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                         if (page->mapping)
  85                                 set_bit(AS_EIO, &page->mapping->flags);
  86                 }
  87                 end_page_writeback(page);
  88         } while (bvec >= bio->bi_io_vec);
  89         bio_put(bio);
  90         return 0;
  91 }
  92
  93 static struct bio *mpage_bio_submit(int rw, struct bio *bio)
  94 {
  95         bio->bi_end_io = mpage_end_io_read;
  96         if (rw == WRITE)
  97                 bio->bi_end_io = mpage_end_io_write;
  98         submit_bio(rw, bio);
  99         return NULL;
 100 }
 101
 102 static struct bio *
 103 mpage_alloc(struct block_device *bdev,
 104                 sector_t first_sector, int nr_vecs,
 105                 gfp_t gfp_flags)
 106 {
 107         struct bio *bio;
 108
 109         bio = bio_alloc(gfp_flags, nr_vecs);
 110
 111         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
 112                 while (!bio && (nr_vecs /= 2))
 113                         bio = bio_alloc(gfp_flags, nr_vecs);
 114         }
 115
 116         if (bio) {
 117                 bio->bi_bdev = bdev;
 118                 bio->bi_sector = first_sector;
 119         }
 120         return bio;
 121 }
 122
 123 /*
 124  * support function for mpage_readpages.  The fs supplied get_block might
 125  * return an up to date buffer.  This is used to map that buffer into
 126  * the page, which allows readpage to avoid triggering a duplicate call
 127  * to get_block.
 128  *
 129  * The idea is to avoid adding buffers to pages that don't already have
 130  * them.  So when the buffer is up to date and the page size == block size,
 131  * this marks the page up to date instead of adding new buffers.
 132  */
 133 static void
 134 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
 135 {
 136         struct inode *inode = page->mapping->host;
 137         struct buffer_head *page_bh, *head;
 138         int block = 0;
 139
 140         if (!page_has_buffers(page)) {
 141                 /*
 142                  * don't make any buffers if there is only one buffer on
 143                  * the page and the page just needs to be set up to date
 144                  */
 145                 if (inode->i_blkbits == PAGE_CACHE_SHIFT &&
 146                     buffer_uptodate(bh)) {
 147                         SetPageUptodate(page);
 148                         return;
 149                 }
 150                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
 151         }
 152         head = page_buffers(page);
 153         page_bh = head;
 154         do {
 155                 if (block == page_block) {
 156                         page_bh->b_state = bh->b_state;
 157                         page_bh->b_bdev = bh->b_bdev;
 158                         page_bh->b_blocknr = bh->b_blocknr;
 159                         break;
 160                 }
 161                 page_bh = page_bh->b_this_page;
 162                 block++;
 163         } while (page_bh != head);
 164 }
 165
 166 static struct bio *
 167 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
 168                         sector_t *last_block_in_bio, get_block_t get_block)
 169 {
 170         struct inode *inode = page->mapping->host;
 171         const unsigned blkbits = inode->i_blkbits;
 172         const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
 173         const unsigned blocksize = 1 << blkbits;
 174         sector_t block_in_file;
 175         sector_t last_block;
 176         sector_t blocks[MAX_BUF_PER_PAGE];
 177         unsigned page_block;
 178         unsigned first_hole = blocks_per_page;
 179         struct block_device *bdev = NULL;
 180         struct buffer_head bh;
 181         int length;
 182         int fully_mapped = 1;
 183
 184         if (page_has_buffers(page))
 185                 goto confused;
 186
 187         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
 188         last_block = (i_size_read(inode) + blocksize - 1) >> blkbits;
 189
 190         bh.b_page = page;
 191         for (page_block = 0; page_block < blocks_per_page;
 192                                 page_block++, block_in_file++) {
 193                 bh.b_state = 0;
 194                 if (block_in_file < last_block) {
 195                         bh.b_size = blocksize;
 196                         if (get_block(inode, block_in_file, &bh, 0))
 197                                 goto confused;
 198                 }
 199
 200                 if (!buffer_mapped(&bh)) {
 201                         fully_mapped = 0;
 202                         if (first_hole == blocks_per_page)
 203                                 first_hole = page_block;
 204                         continue;
 205                 }
 206
