fs/mpage.c

   1 /*
   2  * fs/mpage.c
   3  *
   4  * Copyright (C) 2002, Linus Torvalds.
   5  *
   6  * Contains functions related to preparing and submitting BIOs which contain
   7  * multiple pagecache pages.
   8  *
   9  * 15May2002    akpm@zip.com.au
  10  *              Initial version
  11  * 27Jun2002    axboe@suse.de
  12  *              use bio_add_page() to build bio's just the right size
  13  */
  14
  15 #include <linux/kernel.h>
  16 #include <linux/module.h>
  17 #include <linux/mm.h>
  18 #include <linux/kdev_t.h>
  19 #include <linux/bio.h>
  20 #include <linux/fs.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/highmem.h>
  24 #include <linux/prefetch.h>
  25 #include <linux/mpage.h>
  26 #include <linux/writeback.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/pagevec.h>
  29
  30 /*
  31  * I/O completion handler for multipage BIOs.
  32  *
  33  * The mpage code never puts partial pages into a BIO (except for end-of-file).
  34  * If a page does not map to a contiguous run of blocks then it simply falls
  35  * back to block_read_full_page().
  36  *
  37  * Why is this?  If a page's completion depends on a number of different BIOs
  38  * which can complete in any order (or at the same time) then determining the
  39  * status of that page is hard.  See end_buffer_async_read() for the details.
  40  * There is no point in duplicating all that complexity.
  41  */
  42 static int mpage_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
  43 {
  44         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
  45         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
  46
  47         if (bio->bi_size)
  48                 return 1;
  49
  50         do {
  51                 struct page *page = bvec->bv_page;
  52
  53                 if (--bvec >= bio->bi_io_vec)
  54                         prefetchw(&bvec->bv_page->flags);
  55
  56                 if (uptodate) {
  57                         SetPageUptodate(page);
  58                 } else {
  59                         ClearPageUptodate(page);
  60                         SetPageError(page);
  61                 }
  62                 unlock_page(page);
  63         } while (bvec >= bio->bi_io_vec);
  64         bio_put(bio);
  65         return 0;
  66 }
  67
  68 static int mpage_end_io_write(struct bio *bio, unsigned int bytes_done, int err)
  69 {
  70         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
  71         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
  72
  73         if (bio->bi_size)
  74                 return 1;
  75
  76         do {
  77                 struct page *page = bvec->bv_page;
  78
  79                 if (--bvec >= bio->bi_io_vec)
  80                         prefetchw(&bvec->bv_page->flags);
  81
  82                 if (!uptodate)
  83                         SetPageError(page);
  84                 end_page_writeback(page);
  85         } while (bvec >= bio->bi_io_vec);
  86         bio_put(bio);
  87         return 0;
  88 }
  89
  90 struct bio *mpage_bio_submit(int rw, struct bio *bio)
  91 {
  92         bio->bi_end_io = mpage_end_io_read;
  93         if (rw == WRITE)
  94                 bio->bi_end_io = mpage_end_io_write;
  95         submit_bio(rw, bio);
  96         return NULL;
  97 }
  98
  99 static struct bio *
 100 mpage_alloc(struct block_device *bdev,
 101                 sector_t first_sector, int nr_vecs, int gfp_flags)
 102 {
 103         struct bio *bio;
 104
 105         bio = bio_alloc(gfp_flags, nr_vecs);
 106
 107         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
 108                 while (!bio && (nr_vecs /= 2))
 109                         bio = bio_alloc(gfp_flags, nr_vecs);
 110         }
 111
 112         if (bio) {
 113                 bio->bi_bdev = bdev;
 114                 bio->bi_sector = first_sector;
 115         }
 116         return bio;
 117 }
 118
 119 /*
 120  * support function for mpage_readpages.  The fs supplied get_block might
 121  * return an up to date buffer.  This is used to map that buffer into
 122  * the page, which allows readpage to avoid triggering a duplicate call
 123  * to get_block.
 124  *
 125  * The idea is to avoid adding buffers to pages that don't already have
 126  * them.  So when the buffer is up to date and the page size == block size,
 127  * this marks the page up to date instead of adding new buffers.
