mm/filemap.c

   1 /*
   2  *      linux/mm/filemap.c
   3  *
   4  * Copyright (C) 1994-1999  Linus Torvalds
   5  */
   6
   7 /*
   8  * This file handles the generic file mmap semantics used by
   9  * most "normal" filesystems (but you don't /have/ to use this:
  10  * the NFS filesystem used to do this differently, for example)
  11  */
  12 #include <linux/config.h>
  13 #include <linux/module.h>
  14 #include <linux/slab.h>
  15 #include <linux/compiler.h>
  16 #include <linux/fs.h>
  17 #include <linux/aio.h>
  18 #include <linux/kernel_stat.h>
  19 #include <linux/mm.h>
  20 #include <linux/swap.h>
  21 #include <linux/mman.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/file.h>
  24 #include <linux/uio.h>
  25 #include <linux/hash.h>
  26 #include <linux/writeback.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/security.h>
  30 /*
  31  * This is needed for the following functions:
  32  *  - try_to_release_page
  33  *  - block_invalidatepage
  34  *  - generic_osync_inode
  35  *
  36  * FIXME: remove all knowledge of the buffer layer from the core VM
  37  */
  38 #include <linux/buffer_head.h> /* for generic_osync_inode */
  39
  40 #include <asm/uaccess.h>
  41 #include <asm/mman.h>
  42
  43 /*
  44  * Shared mappings implemented 30.11.1994. It's not fully working yet,
  45  * though.
  46  *
  47  * Shared mappings now work. 15.8.1995  Bruno.
  48  *
  49  * finished 'unifying' the page and buffer cache and SMP-threaded the
  50  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
  51  *
  52  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
  53  */
  54
  55 /*
  56  * Lock ordering:
  57  *
  58  *  ->i_mmap_lock               (vmtruncate)
  59  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
  60  *      ->swap_list_lock
  61  *        ->swap_device_lock    (exclusive_swap_page, others)
  62  *          ->mapping->tree_lock
  63  *    ->page_map_lock()         (try_to_unmap_file)
  64  *
  65  *  ->i_sem
  66  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
  67  *
  68  *  ->mmap_sem
  69  *    ->i_mmap_lock
  70  *      ->page_table_lock       (various places, mainly in mmap.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_sem                   (msync)
  78  *
  79  *  ->i_sem
  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  *  ->page_table_lock
  87  *    ->swap_device_lock        (try_to_unmap_one)
  88  *    ->private_lock            (try_to_unmap_one)
  89  *    ->tree_lock               (try_to_unmap_one)
  90  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
  91  *    ->page_map_lock()         (page_add_anon_rmap)
  92  *      ->tree_lock             (page_remove_rmap->set_page_dirty)
  93  *      ->private_lock          (page_remove_rmap->set_page_dirty)
  94  *      ->inode_lock            (page_remove_rmap->set_page_dirty)
  95  *    ->anon_vma.lock           (anon_vma_prepare)
  96  *    ->inode_lock              (zap_pte_range->set_page_dirty)
  97  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
  98  *
  99  *  ->task->proc_lock
 100  *    ->dcache_lock             (proc_pid_lookup)
 101  */
 102
 103 /*
 104  * Remove a page from the page cache and free it. Caller has to make
 105  * sure the page is locked and that nobody else uses it - or that usage
 106  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
 107  */
 108 void __remove_from_page_cache(struct page *page)
 109 {
 110         struct address_space *mapping = page->mapping;
 111
 112         radix_tree_delete(&mapping->page_tree, page->index);
 113         page->mapping = NULL;
 114         mapping->nrpages--;
 115         pagecache_acct(-1);
 116 }
 117
 118 void remove_from_page_cache(struct page *page)
 119 {
 120         struct address_space *mapping = page->mapping;
 121
 122         if (unlikely(!PageLocked(page)))
 123                 PAGE_BUG(page);
 124
 125         spin_lock_irq(&mapping->tree_lock);
 126         __remove_from_page_cache(page);
 127         spin_unlock_irq(&mapping->tree_lock);
 128 }
 129
 130 static inline int sync_page(struct page *page)
 131 {
 132         struct address_space *mapping;
 133
 134         /*
 135          * FIXME, fercrissake.  What is this barrier here for?
 136          */
 137         smp_mb();
 138         mapping = page_mapping(page);
 139         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 140                 return mapping->a_ops->sync_page(page);
 141         return 0;
 142 }
 143
 144 /**
 145  * filemap_fdatawrite - start writeback against all of a mapping's dirty pages
 146  * @mapping: address space structure to write
 147  *
 148  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 149  * opposed to a regular memory * cleansing writeback.  The difference between
 150  * these two operations is that if a dirty page/buffer is encountered, it must
 151  * be waited upon, and not just skipped over.
 152  */
 153 static int __filemap_fdatawrite(struct address_space *mapping, int sync_mode)
 154 {
 155         int ret;
 156         struct writeback_control wbc = {
 157                 .sync_mode = sync_mode,
 158                 .nr_to_write = mapping->nrpages * 2,
 159         };
 160
 161         if (mapping->backing_dev_info->memory_backed)
 162                 return 0;
 163
 164         ret = do_writepages(mapping, &wbc);
 165         return ret;
 166 }
 167
 168 int filemap_fdatawrite(struct address_space *mapping)
 169 {
 170         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 171 }
 172 EXPORT_SYMBOL(filemap_fdatawrite);
 173
 174 /*
 175  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 176  * purposes - I/O may not be started against all dirty pages.
 177  */
 178 int filemap_flush(struct address_space *mapping)
 179 {
 180         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 181 }
 182 EXPORT_SYMBOL(filemap_flush);
 183
 184 /*
 185  * Wait for writeback to complete against pages indexed by start->end
 186  * inclusive
 187  */
 188 static int wait_on_page_writeback_range(struct address_space *mapping,
 189                                 pgoff_t start, pgoff_t end)
 190 {
 191         struct pagevec pvec;
 192         int nr_pages;
 193         int ret = 0;
 194         pgoff_t index;
 195
 196         if (end < start)
 197                 return 0;
 198
 199         pagevec_init(&pvec, 0);
 200         index = start;
 201         while ((nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 202                         PAGECACHE_TAG_WRITEBACK,
 203                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 204                 unsigned i;
 205
 206                 for (i = 0; i < nr_pages; i++) {
 207                         struct page *page = pvec.pages[i];
 208
 209                         wait_on_page_writeback(page);
 210                         if (PageError(page))
 211                                 ret = -EIO;
 212                 }
 213                 pagevec_release(&pvec);
 214                 cond_resched();
 215         }
 216
 217         /* Check for outstanding write errors */
 218         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 219                 ret = -ENOSPC;
 220         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 221                 ret = -EIO;
 222
 223         return ret;
 224 }
 225
 226 /**
 227  * filemap_fdatawait - walk the list of under-writeback pages of the given
 228  *     address space and wait for all of them.
 229  *
 230  * @mapping: address space structure to wait for
 231  */
 232 int filemap_fdatawait(struct address_space *mapping)
 233 {
 234         return wait_on_page_writeback_range(mapping, 0, -1);
 235 }
 236
 237 EXPORT_SYMBOL(filemap_fdatawait);
 238
 239 int filemap_write_and_wait(struct address_space *mapping)
 240 {
 241         int retval = 0;
 242
 243         if (mapping->nrpages) {
 244                 retval = filemap_fdatawrite(mapping);
 245                 if (retval == 0)
 246                         retval = filemap_fdatawait(mapping);
 247         }
 248         return retval;
 249 }
 250
 251 /*
 252  * This function is used to add newly allocated pagecache pages:
 253  * the page is new, so we can just run SetPageLocked() against it.
 254  * The other page state flags were set by rmqueue().
 255  *
 256  * This function does not add the page to the LRU.  The caller must do that.
