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    Andrew Morton
  10  *              Initial version
  11  * 27Jun2002    [email protected]
  12  *              use bio_add_page() to build bio's just the right size
  13  */
  14
  15 #include <linux/kernel.h>
  16 #include <linux/export.h>
  17 #include <linux/mm.h>
  18 #include <linux/kdev_t.h>
  19 #include <linux/gfp.h>
  20 #include <linux/bio.h>
  21 #include <linux/fs.h>
  22 #include <linux/buffer_head.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/highmem.h>
  25 #include <linux/prefetch.h>
  26 #include <linux/mpage.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/writeback.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/pagevec.h>
  31 #include <linux/cleancache.h>
  32 #include "internal.h"
  33
  34 /*
  35  * I/O completion handler for multipage BIOs.
  36  *
  37  * The mpage code never puts partial pages into a BIO (except for end-of-file).
  38  * If a page does not map to a contiguous run of blocks then it simply falls
  39  * back to block_read_full_page().
  40  *
  41  * Why is this?  If a page's completion depends on a number of different BIOs
  42  * which can complete in any order (or at the same time) then determining the
  43  * status of that page is hard.  See end_buffer_async_read() for the details.
  44  * There is no point in duplicating all that complexity.
  45  */
  46 static void mpage_end_io(struct bio *bio)
  47 {
  48         struct bio_vec *bv;
  49         int i;
  50
  51         bio_for_each_segment_all(bv, bio, i) {
  52                 struct page *page = bv->bv_page;
  53                 page_endio(page, bio_data_dir(bio), bio->bi_error);
  54         }
  55
  56         bio_put(bio);
  57 }
  58
  59 static struct bio *mpage_bio_submit(int rw, struct bio *bio)
  60 {
  61         bio->bi_end_io = mpage_end_io;
  62         guard_bio_eod(rw, bio);
  63         submit_bio(rw, bio);
  64         return NULL;
  65 }
  66
  67 static struct bio *
  68 mpage_alloc(struct block_device *bdev,
  69                 sector_t first_sector, int nr_vecs,
  70                 gfp_t gfp_flags)
  71 {
  72         struct bio *bio;
  73
  74         /* Restrict the given (page cache) mask for slab allocations */
  75         gfp_flags &= GFP_KERNEL;
  76         bio = bio_alloc(gfp_flags, nr_vecs);
  77
  78         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
  79                 while (!bio && (nr_vecs /= 2))
  80                         bio = bio_alloc(gfp_flags, nr_vecs);
  81         }
  82
  83         if (bio) {
  84                 bio->bi_bdev = bdev;
  85                 bio->bi_iter.bi_sector = first_sector;
  86         }
  87         return bio;
  88 }
  89
  90 /*
  91  * support function for mpage_readpages.  The fs supplied get_block might
  92  * return an up to date buffer.  This is used to map that buffer into
  93  * the page, which allows readpage to avoid triggering a duplicate call
  94  * to get_block.
  95  *
  96  * The idea is to avoid adding buffers to pages that don't already have
  97  * them.  So when the buffer is up to date and the page size == block size,
  98  * this marks the page up to date instead of adding new buffers.
  99  */
 100 static void
 101 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
 102 {
 103         struct inode *inode = page->mapping->host;
 104         struct buffer_head *page_bh, *head;
 105         int block = 0;
 106
 107         if (!page_has_buffers(page)) {
 108                 /*
 109                  * don't make any buffers if there is only one buffer on
 110                  * the page and the page just needs to be set up to date
 111                  */
 112                 if (inode->i_blkbits == PAGE_SHIFT &&
 113                     buffer_uptodate(bh)) {
 114                         SetPageUptodate(page);
 115                         return;
 116                 }
 117                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
 118         }
 119         head = page_buffers(page);
 120         page_bh = head;
 121         do {
 122                 if (block == page_block) {
 123                         page_bh->b_state = bh->b_state;
 124                         page_bh->b_bdev = bh->b_bdev;
 125                         page_bh->b_blocknr = bh->b_blocknr;
 126                         break;
 127                 }
 128                 page_bh = page_bh->b_this_page;
 129                 block++;
 130         } while (page_bh != head);
 131 }
 132
 133 /*
 134  * This is the worker routine which does all the work of mapping the disk
 135  * blocks and constructs largest possible bios, submits them for IO if the
 136  * blocks are not contiguous on the disk.