 128  */
 129 static void
 130 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
 131 {
 132         struct inode *inode = page->mapping->host;
 133         struct buffer_head *page_bh, *head;
 134         int block = 0;
 135
 136         if (!page_has_buffers(page)) {
 137                 /*
 138                  * don't make any buffers if there is only one buffer on
 139                  * the page and the page just needs to be set up to date
 140                  */
 141                 if (inode->i_blkbits == PAGE_CACHE_SHIFT &&
 142                     buffer_uptodate(bh)) {
 143                         SetPageUptodate(page);
 144                         return;
 145                 }
 146                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
 147         }
 148         head = page_buffers(page);
 149         page_bh = head;
 150         do {
 151                 if (block == page_block) {
 152                         page_bh->b_state = bh->b_state;
 153                         page_bh->b_bdev = bh->b_bdev;
 154                         page_bh->b_blocknr = bh->b_blocknr;
 155                         break;
 156                 }
 157                 page_bh = page_bh->b_this_page;
 158                 block++;
 159         } while (page_bh != head);
 160 }
 161
 162 /**
 163  * mpage_readpages - populate an address space with some pages, and
 164  *                       start reads against them.
 165  *
 166  * @mapping: the address_space
 167  * @pages: The address of a list_head which contains the target pages.  These
 168  *   pages have their ->index populated and are otherwise uninitialised.
 169  *
 170  *   The page at @pages->prev has the lowest file offset, and reads should be
 171  *   issued in @pages->prev to @pages->next order.
 172  *
 173  * @nr_pages: The number of pages at *@pages
 174  * @get_block: The filesystem's block mapper function.
 175  *
 176  * This function walks the pages and the blocks within each page, building and
 177  * emitting large BIOs.
 178  *
 179  * If anything unusual happens, such as:
 180  *
 181  * - encountering a page which has buffers
 182  * - encountering a page which has a non-hole after a hole
 183  * - encountering a page with non-contiguous blocks
 184  *
 185  * then this code just gives up and calls the buffer_head-based read function.
 186  * It does handle a page which has holes at the end - that is a common case:
 187  * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
 188  *
 189  * BH_Boundary explanation:
 190  *
 191  * There is a problem.  The mpage read code assembles several pages, gets all
 192  * their disk mappings, and then submits them all.  That's fine, but obtaining
 193  * the disk mappings may require I/O.  Reads of indirect blocks, for example.
 194  *
 195  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
 196  * submitted in the following order:
 197  *      12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
 198  * because the indirect block has to be read to get the mappings of blocks
 199  * 13,14,15,16.  Obviously, this impacts performance.
 200  *
 201  * So what we do it to allow the filesystem's get_block() function to set
 202  * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
 203  * after this one will require I/O against a block which is probably close to
 204  * this one.  So you should push what I/O you have currently accumulated.
 205  *
 206  * This all causes the disk requests to be issued in the correct order.
 207  */
 208 static struct bio *
 209 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
 210                         sector_t *last_block_in_bio, get_block_t get_block)
 211 {
 212         struct inode *inode = page->mapping->host;
 213         const unsigned blkbits = inode->i_blkbits;
 214         const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
 215         const unsigned blocksize = 1 << blkbits;
 216         sector_t block_in_file;
 217         sector_t last_block;
 218         sector_t blocks[MAX_BUF_PER_PAGE];
 219         unsigned page_block;
 220         unsigned first_hole = blocks_per_page;
 221         struct block_device *bdev = NULL;
 222         struct buffer_head bh;
 223         int length;
 224         int fully_mapped = 1;
 225
 226         if (page_has_buffers(page))
 227                 goto confused;
 228
 229         block_in_file = page->index << (PAGE_CACHE_SHIFT - blkbits);
 230         last_block = (inode->i_size + blocksize - 1) >> blkbits;
 231
 232         bh.b_page = page;
 233         for (page_block = 0; page_block < blocks_per_page;
 234                                 page_block++, block_in_file++) {
 235                 bh.b_state = 0;
 236                 if (block_in_file < last_block) {
 237                         if (get_block(inode, block_in_file, &bh, 0))
 238                                 goto confused;
 239                 }
 240
 241                 if (!buffer_mapped(&bh)) {
 242                         fully_mapped = 0;
 243                         if (first_hole == blocks_per_page)