 257  */
 258 int add_to_page_cache(struct page *page, struct address_space *mapping,
 259                 pgoff_t offset, int gfp_mask)
 260 {
 261         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 262
 263         if (error == 0) {
 264                 spin_lock_irq(&mapping->tree_lock);
 265                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 266                 if (!error) {
 267                         page_cache_get(page);
 268                         SetPageLocked(page);
 269                         page->mapping = mapping;
 270                         page->index = offset;
 271                         mapping->nrpages++;
 272                         pagecache_acct(1);
 273                 }
 274                 spin_unlock_irq(&mapping->tree_lock);
 275                 radix_tree_preload_end();
 276         }
 277         return error;
 278 }
 279
 280 EXPORT_SYMBOL(add_to_page_cache);
 281
 282 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 283                                 pgoff_t offset, int gfp_mask)
 284 {
 285         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 286         if (ret == 0)
 287                 lru_cache_add(page);
 288         return ret;
 289 }
 290
 291 /*
 292  * In order to wait for pages to become available there must be
 293  * waitqueues associated with pages. By using a hash table of
 294  * waitqueues where the bucket discipline is to maintain all
 295  * waiters on the same queue and wake all when any of the pages
 296  * become available, and for the woken contexts to check to be
 297  * sure the appropriate page became available, this saves space
 298  * at a cost of "thundering herd" phenomena during rare hash
 299  * collisions.
 300  */
 301 struct page_wait_queue {
 302         struct page *page;
 303         int bit;
 304         wait_queue_t wait;
 305 };
 306
 307 static int page_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
 308 {
 309         struct page *page = key;
 310         struct page_wait_queue *wq;
 311
 312         wq = container_of(wait, struct page_wait_queue, wait);
 313         if (wq->page != page || test_bit(wq->bit, &page->flags))
 314                 return 0;
 315         else
 316                 return autoremove_wake_function(wait, mode, sync, NULL);
 317 }
 318
 319 #define __DEFINE_PAGE_WAIT(name, p, b, f)                               \
 320         struct page_wait_queue name = {                                 \
 321                 .page   = p,                                            \
 322                 .bit    = b,                                            \
 323                 .wait   = {                                             \
 324                         .task   = current,                              \
 325                         .func   = page_wake_function,                   \
 326                         .flags  = f,                                    \
 327                         .task_list = LIST_HEAD_INIT(name.wait.task_list),\
 328                 },                                                      \
 329         }
 330
 331 #define DEFINE_PAGE_WAIT(name, p, b)    __DEFINE_PAGE_WAIT(name, p, b, 0)
 332 #define DEFINE_PAGE_WAIT_EXCLUSIVE(name, p, b)                          \
 333                 __DEFINE_PAGE_WAIT(name, p, b, WQ_FLAG_EXCLUSIVE)
 334
 335 static wait_queue_head_t *page_waitqueue(struct page *page)
 336 {
 337         const struct zone *zone = page_zone(page);
 338
 339         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 340 }
 341
 342 static void wake_up_page(struct page *page)
 343 {
 344         const unsigned int mode = TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE;
 345         wait_queue_head_t *waitqueue = page_waitqueue(page);
 346
 347         if (waitqueue_active(waitqueue))
 348                 __wake_up(waitqueue, mode, 1, page);
 349 }
 350
 351 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 352 {
 353         wait_queue_head_t *waitqueue = page_waitqueue(page);
 354         DEFINE_PAGE_WAIT(wait, page, bit_nr);
 355
 356         do {
 357                 prepare_to_wait(waitqueue, &wait.wait, TASK_UNINTERRUPTIBLE);
 358                 if (test_bit(bit_nr, &page->flags)) {
 359                         sync_page(page);
 360                         io_schedule();
 361                 }
 362         } while (test_bit(bit_nr, &page->flags));
 363         finish_wait(waitqueue, &wait.wait);
 364 }
 365
 366 EXPORT_SYMBOL(wait_on_page_bit);
 367
 368 /**
 369  * unlock_page() - unlock a locked page
 370  *
 371  * @page: the page
 372  *
 373  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 374  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 375  * mechananism between PageLocked pages and PageWriteback pages is shared.
 376  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 377  *
 378  * The first mb is necessary to safely close the critical section opened by the
 379  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 380  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 381  * parallel wait_on_page_locked()).
 382  */
 383 void fastcall unlock_page(struct page *page)
 384 {
 385         smp_mb__before_clear_bit();
 386         if (!TestClearPageLocked(page))
 387                 BUG();
 388         smp_mb__after_clear_bit();
 389         wake_up_page(page);
 390 }
 391
 392 EXPORT_SYMBOL(unlock_page);
 393 EXPORT_SYMBOL(lock_page);
 394
 395 /*
 396  * End writeback against a page.
 397  */
 398 void end_page_writeback(struct page *page)
 399 {
 400         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 401                 if (!test_clear_page_writeback(page))
 402                         BUG();
 403                 smp_mb__after_clear_bit();
 404         }
 405         wake_up_page(page);
 406 }
 407
 408 EXPORT_SYMBOL(end_page_writeback);
 409
 410 /*
 411  * Get a lock on the page, assuming we need to sleep to get it.
 412  *
 413  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 414  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 415  * chances are that on the second loop, the block layer's plug list is empty,
 416  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 417  */
 418 void fastcall __lock_page(struct page *page)
 419 {
 420         wait_queue_head_t *wqh = page_waitqueue(page);
 421         DEFINE_PAGE_WAIT_EXCLUSIVE(wait, page, PG_locked);
 422
 423         while (TestSetPageLocked(page)) {
 424                 prepare_to_wait_exclusive(wqh, &wait.wait, TASK_UNINTERRUPTIBLE);
 425                 if (PageLocked(page)) {
 426                         sync_page(page);
 427                         io_schedule();
 428                 }
 429         }
 430         finish_wait(wqh, &wait.wait);
 431 }
 432
 433 EXPORT_SYMBOL(__lock_page);
 434
 435 /*
 436  * a rather lightweight function, finding and getting a reference to a
 437  * hashed page atomically.
 438  */
 439 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 440 {
 441         struct page *page;
 442
 443         spin_lock_irq(&mapping->tree_lock);
 444         page = radix_tree_lookup(&mapping->page_tree, offset);
 445         if (page)
 446                 page_cache_get(page);
 447         spin_unlock_irq(&mapping->tree_lock);
 448         return page;
 449 }
 450
 451 EXPORT_SYMBOL(find_get_page);
 452
 453 /*
 454  * Same as above, but trylock it instead of incrementing the count.
 455  */
 456 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 457 {
 458         struct page *page;
 459
 460         spin_lock_irq(&mapping->tree_lock);
 461         page = radix_tree_lookup(&mapping->page_tree, offset);
 462         if (page && TestSetPageLocked(page))
 463                 page = NULL;
 464         spin_unlock_irq(&mapping->tree_lock);
 465         return page;
 466 }
 467
 468 EXPORT_SYMBOL(find_trylock_page);
 469
 470 /**
 471  * find_lock_page - locate, pin and lock a pagecache page
 472  *
 473  * @mapping - the address_space to search
 474  * @offset - the page index
 475  *
 476  * Locates the desired pagecache page, locks it, increments its reference
 477  * count and returns its address.
 478  *
 479  * Returns zero if the page was not present. find_lock_page() may sleep.
 480  */
 481 struct page *find_lock_page(struct address_space *mapping,
 482                                 unsigned long offset)
 483 {
 484         struct page *page;
 485
 486         spin_lock_irq(&mapping->tree_lock);
 487 repeat:
 488         page = radix_tree_lookup(&mapping->page_tree, offset);
 489         if (page) {
 490                 page_cache_get(page);
 491                 if (TestSetPageLocked(page)) {
 492                         spin_unlock_irq(&mapping->tree_lock);
 493                         lock_page(page);
 494                         spin_lock_irq(&mapping->tree_lock);
 495
 496                         /* Has the page been truncated while we slept? */
 497                         if (page->mapping != mapping || page->index != offset) {
 498                                 unlock_page(page);
 499                                 page_cache_release(page);
 500                                 goto repeat;
 501                         }
 502                 }
 503         }
 504         spin_unlock_irq(&mapping->tree_lock);
 505         return page;
 506 }
 507
 508 EXPORT_SYMBOL(find_lock_page);
 509
 510 /**
 511  * find_or_create_page - locate or add a pagecache page
 512  *
 513  * @mapping - the page's address_space
 514  * @index - the page's index into the mapping
 515  * @gfp_mask - page allocation mode
 516  *
 517  * Locates a page in the pagecache.  If the page is not present, a new page
 518  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 519  * LRU list.  The returned page is locked and has its reference count
 520  * incremented.
 521  *
 522  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 523  * allocation!
