
 617         read_unlock_irq(&mapping->tree_lock);
 618         return page;
 619 }
 620
 621 EXPORT_SYMBOL(find_lock_page);
 622
 623 /**
 624  * find_or_create_page - locate or add a pagecache page
 625  *
 626  * @mapping: the page's address_space
 627  * @index: the page's index into the mapping
 628  * @gfp_mask: page allocation mode
 629  *
 630  * Locates a page in the pagecache.  If the page is not present, a new page
 631  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 632  * LRU list.  The returned page is locked and has its reference count
 633  * incremented.
 634  *
 635  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 636  * allocation!
 637  *
 638  * find_or_create_page() returns the desired page's address, or zero on
 639  * memory exhaustion.
 640  */
 641 struct page *find_or_create_page(struct address_space *mapping,
 642                 unsigned long index, gfp_t gfp_mask)
 643 {
 644         struct page *page, *cached_page = NULL;
 645         int err;
 646 repeat:
 647         page = find_lock_page(mapping, index);
 648         if (!page) {
 649                 if (!cached_page) {
 650                         cached_page = alloc_page(gfp_mask);
 651                         if (!cached_page)
 652                                 return NULL;
 653                 }
 654                 err = add_to_page_cache_lru(cached_page, mapping,
 655                                         index, gfp_mask);
 656                 if (!err) {
 657                         page = cached_page;
 658                         cached_page = NULL;
 659                 } else if (err == -EEXIST)
 660                         goto repeat;
 661         }
 662         if (cached_page)
 663                 page_cache_release(cached_page);
 664         return page;
 665 }
 666
 667 EXPORT_SYMBOL(find_or_create_page);
 668
 669 /**
 670  * find_get_pages - gang pagecache lookup
 671  * @mapping:    The address_space to search
 672  * @start:      The starting page index
 673  * @nr_pages:   The maximum number of pages
 674  * @pages:      Where the resulting pages are placed
 675  *
 676  * find_get_pages() will search for and return a group of up to
 677  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 678  * find_get_pages() takes a reference against the returned pages.
 679  *
 680  * The search returns a group of mapping-contiguous pages with ascending
 681  * indexes.  There may be holes in the indices due to not-present pages.
 682  *
 683  * find_get_pages() returns the number of pages which were found.
 684  */
 685 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 686                             unsigned int nr_pages, struct page **pages)
 687 {
 688         unsigned int i;
 689         unsigned int ret;
 690
 691         read_lock_irq(&mapping->tree_lock);
 692         ret = radix_tree_gang_lookup(&mapping->page_tree,
 693                                 (void **)pages, start, nr_pages);
 694         for (i = 0; i < ret; i++)
 695                 page_cache_get(pages[i]);
 696         read_unlock_irq(&mapping->tree_lock);
 697         return ret;
 698 }
 699
 700 /*
 701  * Like find_get_pages, except we only return pages which are tagged with
 702  * `tag'.   We update *index to index the next page for the traversal.
 703  */
 704 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 705                         int tag, unsigned int nr_pages, struct page **pages)
 706 {
 707         unsigned int i;
 708         unsigned int ret;
 709
 710         read_lock_irq(&mapping->tree_lock);
 711         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 712                                 (void **)pages, *index, nr_pages, tag);
 713         for (i = 0; i < ret; i++)
 714                 page_cache_get(pages[i]);
 715         if (ret)
 716                 *index = pages[ret - 1]->index + 1;
 717         read_unlock_irq(&mapping->tree_lock);
 718         return ret;
 719 }
 720
 721 /*
 722  * Same as grab_cache_page, but do not wait if the page is unavailable.
 723  * This is intended for speculative data generators, where the data can
 724  * be regenerated if the page couldn't be grabbed.  This routine should
 725  * be safe to call while holding the lock for another page.
 726  *
 727  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 728  * and deadlock against the caller's locked page.
 729  */
 730 struct page *
 731 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 732 {
 733         struct page *page = find_get_page(mapping, index);
 734         gfp_t gfp_mask;
 735
 736         if (page) {
 737                 if (!TestSetPageLocked(page))
 738                         return page;
 739                 page_cache_release(page);
 740                 return NULL;
 741         }
 742         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 743         page = alloc_pages(gfp_mask, 0);
 744         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 745                 page_cache_release(page);
 746                 page = NULL;
 747         }
 748         return page;
 749 }
 750
 751 EXPORT_SYMBOL(grab_cache_page_nowait);
 752
 753 /*
 754  * This is a generic file read routine, and uses the
 755  * mapping->a_ops->readpage() function for the actual low-level
 756  * stuff.
 757  *
 758  * This is really ugly. But the goto's actually try to clarify some
 759  * of the logic when it comes to error handling etc.
 760  *
 761  * Note the struct file* is only passed for the use of readpage.  It may be
 762  * NULL.
 763  */
 764 void do_generic_mapping_read(struct address_space *mapping,
 765                              struct file_ra_state *_ra,
 766                              struct file *filp,
 767                              loff_t *ppos,
 768                              read_descriptor_t *desc,
 769                              read_actor_t actor)
 770 {
 771         struct inode *inode = mapping->host;
 772         unsigned long index;
 773         unsigned long end_index;
 774         unsigned long offset;
 775         unsigned long last_index;
 776         unsigned long next_index;
 777         unsigned long prev_index;
 778         loff_t isize;
 779         struct page *cached_page;
 780         int error;
 781         struct file_ra_state ra = *_ra;
 782
 783         cached_page = NULL;
 784         index = *ppos >> PAGE_CACHE_SHIFT;
 785         next_index = index;
 786         prev_index = ra.prev_page;
 787         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 788         offset = *ppos & ~PAGE_CACHE_MASK;
 789
 790         isize = i_size_read(inode);
 791         if (!isize)
 792                 goto out;
 793
 794         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 795         for (;;) {
 796                 struct page *page;
 797                 unsigned long nr, ret;
 798
 799                 /* nr is the maximum number of bytes to copy from this page */
 800                 nr = PAGE_CACHE_SIZE;
 801                 if (index >= end_index) {
 802                         if (index > end_index)
 803                                 goto out;
 804                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 805                         if (nr <= offset) {
 806                                 goto out;
 807                         }
 808                 }
 809                 nr = nr - offset;
 810
 811                 cond_resched();
 812                 if (index == next_index)
 813                         next_index = page_cache_readahead(mapping, &ra, filp,
 814                                         index, last_index - index);
 815
 816 find_page:
 817                 page = find_get_page(mapping, index);
 818                 if (unlikely(page == NULL)) {
 819                         handle_ra_miss(mapping, &ra, index);
 820                         goto no_cached_page;
 821                 }
 822                 if (!PageUptodate(page))
 823                         goto page_not_up_to_date;
 824 page_ok:
 825
 826                 /* If users can be writing to this page using arbitrary
 827                  * virtual addresses, take care about potential aliasing
 828                  * before reading the page on the kernel side.