 207                 /* some filesystems will copy data into the page during
 208                  * the get_block call, in which case we don't want to
 209                  * read it again.  map_buffer_to_page copies the data
 210                  * we just collected from get_block into the page's buffers
 211                  * so readpage doesn't have to repeat the get_block call
 212                  */
 213                 if (buffer_uptodate(&bh)) {
 214                         map_buffer_to_page(page, &bh, page_block);
 215                         goto confused;
 216                 }
 217
 218                 if (first_hole != blocks_per_page)
 219                         goto confused;          /* hole -> non-hole */
 220
 221                 /* Contiguous blocks? */
 222                 if (page_block && blocks[page_block-1] != bh.b_blocknr-1)
 223                         goto confused;
 224                 blocks[page_block] = bh.b_blocknr;
 225                 bdev = bh.b_bdev;
 226         }
 227
 228         if (first_hole != blocks_per_page) {
 229                 char *kaddr = kmap_atomic(page, KM_USER0);
 230                 memset(kaddr + (first_hole << blkbits), 0,
 231                                 PAGE_CACHE_SIZE - (first_hole << blkbits));
 232                 flush_dcache_page(page);
 233                 kunmap_atomic(kaddr, KM_USER0);
 234                 if (first_hole == 0) {
 235                         SetPageUptodate(page);
 236                         unlock_page(page);
 237                         goto out;
 238                 }
 239         } else if (fully_mapped) {
 240                 SetPageMappedToDisk(page);
 241         }
 242
 243         /*
 244          * This page will go to BIO.  Do we need to send this BIO off first?
 245          */
 246         if (bio && (*last_block_in_bio != blocks[0] - 1))
 247                 bio = mpage_bio_submit(READ, bio);
 248
 249 alloc_new:
 250         if (bio == NULL) {
 251                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 252                                 min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
 253                                 GFP_KERNEL);
 254                 if (bio == NULL)
 255                         goto confused;
 256         }
 257
 258         length = first_hole << blkbits;
 259         if (bio_add_page(bio, page, length, 0) < length) {
 260                 bio = mpage_bio_submit(READ, bio);
 261                 goto alloc_new;
 262         }
 263
 264         if (buffer_boundary(&bh) || (first_hole != blocks_per_page))
 265                 bio = mpage_bio_submit(READ, bio);
 266         else
 267                 *last_block_in_bio = blocks[blocks_per_page - 1];
 268 out:
 269         return bio;
 270
 271 confused:
 272         if (bio)
 273                 bio = mpage_bio_submit(READ, bio);
 274         if (!PageUptodate(page))
 275                 block_read_full_page(page, get_block);
 276         else
 277                 unlock_page(page);
 278         goto out;
 279 }
 280
 281 /**
 282  * mpage_readpages - populate an address space with some pages, and
 283  *                       start reads against them.
 284  *
 285  * @mapping: the address_space
 286  * @pages: The address of a list_head which contains the target pages.  These
 287  *   pages have their ->index populated and are otherwise uninitialised.
 288  *
 289  *   The page at @pages->prev has the lowest file offset, and reads should be
 290  *   issued in @pages->prev to @pages->next order.
 291  *
 292  * @nr_pages: The number of pages at *@pages
 293  * @get_block: The filesystem's block mapper function.
 294  *
 295  * This function walks the pages and the blocks within each page, building and
 296  * emitting large BIOs.
 297  *
 298  * If anything unusual happens, such as:
 299  *
 300  * - encountering a page which has buffers
 301  * - encountering a page which has a non-hole after a hole
 302  * - encountering a page with non-contiguous blocks
 303  *
 304  * then this code just gives up and calls the buffer_head-based read function.
 305  * It does handle a page which has holes at the end - that is a common case:
 306  * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
 307  *
 308  * BH_Boundary explanation:
 309  *
 310  * There is a problem.  The mpage read code assembles several pages, gets all
 311  * their disk mappings, and then submits them all.  That's fine, but obtaining
 312  * the disk mappings may require I/O.  Reads of indirect blocks, for example.
 313  *
 314  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
 315  * submitted in the following order:
 316  *      12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
 317  * because the indirect block has to be read to get the mappings of blocks
 318  * 13,14,15,16.  Obviously, this impacts performance.
