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 void mpage_end_io_read(struct bio *bio, 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         do {
  48                 struct page *page = bvec->bv_page;
  49
  50                 if (--bvec >= bio->bi_io_vec)
  51                         prefetchw(&bvec->bv_page->flags);
  52
  53                 if (uptodate) {
  54                         SetPageUptodate(page);
  55                 } else {
  56                         ClearPageUptodate(page);
  57                         SetPageError(page);
  58                 }
  59                 unlock_page(page);
  60         } while (bvec >= bio->bi_io_vec);
  61         bio_put(bio);
  62 }
  63
  64 static void mpage_end_io_write(struct bio *bio, int err)
  65 {
  66         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
  67         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
  68
  69         do {
  70                 struct page *page = bvec->bv_page;
  71
  72                 if (--bvec >= bio->bi_io_vec)
  73                         prefetchw(&bvec->bv_page->flags);
  74
  75                 if (!uptodate){
  76                         SetPageError(page);
  77                         if (page->mapping)
  78                                 set_bit(AS_EIO, &page->mapping->flags);
  79                 }
  80                 end_page_writeback(page);
  81         } while (bvec >= bio->bi_io_vec);
  82         bio_put(bio);
  83 }
  84
  85 static struct bio *mpage_bio_submit(int rw, struct bio *bio)
  86 {
  87         bio->bi_end_io = mpage_end_io_read;
  88         if (rw == WRITE)
  89                 bio->bi_end_io = mpage_end_io_write;
  90         submit_bio(rw, bio);
  91         return NULL;
  92 }
  93
  94 static struct bio *
  95 mpage_alloc(struct block_device *bdev,
  96                 sector_t first_sector, int nr_vecs,
  97                 gfp_t gfp_flags)
  98 {
  99         struct bio *bio;
 100
 101         bio = bio_alloc(gfp_flags, nr_vecs);
 102
 103         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
 104                 while (!bio && (nr_vecs /= 2))
 105                         bio = bio_alloc(gfp_flags, nr_vecs);
 106         }
 107
 108         if (bio) {
 109                 bio->bi_bdev = bdev;
 110                 bio->bi_sector = first_sector;
 111         }
 112         return bio;
 113 }
 114
 115 /*
 116  * support function for mpage_readpages.  The fs supplied get_block might
 117  * return an up to date buffer.  This is used to map that buffer into
 118  * the page, which allows readpage to avoid triggering a duplicate call
 119  * to get_block.
 120  *
 121  * The idea is to avoid adding buffers to pages that don't already have
 122  * them.  So when the buffer is up to date and the page size == block size,
 123  * this marks the page up to date instead of adding new buffers.
 124  */
 125 static void
 126 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
 127 {
 128         struct inode *inode = page->mapping->host;
 129         struct buffer_head *page_bh, *head;
 130         int block = 0;
 131
 132         if (!page_has_buffers(page)) {
 133                 /*
 134                  * don't make any buffers if there is only one buffer on
 135                  * the page and the page just needs to be set up to date
 136                  */
 137                 if (inode->i_blkbits == PAGE_CACHE_SHIFT &&
 138                     buffer_uptodate(bh)) {
 139                         SetPageUptodate(page);
 140                         return;
 141                 }
 142                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
 143         }
 144         head = page_buffers(page);
 145         page_bh = head;
 146         do {
 147                 if (block == page_block) {
 148                         page_bh->b_state = bh->b_state;
 149                         page_bh->b_bdev = bh->b_bdev;
 150                         page_bh->b_blocknr = bh->b_blocknr;
 151                         break;
 152                 }
 153                 page_bh = page_bh->b_this_page;
 154                 block++;
 155         } while (page_bh != head);
 156 }
 157
 158 /*
 159  * This is the worker routine which does all the work of mapping the disk
 160  * blocks and constructs largest possible bios, submits them for IO if the
 161  * blocks are not contiguous on the disk.
 162  *
 163  * We pass a buffer_head back and forth and use its buffer_mapped() flag to
 164  * represent the validity of its disk mapping and to decide when to do the next
 165  * get_block() call.