 829                  */
 830                 if (mapping_writably_mapped(mapping))
 831                         flush_dcache_page(page);
 832
 833                 /*
 834                  * When (part of) the same page is read multiple times
 835                  * in succession, only mark it as accessed the first time.
 836                  */
 837                 if (prev_index != index)
 838                         mark_page_accessed(page);
 839                 prev_index = index;
 840
 841                 /*
 842                  * Ok, we have the page, and it's up-to-date, so
 843                  * now we can copy it to user space...
 844                  *
 845                  * The actor routine returns how many bytes were actually used..
 846                  * NOTE! This may not be the same as how much of a user buffer
 847                  * we filled up (we may be padding etc), so we can only update
 848                  * "pos" here (the actor routine has to update the user buffer
 849                  * pointers and the remaining count).
 850                  */
 851                 ret = actor(desc, page, offset, nr);
 852                 offset += ret;
 853                 index += offset >> PAGE_CACHE_SHIFT;
 854                 offset &= ~PAGE_CACHE_MASK;
 855
 856                 page_cache_release(page);
 857                 if (ret == nr && desc->count)
 858                         continue;
 859                 goto out;
 860
 861 page_not_up_to_date:
 862                 /* Get exclusive access to the page ... */
 863                 lock_page(page);
 864
 865                 /* Did it get unhashed before we got the lock? */
 866                 if (!page->mapping) {
 867                         unlock_page(page);
 868                         page_cache_release(page);
 869                         continue;
 870                 }
 871
 872                 /* Did somebody else fill it already? */
 873                 if (PageUptodate(page)) {
 874                         unlock_page(page);
 875                         goto page_ok;
 876                 }
 877
 878 readpage:
 879                 /* Start the actual read. The read will unlock the page. */
 880                 error = mapping->a_ops->readpage(filp, page);
 881
 882                 if (unlikely(error)) {
 883                         if (error == AOP_TRUNCATED_PAGE) {
 884                                 page_cache_release(page);
 885                                 goto find_page;
 886                         }
 887                         goto readpage_error;
 888                 }
 889
 890                 if (!PageUptodate(page)) {
 891                         lock_page(page);
 892                         if (!PageUptodate(page)) {
 893                                 if (page->mapping == NULL) {
 894                                         /*
 895                                          * invalidate_inode_pages got it
 896                                          */
 897                                         unlock_page(page);
 898                                         page_cache_release(page);
 899                                         goto find_page;
 900                                 }
 901                                 unlock_page(page);
 902                                 error = -EIO;
 903                                 goto readpage_error;
 904                         }
 905                         unlock_page(page);
 906                 }
 907
 908                 /*
 909                  * i_size must be checked after we have done ->readpage.
 910                  *
 911                  * Checking i_size after the readpage allows us to calculate
 912                  * the correct value for "nr", which means the zero-filled
 913                  * part of the page is not copied back to userspace (unless
 914                  * another truncate extends the file - this is desired though).
 915                  */
 916                 isize = i_size_read(inode);
 917                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 918                 if (unlikely(!isize || index > end_index)) {
 919                         page_cache_release(page);
 920                         goto out;
 921                 }
 922
 923                 /* nr is the maximum number of bytes to copy from this page */
 924                 nr = PAGE_CACHE_SIZE;
 925                 if (index == end_index) {
 926                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 927                         if (nr <= offset) {
 928                                 page_cache_release(page);
 929                                 goto out;
 930                         }
 931                 }
 932                 nr = nr - offset;
 933                 goto page_ok;
 934
 935 readpage_error:
 936                 /* UHHUH! A synchronous read error occurred. Report it */
 937                 desc->error = error;
 938                 page_cache_release(page);
 939                 goto out;
 940
 941 no_cached_page:
 942                 /*
 943                  * Ok, it wasn't cached, so we need to create a new
 944                  * page..
 945                  */
 946                 if (!cached_page) {
 947                         cached_page = page_cache_alloc_cold(mapping);
 948                         if (!cached_page) {
 949                                 desc->error = -ENOMEM;
 950                                 goto out;
 951                         }
 952                 }
 953                 error = add_to_page_cache_lru(cached_page, mapping,
 954                                                 index, GFP_KERNEL);
 955                 if (error) {
 956                         if (error == -EEXIST)
 957                                 goto find_page;
 958                         desc->error = error;
 959                         goto out;
 960                 }
 961                 page = cached_page;
 962                 cached_page = NULL;
 963                 goto readpage;
 964         }
 965
 966 out:
 967         *_ra = ra;
 968
 969         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 970         if (cached_page)
 971                 page_cache_release(cached_page);
 972         if (filp)
 973                 file_accessed(filp);
 974 }
 975
 976 EXPORT_SYMBOL(do_generic_mapping_read);
 977
 978 int file_read_actor(read_descriptor_t *desc, struct page *page,
 979                         unsigned long offset, unsigned long size)
 980 {
 981         char *kaddr;
 982         unsigned long left, count = desc->count;
 983
 984         if (size > count)
 985                 size = count;
 986
 987         /*
 988          * Faults on the destination of a read are common, so do it before
 989          * taking the kmap.
 990          */
 991         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 992                 kaddr = kmap_atomic(page, KM_USER0);
 993                 left = __copy_to_user_inatomic(desc->arg.buf,
 994                                                 kaddr + offset, size);
 995                 kunmap_atomic(kaddr, KM_USER0);
 996                 if (left == 0)
 997                         goto success;
 998         }
 999
1000         /* Do it the slow way */
1001         kaddr = kmap(page);
1002         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1003         kunmap(page);
1004
1005         if (left) {
1006                 size -= left;
1007                 desc->error = -EFAULT;
1008         }
1009 success:
1010         desc->count = count - size;
1011         desc->written += size;
1012         desc->arg.buf += size;
1013         return size;
1014 }
1015
1016 /*
1017  * This is the "read()" routine for all filesystems
1018  * that can use the page cache directly.
1019  */
1020 ssize_t
1021 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1022                 unsigned long nr_segs, loff_t *ppos)
1023 {
1024         struct file *filp = iocb->ki_filp;
1025         ssize_t retval;
1026         unsigned long seg;
1027         size_t count;
1028
1029         count = 0;
1030         for (seg = 0; seg < nr_segs; seg++) {
1031                 const struct iovec *iv = &iov[seg];
1032
1033                 /*
1034                  * If any segment has a negative length, or the cumulative
1035                  * length ever wraps negative then return -EINVAL.