 524  *
 525  * find_or_create_page() returns the desired page's address, or zero on
 526  * memory exhaustion.
 527  */
 528 struct page *find_or_create_page(struct address_space *mapping,
 529                 unsigned long index, unsigned int gfp_mask)
 530 {
 531         struct page *page, *cached_page = NULL;
 532         int err;
 533 repeat:
 534         page = find_lock_page(mapping, index);
 535         if (!page) {
 536                 if (!cached_page) {
 537                         cached_page = alloc_page(gfp_mask);
 538                         if (!cached_page)
 539                                 return NULL;
 540                 }
 541                 err = add_to_page_cache_lru(cached_page, mapping,
 542                                         index, gfp_mask);
 543                 if (!err) {
 544                         page = cached_page;
 545                         cached_page = NULL;
 546                 } else if (err == -EEXIST)
 547                         goto repeat;
 548         }
 549         if (cached_page)
 550                 page_cache_release(cached_page);
 551         return page;
 552 }
 553
 554 EXPORT_SYMBOL(find_or_create_page);
 555
 556 /**
 557  * find_get_pages - gang pagecache lookup
 558  * @mapping:    The address_space to search
 559  * @start:      The starting page index
 560  * @nr_pages:   The maximum number of pages
 561  * @pages:      Where the resulting pages are placed
 562  *
 563  * find_get_pages() will search for and return a group of up to
 564  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 565  * find_get_pages() takes a reference against the returned pages.
 566  *
 567  * The search returns a group of mapping-contiguous pages with ascending
 568  * indexes.  There may be holes in the indices due to not-present pages.
 569  *
 570  * find_get_pages() returns the number of pages which were found.
 571  */
 572 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 573                             unsigned int nr_pages, struct page **pages)
 574 {
 575         unsigned int i;
 576         unsigned int ret;
 577
 578         spin_lock_irq(&mapping->tree_lock);
 579         ret = radix_tree_gang_lookup(&mapping->page_tree,
 580                                 (void **)pages, start, nr_pages);
 581         for (i = 0; i < ret; i++)
 582                 page_cache_get(pages[i]);
 583         spin_unlock_irq(&mapping->tree_lock);
 584         return ret;
 585 }
 586
 587 /*
 588  * Like find_get_pages, except we only return pages which are tagged with
 589  * `tag'.   We update *index to index the next page for the traversal.
 590  */
 591 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 592                         int tag, unsigned int nr_pages, struct page **pages)
 593 {
 594         unsigned int i;
 595         unsigned int ret;
 596
 597         spin_lock_irq(&mapping->tree_lock);
 598         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 599                                 (void **)pages, *index, nr_pages, tag);
 600         for (i = 0; i < ret; i++)
 601                 page_cache_get(pages[i]);
 602         if (ret)
 603                 *index = pages[ret - 1]->index + 1;
 604         spin_unlock_irq(&mapping->tree_lock);
 605         return ret;
 606 }
 607
 608 /*
 609  * Same as grab_cache_page, but do not wait if the page is unavailable.
 610  * This is intended for speculative data generators, where the data can
 611  * be regenerated if the page couldn't be grabbed.  This routine should
 612  * be safe to call while holding the lock for another page.
 613  *
 614  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 615  * and deadlock against the caller's locked page.
 616  */
 617 struct page *
 618 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 619 {
 620         struct page *page = find_get_page(mapping, index);
 621         int gfp_mask;
 622
 623         if (page) {
 624                 if (!TestSetPageLocked(page))
 625                         return page;
 626                 page_cache_release(page);
 627                 return NULL;
 628         }
 629         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 630         page = alloc_pages(gfp_mask, 0);
 631         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 632                 page_cache_release(page);
 633                 page = NULL;
 634         }
 635         return page;
 636 }
 637
 638 EXPORT_SYMBOL(grab_cache_page_nowait);
 639
 640 /*
 641  * This is a generic file read routine, and uses the
 642  * mapping->a_ops->readpage() function for the actual low-level
 643  * stuff.
 644  *
 645  * This is really ugly. But the goto's actually try to clarify some
 646  * of the logic when it comes to error handling etc.
 647  * - note the struct file * is only passed for the use of readpage
 648  */
 649 void do_generic_mapping_read(struct address_space *mapping,
 650                              struct file_ra_state *_ra,
 651                              struct file * filp,
 652                              loff_t *ppos,
 653                              read_descriptor_t * desc,
 654                              read_actor_t actor)
 655 {
 656         struct inode *inode = mapping->host;
 657         unsigned long index, end_index, offset;
 658         loff_t isize;
 659         struct page *cached_page;
 660         int error;
 661         struct file_ra_state ra = *_ra;
 662
 663         cached_page = NULL;
 664         index = *ppos >> PAGE_CACHE_SHIFT;
 665         offset = *ppos & ~PAGE_CACHE_MASK;
 666
 667         isize = i_size_read(inode);
 668         end_index = isize >> PAGE_CACHE_SHIFT;
 669         if (index > end_index)
 670                 goto out;
 671
 672         for (;;) {
 673                 struct page *page;
 674                 unsigned long nr, ret;
 675
 676                 cond_resched();
 677                 page_cache_readahead(mapping, &ra, filp, index);
 678
 679 find_page:
 680                 page = find_get_page(mapping, index);
 681                 if (unlikely(page == NULL)) {
 682                         handle_ra_miss(mapping, &ra, index);
 683                         goto no_cached_page;
 684                 }
 685                 if (!PageUptodate(page))
 686                         goto page_not_up_to_date;
 687 page_ok:
 688                 /* nr is the maximum number of bytes to copy from this page */
 689                 nr = PAGE_CACHE_SIZE;
 690                 if (index == end_index) {
 691                         nr = isize & ~PAGE_CACHE_MASK;
 692                         if (nr <= offset) {
 693                                 page_cache_release(page);
 694                                 goto out;
 695                         }
 696                 }
 697                 nr = nr - offset;
 698
 699                 /* If users can be writing to this page using arbitrary
 700                  * virtual addresses, take care about potential aliasing
 701                  * before reading the page on the kernel side.
 702                  */
 703                 if (mapping_writably_mapped(mapping))
 704                         flush_dcache_page(page);
 705
 706                 /*
 707                  * Mark the page accessed if we read the beginning.
 708                  */
 709                 if (!offset)
 710                         mark_page_accessed(page);
 711
 712                 /*
 713                  * Ok, we have the page, and it's up-to-date, so
 714                  * now we can copy it to user space...
 715                  *
 716                  * The actor routine returns how many bytes were actually used..
 717                  * NOTE! This may not be the same as how much of a user buffer
 718                  * we filled up (we may be padding etc), so we can only update
 719                  * "pos" here (the actor routine has to update the user buffer
 720                  * pointers and the remaining count).
 721                  */
 722                 ret = actor(desc, page, offset, nr);
 723                 offset += ret;
 724                 index += offset >> PAGE_CACHE_SHIFT;
 725                 offset &= ~PAGE_CACHE_MASK;
 726
 727                 page_cache_release(page);
 728                 if (ret == nr && desc->count)
 729                         continue;
 730                 goto out;
 731
 732 page_not_up_to_date:
 733                 /* Get exclusive access to the page ... */
 734                 lock_page(page);
 735
 736                 /* Did it get unhashed before we got the lock? */
 737                 if (!page->mapping) {
 738                         unlock_page(page);
 739                         page_cache_release(page);
 740                         continue;
 741                 }
 742
 743                 /* Did somebody else fill it already? */
 744                 if (PageUptodate(page)) {
 745                         unlock_page(page);
 746                         goto page_ok;
 747                 }
 748
 749 readpage:
 750                 /* Start the actual read. The read will unlock the page. */
 751                 error = mapping->a_ops->readpage(filp, page);
 752
 753                 if (unlikely(error))
 754                         goto readpage_error;
 755
 756                 if (!PageUptodate(page)) {
 757                         wait_on_page_locked(page);
 758                         if (!PageUptodate(page)) {
 759                                 error = -EIO;
 760                                 goto readpage_error;
 761                         }
 762                 }
 763
 764                 /*
 765                  * i_size must be checked after we have done ->readpage.
 766                  *
 767                  * Checking i_size after the readpage allows us to calculate
 768                  * the correct value for "nr", which means the zero-filled
 769                  * part of the page is not copied back to userspace (unless
 770                  * another truncate extends the file - this is desired though).