 166  */
 167 static struct bio *
 168 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
 169                 sector_t *last_block_in_bio, struct buffer_head *map_bh,
 170                 unsigned long *first_logical_block, get_block_t get_block)
 171 {
 172         struct inode *inode = page->mapping->host;
 173         const unsigned blkbits = inode->i_blkbits;
 174         const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
 175         const unsigned blocksize = 1 << blkbits;
 176         sector_t block_in_file;
 177         sector_t last_block;
 178         sector_t last_block_in_file;
 179         sector_t blocks[MAX_BUF_PER_PAGE];
 180         unsigned page_block;
 181         unsigned first_hole = blocks_per_page;
 182         struct block_device *bdev = NULL;
 183         int length;
 184         int fully_mapped = 1;
 185         unsigned nblocks;
 186         unsigned relative_block;
 187
 188         if (page_has_buffers(page))
 189                 goto confused;
 190
 191         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
 192         last_block = block_in_file + nr_pages * blocks_per_page;
 193         last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
 194         if (last_block > last_block_in_file)
 195                 last_block = last_block_in_file;
 196         page_block = 0;
 197
 198         /*
 199          * Map blocks using the result from the previous get_blocks call first.
 200          */
 201         nblocks = map_bh->b_size >> blkbits;
 202         if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
 203                         block_in_file < (*first_logical_block + nblocks)) {
 204                 unsigned map_offset = block_in_file - *first_logical_block;
 205                 unsigned last = nblocks - map_offset;
 206
 207                 for (relative_block = 0; ; relative_block++) {
 208                         if (relative_block == last) {
 209                                 clear_buffer_mapped(map_bh);
 210                                 break;
 211                         }
 212                         if (page_block == blocks_per_page)
 213                                 break;
 214                         blocks[page_block] = map_bh->b_blocknr + map_offset +
 215                                                 relative_block;
 216                         page_block++;
 217                         block_in_file++;
 218                 }
 219                 bdev = map_bh->b_bdev;
 220         }
 221
 222         /*
 223          * Then do more get_blocks calls until we are done with this page.
 224          */
 225         map_bh->b_page = page;
 226         while (page_block < blocks_per_page) {
 227                 map_bh->b_state = 0;
 228                 map_bh->b_size = 0;
 229
 230                 if (block_in_file < last_block) {
 231                         map_bh->b_size = (last_block-block_in_file) << blkbits;
 232                         if (get_block(inode, block_in_file, map_bh, 0))
 233                                 goto confused;
 234                         *first_logical_block = block_in_file;
 235                 }
 236
 237                 if (!buffer_mapped(map_bh)) {
 238                         fully_mapped = 0;
 239                         if (first_hole == blocks_per_page)
 240                                 first_hole = page_block;
 241                         page_block++;
 242                         block_in_file++;
 243                         clear_buffer_mapped(map_bh);
 244                         continue;
 245                 }
 246
 247                 /* some filesystems will copy data into the page during
 248                  * the get_block call, in which case we don't want to
 249                  * read it again.  map_buffer_to_page copies the data
 250                  * we just collected from get_block into the page's buffers
 251                  * so readpage doesn't have to repeat the get_block call
 252                  */
 253                 if (buffer_uptodate(map_bh)) {
 254                         map_buffer_to_page(page, map_bh, page_block);
 255                         goto confused;
 256                 }
 257
 258                 if (first_hole != blocks_per_page)
 259                         goto confused;          /* hole -> non-hole */
 260
 261                 /* Contiguous blocks? */
 262                 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
 263                         goto confused;
 264                 nblocks = map_bh->b_size >> blkbits;
 265                 for (relative_block = 0; ; relative_block++) {
 266                         if (relative_block == nblocks) {
 267                                 clear_buffer_mapped(map_bh);
 268                                 break;
 269                         } else if (page_block == blocks_per_page)
 270                                 break;
 271                         blocks[page_block] = map_bh->b_blocknr+relative_block;
 272                         page_block++;
 273                         block_in_file++;
 274                 }
 275                 bdev = map_bh->b_bdev;
 276         }
 277
 278         if (first_hole != blocks_per_page) {
 279                 zero_user_segment(page, first_hole << blkbits, PAGE_CACHE_SIZE);
 280                 if (first_hole == 0) {
 281                         SetPageUptodate(page);
 282                         unlock_page(page);
 283                         goto out;
 284                 }
 285         } else if (fully_mapped) {
 286                 SetPageMappedToDisk(page);
 287         }
 288
 289         /*
 290          * This page will go to BIO.  Do we need to send this BIO off first?