1036                  */
1037                 count += iv->iov_len;
1038                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1039                         return -EINVAL;
1040                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1041                         continue;
1042                 if (seg == 0)
1043                         return -EFAULT;
1044                 nr_segs = seg;
1045                 count -= iv->iov_len;   /* This segment is no good */
1046                 break;
1047         }
1048
1049         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1050         if (filp->f_flags & O_DIRECT) {
1051                 loff_t pos = *ppos, size;
1052                 struct address_space *mapping;
1053                 struct inode *inode;
1054
1055                 mapping = filp->f_mapping;
1056                 inode = mapping->host;
1057                 retval = 0;
1058                 if (!count)
1059                         goto out; /* skip atime */
1060                 size = i_size_read(inode);
1061                 if (pos < size) {
1062                         retval = generic_file_direct_IO(READ, iocb,
1063                                                 iov, pos, nr_segs);
1064                         if (retval > 0 && !is_sync_kiocb(iocb))
1065                                 retval = -EIOCBQUEUED;
1066                         if (retval > 0)
1067                                 *ppos = pos + retval;
1068                 }
1069                 file_accessed(filp);
1070                 goto out;
1071         }
1072
1073         retval = 0;
1074         if (count) {
1075                 for (seg = 0; seg < nr_segs; seg++) {
1076                         read_descriptor_t desc;
1077
1078                         desc.written = 0;
1079                         desc.arg.buf = iov[seg].iov_base;
1080                         desc.count = iov[seg].iov_len;
1081                         if (desc.count == 0)
1082                                 continue;
1083                         desc.error = 0;
1084                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1085                         retval += desc.written;
1086                         if (desc.error) {
1087                                 retval = retval ?: desc.error;
1088                                 break;
1089                         }
1090                 }
1091         }
1092 out:
1093         return retval;
1094 }
1095
1096 EXPORT_SYMBOL(__generic_file_aio_read);
1097
1098 ssize_t
1099 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1100 {
1101         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1102
1103         BUG_ON(iocb->ki_pos != pos);
1104         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1105 }
1106
1107 EXPORT_SYMBOL(generic_file_aio_read);
1108
1109 ssize_t
1110 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1111 {
1112         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1113         struct kiocb kiocb;
1114         ssize_t ret;
1115
1116         init_sync_kiocb(&kiocb, filp);
1117         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1118         if (-EIOCBQUEUED == ret)
1119                 ret = wait_on_sync_kiocb(&kiocb);
1120         return ret;
1121 }
1122
1123 EXPORT_SYMBOL(generic_file_read);
1124
1125 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1126 {
1127         ssize_t written;
1128         unsigned long count = desc->count;
1129         struct file *file = desc->arg.data;
1130
1131         if (size > count)
1132                 size = count;
1133
1134         written = file->f_op->sendpage(file, page, offset,
1135                                        size, &file->f_pos, size<count);
1136         if (written < 0) {
1137                 desc->error = written;
1138                 written = 0;
1139         }
1140         desc->count = count - written;
1141         desc->written += written;
1142         return written;
1143 }
1144
1145 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1146                          size_t count, read_actor_t actor, void *target)
1147 {
1148         read_descriptor_t desc;
1149
1150         if (!count)
1151                 return 0;
1152
1153         desc.written = 0;
1154         desc.count = count;
1155         desc.arg.data = target;
1156         desc.error = 0;
1157
1158         do_generic_file_read(in_file, ppos, &desc, actor);
1159         if (desc.written)
1160                 return desc.written;
1161         return desc.error;
1162 }
1163
1164 EXPORT_SYMBOL(generic_file_sendfile);
1165
1166 static ssize_t
1167 do_readahead(struct address_space *mapping, struct file *filp,
1168              unsigned long index, unsigned long nr)
1169 {
1170         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1171                 return -EINVAL;
1172
1173         force_page_cache_readahead(mapping, filp, index,
1174                                         max_sane_readahead(nr));
1175         return 0;
1176 }
1177
1178 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1179 {
1180         ssize_t ret;
1181         struct file *file;
1182
1183         ret = -EBADF;
1184         file = fget(fd);
1185         if (file) {
1186                 if (file->f_mode & FMODE_READ) {
1187                         struct address_space *mapping = file->f_mapping;
1188                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1189                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1190                         unsigned long len = end - start + 1;
1191                         ret = do_readahead(mapping, file, start, len);
1192                 }
1193                 fput(file);
1194         }
1195         return ret;
1196 }
1197
1198 #ifdef CONFIG_MMU
1199 /*
1200  * This adds the requested page to the page cache if it isn't already there,
1201  * and schedules an I/O to read in its contents from disk.
1202  */
1203 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1204 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1205 {
1206         struct address_space *mapping = file->f_mapping;
1207         struct page *page;
1208         int ret;
1209
1210         do {
1211                 page = page_cache_alloc_cold(mapping);
1212                 if (!page)
1213                         return -ENOMEM;
1214
1215                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1216                 if (ret == 0)
1217                         ret = mapping->a_ops->readpage(file, page);
1218                 else if (ret == -EEXIST)
1219                         ret = 0; /* losing race to add is OK */
1220
1221                 page_cache_release(page);
1222
1223         } while (ret == AOP_TRUNCATED_PAGE);
1224
1225         return ret;
1226 }
1227
1228 #define MMAP_LOTSAMISS  (100)
1229
1230 /*
1231  * filemap_nopage() is invoked via the vma operations vector for a
1232  * mapped memory region to read in file data during a page fault.
1233  *
1234  * The goto's are kind of ugly, but this streamlines the normal case of having
1235  * it in the page cache, and handles the special cases reasonably without
1236  * having a lot of duplicated code.
1237  */
1238 struct page *filemap_nopage(struct vm_area_struct *area,
1239                                 unsigned long address, int *type)
1240 {
1241         int error;
1242         struct file *file = area->vm_file;
1243         struct address_space *mapping = file->f_mapping;
1244         struct file_ra_state *ra = &file->f_ra;
1245         struct inode *inode = mapping->host;
1246         struct page *page;
1247         unsigned long size, pgoff;
1248         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1249
1250         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1251
1252 retry_all:
1253         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1254         if (pgoff >= size)
1255                 goto outside_data_content;
1256
1257         /* If we don't want any read-ahead, don't bother */
1258         if (VM_RandomReadHint(area))
1259                 goto no_cached_page;
1260
1261         /*
1262          * The readahead code wants to be told about each and every page
1263          * so it can build and shrink its windows appropriately
1264          *
1265          * For sequential accesses, we use the generic readahead logic.