 319  *
 320  * So what we do it to allow the filesystem's get_block() function to set
 321  * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
 322  * after this one will require I/O against a block which is probably close to
 323  * this one.  So you should push what I/O you have currently accumulated.
 324  *
 325  * This all causes the disk requests to be issued in the correct order.
 326  */
 327 int
 328 mpage_readpages(struct address_space *mapping, struct list_head *pages,
 329                                 unsigned nr_pages, get_block_t get_block)
 330 {
 331         struct bio *bio = NULL;
 332         unsigned page_idx;
 333         sector_t last_block_in_bio = 0;
 334         struct pagevec lru_pvec;
 335
 336         pagevec_init(&lru_pvec, 0);
 337         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 338                 struct page *page = list_entry(pages->prev, struct page, lru);
 339
 340                 prefetchw(&page->flags);
 341                 list_del(&page->lru);
 342                 if (!add_to_page_cache(page, mapping,
 343                                         page->index, GFP_KERNEL)) {
 344                         bio = do_mpage_readpage(bio, page,
 345                                         nr_pages - page_idx,
 346                                         &last_block_in_bio, get_block);
 347                         if (!pagevec_add(&lru_pvec, page))
 348                                 __pagevec_lru_add(&lru_pvec);
 349                 } else {
 350                         page_cache_release(page);
 351                 }
 352         }
 353         pagevec_lru_add(&lru_pvec);
 354         BUG_ON(!list_empty(pages));
 355         if (bio)
 356                 mpage_bio_submit(READ, bio);
 357         return 0;
 358 }
 359 EXPORT_SYMBOL(mpage_readpages);
 360
 361 /*
 362  * This isn't called much at all
 363  */
 364 int mpage_readpage(struct page *page, get_block_t get_block)
 365 {
 366         struct bio *bio = NULL;
 367         sector_t last_block_in_bio = 0;
 368
 369         bio = do_mpage_readpage(bio, page, 1,
 370                         &last_block_in_bio, get_block);
 371         if (bio)
 372                 mpage_bio_submit(READ, bio);
 373         return 0;
 374 }
 375 EXPORT_SYMBOL(mpage_readpage);
 376
 377 /*
 378  * Writing is not so simple.
 379  *
 380  * If the page has buffers then they will be used for obtaining the disk
 381  * mapping.  We only support pages which are fully mapped-and-dirty, with a
 382  * special case for pages which are unmapped at the end: end-of-file.
 383  *
 384  * If the page has no buffers (preferred) then the page is mapped here.
 385  *
 386  * If all blocks are found to be contiguous then the page can go into the
 387  * BIO.  Otherwise fall back to the mapping's writepage().
 388  *
 389  * FIXME: This code wants an estimate of how many pages are still to be
 390  * written, so it can intelligently allocate a suitably-sized BIO.  For now,
 391  * just allocate full-size (16-page) BIOs.