 291          */
 292         if (bio && (*last_block_in_bio != blocks[0] - 1))
 293                 bio = mpage_bio_submit(READ, bio);
 294
 295 alloc_new:
 296         if (bio == NULL) {
 297                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 298                                 min_t(int, nr_pages, bio_get_nr_vecs(bdev)),
 299                                 GFP_KERNEL);
 300                 if (bio == NULL)
 301                         goto confused;
 302         }
 303
 304         length = first_hole << blkbits;
 305         if (bio_add_page(bio, page, length, 0) < length) {
 306                 bio = mpage_bio_submit(READ, bio);
 307                 goto alloc_new;
 308         }
 309
 310         if (buffer_boundary(map_bh) || (first_hole != blocks_per_page))
 311                 bio = mpage_bio_submit(READ, bio);
 312         else
 313                 *last_block_in_bio = blocks[blocks_per_page - 1];
 314 out:
 315         return bio;
 316
 317 confused:
 318         if (bio)
 319                 bio = mpage_bio_submit(READ, bio);
 320         if (!PageUptodate(page))
 321                 block_read_full_page(page, get_block);
 322         else
 323                 unlock_page(page);
 324         goto out;
 325 }
 326
 327 /**
 328  * mpage_readpages - populate an address space with some pages & start reads against them
 329  * @mapping: the address_space
 330  * @pages: The address of a list_head which contains the target pages.  These
 331  *   pages have their ->index populated and are otherwise uninitialised.
 332  *   The page at @pages->prev has the lowest file offset, and reads should be
 333  *   issued in @pages->prev to @pages->next order.
 334  * @nr_pages: The number of pages at *@pages
 335  * @get_block: The filesystem's block mapper function.
 336  *
 337  * This function walks the pages and the blocks within each page, building and
 338  * emitting large BIOs.
 339  *
 340  * If anything unusual happens, such as:
 341  *
 342  * - encountering a page which has buffers
 343  * - encountering a page which has a non-hole after a hole
 344  * - encountering a page with non-contiguous blocks
 345  *
 346  * then this code just gives up and calls the buffer_head-based read function.
 347  * It does handle a page which has holes at the end - that is a common case:
 348  * the end-of-file on blocksize < PAGE_CACHE_SIZE setups.
 349  *
 350  * BH_Boundary explanation:
 351  *
 352  * There is a problem.  The mpage read code assembles several pages, gets all
 353  * their disk mappings, and then submits them all.  That's fine, but obtaining
 354  * the disk mappings may require I/O.  Reads of indirect blocks, for example.
 355  *
 356  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
 357  * submitted in the following order:
 358  *      12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
 359  *
 360  * because the indirect block has to be read to get the mappings of blocks
 361  * 13,14,15,16.  Obviously, this impacts performance.
 362  *
 363  * So what we do it to allow the filesystem's get_block() function to set
 364  * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
 365  * after this one will require I/O against a block which is probably close to
 366  * this one.  So you should push what I/O you have currently accumulated.
 367  *
 368  * This all causes the disk requests to be issued in the correct order.