 771                  */
 772                 isize = i_size_read(inode);
 773                 end_index = isize >> PAGE_CACHE_SHIFT;
 774                 if (index > end_index) {
 775                         page_cache_release(page);
 776                         goto out;
 777                 }
 778                 goto page_ok;
 779
 780 readpage_error:
 781                 /* UHHUH! A synchronous read error occurred. Report it */
 782                 desc->error = error;
 783                 page_cache_release(page);
 784                 goto out;
 785
 786 no_cached_page:
 787                 /*
 788                  * Ok, it wasn't cached, so we need to create a new
 789                  * page..
 790                  */
 791                 if (!cached_page) {
 792                         cached_page = page_cache_alloc_cold(mapping);
 793                         if (!cached_page) {
 794                                 desc->error = -ENOMEM;
 795                                 goto out;
 796                         }
 797                 }
 798                 error = add_to_page_cache_lru(cached_page, mapping,
 799                                                 index, GFP_KERNEL);
 800                 if (error) {
 801                         if (error == -EEXIST)
 802                                 goto find_page;
 803                         desc->error = error;
 804                         goto out;
 805                 }
 806                 page = cached_page;
 807                 cached_page = NULL;
 808                 goto readpage;
 809         }
 810
 811 out:
 812         *_ra = ra;
 813
 814         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 815         if (cached_page)
 816                 page_cache_release(cached_page);
 817         file_accessed(filp);
 818 }
 819
 820 EXPORT_SYMBOL(do_generic_mapping_read);
 821
 822 int file_read_actor(read_descriptor_t *desc, struct page *page,
 823                         unsigned long offset, unsigned long size)
 824 {
 825         char *kaddr;
 826         unsigned long left, count = desc->count;
 827
 828         if (size > count)
 829                 size = count;
 830
 831         /*
 832          * Faults on the destination of a read are common, so do it before
 833          * taking the kmap.
 834          */
 835         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 836                 kaddr = kmap_atomic(page, KM_USER0);
 837                 left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 838                 kunmap_atomic(kaddr, KM_USER0);
 839                 if (left == 0)
 840                         goto success;
 841         }
 842
 843         /* Do it the slow way */
 844         kaddr = kmap(page);
 845         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 846         kunmap(page);
 847
 848         if (left) {
 849                 size -= left;
 850                 desc->error = -EFAULT;
 851         }
 852 success:
 853         desc->count = count - size;
 854         desc->written += size;
 855         desc->arg.buf += size;
 856         return size;
 857 }
 858
 859 /*
 860  * This is the "read()" routine for all filesystems
 861  * that can use the page cache directly.
 862  */
 863 ssize_t
 864 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 865                 unsigned long nr_segs, loff_t *ppos)
 866 {
 867         struct file *filp = iocb->ki_filp;
 868         ssize_t retval;
 869         unsigned long seg;
 870         size_t count;
 871
 872         count = 0;
 873         for (seg = 0; seg < nr_segs; seg++) {
 874                 const struct iovec *iv = &iov[seg];
 875
 876                 /*
 877                  * If any segment has a negative length, or the cumulative
 878                  * length ever wraps negative then return -EINVAL.
 879                  */
 880                 count += iv->iov_len;
 881                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
 882                         return -EINVAL;
 883                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
 884                         continue;
 885                 if (seg == 0)
 886                         return -EFAULT;
 887                 nr_segs = seg;
 888                 count -= iv->iov_len;   /* This segment is no good */
 889                 break;
 890         }
 891
 892         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
 893         if (filp->f_flags & O_DIRECT) {
 894                 loff_t pos = *ppos, size;
 895                 struct address_space *mapping;
 896                 struct inode *inode;
 897
 898                 mapping = filp->f_mapping;
 899                 inode = mapping->host;
 900                 retval = 0;
 901                 if (!count)
 902                         goto out; /* skip atime */
 903                 size = i_size_read(inode);
 904                 if (pos < size) {
 905                         retval = generic_file_direct_IO(READ, iocb,
 906                                                 iov, pos, nr_segs);
 907                         if (retval >= 0 && !is_sync_kiocb(iocb))
 908                                 retval = -EIOCBQUEUED;
 909                         if (retval > 0)
 910                                 *ppos = pos + retval;
 911                 }
 912                 file_accessed(filp);
 913                 goto out;
 914         }
 915
 916         retval = 0;
 917         if (count) {
 918                 for (seg = 0; seg < nr_segs; seg++) {
 919                         read_descriptor_t desc;
 920
 921                         desc.written = 0;
 922                         desc.arg.buf = iov[seg].iov_base;
 923                         desc.count = iov[seg].iov_len;
 924                         if (desc.count == 0)
 925                                 continue;
 926                         desc.error = 0;
 927                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
 928                         retval += desc.written;
 929                         if (!retval) {
 930                                 retval = desc.error;
 931                                 break;
 932                         }
 933                 }
 934         }
 935 out:
 936         return retval;
 937 }
 938
 939 EXPORT_SYMBOL(__generic_file_aio_read);
 940
 941 ssize_t
 942 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 943 {
 944         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
 945
 946         BUG_ON(iocb->ki_pos != pos);
 947         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
 948 }
 949
 950 EXPORT_SYMBOL(generic_file_aio_read);
 951
 952 ssize_t
 953 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 954 {
 955         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
 956         struct kiocb kiocb;
 957         ssize_t ret;
 958
 959         init_sync_kiocb(&kiocb, filp);
 960         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
 961         if (-EIOCBQUEUED == ret)
 962                 ret = wait_on_sync_kiocb(&kiocb);
 963         return ret;
 964 }
 965
 966 EXPORT_SYMBOL(generic_file_read);
 967
 968 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
 969 {
 970         ssize_t written;
 971         unsigned long count = desc->count;
 972         struct file *file = desc->arg.data;
 973
 974         if (size > count)
 975                 size = count;
 976
 977         written = file->f_op->sendpage(file, page, offset,
 978                                        size, &file->f_pos, size<count);
 979         if (written < 0) {
 980                 desc->error = written;
 981                 written = 0;
 982         }
 983         desc->count = count - written;
 984         desc->written += written;
 985         return written;
 986 }
 987
 988 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
 989                          size_t count, read_actor_t actor, void *target)
 990 {
 991         read_descriptor_t desc;
 992
 993         if (!count)
 994                 return 0;
 995
 996         desc.written = 0;
 997         desc.count = count;
 998         desc.arg.data = target;
 999         desc.error = 0;
1000
1001         do_generic_file_read(in_file, ppos, &desc, actor);
1002         if (desc.written)
1003                 return desc.written;
1004         return desc.error;
1005 }
1006
1007 EXPORT_SYMBOL(generic_file_sendfile);
1008
1009 static ssize_t
1010 do_readahead(struct address_space *mapping, struct file *filp,
1011              unsigned long index, unsigned long nr)
1012 {
1013         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1014                 return -EINVAL;
1015
1016         force_page_cache_readahead(mapping, filp, index,
1017                                         max_sane_readahead(nr));
1018         return 0;
1019 }
1020
1021 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1022 {
1023         ssize_t ret;
1024         struct file *file;
1025
1026         ret = -EBADF;
1027         file = fget(fd);
1028         if (file) {
1029                 if (file->f_mode & FMODE_READ) {
1030                         struct address_space *mapping = file->f_mapping;
1031                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1032                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1033                         unsigned long len = end - start + 1;
1034                         ret = do_readahead(mapping, file, start, len);
1035                 }
1036                 fput(file);
1037         }
1038         return ret;
1039 }
1040
1041 #ifdef CONFIG_MMU
1042 /*
1043  * This adds the requested page to the page cache if it isn't already there,
1044  * and schedules an I/O to read in its contents from disk.
1045  */
1046 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1047 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1048 {
1049         struct address_space *mapping = file->f_mapping;
1050         struct page *page;
1051         int error;
1052
1053         page = page_cache_alloc_cold(mapping);
1054         if (!page)
1055                 return -ENOMEM;
1056
1057         error = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1058         if (!error) {
1059                 error = mapping->a_ops->readpage(file, page);
1060                 page_cache_release(page);
1061                 return error;
1062         }
1063
1064         /*
1065          * We arrive here in the unlikely event that someone
1066          * raced with us and added our page to the cache first
1067          * or we are out of memory for radix-tree nodes.