 392  */
 393 static struct bio *
 394 __mpage_writepage(struct bio *bio, struct page *page, get_block_t get_block,
 395         sector_t *last_block_in_bio, int *ret, struct writeback_control *wbc,
 396         writepage_t writepage_fn)
 397 {
 398         struct address_space *mapping = page->mapping;
 399         struct inode *inode = page->mapping->host;
 400         const unsigned blkbits = inode->i_blkbits;
 401         unsigned long end_index;
 402         const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
 403         sector_t last_block;
 404         sector_t block_in_file;
 405         sector_t blocks[MAX_BUF_PER_PAGE];
 406         unsigned page_block;
 407         unsigned first_unmapped = blocks_per_page;
 408         struct block_device *bdev = NULL;
 409         int boundary = 0;
 410         sector_t boundary_block = 0;
 411         struct block_device *boundary_bdev = NULL;
 412         int length;
 413         struct buffer_head map_bh;
 414         loff_t i_size = i_size_read(inode);
 415
 416         if (page_has_buffers(page)) {
 417                 struct buffer_head *head = page_buffers(page);
 418                 struct buffer_head *bh = head;
 419
 420                 /* If they're all mapped and dirty, do it */
 421                 page_block = 0;
 422                 do {
 423                         BUG_ON(buffer_locked(bh));
 424                         if (!buffer_mapped(bh)) {
 425                                 /*
 426                                  * unmapped dirty buffers are created by
 427                                  * __set_page_dirty_buffers -> mmapped data
 428                                  */
 429                                 if (buffer_dirty(bh))
 430                                         goto confused;
 431                                 if (first_unmapped == blocks_per_page)
 432                                         first_unmapped = page_block;
 433                                 continue;
 434                         }
 435
 436                         if (first_unmapped != blocks_per_page)
 437                                 goto confused;  /* hole -> non-hole */
 438
 439                         if (!buffer_dirty(bh) || !buffer_uptodate(bh))
 440                                 goto confused;
 441                         if (page_block) {
 442                                 if (bh->b_blocknr != blocks[page_block-1] + 1)
 443                                         goto confused;
 444                         }
 445                         blocks[page_block++] = bh->b_blocknr;
 446                         boundary = buffer_boundary(bh);
 447                         if (boundary) {
 448                                 boundary_block = bh->b_blocknr;
 449                                 boundary_bdev = bh->b_bdev;
 450                         }
 451                         bdev = bh->b_bdev;
 452                 } while ((bh = bh->b_this_page) != head);
 453
 454                 if (first_unmapped)
 455                         goto page_is_mapped;
 456
 457                 /*
 458                  * Page has buffers, but they are all unmapped. The page was
 459                  * created by pagein or read over a hole which was handled by
 460                  * block_read_full_page().  If this address_space is also
 461                  * using mpage_readpages then this can rarely happen.
 462                  */
 463                 goto confused;
 464         }
 465
 466         /*
 467          * The page has no buffers: map it to disk
 468          */
 469         BUG_ON(!PageUptodate(page));
 470         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
 471         last_block = (i_size - 1) >> blkbits;
 472         map_bh.b_page = page;
 473         for (page_block = 0; page_block < blocks_per_page; ) {
 474
 475                 map_bh.b_state = 0;
 476                 map_bh.b_size = 1 << blkbits;
 477                 if (get_block(inode, block_in_file, &map_bh, 1))
 478                         goto confused;
 479                 if (buffer_new(&map_bh))
 480                         unmap_underlying_metadata(map_bh.b_bdev,
 481                                                 map_bh.b_blocknr);
 482                 if (buffer_boundary(&map_bh)) {
 483                         boundary_block = map_bh.b_blocknr;
 484                         boundary_bdev = map_bh.b_bdev;
 485                 }
 486                 if (page_block) {
 487                         if (map_bh.b_blocknr != blocks[page_block-1] + 1)
 488                                 goto confused;
 489                 }
 490                 blocks[page_block++] = map_bh.b_blocknr;
 491                 boundary = buffer_boundary(&map_bh);
 492                 bdev = map_bh.b_bdev;
 493                 if (block_in_file == last_block)
 494                         break;
 495                 block_in_file++;
 496         }
 497         BUG_ON(page_block == 0);
 498
 499         first_unmapped = page_block;
 500
 501 page_is_mapped:
 502         end_index = i_size >> PAGE_CACHE_SHIFT;
 503         if (page->index >= end_index) {
 504                 /*
 505                  * The page straddles i_size.  It must be zeroed out on each
 506                  * and every writepage invokation because it may be mmapped.
 507                  * "A file is mapped in multiples of the page size.  For a file
 508                  * that is not a multiple of the page size, the remaining memory
 509                  * is zeroed when mapped, and writes to that region are not
 510                  * written out to the file."
 511                  */
 512                 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
 513                 char *kaddr;
 514
 515                 if (page->index > end_index || !offset)
 516                         goto confused;
 517                 kaddr = kmap_atomic(page, KM_USER0);
 518                 memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
 519                 flush_dcache_page(page);
 520                 kunmap_atomic(kaddr, KM_USER0);
 521         }
 522
 523         /*
 524          * This page will go to BIO.  Do we need to send this BIO off first?