 369  */
 370 int
 371 mpage_readpages(struct address_space *mapping, struct list_head *pages,
 372                                 unsigned nr_pages, get_block_t get_block)
 373 {
 374         struct bio *bio = NULL;
 375         unsigned page_idx;
 376         sector_t last_block_in_bio = 0;
 377         struct buffer_head map_bh;
 378         unsigned long first_logical_block = 0;
 379
 380         clear_buffer_mapped(&map_bh);
 381         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 382                 struct page *page = list_entry(pages->prev, struct page, lru);
 383
 384                 prefetchw(&page->flags);
 385                 list_del(&page->lru);
 386                 if (!add_to_page_cache_lru(page, mapping,
 387                                         page->index, GFP_KERNEL)) {
 388                         bio = do_mpage_readpage(bio, page,
 389                                         nr_pages - page_idx,
 390                                         &last_block_in_bio, &map_bh,
 391                                         &first_logical_block,
 392                                         get_block);
 393                 }
 394                 page_cache_release(page);
 395         }
 396         BUG_ON(!list_empty(pages));
 397         if (bio)
 398                 mpage_bio_submit(READ, bio);
 399         return 0;
 400 }
 401 EXPORT_SYMBOL(mpage_readpages);
 402
 403 /*
 404  * This isn't called much at all
 405  */
 406 int mpage_readpage(struct page *page, get_block_t get_block)
 407 {
 408         struct bio *bio = NULL;
 409         sector_t last_block_in_bio = 0;
 410         struct buffer_head map_bh;
 411         unsigned long first_logical_block = 0;
 412
 413         clear_buffer_mapped(&map_bh);
 414         bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
 415                         &map_bh, &first_logical_block, get_block);
 416         if (bio)
 417                 mpage_bio_submit(READ, bio);
 418         return 0;
 419 }
 420 EXPORT_SYMBOL(mpage_readpage);
 421
 422 /*
 423  * Writing is not so simple.
 424  *
 425  * If the page has buffers then they will be used for obtaining the disk
 426  * mapping.  We only support pages which are fully mapped-and-dirty, with a
 427  * special case for pages which are unmapped at the end: end-of-file.
 428  *
 429  * If the page has no buffers (preferred) then the page is mapped here.
 430  *
 431  * If all blocks are found to be contiguous then the page can go into the
 432  * BIO.  Otherwise fall back to the mapping's writepage().
 433  *
 434  * FIXME: This code wants an estimate of how many pages are still to be
 435  * written, so it can intelligently allocate a suitably-sized BIO.  For now,
 436  * just allocate full-size (16-page) BIOs.
 437  */
 438 struct mpage_data {
 439         struct bio *bio;
 440         sector_t last_block_in_bio;
 441         get_block_t *get_block;
 442         unsigned use_writepage;
 443 };
 444
 445 static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
 446                              void *data)
 447 {
 448         struct mpage_data *mpd = data;
 449         struct bio *bio = mpd->bio;
 450         struct address_space *mapping = page->mapping;
 451         struct inode *inode = page->mapping->host;
 452         const unsigned blkbits = inode->i_blkbits;
 453         unsigned long end_index;
 454         const unsigned blocks_per_page = PAGE_CACHE_SIZE >> blkbits;
 455         sector_t last_block;
 456         sector_t block_in_file;
 457         sector_t blocks[MAX_BUF_PER_PAGE];
 458         unsigned page_block;
 459         unsigned first_unmapped = blocks_per_page;
 460         struct block_device *bdev = NULL;
 461         int boundary = 0;
 462         sector_t boundary_block = 0;
 463         struct block_device *boundary_bdev = NULL;
 464         int length;
 465         struct buffer_head map_bh;
 466         loff_t i_size = i_size_read(inode);
 467         int ret = 0;
 468
 469         if (page_has_buffers(page)) {
 470                 struct buffer_head *head = page_buffers(page);
 471                 struct buffer_head *bh = head;
 472
 473                 /* If they're all mapped and dirty, do it */
 474                 page_block = 0;
 475                 do {
 476                         BUG_ON(buffer_locked(bh));
 477                         if (!buffer_mapped(bh)) {
 478                                 /*
 479                                  * unmapped dirty buffers are created by
 480                                  * __set_page_dirty_buffers -> mmapped data
 481                                  */
 482                                 if (buffer_dirty(bh))
 483                                         goto confused;
 484                                 if (first_unmapped == blocks_per_page)
 485                                         first_unmapped = page_block;
 486                                 continue;
 487                         }
 488
 489                         if (first_unmapped != blocks_per_page)
 490                                 goto confused;  /* hole -> non-hole */
 491
 492                         if (!buffer_dirty(bh) || !buffer_uptodate(bh))
 493                                 goto confused;
 494                         if (page_block) {
 495                                 if (bh->b_blocknr != blocks[page_block-1] + 1)
 496                                         goto confused;
 497                         }
 498                         blocks[page_block++] = bh->b_blocknr;
 499                         boundary = buffer_boundary(bh);
 500                         if (boundary) {
 501                                 boundary_block = bh->b_blocknr;
 502                                 boundary_bdev = bh->b_bdev;
 503                         }
 504                         bdev = bh->b_bdev;
 505                 } while ((bh = bh->b_this_page) != head);
 506
 507                 if (first_unmapped)
 508                         goto page_is_mapped;
 509
 510                 /*
 511                  * Page has buffers, but they are all unmapped. The page was
 512                  * created by pagein or read over a hole which was handled by
 513                  * block_read_full_page().  If this address_space is also
 514                  * using mpage_readpages then this can rarely happen.