1068          */
1069         page_cache_release(page);
1070         return error == -EEXIST ? 0 : error;
1071 }
1072
1073 #define MMAP_LOTSAMISS  (100)
1074
1075 /*
1076  * filemap_nopage() is invoked via the vma operations vector for a
1077  * mapped memory region to read in file data during a page fault.
1078  *
1079  * The goto's are kind of ugly, but this streamlines the normal case of having
1080  * it in the page cache, and handles the special cases reasonably without
1081  * having a lot of duplicated code.
1082  */
1083 struct page * filemap_nopage(struct vm_area_struct * area, unsigned long address, int *type)
1084 {
1085         int error;
1086         struct file *file = area->vm_file;
1087         struct address_space *mapping = file->f_mapping;
1088         struct file_ra_state *ra = &file->f_ra;
1089         struct inode *inode = mapping->host;
1090         struct page *page;
1091         unsigned long size, pgoff, endoff;
1092         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1093
1094         pgoff = ((address - area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1095         endoff = ((area->vm_end - area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1096
1097 retry_all:
1098         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1099         if (pgoff >= size)
1100                 goto outside_data_content;
1101
1102         /* If we don't want any read-ahead, don't bother */
1103         if (VM_RandomReadHint(area))
1104                 goto no_cached_page;
1105
1106         /*
1107          * The "size" of the file, as far as mmap is concerned, isn't bigger
1108          * than the mapping
1109          */
1110         if (size > endoff)
1111                 size = endoff;
1112
1113         /*
1114          * The readahead code wants to be told about each and every page
1115          * so it can build and shrink its windows appropriately
1116          *
1117          * For sequential accesses, we use the generic readahead logic.
1118          */
1119         if (VM_SequentialReadHint(area))
1120                 page_cache_readahead(mapping, ra, file, pgoff);
1121
1122         /*
1123          * Do we have something in the page cache already?
1124          */
1125 retry_find:
1126         page = find_get_page(mapping, pgoff);
1127         if (!page) {
1128                 unsigned long ra_pages;
1129
1130                 if (VM_SequentialReadHint(area)) {
1131                         handle_ra_miss(mapping, ra, pgoff);
1132                         goto no_cached_page;
1133                 }
1134                 ra->mmap_miss++;
1135
1136                 /*
1137                  * Do we miss much more than hit in this file? If so,
1138                  * stop bothering with read-ahead. It will only hurt.
1139                  */
1140                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1141                         goto no_cached_page;
1142
1143                 /*
1144                  * To keep the pgmajfault counter straight, we need to
1145                  * check did_readaround, as this is an inner loop.
1146                  */
1147                 if (!did_readaround) {
1148                         majmin = VM_FAULT_MAJOR;
1149                         inc_page_state(pgmajfault);
1150                 }
1151                 did_readaround = 1;
1152                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1153                 if (ra_pages) {
1154                         pgoff_t start = 0;
1155
1156                         if (pgoff > ra_pages / 2)
1157                                 start = pgoff - ra_pages / 2;
1158                         do_page_cache_readahead(mapping, file, start, ra_pages);
1159                 }
1160                 page = find_get_page(mapping, pgoff);
1161                 if (!page)
1162                         goto no_cached_page;
1163         }
1164
1165         if (!did_readaround)
1166                 ra->mmap_hit++;
1167
1168         /*
1169          * Ok, found a page in the page cache, now we need to check
1170          * that it's up-to-date.
1171          */
1172         if (!PageUptodate(page))
1173                 goto page_not_uptodate;
1174
1175 success:
1176         /*
1177          * Found the page and have a reference on it.
1178          */
1179         mark_page_accessed(page);
1180         if (type)
1181                 *type = majmin;
1182         return page;
1183
1184 outside_data_content:
1185         /*
1186          * An external ptracer can access pages that normally aren't
1187          * accessible..
1188          */
1189         if (area->vm_mm == current->mm)
1190                 return NULL;
1191         /* Fall through to the non-read-ahead case */
1192 no_cached_page:
1193         /*
1194          * We're only likely to ever get here if MADV_RANDOM is in
1195          * effect.
1196          */
1197         error = page_cache_read(file, pgoff);
1198
1199         /*
1200          * The page we want has now been added to the page cache.
1201          * In the unlikely event that someone removed it in the
1202          * meantime, we'll just come back here and read it again.
1203          */
1204         if (error >= 0)
1205                 goto retry_find;
1206
1207         /*
1208          * An error return from page_cache_read can result if the
1209          * system is low on memory, or a problem occurs while trying
1210          * to schedule I/O.
1211          */
1212         if (error == -ENOMEM)
1213                 return NOPAGE_OOM;
1214         return NULL;
1215
1216 page_not_uptodate:
1217         if (!did_readaround) {
1218                 majmin = VM_FAULT_MAJOR;
1219                 inc_page_state(pgmajfault);
1220         }
1221         lock_page(page);
1222
1223         /* Did it get unhashed while we waited for it? */
1224         if (!page->mapping) {
1225                 unlock_page(page);
1226                 page_cache_release(page);
1227                 goto retry_all;
1228         }
1229
1230         /* Did somebody else get it up-to-date? */
1231         if (PageUptodate(page)) {
1232                 unlock_page(page);
1233                 goto success;
1234         }
1235
1236         if (!mapping->a_ops->readpage(file, page)) {
1237                 wait_on_page_locked(page);
1238                 if (PageUptodate(page))
1239                         goto success;
1240         }
1241
1242         /*
1243          * Umm, take care of errors if the page isn't up-to-date.
1244          * Try to re-read it _once_. We do this synchronously,
1245          * because there really aren't any performance issues here
1246          * and we need to check for errors.
1247          */
1248         lock_page(page);
1249
1250         /* Somebody truncated the page on us? */
1251         if (!page->mapping) {
1252                 unlock_page(page);
1253                 page_cache_release(page);
1254                 goto retry_all;
1255         }
1256
1257         /* Somebody else successfully read it in? */
1258         if (PageUptodate(page)) {
1259                 unlock_page(page);
1260                 goto success;
1261         }
1262         ClearPageError(page);
1263         if (!mapping->a_ops->readpage(file, page)) {
1264                 wait_on_page_locked(page);
1265                 if (PageUptodate(page))
1266                         goto success;
1267         }
1268
1269         /*
1270          * Things didn't work out. Return zero to tell the
1271          * mm layer so, possibly freeing the page cache page first.
1272          */
1273         page_cache_release(page);
1274         return NULL;
1275 }
1276
1277 EXPORT_SYMBOL(filemap_nopage);
1278
1279 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1280                                         int nonblock)
1281 {
1282         struct address_space *mapping = file->f_mapping;
1283         struct page *page;
1284         int error;
1285
1286         /*
1287          * Do we have something in the page cache already?
1288          */
1289 retry_find:
1290         page = find_get_page(mapping, pgoff);
1291         if (!page) {
1292                 if (nonblock)
1293                         return NULL;
1294                 goto no_cached_page;
1295         }
1296
1297         /*
1298          * Ok, found a page in the page cache, now we need to check
1299          * that it's up-to-date.
1300          */
1301         if (!PageUptodate(page))
1302                 goto page_not_uptodate;
1303
1304 success:
1305         /*
1306          * Found the page and have a reference on it.
1307          */
1308         mark_page_accessed(page);
1309         return page;
1310
1311 no_cached_page:
1312         error = page_cache_read(file, pgoff);
1313
1314         /*
1315          * The page we want has now been added to the page cache.
1316          * In the unlikely event that someone removed it in the
1317          * meantime, we'll just come back here and read it again.
1318          */
1319         if (error >= 0)
1320                 goto retry_find;
1321
1322         /*
1323          * An error return from page_cache_read can result if the
1324          * system is low on memory, or a problem occurs while trying
1325          * to schedule I/O.
1326          */
1327         return NULL;
1328
1329 page_not_uptodate:
1330         lock_page(page);
1331
1332         /* Did it get unhashed while we waited for it? */
1333         if (!page->mapping) {
1334                 unlock_page(page);
1335                 goto err;
1336         }
1337
1338         /* Did somebody else get it up-to-date? */
1339         if (PageUptodate(page)) {
1340                 unlock_page(page);
1341                 goto success;
1342         }
1343
1344         if (!mapping->a_ops->readpage(file, page)) {
1345                 wait_on_page_locked(page);
1346                 if (PageUptodate(page))
1347                         goto success;
1348         }
1349
1350         /*
1351          * Umm, take care of errors if the page isn't up-to-date.