 244                                 first_hole = page_block;
 245                         continue;
 246                 }
 247
 248                 /* some filesystems will copy data into the page during
 249                  * the get_block call, in which case we don't want to
 250                  * read it again.  map_buffer_to_page copies the data
 251                  * we just collected from get_block into the page's buffers
 252                  * so readpage doesn't have to repeat the get_block call
 253                  */
 254                 if (buffer_uptodate(&bh)) {
 255                         map_buffer_to_page(page, &bh, page_block);
 256                         goto confused;
 257                 }
 258
 259                 if (first_hole != blocks_per_page)
 260                         goto confused;          /* hole -> non-hole */
 261
 262                 /* Contiguous blocks? */
 263                 if (page_block && blocks[page_block-1] != bh.b_blocknr-1)
 264                         goto confused;
 265                 blocks[page_block] = bh.b_blocknr;
 266                 bdev = bh.b_bdev;
 267         }
 268
 269         if (first_hole != blocks_per_page) {
 270                 char *kaddr = kmap_atomic(page, KM_USER0);
 271                 memset(kaddr + (first_hole << blkbits), 0,
 272                                 PAGE_CACHE_SIZE - (first_hole << blkbits));
 273                 flush_dcache_page(page);
 274                 kunmap_atomic(kaddr, KM_USER0);
 275                 if (first_hole == 0) {
 276                         SetPageUptodate(page);
 277                         unlock_page(page);
 278                         goto out;
 279                 }
 280         } else if (fully_mapped) {
 281                 SetPageMappedToDisk(page);
 282         }
 283
 284         /*
 285          * This page will go to BIO.  Do we need to send this BIO off first?
 286          */
 287         if (bio && (*last_block_in_bio != blocks[0] - 1))
 288                 bio = mpage_bio_submit(READ, bio);
 289
 290 alloc_new:
 291         if (bio == NULL) {
 292                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 293                                         nr_pages, GFP_KERNEL);
 294                 if (bio == NULL)
 295                         goto confused;
 296         }
 297
 298         length = first_hole << blkbits;
 299         if (bio_add_page(bio, page, length, 0) < length) {
 300                 bio = mpage_bio_submit(READ, bio);
 301                 goto alloc_new;
 302         }
 303
 304         if (buffer_boundary(&bh) || (first_hole != blocks_per_page))
 305                 bio = mpage_bio_submit(READ, bio);
 306         else
 307                 *last_block_in_bio = blocks[blocks_per_page - 1];
 308 out:
 309         return bio;
 310
 311 confused:
 312         if (bio)
 313                 bio = mpage_bio_submit(READ, bio);
 314         if (!PageUptodate(page))
 315                 block_read_full_page(page, get_block);
 316         else
 317                 unlock_page(page);
 318         goto out;
 319 }
 320
 321 int
 322 mpage_readpages(struct address_space *mapping, struct list_head *pages,
 323                                 unsigned nr_pages, get_block_t get_block)
 324 {
 325         struct bio *bio = NULL;
 326         unsigned page_idx;
 327         sector_t last_block_in_bio = 0;
 328         struct pagevec lru_pvec;
 329
 330         pagevec_init(&lru_pvec, 0);
 331         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 332                 struct page *page = list_entry(pages->prev, struct page, list);
 333
 334                 prefetchw(&page->flags);
 335                 list_del(&page->list);
 336                 if (!add_to_page_cache(page, mapping,
 337                                         page->index, GFP_KERNEL)) {
 338                         bio = do_mpage_readpage(bio, page,
 339                                         nr_pages - page_idx,
 340                                         &last_block_in_bio, get_block);
 341                         if (!pagevec_add(&lru_pvec, page))
 342                                 __pagevec_lru_add(&lru_pvec);
 343                 } else {
 344                         page_cache_release(page);
 345                 }
 346         }
 347         pagevec_lru_add(&lru_pvec);
 348         BUG_ON(!list_empty(pages));
 349         if (bio)
 350                 mpage_bio_submit(READ, bio);
 351         return 0;
 352 }
 353 EXPORT_SYMBOL(mpage_readpages);
 354
 355 /*
 356  * This isn't called much at all
 357  */
 358 int mpage_readpage(struct page *page, get_block_t get_block)
 359 {
 360         struct bio *bio = NULL;
 361         sector_t last_block_in_bio = 0;
 362
 363         bio = do_mpage_readpage(bio, page, 1,
 364                         &last_block_in_bio, get_block);
 365         if (bio)
 366                 mpage_bio_submit(READ, bio);
 367         return 0;
 368 }
 369 EXPORT_SYMBOL(mpage_readpage);
 370
 371 /*
 372  * Writing is not so simple.