1266          */
1267         if (VM_SequentialReadHint(area))
1268                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1269
1270         /*
1271          * Do we have something in the page cache already?
1272          */
1273 retry_find:
1274         page = find_get_page(mapping, pgoff);
1275         if (!page) {
1276                 unsigned long ra_pages;
1277
1278                 if (VM_SequentialReadHint(area)) {
1279                         handle_ra_miss(mapping, ra, pgoff);
1280                         goto no_cached_page;
1281                 }
1282                 ra->mmap_miss++;
1283
1284                 /*
1285                  * Do we miss much more than hit in this file? If so,
1286                  * stop bothering with read-ahead. It will only hurt.
1287                  */
1288                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1289                         goto no_cached_page;
1290
1291                 /*
1292                  * To keep the pgmajfault counter straight, we need to
1293                  * check did_readaround, as this is an inner loop.
1294                  */
1295                 if (!did_readaround) {
1296                         majmin = VM_FAULT_MAJOR;
1297                         inc_page_state(pgmajfault);
1298                 }
1299                 did_readaround = 1;
1300                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1301                 if (ra_pages) {
1302                         pgoff_t start = 0;
1303
1304                         if (pgoff > ra_pages / 2)
1305                                 start = pgoff - ra_pages / 2;
1306                         do_page_cache_readahead(mapping, file, start, ra_pages);
1307                 }
1308                 page = find_get_page(mapping, pgoff);
1309                 if (!page)
1310                         goto no_cached_page;
1311         }
1312
1313         if (!did_readaround)
1314                 ra->mmap_hit++;
1315
1316         /*
1317          * Ok, found a page in the page cache, now we need to check
1318          * that it's up-to-date.
1319          */
1320         if (!PageUptodate(page))
1321                 goto page_not_uptodate;
1322
1323 success:
1324         /*
1325          * Found the page and have a reference on it.
1326          */
1327         mark_page_accessed(page);
1328         if (type)
1329                 *type = majmin;
1330         return page;
1331
1332 outside_data_content:
1333         /*
1334          * An external ptracer can access pages that normally aren't
1335          * accessible..
1336          */
1337         if (area->vm_mm == current->mm)
1338                 return NULL;
1339         /* Fall through to the non-read-ahead case */
1340 no_cached_page:
1341         /*
1342          * We're only likely to ever get here if MADV_RANDOM is in
1343          * effect.
1344          */
1345         error = page_cache_read(file, pgoff);
1346         grab_swap_token();
1347
1348         /*
1349          * The page we want has now been added to the page cache.
1350          * In the unlikely event that someone removed it in the
1351          * meantime, we'll just come back here and read it again.
1352          */
1353         if (error >= 0)
1354                 goto retry_find;
1355
1356         /*
1357          * An error return from page_cache_read can result if the
1358          * system is low on memory, or a problem occurs while trying
1359          * to schedule I/O.
1360          */
1361         if (error == -ENOMEM)
1362                 return NOPAGE_OOM;
1363         return NULL;
1364
1365 page_not_uptodate:
1366         if (!did_readaround) {
1367                 majmin = VM_FAULT_MAJOR;
1368                 inc_page_state(pgmajfault);
1369         }
1370         lock_page(page);
1371
1372         /* Did it get unhashed while we waited for it? */
1373         if (!page->mapping) {
1374                 unlock_page(page);
1375                 page_cache_release(page);
1376                 goto retry_all;
1377         }
1378
1379         /* Did somebody else get it up-to-date? */
1380         if (PageUptodate(page)) {
1381                 unlock_page(page);
1382                 goto success;
1383         }
1384
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          * Umm, take care of errors if the page isn't up-to-date.
1397          * Try to re-read it _once_. We do this synchronously,
1398          * because there really aren't any performance issues here
1399          * and we need to check for errors.
1400          */
1401         lock_page(page);
1402
1403         /* Somebody truncated the page on us? */
1404         if (!page->mapping) {
1405                 unlock_page(page);
1406                 page_cache_release(page);
1407                 goto retry_all;
1408         }
1409
1410         /* Somebody else successfully read it in? */
1411         if (PageUptodate(page)) {
1412                 unlock_page(page);
1413                 goto success;
1414         }
1415         ClearPageError(page);
1416         error = mapping->a_ops->readpage(file, page);
1417         if (!error) {
1418                 wait_on_page_locked(page);
1419                 if (PageUptodate(page))
1420                         goto success;
1421         } else if (error == AOP_TRUNCATED_PAGE) {
1422                 page_cache_release(page);
1423                 goto retry_find;
1424         }
1425
1426         /*
1427          * Things didn't work out. Return zero to tell the
1428          * mm layer so, possibly freeing the page cache page first.
1429          */
1430         page_cache_release(page);
1431         return NULL;
1432 }
1433
1434 EXPORT_SYMBOL(filemap_nopage);
1435
1436 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1437                                         int nonblock)
1438 {
1439         struct address_space *mapping = file->f_mapping;
1440         struct page *page;
1441         int error;
1442
1443         /*
1444          * Do we have something in the page cache already?
1445          */
1446 retry_find:
1447         page = find_get_page(mapping, pgoff);
1448         if (!page) {
1449                 if (nonblock)
1450                         return NULL;
1451                 goto no_cached_page;
1452         }
1453
1454         /*
1455          * Ok, found a page in the page cache, now we need to check
1456          * that it's up-to-date.
1457          */
1458         if (!PageUptodate(page)) {
1459                 if (nonblock) {
1460                         page_cache_release(page);
1461                         return NULL;
1462                 }
1463                 goto page_not_uptodate;
1464         }
1465
1466 success:
1467         /*
1468          * Found the page and have a reference on it.
1469          */
1470         mark_page_accessed(page);
1471         return page;
1472
1473 no_cached_page:
1474         error = page_cache_read(file, pgoff);
1475
1476         /*
1477          * The page we want has now been added to the page cache.
1478          * In the unlikely event that someone removed it in the
1479          * meantime, we'll just come back here and read it again.
1480          */
1481         if (error >= 0)
1482                 goto retry_find;
1483
1484         /*
1485          * An error return from page_cache_read can result if the
1486          * system is low on memory, or a problem occurs while trying
1487          * to schedule I/O.