 515                  */
 516                 goto confused;
 517         }
 518
 519         /*
 520          * The page has no buffers: map it to disk
 521          */
 522         BUG_ON(!PageUptodate(page));
 523         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
 524         last_block = (i_size - 1) >> blkbits;
 525         map_bh.b_page = page;
 526         for (page_block = 0; page_block < blocks_per_page; ) {
 527
 528                 map_bh.b_state = 0;
 529                 map_bh.b_size = 1 << blkbits;
 530                 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
 531                         goto confused;
 532                 if (buffer_new(&map_bh))
 533                         unmap_underlying_metadata(map_bh.b_bdev,
 534                                                 map_bh.b_blocknr);
 535                 if (buffer_boundary(&map_bh)) {
 536                         boundary_block = map_bh.b_blocknr;
 537                         boundary_bdev = map_bh.b_bdev;
 538                 }
 539                 if (page_block) {
 540                         if (map_bh.b_blocknr != blocks[page_block-1] + 1)
 541                                 goto confused;
 542                 }
 543                 blocks[page_block++] = map_bh.b_blocknr;
 544                 boundary = buffer_boundary(&map_bh);
 545                 bdev = map_bh.b_bdev;
 546                 if (block_in_file == last_block)
 547                         break;
 548                 block_in_file++;
 549         }
 550         BUG_ON(page_block == 0);
 551
 552         first_unmapped = page_block;
 553
 554 page_is_mapped:
 555         end_index = i_size >> PAGE_CACHE_SHIFT;
 556         if (page->index >= end_index) {
 557                 /*
 558                  * The page straddles i_size.  It must be zeroed out on each
 559                  * and every writepage invokation because it may be mmapped.
 560                  * "A file is mapped in multiples of the page size.  For a file
 561                  * that is not a multiple of the page size, the remaining memory
 562                  * is zeroed when mapped, and writes to that region are not
 563                  * written out to the file."
 564                  */
 565                 unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
 566
 567                 if (page->index > end_index || !offset)
 568                         goto confused;
 569                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
 570         }
 571
 572         /*
 573          * This page will go to BIO.  Do we need to send this BIO off first?
 574          */
 575         if (bio && mpd->last_block_in_bio != blocks[0] - 1)
 576                 bio = mpage_bio_submit(WRITE, bio);
 577
 578 alloc_new:
 579         if (bio == NULL) {
 580                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 581                                 bio_get_nr_vecs(bdev), GFP_NOFS|__GFP_HIGH);
 582                 if (bio == NULL)
 583                         goto confused;
 584         }
 585
 586         /*
 587          * Must try to add the page before marking the buffer clean or
 588          * the confused fail path above (OOM) will be very confused when
 589          * it finds all bh marked clean (i.e. it will not write anything)
 590          */
 591         length = first_unmapped << blkbits;
 592         if (bio_add_page(bio, page, length, 0) < length) {
 593                 bio = mpage_bio_submit(WRITE, bio);
 594                 goto alloc_new;
 595         }
 596
 597         /*
 598          * OK, we have our BIO, so we can now mark the buffers clean.  Make
 599          * sure to only clean buffers which we know we'll be writing.