 373  *
 374  * If the page has buffers then they will be used for obtaining the disk
 375  * mapping.  We only support pages which are fully mapped-and-dirty, with a
 376  * special case for pages which are unmapped at the end: end-of-file.
 377  *
 378  * If the page has no buffers (preferred) then the page is mapped here.
 379  *
 380  * If all blocks are found to be contiguous then the page can go into the
 381  * BIO.  Otherwise fall back to the mapping's writepage().
 382  *
 383  * FIXME: This code wants an estimate of how many pages are still to be
 384  * written, so it can intelligently allocate a suitably-sized BIO.  For now,
 385  * just allocate full-size (16-page) BIOs.
 386  */
 387 static struct bio *
 388 mpage_writepage(struct bio *bio, struct page *page, get_block_t get_block,
 389         sector_t *last_block_in_bio, int *ret, struct writeback_control *wbc)
 390 {
 391         struct inode *inode = page->mapping->host;
 392         const unsigned blkbits = inode->i_blkbits;
 393         unsigned long end_index;
 394         const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
 395         sector_t last_block;
 396         sector_t block_in_file;
 397         sector_t blocks[MAX_BUF_PER_PAGE];
 398         unsigned page_block;
 399         unsigned first_unmapped = blocks_per_page;
 400         struct block_device *bdev = NULL;
 401         int boundary = 0;
 402         sector_t boundary_block = 0;
 403         struct block_device *boundary_bdev = NULL;
 404         int length;
 405         struct buffer_head map_bh;
 406
 407         if (page_has_buffers(page)) {
 408                 struct buffer_head *head = page_buffers(page);
 409                 struct buffer_head *bh = head;
 410
 411                 /* If they're all mapped and dirty, do it */
 412                 page_block = 0;
 413                 do {
 414                         BUG_ON(buffer_locked(bh));
 415                         if (!buffer_mapped(bh)) {
 416                                 /*
 417                                  * unmapped dirty buffers are created by
 418                                  * __set_page_dirty_buffers -> mmapped data
 419                                  */
 420                                 if (buffer_dirty(bh))
 421                                         goto confused;
 422                                 if (first_unmapped == blocks_per_page)
 423                                         first_unmapped = page_block;
 424                                 continue;
 425                         }
 426
 427                         if (first_unmapped != blocks_per_page)
 428                                 goto confused;  /* hole -> non-hole */
 429
 430                         if (!buffer_dirty(bh) || !buffer_uptodate(bh))
 431                                 goto confused;
 432                         if (page_block) {
 433                                 if (bh->b_blocknr != blocks[page_block-1] + 1)
 434                                         goto confused;
 435                         }
 436                         blocks[page_block++] = bh->b_blocknr;
 437                         boundary = buffer_boundary(bh);
 438                         if (boundary) {
 439                                 boundary_block = bh->b_blocknr;
 440                                 boundary_bdev = bh->b_bdev;
 441                         }
 442                         bdev = bh->b_bdev;
 443                 } while ((bh = bh->b_this_page) != head);
 444
 445                 if (first_unmapped)
 446                         goto page_is_mapped;
 447
 448                 /*
 449                  * Page has buffers, but they are all unmapped. The page was
 450                  * created by pagein or read over a hole which was handled by
 451                  * block_read_full_page().  If this address_space is also
 452                  * using mpage_readpages then this can rarely happen.