1352          * Try to re-read it _once_. We do this synchronously,
1353          * because there really aren't any performance issues here
1354          * and we need to check for errors.
1355          */
1356         lock_page(page);
1357
1358         /* Somebody truncated the page on us? */
1359         if (!page->mapping) {
1360                 unlock_page(page);
1361                 goto err;
1362         }
1363         /* Somebody else successfully read it in? */
1364         if (PageUptodate(page)) {
1365                 unlock_page(page);
1366                 goto success;
1367         }
1368
1369         ClearPageError(page);
1370         if (!mapping->a_ops->readpage(file, page)) {
1371                 wait_on_page_locked(page);
1372                 if (PageUptodate(page))
1373                         goto success;
1374         }
1375
1376         /*
1377          * Things didn't work out. Return zero to tell the
1378          * mm layer so, possibly freeing the page cache page first.
1379          */
1380 err:
1381         page_cache_release(page);
1382
1383         return NULL;
1384 }
1385
1386 static int filemap_populate(struct vm_area_struct *vma,
1387                         unsigned long addr,
1388                         unsigned long len,
1389                         pgprot_t prot,
1390                         unsigned long pgoff,
1391                         int nonblock)
1392 {
1393         struct file *file = vma->vm_file;
1394         struct address_space *mapping = file->f_mapping;
1395         struct inode *inode = mapping->host;
1396         unsigned long size;
1397         struct mm_struct *mm = vma->vm_mm;
1398         struct page *page;
1399         int err;
1400
1401         if (!nonblock)
1402                 force_page_cache_readahead(mapping, vma->vm_file,
1403                                         pgoff, len >> PAGE_CACHE_SHIFT);
1404
1405 repeat:
1406         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1407         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1408                 return -EINVAL;
1409
1410         page = filemap_getpage(file, pgoff, nonblock);
1411         if (!page && !nonblock)
1412                 return -ENOMEM;
1413         if (page) {
1414                 err = install_page(mm, vma, addr, page, prot);
1415                 if (err) {
1416                         page_cache_release(page);
1417                         return err;
1418                 }
1419         } else {
1420                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1421                 if (err)
1422                         return err;
1423         }
1424
1425         len -= PAGE_SIZE;
1426         addr += PAGE_SIZE;
1427         pgoff++;
1428         if (len)
1429                 goto repeat;
1430
1431         return 0;
1432 }
1433
1434 static struct vm_operations_struct generic_file_vm_ops = {
1435         .nopage         = filemap_nopage,
1436         .populate       = filemap_populate,
1437 };
1438
1439 /* This is used for a general mmap of a disk file */
1440
1441 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1442 {
1443         struct address_space *mapping = file->f_mapping;
1444
1445         if (!mapping->a_ops->readpage)
1446                 return -ENOEXEC;
1447         file_accessed(file);
1448         vma->vm_ops = &generic_file_vm_ops;
1449         return 0;
1450 }
1451
1452 /*
1453  * This is for filesystems which do not implement ->writepage.
1454  */
1455 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1456 {
1457         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1458                 return -EINVAL;
1459         return generic_file_mmap(file, vma);
1460 }
1461 #else
1462 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1463 {
1464         return -ENOSYS;
1465 }
1466 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1467 {
1468         return -ENOSYS;
1469 }
1470 #endif /* CONFIG_MMU */
1471
1472 EXPORT_SYMBOL(generic_file_mmap);
1473 EXPORT_SYMBOL(generic_file_readonly_mmap);
1474
1475 static inline struct page *__read_cache_page(struct address_space *mapping,
1476                                 unsigned long index,
1477                                 int (*filler)(void *,struct page*),
1478                                 void *data)
1479 {
1480         struct page *page, *cached_page = NULL;
1481         int err;
1482 repeat:
1483         page = find_get_page(mapping, index);
1484         if (!page) {
1485                 if (!cached_page) {
1486                         cached_page = page_cache_alloc_cold(mapping);
1487                         if (!cached_page)
1488                                 return ERR_PTR(-ENOMEM);
1489                 }
1490                 err = add_to_page_cache_lru(cached_page, mapping,
1491                                         index, GFP_KERNEL);
1492                 if (err == -EEXIST)
1493                         goto repeat;
1494                 if (err < 0) {
1495                         /* Presumably ENOMEM for radix tree node */
1496                         page_cache_release(cached_page);
1497                         return ERR_PTR(err);
1498                 }
1499                 page = cached_page;
1500                 cached_page = NULL;
1501                 err = filler(data, page);
1502                 if (err < 0) {
1503                         page_cache_release(page);
1504                         page = ERR_PTR(err);
1505                 }
1506         }
1507         if (cached_page)
1508                 page_cache_release(cached_page);
1509         return page;
1510 }
1511
1512 /*
1513  * Read into the page cache. If a page already exists,
1514  * and PageUptodate() is not set, try to fill the page.
1515  */
1516 struct page *read_cache_page(struct address_space *mapping,
1517                                 unsigned long index,
1518                                 int (*filler)(void *,struct page*),
1519                                 void *data)
1520 {
1521         struct page *page;
1522         int err;
1523
1524 retry:
1525         page = __read_cache_page(mapping, index, filler, data);
1526         if (IS_ERR(page))
1527                 goto out;
1528         mark_page_accessed(page);
1529         if (PageUptodate(page))
1530                 goto out;
1531
1532         lock_page(page);
1533         if (!page->mapping) {
1534                 unlock_page(page);
1535                 page_cache_release(page);
1536                 goto retry;
1537         }
1538         if (PageUptodate(page)) {
1539                 unlock_page(page);
1540                 goto out;
1541         }
1542         err = filler(data, page);
1543         if (err < 0) {
1544                 page_cache_release(page);
1545                 page = ERR_PTR(err);
1546         }
1547  out:
1548         return page;
1549 }
1550
1551 EXPORT_SYMBOL(read_cache_page);
1552
1553 /*
1554  * If the page was newly created, increment its refcount and add it to the
1555  * caller's lru-buffering pagevec.  This function is specifically for
1556  * generic_file_write().
1557  */
1558 static inline struct page *
1559 __grab_cache_page(struct address_space *mapping, unsigned long index,
1560                         struct page **cached_page, struct pagevec *lru_pvec)
1561 {
1562         int err;
1563         struct page *page;
1564 repeat:
1565         page = find_lock_page(mapping, index);
1566         if (!page) {
1567                 if (!*cached_page) {
1568                         *cached_page = page_cache_alloc(mapping);
1569                         if (!*cached_page)
1570                                 return NULL;
1571                 }
1572                 err = add_to_page_cache(*cached_page, mapping,
1573                                         index, GFP_KERNEL);
1574                 if (err == -EEXIST)
1575                         goto repeat;
1576                 if (err == 0) {
1577                         page = *cached_page;
1578                         page_cache_get(page);
1579                         if (!pagevec_add(lru_pvec, page))
1580                                 __pagevec_lru_add(lru_pvec);
1581                         *cached_page = NULL;
1582                 }
1583         }
1584         return page;
1585 }
1586
1587 /*
1588  * The logic we want is
1589  *
1590  *      if suid or (sgid and xgrp)
1591  *              remove privs
1592  */
1593 int remove_suid(struct dentry *dentry)
1594 {
1595         mode_t mode = dentry->d_inode->i_mode;
1596         int kill = 0;
1597         int result = 0;
1598
1599         /* suid always must be killed */
1600         if (unlikely(mode & S_ISUID))
1601                 kill = ATTR_KILL_SUID;
1602
1603         /*
1604          * sgid without any exec bits is just a mandatory locking mark; leave
1605          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1606          */
1607         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1608                 kill |= ATTR_KILL_SGID;
1609
1610         if (unlikely(kill && !capable(CAP_FSETID))) {
1611                 struct iattr newattrs;
1612
1613                 newattrs.ia_valid = ATTR_FORCE | kill;
1614                 result = notify_change(dentry, &newattrs);
1615         }
1616         return result;
1617 }
1618 EXPORT_SYMBOL(remove_suid);
1619
1620 /*
1621  * Copy as much as we can into the page and return the number of bytes which
1622  * were sucessfully copied.  If a fault is encountered then clear the page
1623  * out to (offset+bytes) and return the number of bytes which were copied.