1488          */
1489         return NULL;
1490
1491 page_not_uptodate:
1492         lock_page(page);
1493
1494         /* Did it get unhashed while we waited for it? */
1495         if (!page->mapping) {
1496                 unlock_page(page);
1497                 goto err;
1498         }
1499
1500         /* Did somebody else get it up-to-date? */
1501         if (PageUptodate(page)) {
1502                 unlock_page(page);
1503                 goto success;
1504         }
1505
1506         error = mapping->a_ops->readpage(file, page);
1507         if (!error) {
1508                 wait_on_page_locked(page);
1509                 if (PageUptodate(page))
1510                         goto success;
1511         } else if (error == AOP_TRUNCATED_PAGE) {
1512                 page_cache_release(page);
1513                 goto retry_find;
1514         }
1515
1516         /*
1517          * Umm, take care of errors if the page isn't up-to-date.
1518          * Try to re-read it _once_. We do this synchronously,
1519          * because there really aren't any performance issues here
1520          * and we need to check for errors.
1521          */
1522         lock_page(page);
1523
1524         /* Somebody truncated the page on us? */
1525         if (!page->mapping) {
1526                 unlock_page(page);
1527                 goto err;
1528         }
1529         /* Somebody else successfully read it in? */
1530         if (PageUptodate(page)) {
1531                 unlock_page(page);
1532                 goto success;
1533         }
1534
1535         ClearPageError(page);
1536         error = mapping->a_ops->readpage(file, page);
1537         if (!error) {
1538                 wait_on_page_locked(page);
1539                 if (PageUptodate(page))
1540                         goto success;
1541         } else if (error == AOP_TRUNCATED_PAGE) {
1542                 page_cache_release(page);
1543                 goto retry_find;
1544         }
1545
1546         /*
1547          * Things didn't work out. Return zero to tell the
1548          * mm layer so, possibly freeing the page cache page first.
1549          */
1550 err:
1551         page_cache_release(page);
1552
1553         return NULL;
1554 }
1555
1556 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1557                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1558                 int nonblock)
1559 {
1560         struct file *file = vma->vm_file;
1561         struct address_space *mapping = file->f_mapping;
1562         struct inode *inode = mapping->host;
1563         unsigned long size;
1564         struct mm_struct *mm = vma->vm_mm;
1565         struct page *page;
1566         int err;
1567
1568         if (!nonblock)
1569                 force_page_cache_readahead(mapping, vma->vm_file,
1570                                         pgoff, len >> PAGE_CACHE_SHIFT);
1571
1572 repeat:
1573         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1574         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1575                 return -EINVAL;
1576
1577         page = filemap_getpage(file, pgoff, nonblock);
1578
1579         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1580          * done in shmem_populate calling shmem_getpage */
1581         if (!page && !nonblock)
1582                 return -ENOMEM;
1583
1584         if (page) {
1585                 err = install_page(mm, vma, addr, page, prot);
1586                 if (err) {
1587                         page_cache_release(page);
1588                         return err;
1589                 }
1590         } else if (vma->vm_flags & VM_NONLINEAR) {
1591                 /* No page was found just because we can't read it in now (being
1592                  * here implies nonblock != 0), but the page may exist, so set
1593                  * the PTE to fault it in later. */
1594                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1595                 if (err)
1596                         return err;
1597         }
1598
1599         len -= PAGE_SIZE;
1600         addr += PAGE_SIZE;
1601         pgoff++;
1602         if (len)
1603                 goto repeat;
1604
1605         return 0;
1606 }
1607 EXPORT_SYMBOL(filemap_populate);
1608
1609 struct vm_operations_struct generic_file_vm_ops = {
1610         .nopage         = filemap_nopage,
1611         .populate       = filemap_populate,
1612 };
1613
1614 /* This is used for a general mmap of a disk file */
1615
1616 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1617 {
1618         struct address_space *mapping = file->f_mapping;
1619
1620         if (!mapping->a_ops->readpage)
1621                 return -ENOEXEC;
1622         file_accessed(file);
1623         vma->vm_ops = &generic_file_vm_ops;
1624         return 0;
1625 }
1626
1627 /*
1628  * This is for filesystems which do not implement ->writepage.
1629  */
1630 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1631 {
1632         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1633                 return -EINVAL;
1634         return generic_file_mmap(file, vma);
1635 }
1636 #else
1637 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1638 {
1639         return -ENOSYS;
1640 }
1641 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1642 {
1643         return -ENOSYS;
1644 }
1645 #endif /* CONFIG_MMU */
1646
1647 EXPORT_SYMBOL(generic_file_mmap);
1648 EXPORT_SYMBOL(generic_file_readonly_mmap);
1649
1650 static inline struct page *__read_cache_page(struct address_space *mapping,
1651                                 unsigned long index,
1652                                 int (*filler)(void *,struct page*),
1653                                 void *data)
1654 {
1655         struct page *page, *cached_page = NULL;
1656         int err;
1657 repeat:
1658         page = find_get_page(mapping, index);
1659         if (!page) {
1660                 if (!cached_page) {
1661                         cached_page = page_cache_alloc_cold(mapping);
1662                         if (!cached_page)
1663                                 return ERR_PTR(-ENOMEM);
1664                 }
1665                 err = add_to_page_cache_lru(cached_page, mapping,
1666                                         index, GFP_KERNEL);
1667                 if (err == -EEXIST)
1668                         goto repeat;
1669                 if (err < 0) {
1670                         /* Presumably ENOMEM for radix tree node */
1671                         page_cache_release(cached_page);
1672                         return ERR_PTR(err);
1673                 }
1674                 page = cached_page;
1675                 cached_page = NULL;
1676                 err = filler(data, page);
1677                 if (err < 0) {
1678                         page_cache_release(page);
1679                         page = ERR_PTR(err);
1680                 }
1681         }
1682         if (cached_page)
1683                 page_cache_release(cached_page);
1684         return page;
1685 }
1686
1687 /*
1688  * Read into the page cache. If a page already exists,
1689  * and PageUptodate() is not set, try to fill the page.
1690  */
1691 struct page *read_cache_page(struct address_space *mapping,
1692                                 unsigned long index,
1693                                 int (*filler)(void *,struct page*),
1694                                 void *data)
1695 {
1696         struct page *page;
1697         int err;
1698
1699 retry:
1700         page = __read_cache_page(mapping, index, filler, data);
1701         if (IS_ERR(page))
1702                 goto out;
1703         mark_page_accessed(page);
1704         if (PageUptodate(page))
1705                 goto out;
1706
1707         lock_page(page);
1708         if (!page->mapping) {
1709                 unlock_page(page);
1710                 page_cache_release(page);
1711                 goto retry;
1712         }
1713         if (PageUptodate(page)) {
1714                 unlock_page(page);
1715                 goto out;
1716         }
1717         err = filler(data, page);
1718         if (err < 0) {
1719                 page_cache_release(page);
1720                 page = ERR_PTR(err);
1721         }
1722  out:
1723         return page;
1724 }
1725
1726 EXPORT_SYMBOL(read_cache_page);
1727
1728 /*
1729  * If the page was newly created, increment its refcount and add it to the
1730  * caller's lru-buffering pagevec.  This function is specifically for
1731  * generic_file_write().