 600          */
 601         if (page_has_buffers(page)) {
 602                 struct buffer_head *head = page_buffers(page);
 603                 struct buffer_head *bh = head;
 604                 unsigned buffer_counter = 0;
 605
 606                 do {
 607                         if (buffer_counter++ == first_unmapped)
 608                                 break;
 609                         clear_buffer_dirty(bh);
 610                         bh = bh->b_this_page;
 611                 } while (bh != head);
 612
 613                 /*
 614                  * we cannot drop the bh if the page is not uptodate
 615                  * or a concurrent readpage would fail to serialize with the bh
 616                  * and it would read from disk before we reach the platter.
 617                  */
 618                 if (buffer_heads_over_limit && PageUptodate(page))
 619                         try_to_free_buffers(page);
 620         }
 621
 622         BUG_ON(PageWriteback(page));
 623         set_page_writeback(page);
 624         unlock_page(page);
 625         if (boundary || (first_unmapped != blocks_per_page)) {
 626                 bio = mpage_bio_submit(WRITE, bio);
 627                 if (boundary_block) {
 628                         write_boundary_block(boundary_bdev,
 629                                         boundary_block, 1 << blkbits);
 630                 }
 631         } else {
 632                 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
 633         }
 634         goto out;
 635
 636 confused:
 637         if (bio)
 638                 bio = mpage_bio_submit(WRITE, bio);
 639
 640         if (mpd->use_writepage) {
 641                 ret = mapping->a_ops->writepage(page, wbc);
 642         } else {
 643                 ret = -EAGAIN;
 644                 goto out;
 645         }
 646         /*
 647          * The caller has a ref on the inode, so *mapping is stable
 648          */
 649         mapping_set_error(mapping, ret);
 650 out:
 651         mpd->bio = bio;
 652         return ret;
 653 }
 654
 655 /**
 656  * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
 657  * @mapping: address space structure to write
 658  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
 659  * @get_block: the filesystem's block mapper function.
 660  *             If this is NULL then use a_ops->writepage.  Otherwise, go
 661  *             direct-to-BIO.
 662  *
 663  * This is a library function, which implements the writepages()
 664  * address_space_operation.
 665  *
 666  * If a page is already under I/O, generic_writepages() skips it, even
 667  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
 668  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
 669  * and msync() need to guarantee that all the data which was dirty at the time
 670  * the call was made get new I/O started against them.  If wbc->sync_mode is
 671  * WB_SYNC_ALL then we were called for data integrity and we must wait for
 672  * existing IO to complete.
 673  */
 674 int
 675 mpage_writepages(struct address_space *mapping,
 676                 struct writeback_control *wbc, get_block_t get_block)
 677 {
 678         int ret;
 679
 680         if (!get_block)
 681                 ret = generic_writepages(mapping, wbc);
 682         else {
 683                 struct mpage_data mpd = {
 684                         .bio = NULL,
 685                         .last_block_in_bio = 0,
 686                         .get_block = get_block,
 687                         .use_writepage = 1,
 688                 };
 689
 690                 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
 691                 if (mpd.bio)
 692                         mpage_bio_submit(WRITE, mpd.bio);
 693         }
 694         return ret;
 695 }
 696 EXPORT_SYMBOL(mpage_writepages);
 697
 698 int mpage_writepage(struct page *page, get_block_t get_block,
 699         struct writeback_control *wbc)
 700 {
 701         struct mpage_data mpd = {
 702                 .bio = NULL,
 703                 .last_block_in_bio = 0,
 704                 .get_block = get_block,
 705                 .use_writepage = 0,
 706         };
 707         int ret = __mpage_writepage(page, wbc, &mpd);
 708         if (mpd.bio)
 709                 mpage_bio_submit(WRITE, mpd.bio);
 710         return ret;
 711 }
 712 EXPORT_SYMBOL(mpage_writepage);