1624  */
1625 static inline size_t
1626 filemap_copy_from_user(struct page *page, unsigned long offset,
1627                         const char __user *buf, unsigned bytes)
1628 {
1629         char *kaddr;
1630         int left;
1631
1632         kaddr = kmap_atomic(page, KM_USER0);
1633         left = __copy_from_user(kaddr + offset, buf, bytes);
1634         kunmap_atomic(kaddr, KM_USER0);
1635
1636         if (left != 0) {
1637                 /* Do it the slow way */
1638                 kaddr = kmap(page);
1639                 left = __copy_from_user(kaddr + offset, buf, bytes);
1640                 kunmap(page);
1641         }
1642         return bytes - left;
1643 }
1644
1645 static size_t
1646 __filemap_copy_from_user_iovec(char *vaddr,
1647                         const struct iovec *iov, size_t base, size_t bytes)
1648 {
1649         size_t copied = 0, left = 0;
1650
1651         while (bytes) {
1652                 char __user *buf = iov->iov_base + base;
1653                 int copy = min(bytes, iov->iov_len - base);
1654
1655                 base = 0;
1656                 left = __copy_from_user(vaddr, buf, copy);
1657                 copied += copy;
1658                 bytes -= copy;
1659                 vaddr += copy;
1660                 iov++;
1661
1662                 if (unlikely(left)) {
1663                         /* zero the rest of the target like __copy_from_user */
1664                         if (bytes)
1665                                 memset(vaddr, 0, bytes);
1666                         break;
1667                 }
1668         }
1669         return copied - left;
1670 }
1671
1672 /*
1673  * This has the same sideeffects and return value as filemap_copy_from_user().
1674  * The difference is that on a fault we need to memset the remainder of the
1675  * page (out to offset+bytes), to emulate filemap_copy_from_user()'s
1676  * single-segment behaviour.
1677  */
1678 static inline size_t
1679 filemap_copy_from_user_iovec(struct page *page, unsigned long offset,
1680                         const struct iovec *iov, size_t base, size_t bytes)
1681 {
1682         char *kaddr;
1683         size_t copied;
1684
1685         kaddr = kmap_atomic(page, KM_USER0);
1686         copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1687                                                 base, bytes);
1688         kunmap_atomic(kaddr, KM_USER0);
1689         if (copied != bytes) {
1690                 kaddr = kmap(page);
1691                 copied = __filemap_copy_from_user_iovec(kaddr + offset, iov,
1692                                                         base, bytes);
1693                 kunmap(page);
1694         }
1695         return copied;
1696 }
1697
1698 static inline void
1699 filemap_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1700 {
1701         const struct iovec *iov = *iovp;
1702         size_t base = *basep;
1703
1704         while (bytes) {
1705                 int copy = min(bytes, iov->iov_len - base);
1706
1707                 bytes -= copy;
1708                 base += copy;
1709                 if (iov->iov_len == base) {
1710                         iov++;
1711                         base = 0;
1712                 }
1713         }
1714         *iovp = iov;
1715         *basep = base;
1716 }
1717
1718 /*
1719  * Performs necessary checks before doing a write
1720  *
1721  * Can adjust writing position aor amount of bytes to write.
1722  * Returns appropriate error code that caller should return or
1723  * zero in case that write should be allowed.
1724  */
1725 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1726 {
1727         struct inode *inode = file->f_mapping->host;
1728         unsigned long limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1729
1730         if (unlikely(*pos < 0))
1731                 return -EINVAL;
1732
1733         if (unlikely(file->f_error)) {
1734                 int err = file->f_error;
1735                 file->f_error = 0;
1736                 return err;
1737         }
1738
1739         if (!isblk) {
1740                 /* FIXME: this is for backwards compatibility with 2.4 */
1741                 if (file->f_flags & O_APPEND)
1742                         *pos = i_size_read(inode);
1743
1744                 if (limit != RLIM_INFINITY) {
1745                         if (*pos >= limit) {
1746                                 send_sig(SIGXFSZ, current, 0);
1747                                 return -EFBIG;
1748                         }
1749                         if (*count > limit - (typeof(limit))*pos) {
1750                                 *count = limit - (typeof(limit))*pos;
1751                         }
1752                 }
1753         }
1754
1755         /*
1756          * LFS rule
1757          */
1758         if (unlikely(*pos + *count > MAX_NON_LFS &&
1759                                 !(file->f_flags & O_LARGEFILE))) {
1760                 if (*pos >= MAX_NON_LFS) {
1761                         send_sig(SIGXFSZ, current, 0);
1762                         return -EFBIG;
1763                 }
1764                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1765                         *count = MAX_NON_LFS - (unsigned long)*pos;
1766                 }
1767         }
1768
1769         /*
1770          * Are we about to exceed the fs block limit ?
1771          *
1772          * If we have written data it becomes a short write.  If we have
1773          * exceeded without writing data we send a signal and return EFBIG.
1774          * Linus frestrict idea will clean these up nicely..
1775          */
1776         if (likely(!isblk)) {
1777                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1778                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1779                                 send_sig(SIGXFSZ, current, 0);
1780                                 return -EFBIG;
1781                         }
1782                         /* zero-length writes at ->s_maxbytes are OK */
1783                 }
1784
1785                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1786                         *count = inode->i_sb->s_maxbytes - *pos;
1787         } else {
1788                 loff_t isize;
1789                 if (bdev_read_only(I_BDEV(inode)))
1790                         return -EPERM;
1791                 isize = i_size_read(inode);
1792                 if (*pos >= isize) {
1793                         if (*count || *pos > isize)
1794                                 return -ENOSPC;
1795                 }
1796
1797                 if (*pos + *count > isize)
1798                         *count = isize - *pos;
1799         }
1800         return 0;
1801 }
1802
1803 EXPORT_SYMBOL(generic_write_checks);
1804
1805 /*
1806  * Write to a file through the page cache.
1807  * Called under i_sem for S_ISREG files.
1808  *
1809  * We put everything into the page cache prior to writing it. This is not a
1810  * problem when writing full pages. With partial pages, however, we first have
1811  * to read the data into the cache, then dirty the page, and finally schedule
1812  * it for writing by marking it dirty.
1813  *                                                      okir@monad.swb.de
1814  */
1815 ssize_t
1816 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
1817                                 unsigned long nr_segs, loff_t *ppos)
1818 {
1819         struct file *file = iocb->ki_filp;
1820         struct address_space * mapping = file->f_mapping;
1821         struct address_space_operations *a_ops = mapping->a_ops;
1822         size_t ocount;          /* original count */
1823         size_t count;           /* after file limit checks */
1824         struct inode    *inode = mapping->host;
1825         long            status = 0;
1826         loff_t          pos;
1827         struct page     *page;
1828         struct page     *cached_page = NULL;
1829         const int       isblk = S_ISBLK(inode->i_mode);
1830         ssize_t         written;
1831         ssize_t         err;
1832         size_t          bytes;
1833         struct pagevec  lru_pvec;
1834         const struct iovec *cur_iov = iov; /* current iovec */
1835         size_t          iov_base = 0;      /* offset in the current iovec */
1836         unsigned long   seg;
1837         char __user     *buf;
1838
1839         ocount = 0;
1840         for (seg = 0; seg < nr_segs; seg++) {
1841                 const struct iovec *iv = &iov[seg];
1842
1843                 /*
1844                  * If any segment has a negative length, or the cumulative
1845                  * length ever wraps negative then return -EINVAL.