1732  */
1733 static inline struct page *
1734 __grab_cache_page(struct address_space *mapping, unsigned long index,
1735                         struct page **cached_page, struct pagevec *lru_pvec)
1736 {
1737         int err;
1738         struct page *page;
1739 repeat:
1740         page = find_lock_page(mapping, index);
1741         if (!page) {
1742                 if (!*cached_page) {
1743                         *cached_page = page_cache_alloc(mapping);
1744                         if (!*cached_page)
1745                                 return NULL;
1746                 }
1747                 err = add_to_page_cache(*cached_page, mapping,
1748                                         index, GFP_KERNEL);
1749                 if (err == -EEXIST)
1750                         goto repeat;
1751                 if (err == 0) {
1752                         page = *cached_page;
1753                         page_cache_get(page);
1754                         if (!pagevec_add(lru_pvec, page))
1755                                 __pagevec_lru_add(lru_pvec);
1756                         *cached_page = NULL;
1757                 }
1758         }
1759         return page;
1760 }
1761
1762 /*
1763  * The logic we want is
1764  *
1765  *      if suid or (sgid and xgrp)
1766  *              remove privs
1767  */
1768 int remove_suid(struct dentry *dentry)
1769 {
1770         mode_t mode = dentry->d_inode->i_mode;
1771         int kill = 0;
1772         int result = 0;
1773
1774         /* suid always must be killed */
1775         if (unlikely(mode & S_ISUID))
1776                 kill = ATTR_KILL_SUID;
1777
1778         /*
1779          * sgid without any exec bits is just a mandatory locking mark; leave
1780          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1781          */
1782         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1783                 kill |= ATTR_KILL_SGID;
1784
1785         if (unlikely(kill && !capable(CAP_FSETID))) {
1786                 struct iattr newattrs;
1787
1788                 newattrs.ia_valid = ATTR_FORCE | kill;
1789                 result = notify_change(dentry, &newattrs);
1790         }
1791         return result;
1792 }
1793 EXPORT_SYMBOL(remove_suid);
1794
1795 size_t
1796 __filemap_copy_from_user_iovec(char *vaddr,
1797                         const struct iovec *iov, size_t base, size_t bytes)
1798 {
1799         size_t copied = 0, left = 0;
1800
1801         while (bytes) {
1802                 char __user *buf = iov->iov_base + base;
1803                 int copy = min(bytes, iov->iov_len - base);
1804
1805                 base = 0;
1806                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1807                 copied += copy;
1808                 bytes -= copy;
1809                 vaddr += copy;
1810                 iov++;
1811
1812                 if (unlikely(left)) {
1813                         /* zero the rest of the target like __copy_from_user */
1814                         if (bytes)
1815                                 memset(vaddr, 0, bytes);
1816                         break;
1817                 }
1818         }
1819         return copied - left;
1820 }
1821
1822 /*
1823  * Performs necessary checks before doing a write
1824  *
1825  * Can adjust writing position aor amount of bytes to write.
1826  * Returns appropriate error code that caller should return or
1827  * zero in case that write should be allowed.
1828  */
1829 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1830 {
1831         struct inode *inode = file->f_mapping->host;
1832         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1833
1834         if (unlikely(*pos < 0))
1835                 return -EINVAL;
1836
1837         if (!isblk) {
1838                 /* FIXME: this is for backwards compatibility with 2.4 */
1839                 if (file->f_flags & O_APPEND)
1840                         *pos = i_size_read(inode);
1841
1842                 if (limit != RLIM_INFINITY) {
1843                         if (*pos >= limit) {
1844                                 send_sig(SIGXFSZ, current, 0);
1845                                 return -EFBIG;
1846                         }
1847                         if (*count > limit - (typeof(limit))*pos) {
1848                                 *count = limit - (typeof(limit))*pos;
1849                         }
1850                 }
1851         }
1852
1853         /*
1854          * LFS rule
1855          */
1856         if (unlikely(*pos + *count > MAX_NON_LFS &&
1857                                 !(file->f_flags & O_LARGEFILE))) {
1858                 if (*pos >= MAX_NON_LFS) {
1859                         send_sig(SIGXFSZ, current, 0);
1860                         return -EFBIG;
1861                 }
1862                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1863                         *count = MAX_NON_LFS - (unsigned long)*pos;
1864                 }
1865         }
1866
1867         /*
1868          * Are we about to exceed the fs block limit ?
1869          *
1870          * If we have written data it becomes a short write.  If we have
1871          * exceeded without writing data we send a signal and return EFBIG.
1872          * Linus frestrict idea will clean these up nicely..
1873          */
1874         if (likely(!isblk)) {
1875                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1876                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1877                                 send_sig(SIGXFSZ, current, 0);
1878                                 return -EFBIG;
1879                         }
1880                         /* zero-length writes at ->s_maxbytes are OK */
1881                 }
1882
1883                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1884                         *count = inode->i_sb->s_maxbytes - *pos;
1885         } else {
1886                 loff_t isize;
1887                 if (bdev_read_only(I_BDEV(inode)))
1888                         return -EPERM;
1889                 isize = i_size_read(inode);
1890                 if (*pos >= isize) {
1891                         if (*count || *pos > isize)
1892                                 return -ENOSPC;
1893                 }
1894
1895                 if (*pos + *count > isize)
1896                         *count = isize - *pos;
1897         }
1898         return 0;
1899 }
1900 EXPORT_SYMBOL(generic_write_checks);
1901
1902 ssize_t
1903 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1904                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1905                 size_t count, size_t ocount)
1906 {
1907         struct file     *file = iocb->ki_filp;
1908         struct address_space *mapping = file->f_mapping;
1909         struct inode    *inode = mapping->host;
1910         ssize_t         written;
1911
1912         if (count != ocount)
1913                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1914
1915         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1916         if (written > 0) {
1917                 loff_t end = pos + written;
1918                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1919                         i_size_write(inode,  end);
1920                         mark_inode_dirty(inode);
1921                 }
1922                 *ppos = end;
1923         }
1924
1925         /*
1926          * Sync the fs metadata but not the minor inode changes and
1927          * of course not the data as we did direct DMA for the IO.