 453                  */
 454                 goto confused;
 455         }
 456
 457         /*
 458          * The page has no buffers: map it to disk
 459          */
 460         BUG_ON(!PageUptodate(page));
 461         block_in_file = page->index << (PAGE_CACHE_SHIFT - blkbits);
 462         last_block = (inode->i_size - 1) >> blkbits;
 463         map_bh.b_page = page;
 464         for (page_block = 0; page_block < blocks_per_page; ) {
 465
 466                 map_bh.b_state = 0;
 467                 if (get_block(inode, block_in_file, &map_bh, 1))
 468                         goto confused;
 469                 if (buffer_new(&map_bh))
 470                         unmap_underlying_metadata(map_bh.b_bdev,
 471                                                 map_bh.b_blocknr);
 472                 if (buffer_boundary(&map_bh)) {
 473                         boundary_block = map_bh.b_blocknr;
 474                         boundary_bdev = map_bh.b_bdev;
 475                 }
 476                 if (page_block) {
 477                         if (map_bh.b_blocknr != blocks[page_block-1] + 1)
 478                                 goto confused;
 479                 }
 480                 blocks[page_block++] = map_bh.b_blocknr;
 481                 boundary = buffer_boundary(&map_bh);
 482                 bdev = map_bh.b_bdev;
 483                 if (block_in_file == last_block)
 484                         break;
 485                 block_in_file++;
 486         }
 487         if (page_block == 0)
 488                 buffer_error();
 489
 490         first_unmapped = page_block;
 491
 492         end_index = inode->i_size >> PAGE_CACHE_SHIFT;
 493         if (page->index >= end_index) {
 494                 unsigned offset = inode->i_size & (PAGE_CACHE_SIZE - 1);
 495                 char *kaddr;
 496
 497                 if (page->index > end_index || !offset)
 498                         goto confused;
 499                 kaddr = kmap_atomic(page, KM_USER0);
 500                 memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
 501                 flush_dcache_page(page);
 502                 kunmap_atomic(kaddr, KM_USER0);
 503         }
 504
 505 page_is_mapped:
 506
 507         /*
 508          * This page will go to BIO.  Do we need to send this BIO off first?
 509          */
 510         if (bio && *last_block_in_bio != blocks[0] - 1)
 511                 bio = mpage_bio_submit(WRITE, bio);
 512
 513 alloc_new:
 514         if (bio == NULL) {
 515                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 516                                 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
 517                 if (bio == NULL)
 518                         goto confused;
 519         }
 520
 521         /*
 522          * OK, we have our BIO, so we can now mark the buffers clean.  Make
 523          * sure to only clean buffers which we know we'll be writing.
 524          */
 525         if (page_has_buffers(page)) {
 526                 struct buffer_head *head = page_buffers(page);
 527                 struct buffer_head *bh = head;
 528                 unsigned buffer_counter = 0;
 529
 530                 do {
 531                         if (buffer_counter++ == first_unmapped)
 532                                 break;
 533                         clear_buffer_dirty(bh);
 534                         bh = bh->b_this_page;
 535                 } while (bh != head);
 536
 537                 if (buffer_heads_over_limit)
 538                         try_to_free_buffers(page);
 539         }
 540
 541         length = first_unmapped << blkbits;
 542         if (bio_add_page(bio, page, length, 0) < length) {
 543                 bio = mpage_bio_submit(WRITE, bio);
 544                 goto alloc_new;
 545         }
 546
 547         BUG_ON(PageWriteback(page));
 548         SetPageWriteback(page);
 549         unlock_page(page);
 550         if (boundary || (first_unmapped != blocks_per_page)) {
 551                 bio = mpage_bio_submit(WRITE, bio);
 552                 if (boundary_block) {
 553                         write_boundary_block(boundary_bdev,
 554                                         boundary_block, 1 << blkbits);
 555                 }
 556         } else {
 557                 *last_block_in_bio = blocks[blocks_per_page - 1];
 558         }
 559         goto out;
 560
 561 confused:
 562         if (bio)
 563                 bio = mpage_bio_submit(WRITE, bio);
 564         *ret = page->mapping->a_ops->writepage(page, wbc);
 565 out:
 566         return bio;
 567 }
 568
 569 /**
 570  * mpage_writepages - walk the list of dirty pages of the given
 571  * address space and writepage() all of them.
 572  *
 573  * @mapping: address space structure to write
 574  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
 575  * @get_block: the filesystem's block mapper function.
 576  *             If this is NULL then use a_ops->writepage.  Otherwise, go
 577  *             direct-to-BIO.
 578  *
 579  * This is a library function, which implements the writepages()
 580  * address_space_operation.
 581  *
 582  * (The next two paragraphs refer to code which isn't here yet, but they
 583  *  explain the presence of address_space.io_pages)
 584  *
 585  * Pages can be moved from clean_pages or locked_pages onto dirty_pages
 586  * at any time - it's not possible to lock against that.  So pages which
 587  * have already been added to a BIO may magically reappear on the dirty_pages
 588  * list.  And mpage_writepages() will again try to lock those pages.