 525          */
 526         if (bio && *last_block_in_bio != blocks[0] - 1)
 527                 bio = mpage_bio_submit(WRITE, bio);
 528
 529 alloc_new:
 530         if (bio == NULL) {
 531                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 532                                 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
 533                 if (bio == NULL)
 534                         goto confused;
 535         }
 536
 537         /*
 538          * Must try to add the page before marking the buffer clean or
 539          * the confused fail path above (OOM) will be very confused when
 540          * it finds all bh marked clean (i.e. it will not write anything)
 541          */
 542         length = first_unmapped << blkbits;
 543         if (bio_add_page(bio, page, length, 0) < length) {
 544                 bio = mpage_bio_submit(WRITE, bio);
 545                 goto alloc_new;
 546         }
 547
 548         /*
 549          * OK, we have our BIO, so we can now mark the buffers clean.  Make
 550          * sure to only clean buffers which we know we'll be writing.
 551          */
 552         if (page_has_buffers(page)) {
 553                 struct buffer_head *head = page_buffers(page);
 554                 struct buffer_head *bh = head;
 555                 unsigned buffer_counter = 0;
 556
 557                 do {
 558                         if (buffer_counter++ == first_unmapped)
 559                                 break;
 560                         clear_buffer_dirty(bh);
 561                         bh = bh->b_this_page;
 562                 } while (bh != head);
 563
 564                 /*
 565                  * we cannot drop the bh if the page is not uptodate
 566                  * or a concurrent readpage would fail to serialize with the bh
 567                  * and it would read from disk before we reach the platter.
 568                  */
 569                 if (buffer_heads_over_limit && PageUptodate(page))
 570                         try_to_free_buffers(page);
 571         }
 572
 573         BUG_ON(PageWriteback(page));
 574         set_page_writeback(page);
 575         unlock_page(page);
 576         if (boundary || (first_unmapped != blocks_per_page)) {
 577                 bio = mpage_bio_submit(WRITE, bio);
 578                 if (boundary_block) {
 579                         write_boundary_block(boundary_bdev,
 580                                         boundary_block, 1 << blkbits);
 581                 }
 582         } else {
 583                 *last_block_in_bio = blocks[blocks_per_page - 1];
 584         }
 585         goto out;
 586
 587 confused:
 588         if (bio)
 589                 bio = mpage_bio_submit(WRITE, bio);
 590
 591         if (writepage_fn) {
 592                 *ret = (*writepage_fn)(page, wbc);
 593         } else {
 594                 *ret = -EAGAIN;
 595                 goto out;
 596         }
 597         /*
 598          * The caller has a ref on the inode, so *mapping is stable
 599          */
 600         if (*ret) {
 601                 if (*ret == -ENOSPC)
 602                         set_bit(AS_ENOSPC, &mapping->flags);
 603                 else
 604                         set_bit(AS_EIO, &mapping->flags);
 605         }
 606 out:
 607         return bio;
 608 }
 609
 610 /**
 611  * mpage_writepages - walk the list of dirty pages of the given
 612  * address space and writepage() all of them.
 613  *
 614  * @mapping: address space structure to write
 615  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
 616  * @get_block: the filesystem's block mapper function.
 617  *             If this is NULL then use a_ops->writepage.  Otherwise, go
 618  *             direct-to-BIO.
 619  *
 620  * This is a library function, which implements the writepages()
 621  * address_space_operation.
 622  *
 623  * If a page is already under I/O, generic_writepages() skips it, even
 624  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
 625  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
 626  * and msync() need to guarantee that all the data which was dirty at the time
 627  * the call was made get new I/O started against them.  If wbc->sync_mode is
 628  * WB_SYNC_ALL then we were called for data integrity and we must wait for
 629  * existing IO to complete.