1928          * i_mutex is held, which protects generic_osync_inode() from
1929          * livelocking.
1930          */
1931         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1932                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1933                 if (err < 0)
1934                         written = err;
1935         }
1936         if (written == count && !is_sync_kiocb(iocb))
1937                 written = -EIOCBQUEUED;
1938         return written;
1939 }
1940 EXPORT_SYMBOL(generic_file_direct_write);
1941
1942 ssize_t
1943 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1944                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1945                 size_t count, ssize_t written)
1946 {
1947         struct file *file = iocb->ki_filp;
1948         struct address_space * mapping = file->f_mapping;
1949         struct address_space_operations *a_ops = mapping->a_ops;
1950         struct inode    *inode = mapping->host;
1951         long            status = 0;
1952         struct page     *page;
1953         struct page     *cached_page = NULL;
1954         size_t          bytes;
1955         struct pagevec  lru_pvec;
1956         const struct iovec *cur_iov = iov; /* current iovec */
1957         size_t          iov_base = 0;      /* offset in the current iovec */
1958         char __user     *buf;
1959
1960         pagevec_init(&lru_pvec, 0);
1961
1962         /*
1963          * handle partial DIO write.  Adjust cur_iov if needed.
1964          */
1965         if (likely(nr_segs == 1))
1966                 buf = iov->iov_base + written;
1967         else {
1968                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1969                 buf = cur_iov->iov_base + iov_base;
1970         }
1971
1972         do {
1973                 unsigned long index;
1974                 unsigned long offset;
1975                 unsigned long maxlen;
1976                 size_t copied;
1977
1978                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1979                 index = pos >> PAGE_CACHE_SHIFT;
1980                 bytes = PAGE_CACHE_SIZE - offset;
1981                 if (bytes > count)
1982                         bytes = count;
1983
1984                 /*
1985                  * Bring in the user page that we will copy from _first_.
1986                  * Otherwise there's a nasty deadlock on copying from the
1987                  * same page as we're writing to, without it being marked
1988                  * up-to-date.
1989                  */
1990                 maxlen = cur_iov->iov_len - iov_base;
1991                 if (maxlen > bytes)
1992                         maxlen = bytes;
1993                 fault_in_pages_readable(buf, maxlen);
1994
1995                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1996                 if (!page) {
1997                         status = -ENOMEM;
1998                         break;
1999                 }
2000
2001                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2002                 if (unlikely(status)) {
2003                         loff_t isize = i_size_read(inode);
2004
2005                         if (status != AOP_TRUNCATED_PAGE)
2006                                 unlock_page(page);
2007                         page_cache_release(page);
2008                         if (status == AOP_TRUNCATED_PAGE)
2009                                 continue;
2010                         /*
2011                          * prepare_write() may have instantiated a few blocks
2012                          * outside i_size.  Trim these off again.
2013                          */
2014                         if (pos + bytes > isize)
2015                                 vmtruncate(inode, isize);
2016                         break;
2017                 }
2018                 if (likely(nr_segs == 1))
2019                         copied = filemap_copy_from_user(page, offset,
2020                                                         buf, bytes);
2021                 else
2022                         copied = filemap_copy_from_user_iovec(page, offset,
2023                                                 cur_iov, iov_base, bytes);
2024                 flush_dcache_page(page);
2025                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2026                 if (status == AOP_TRUNCATED_PAGE) {
2027                         page_cache_release(page);
2028                         continue;
2029                 }
2030                 if (likely(copied > 0)) {
2031                         if (!status)
2032                                 status = copied;
2033
2034                         if (status >= 0) {
2035                                 written += status;
2036                                 count -= status;
2037                                 pos += status;
2038                                 buf += status;
2039                                 if (unlikely(nr_segs > 1)) {
2040                                         filemap_set_next_iovec(&cur_iov,
2041                                                         &iov_base, status);
2042                                         if (count)
2043                                                 buf = cur_iov->iov_base +
2044                                                         iov_base;
2045                                 } else {
2046                                         iov_base += status;
2047                                 }
2048                         }
2049                 }
2050                 if (unlikely(copied != bytes))
2051                         if (status >= 0)
2052                                 status = -EFAULT;
2053                 unlock_page(page);
2054                 mark_page_accessed(page);
2055                 page_cache_release(page);
2056                 if (status < 0)
2057                         break;
2058                 balance_dirty_pages_ratelimited(mapping);
2059                 cond_resched();
2060         } while (count);
2061         *ppos = pos;
2062
2063         if (cached_page)
2064                 page_cache_release(cached_page);
2065
2066         /*
2067          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2068          */
2069         if (likely(status >= 0)) {
2070                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2071                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2072                                 status = generic_osync_inode(inode, mapping,
2073                                                 OSYNC_METADATA|OSYNC_DATA);
2074                 }
2075         }
2076
2077         /*
2078          * If we get here for O_DIRECT writes then we must have fallen through
2079          * to buffered writes (block instantiation inside i_size).  So we sync
2080          * the file data here, to try to honour O_DIRECT expectations.
2081          */
2082         if (unlikely(file->f_flags & O_DIRECT) && written)
2083                 status = filemap_write_and_wait(mapping);
2084
2085         pagevec_lru_add(&lru_pvec);
2086         return written ? written : status;
2087 }
2088 EXPORT_SYMBOL(generic_file_buffered_write);
2089
2090 static ssize_t
2091 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2092                                 unsigned long nr_segs, loff_t *ppos)
2093 {
2094         struct file *file = iocb->ki_filp;
2095         struct address_space * mapping = file->f_mapping;
2096         size_t ocount;          /* original count */
2097         size_t count;           /* after file limit checks */
2098         struct inode    *inode = mapping->host;
2099         unsigned long   seg;
2100         loff_t          pos;
2101         ssize_t         written;
2102         ssize_t         err;
2103
2104         ocount = 0;
2105         for (seg = 0; seg < nr_segs; seg++) {
2106                 const struct iovec *iv = &iov[seg];
2107
2108                 /*
2109                  * If any segment has a negative length, or the cumulative
2110                  * length ever wraps negative then return -EINVAL.