 589  * But I/O has not yet been started against the page.  Thus deadlock.
 590  *
 591  * To avoid this, mpage_writepages() will only write pages from io_pages. The
 592  * caller must place them there.  We walk io_pages, locking the pages and
 593  * submitting them for I/O, moving them to locked_pages.
 594  *
 595  * This has the added benefit of preventing a livelock which would otherwise
 596  * occur if pages are being dirtied faster than we can write them out.
 597  *
 598  * If a page is already under I/O, generic_writepages() skips it, even
 599  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
 600  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
 601  * and msync() need to guarantee that all the data which was dirty at the time
 602  * the call was made get new I/O started against them.  So if called_for_sync()
 603  * is true, we must wait for existing IO to complete.
 604  *
 605  * It's fairly rare for PageWriteback pages to be on ->dirty_pages.  It
 606  * means that someone redirtied the page while it was under I/O.
 607  */
 608 int
 609 mpage_writepages(struct address_space *mapping,
 610                 struct writeback_control *wbc, get_block_t get_block)
 611 {
 612         struct backing_dev_info *bdi = mapping->backing_dev_info;
 613         struct bio *bio = NULL;
 614         sector_t last_block_in_bio = 0;
 615         int ret = 0;
 616         int done = 0;
 617         struct pagevec pvec;
 618         int (*writepage)(struct page *page, struct writeback_control *wbc);
 619
 620         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 621                 wbc->encountered_congestion = 1;
 622                 return 0;
 623         }
 624
 625         writepage = NULL;
 626         if (get_block == NULL)
 627                 writepage = mapping->a_ops->writepage;
 628
 629         pagevec_init(&pvec, 0);
 630         spin_lock(&mapping->page_lock);
 631         while (!list_empty(&mapping->io_pages) && !done) {
 632                 struct page *page = list_entry(mapping->io_pages.prev,
 633                                         struct page, list);
 634                 list_del(&page->list);
 635                 if (PageWriteback(page) && wbc->sync_mode == WB_SYNC_NONE) {
 636                         if (PageDirty(page)) {
 637                                 list_add(&page->list, &mapping->dirty_pages);
 638                                 continue;
 639                         }
 640                         list_add(&page->list, &mapping->locked_pages);
 641                         continue;
 642                 }
 643                 if (!PageDirty(page)) {
 644                         list_add(&page->list, &mapping->clean_pages);
 645                         continue;
 646                 }
 647                 list_add(&page->list, &mapping->locked_pages);
 648
 649                 page_cache_get(page);
 650                 spin_unlock(&mapping->page_lock);
 651
 652                 /*
 653                  * At this point we hold neither mapping->page_lock nor
 654                  * lock on the page itself: the page may be truncated or
 655                  * invalidated (changing page->mapping to NULL), or even
 656                  * swizzled back from swapper_space to tmpfs file mapping.
 657                  */
 658
 659                 lock_page(page);
 660
 661                 if (wbc->sync_mode != WB_SYNC_NONE)
 662                         wait_on_page_writeback(page);
 663
 664                 if (page->mapping == mapping && !PageWriteback(page) &&
 665                                         test_clear_page_dirty(page)) {
 666                         if (writepage) {
 667                                 ret = (*writepage)(page, wbc);
 668                         } else {
 669                                 bio = mpage_writepage(bio, page, get_block,
 670                                         &last_block_in_bio, &ret, wbc);
 671                         }
 672                         if (ret || (--(wbc->nr_to_write) <= 0))
 673                                 done = 1;
 674                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 675                                 wbc->encountered_congestion = 1;
 676                                 done = 1;
 677                         }
 678                 } else {
 679                         unlock_page(page);
 680                 }
 681                 page_cache_release(page);
 682                 spin_lock(&mapping->page_lock);
 683         }
 684         /*
 685          * Leave any remaining dirty pages on ->io_pages
 686          */
 687         spin_unlock(&mapping->page_lock);
 688         if (bio)
 689                 mpage_bio_submit(WRITE, bio);
 690         return ret;
 691 }
 692 EXPORT_SYMBOL(mpage_writepages);