 630  */
 631 int
 632 mpage_writepages(struct address_space *mapping,
 633                 struct writeback_control *wbc, get_block_t get_block)
 634 {
 635         struct backing_dev_info *bdi = mapping->backing_dev_info;
 636         struct bio *bio = NULL;
 637         sector_t last_block_in_bio = 0;
 638         int ret = 0;
 639         int done = 0;
 640         int (*writepage)(struct page *page, struct writeback_control *wbc);
 641         struct pagevec pvec;
 642         int nr_pages;
 643         pgoff_t index;
 644         pgoff_t end = -1;               /* Inclusive */
 645         int scanned = 0;
 646         int is_range = 0;
 647
 648         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 649                 wbc->encountered_congestion = 1;
 650                 return 0;
 651         }
 652
 653         writepage = NULL;
 654         if (get_block == NULL)
 655                 writepage = mapping->a_ops->writepage;
 656
 657         pagevec_init(&pvec, 0);
 658         if (wbc->sync_mode == WB_SYNC_NONE) {
 659                 index = mapping->writeback_index; /* Start from prev offset */
 660         } else {
 661                 index = 0;                        /* whole-file sweep */
 662                 scanned = 1;
 663         }
 664         if (wbc->start || wbc->end) {
 665                 index = wbc->start >> PAGE_CACHE_SHIFT;
 666                 end = wbc->end >> PAGE_CACHE_SHIFT;
 667                 is_range = 1;
 668                 scanned = 1;
 669         }
 670 retry:
 671         while (!done && (index <= end) &&
 672                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 673                         PAGECACHE_TAG_DIRTY,
 674                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
 675                 unsigned i;
 676
 677                 scanned = 1;
 678                 for (i = 0; i < nr_pages; i++) {
 679                         struct page *page = pvec.pages[i];
 680
 681                         /*
 682                          * At this point we hold neither mapping->tree_lock nor
 683                          * lock on the page itself: the page may be truncated or
 684                          * invalidated (changing page->mapping to NULL), or even
 685                          * swizzled back from swapper_space to tmpfs file
 686                          * mapping
 687                          */
 688
 689                         lock_page(page);
 690
 691                         if (unlikely(page->mapping != mapping)) {
 692                                 unlock_page(page);
 693                                 continue;
 694                         }
 695
 696                         if (unlikely(is_range) && page->index > end) {
 697                                 done = 1;
 698                                 unlock_page(page);
 699                                 continue;
 700                         }
 701
 702                         if (wbc->sync_mode != WB_SYNC_NONE)
 703                                 wait_on_page_writeback(page);
 704
 705                         if (PageWriteback(page) ||
 706                                         !clear_page_dirty_for_io(page)) {
 707                                 unlock_page(page);
 708                                 continue;
 709                         }
 710
 711                         if (writepage) {
 712                                 ret = (*writepage)(page, wbc);
 713                                 if (ret) {
 714                                         if (ret == -ENOSPC)
 715                                                 set_bit(AS_ENOSPC,
 716                                                         &mapping->flags);
 717                                         else
 718                                                 set_bit(AS_EIO,
 719                                                         &mapping->flags);
 720                                 }
 721                         } else {
 722                                 bio = __mpage_writepage(bio, page, get_block,
 723                                                 &last_block_in_bio, &ret, wbc,
 724                                                 page->mapping->a_ops->writepage);
 725                         }
 726                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE))
 727                                 unlock_page(page);
 728                         if (ret || (--(wbc->nr_to_write) <= 0))
 729                                 done = 1;
 730                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 731                                 wbc->encountered_congestion = 1;
 732                                 done = 1;
 733                         }
 734                 }
 735                 pagevec_release(&pvec);
 736                 cond_resched();
 737         }
 738         if (!scanned && !done) {
 739                 /*
 740                  * We hit the last page and there is more work to be done: wrap
 741                  * back to the start of the file
 742                  */
 743                 scanned = 1;
 744                 index = 0;
 745                 goto retry;
 746         }
 747         if (!is_range)
 748                 mapping->writeback_index = index;
 749         if (bio)
 750                 mpage_bio_submit(WRITE, bio);
 751         return ret;
 752 }
 753 EXPORT_SYMBOL(mpage_writepages);
 754
 755 int mpage_writepage(struct page *page, get_block_t get_block,
 756         struct writeback_control *wbc)
 757 {
 758         int ret = 0;
 759         struct bio *bio;
 760         sector_t last_block_in_bio = 0;
 761
 762         bio = __mpage_writepage(NULL, page, get_block,
 763                         &last_block_in_bio, &ret, wbc, NULL);
 764         if (bio)
 765                 mpage_bio_submit(WRITE, bio);
 766
 767         return ret;
 768 }
 769 EXPORT_SYMBOL(mpage_writepage);