2111                  */
2112                 ocount += iv->iov_len;
2113                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2114                         return -EINVAL;
2115                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2116                         continue;
2117                 if (seg == 0)
2118                         return -EFAULT;
2119                 nr_segs = seg;
2120                 ocount -= iv->iov_len;  /* This segment is no good */
2121                 break;
2122         }
2123
2124         count = ocount;
2125         pos = *ppos;
2126
2127         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2128
2129         /* We can write back this queue in page reclaim */
2130         current->backing_dev_info = mapping->backing_dev_info;
2131         written = 0;
2132
2133         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2134         if (err)
2135                 goto out;
2136
2137         if (count == 0)
2138                 goto out;
2139
2140         err = remove_suid(file->f_dentry);
2141         if (err)
2142                 goto out;
2143
2144         file_update_time(file);
2145
2146         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2147         if (unlikely(file->f_flags & O_DIRECT)) {
2148                 written = generic_file_direct_write(iocb, iov,
2149                                 &nr_segs, pos, ppos, count, ocount);
2150                 if (written < 0 || written == count)
2151                         goto out;
2152                 /*
2153                  * direct-io write to a hole: fall through to buffered I/O
2154                  * for completing the rest of the request.
2155                  */
2156                 pos += written;
2157                 count -= written;
2158         }
2159
2160         written = generic_file_buffered_write(iocb, iov, nr_segs,
2161                         pos, ppos, count, written);
2162 out:
2163         current->backing_dev_info = NULL;
2164         return written ? written : err;
2165 }
2166 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2167
2168 ssize_t
2169 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2170                                 unsigned long nr_segs, loff_t *ppos)
2171 {
2172         struct file *file = iocb->ki_filp;
2173         struct address_space *mapping = file->f_mapping;
2174         struct inode *inode = mapping->host;
2175         ssize_t ret;
2176         loff_t pos = *ppos;
2177
2178         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2179
2180         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2181                 int err;
2182
2183                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2184                 if (err < 0)
2185                         ret = err;
2186         }
2187         return ret;
2188 }
2189
2190 static ssize_t
2191 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2192                                 unsigned long nr_segs, loff_t *ppos)
2193 {
2194         struct kiocb kiocb;
2195         ssize_t ret;
2196
2197         init_sync_kiocb(&kiocb, file);
2198         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2199         if (ret == -EIOCBQUEUED)
2200                 ret = wait_on_sync_kiocb(&kiocb);
2201         return ret;
2202 }
2203
2204 ssize_t
2205 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2206                                 unsigned long nr_segs, loff_t *ppos)
2207 {
2208         struct kiocb kiocb;
2209         ssize_t ret;
2210
2211         init_sync_kiocb(&kiocb, file);
2212         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2213         if (-EIOCBQUEUED == ret)
2214                 ret = wait_on_sync_kiocb(&kiocb);
2215         return ret;
2216 }
2217 EXPORT_SYMBOL(generic_file_write_nolock);
2218
2219 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2220                                size_t count, loff_t pos)
2221 {
2222         struct file *file = iocb->ki_filp;
2223         struct address_space *mapping = file->f_mapping;
2224         struct inode *inode = mapping->host;
2225         ssize_t ret;
2226         struct iovec local_iov = { .iov_base = (void __user *)buf,
2227                                         .iov_len = count };
2228
2229         BUG_ON(iocb->ki_pos != pos);
2230
2231         mutex_lock(&inode->i_mutex);
2232         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2233                                                 &iocb->ki_pos);
2234         mutex_unlock(&inode->i_mutex);
2235
2236         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2237                 ssize_t err;
2238
2239                 err = sync_page_range(inode, mapping, pos, ret);
2240                 if (err < 0)
2241                         ret = err;
2242         }
2243         return ret;
2244 }
2245 EXPORT_SYMBOL(generic_file_aio_write);
2246
2247 ssize_t generic_file_write(struct file *file, const char __user *buf,
2248                            size_t count, loff_t *ppos)
2249 {
2250         struct address_space *mapping = file->f_mapping;
2251         struct inode *inode = mapping->host;
2252         ssize_t ret;
2253         struct iovec local_iov = { .iov_base = (void __user *)buf,
2254                                         .iov_len = count };
2255
2256         mutex_lock(&inode->i_mutex);
2257         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2258         mutex_unlock(&inode->i_mutex);
2259
2260         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2261                 ssize_t err;
2262
2263                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2264                 if (err < 0)
2265                         ret = err;
2266         }
2267         return ret;
2268 }
2269 EXPORT_SYMBOL(generic_file_write);
2270
2271 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2272                         unsigned long nr_segs, loff_t *ppos)
2273 {
2274         struct kiocb kiocb;
2275         ssize_t ret;
2276
2277         init_sync_kiocb(&kiocb, filp);
2278         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2279         if (-EIOCBQUEUED == ret)
2280                 ret = wait_on_sync_kiocb(&kiocb);
2281         return ret;
2282 }
2283 EXPORT_SYMBOL(generic_file_readv);
2284
2285 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2286                         unsigned long nr_segs, loff_t *ppos)
2287 {
2288         struct address_space *mapping = file->f_mapping;
2289         struct inode *inode = mapping->host;
2290         ssize_t ret;
2291
2292         mutex_lock(&inode->i_mutex);
2293         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2294         mutex_unlock(&inode->i_mutex);
2295
2296         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2297                 int err;
2298
2299                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2300                 if (err < 0)
2301                         ret = err;
2302         }
2303         return ret;
2304 }
2305 EXPORT_SYMBOL(generic_file_writev);
2306
2307 /*
2308  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2309  * went wrong during pagecache shootdown.
2310  */
2311 static ssize_t
2312 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2313         loff_t offset, unsigned long nr_segs)
2314 {
2315         struct file *file = iocb->ki_filp;
2316         struct address_space *mapping = file->f_mapping;
2317         ssize_t retval;
2318         size_t write_len = 0;
2319
2320         /*
2321          * If it's a write, unmap all mmappings of the file up-front.  This
2322          * will cause any pte dirty bits to be propagated into the pageframes
2323          * for the subsequent filemap_write_and_wait().
2324          */
2325         if (rw == WRITE) {
2326                 write_len = iov_length(iov, nr_segs);
2327                 if (mapping_mapped(mapping))
2328                         unmap_mapping_range(mapping, offset, write_len, 0);
2329         }
2330
2331         retval = filemap_write_and_wait(mapping);
2332         if (retval == 0) {
2333                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2334                                                 offset, nr_segs);
2335                 if (rw == WRITE && mapping->nrpages) {
2336                         pgoff_t end = (offset + write_len - 1)
2337                                                 >> PAGE_CACHE_SHIFT;
2338                         int err = invalidate_inode_pages2_range(mapping,
2339                                         offset >> PAGE_CACHE_SHIFT, end);
2340                         if (err)
2341                                 retval = err;
2342                 }
2343         }
2344         return retval;
2345 }