1846                  */
1847                 ocount += iv->iov_len;
1848                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
1849                         return -EINVAL;
1850                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
1851                         continue;
1852                 if (seg == 0)
1853                         return -EFAULT;
1854                 nr_segs = seg;
1855                 ocount -= iv->iov_len;  /* This segment is no good */
1856                 break;
1857         }
1858
1859         count = ocount;
1860         pos = *ppos;
1861         pagevec_init(&lru_pvec, 0);
1862
1863         /* We can write back this queue in page reclaim */
1864         current->backing_dev_info = mapping->backing_dev_info;
1865         written = 0;
1866
1867         err = generic_write_checks(file, &pos, &count, isblk);
1868         if (err)
1869                 goto out;
1870
1871         if (count == 0)
1872                 goto out;
1873
1874         err = remove_suid(file->f_dentry);
1875         if (err)
1876                 goto out;
1877
1878         inode_update_time(inode, 1);
1879
1880         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1881         if (unlikely(file->f_flags & O_DIRECT)) {
1882                 if (count != ocount)
1883                         nr_segs = iov_shorten((struct iovec *)iov,
1884                                                 nr_segs, count);
1885                 written = generic_file_direct_IO(WRITE, iocb,
1886                                         iov, pos, nr_segs);
1887                 if (written > 0) {
1888                         loff_t end = pos + written;
1889                         if (end > i_size_read(inode) && !isblk) {
1890                                 i_size_write(inode,  end);
1891                                 mark_inode_dirty(inode);
1892                         }
1893                         *ppos = end;
1894                 }
1895                 /*
1896                  * Sync the fs metadata but not the minor inode changes and
1897                  * of course not the data as we did direct DMA for the IO.
1898                  * i_sem is held, which protects generic_osync_inode() from
1899                  * livelocking.
1900                  */
1901                 if (written >= 0 && file->f_flags & O_SYNC)
1902                         status = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1903                 if (written == count && !is_sync_kiocb(iocb))
1904                         written = -EIOCBQUEUED;
1905                 if (written < 0 || written == count)
1906                         goto out_status;
1907                 /*
1908                  * direct-io write to a hole: fall through to buffered I/O
1909                  * for completing the rest of the request.
1910                  */
1911                 pos += written;
1912                 count -= written;
1913         }
1914
1915         buf = iov->iov_base + written;  /* handle partial DIO write */
1916         do {
1917                 unsigned long index;
1918                 unsigned long offset;
1919                 size_t copied;
1920
1921                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1922                 index = pos >> PAGE_CACHE_SHIFT;
1923                 bytes = PAGE_CACHE_SIZE - offset;
1924                 if (bytes > count)
1925                         bytes = count;
1926
1927                 /*
1928                  * Bring in the user page that we will copy from _first_.
1929                  * Otherwise there's a nasty deadlock on copying from the
1930                  * same page as we're writing to, without it being marked
1931                  * up-to-date.
1932                  */
1933                 fault_in_pages_readable(buf, bytes);
1934
1935                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1936                 if (!page) {
1937                         status = -ENOMEM;
1938                         break;
1939                 }
1940
1941                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1942                 if (unlikely(status)) {
1943                         loff_t isize = i_size_read(inode);
1944                         /*
1945                          * prepare_write() may have instantiated a few blocks
1946                          * outside i_size.  Trim these off again.
1947                          */
1948                         unlock_page(page);
1949                         page_cache_release(page);
1950                         if (pos + bytes > isize)
1951                                 vmtruncate(inode, isize);
1952                         break;
1953                 }
1954                 if (likely(nr_segs == 1))
1955                         copied = filemap_copy_from_user(page, offset,
1956                                                         buf, bytes);
1957                 else
1958                         copied = filemap_copy_from_user_iovec(page, offset,
1959                                                 cur_iov, iov_base, bytes);
1960                 flush_dcache_page(page);
1961                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1962                 if (likely(copied > 0)) {
1963                         if (!status)
1964                                 status = copied;
1965
1966                         if (status >= 0) {
1967                                 written += status;
1968                                 count -= status;
1969                                 pos += status;
1970                                 buf += status;
1971                                 if (unlikely(nr_segs > 1))
1972                                         filemap_set_next_iovec(&cur_iov,
1973                                                         &iov_base, status);
1974                         }
1975                 }
1976                 if (unlikely(copied != bytes))
1977                         if (status >= 0)
1978                                 status = -EFAULT;
1979                 unlock_page(page);
1980                 mark_page_accessed(page);
1981                 page_cache_release(page);
1982                 if (status < 0)
1983                         break;
1984                 balance_dirty_pages_ratelimited(mapping);
1985                 cond_resched();
1986         } while (count);
1987         *ppos = pos;
1988
1989         if (cached_page)
1990                 page_cache_release(cached_page);
1991
1992         /*
1993          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
1994          */
1995         if (status >= 0) {
1996                 if ((file->f_flags & O_SYNC) || IS_SYNC(inode))
1997                         status = generic_osync_inode(inode, mapping,
1998                                         OSYNC_METADATA|OSYNC_DATA);
1999         }
2000
2001         /*
2002          * If we get here for O_DIRECT writes then we must have fallen through
2003          * to buffered writes (block instantiation inside i_size).  So we sync
2004          * the file data here, to try to honour O_DIRECT expectations.
2005          */
2006         if (unlikely(file->f_flags & O_DIRECT) && written)
2007                 status = filemap_write_and_wait(mapping);
2008
2009 out_status:
2010         err = written ? written : status;
2011 out:
2012         pagevec_lru_add(&lru_pvec);
2013         current->backing_dev_info = NULL;
2014         return err;
2015 }
2016
2017 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2018
2019 ssize_t
2020 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2021                                 unsigned long nr_segs, loff_t *ppos)
2022 {
2023         struct kiocb kiocb;
2024         ssize_t ret;
2025
2026         init_sync_kiocb(&kiocb, file);
2027         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2028         if (-EIOCBQUEUED == ret)
2029                 ret = wait_on_sync_kiocb(&kiocb);
2030         return ret;
2031 }
2032
2033 EXPORT_SYMBOL(generic_file_write_nolock);
2034
2035 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2036                                size_t count, loff_t pos)
2037 {
2038         struct file *file = iocb->ki_filp;
2039         struct inode *inode = file->f_mapping->host;
2040         ssize_t err;
2041         struct iovec local_iov = { .iov_base = (void __user *)buf, .iov_len = count };
2042
2043         BUG_ON(iocb->ki_pos != pos);
2044
2045         down(&inode->i_sem);
2046         err = generic_file_aio_write_nolock(iocb, &local_iov, 1,
2047                                                 &iocb->ki_pos);
2048         up(&inode->i_sem);
2049
2050         return err;
2051 }
2052
2053 EXPORT_SYMBOL(generic_file_aio_write);
2054
2055 ssize_t generic_file_write(struct file *file, const char __user *buf,
2056                            size_t count, loff_t *ppos)
2057 {
2058         struct inode    *inode = file->f_mapping->host;
2059         ssize_t         err;
2060         struct iovec local_iov = { .iov_base = (void __user *)buf, .iov_len = count };
2061
2062         down(&inode->i_sem);
2063         err = generic_file_write_nolock(file, &local_iov, 1, ppos);
2064         up(&inode->i_sem);
2065
2066         return err;
2067 }
2068
2069 EXPORT_SYMBOL(generic_file_write);
2070
2071 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2072                         unsigned long nr_segs, loff_t *ppos)
2073 {
2074         struct kiocb kiocb;
2075         ssize_t ret;
2076
2077         init_sync_kiocb(&kiocb, filp);
2078         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2079         if (-EIOCBQUEUED == ret)
2080                 ret = wait_on_sync_kiocb(&kiocb);
2081         return ret;
2082 }
2083
2084 EXPORT_SYMBOL(generic_file_readv);
2085
2086 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2087                         unsigned long nr_segs, loff_t * ppos)
2088 {
2089         struct inode *inode = file->f_mapping->host;
2090         ssize_t ret;
2091
2092         down(&inode->i_sem);
2093         ret = generic_file_write_nolock(file, iov, nr_segs, ppos);
2094         up(&inode->i_sem);
2095         return ret;
2096 }
2097
2098 EXPORT_SYMBOL(generic_file_writev);
2099
2100 /*
2101  * Called under i_sem for writes to S_ISREG files
2102  */
2103 ssize_t
2104 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2105         loff_t offset, unsigned long nr_segs)
2106 {
2107         struct file *file = iocb->ki_filp;
2108         struct address_space *mapping = file->f_mapping;
2109         ssize_t retval;
2110
2111         retval = filemap_write_and_wait(mapping);
2112         if (retval == 0) {
2113                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2114                                                 offset, nr_segs);
2115                 if (rw == WRITE && mapping->nrpages)
2116                         invalidate_inode_pages2(mapping);
2117         }
2118         return retval;
2119 }
2120
2121 EXPORT_SYMBOL_GPL(generic_file_direct_IO);