 137  *
 138  * We pass a buffer_head back and forth and use its buffer_mapped() flag to
 139  * represent the validity of its disk mapping and to decide when to do the next
 140  * get_block() call.
 141  */
 142 static struct bio *
 143 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
 144                 sector_t *last_block_in_bio, struct buffer_head *map_bh,
 145                 unsigned long *first_logical_block, get_block_t get_block,
 146                 gfp_t gfp)
 147 {
 148         struct inode *inode = page->mapping->host;
 149         const unsigned blkbits = inode->i_blkbits;
 150         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 151         const unsigned blocksize = 1 << blkbits;
 152         sector_t block_in_file;
 153         sector_t last_block;
 154         sector_t last_block_in_file;
 155         sector_t blocks[MAX_BUF_PER_PAGE];
 156         unsigned page_block;
 157         unsigned first_hole = blocks_per_page;
 158         struct block_device *bdev = NULL;
 159         int length;
 160         int fully_mapped = 1;
 161         unsigned nblocks;
 162         unsigned relative_block;
 163
 164         if (page_has_buffers(page))
 165                 goto confused;
 166
 167         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
 168         last_block = block_in_file + nr_pages * blocks_per_page;
 169         last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
 170         if (last_block > last_block_in_file)
 171                 last_block = last_block_in_file;
 172         page_block = 0;
 173
 174         /*
 175          * Map blocks using the result from the previous get_blocks call first.
 176          */
 177         nblocks = map_bh->b_size >> blkbits;
 178         if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
 179                         block_in_file < (*first_logical_block + nblocks)) {
 180                 unsigned map_offset = block_in_file - *first_logical_block;
 181                 unsigned last = nblocks - map_offset;
 182
 183                 for (relative_block = 0; ; relative_block++) {
 184                         if (relative_block == last) {
 185                                 clear_buffer_mapped(map_bh);
 186                                 break;
 187                         }
 188                         if (page_block == blocks_per_page)
 189                                 break;
 190                         blocks[page_block] = map_bh->b_blocknr + map_offset +
 191                                                 relative_block;
 192                         page_block++;
 193                         block_in_file++;
 194                 }
 195                 bdev = map_bh->b_bdev;
 196         }
 197
 198         /*
 199          * Then do more get_blocks calls until we are done with this page.
 200          */
 201         map_bh->b_page = page;
 202         while (page_block < blocks_per_page) {
 203                 map_bh->b_state = 0;
 204                 map_bh->b_size = 0;
 205
 206                 if (block_in_file < last_block) {
 207                         map_bh->b_size = (last_block-block_in_file) << blkbits;
 208                         if (get_block(inode, block_in_file, map_bh, 0))
 209                                 goto confused;
 210                         *first_logical_block = block_in_file;
 211                 }
 212
 213                 if (!buffer_mapped(map_bh)) {
 214                         fully_mapped = 0;
 215                         if (first_hole == blocks_per_page)
 216                                 first_hole = page_block;
 217                         page_block++;
 218                         block_in_file++;
 219                         continue;
 220                 }
 221
 222                 /* some filesystems will copy data into the page during
 223                  * the get_block call, in which case we don't want to
 224                  * read it again.  map_buffer_to_page copies the data
 225                  * we just collected from get_block into the page's buffers
 226                  * so readpage doesn't have to repeat the get_block call
 227                  */
 228                 if (buffer_uptodate(map_bh)) {
 229                         map_buffer_to_page(page, map_bh, page_block);
 230                         goto confused;
 231                 }
 232
 233                 if (first_hole != blocks_per_page)
 234                         goto confused;          /* hole -> non-hole */
 235
 236                 /* Contiguous blocks? */
 237                 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
 238                         goto confused;
 239                 nblocks = map_bh->b_size >> blkbits;
 240                 for (relative_block = 0; ; relative_block++) {
 241                         if (relative_block == nblocks) {
 242                                 clear_buffer_mapped(map_bh);
 243                                 break;
 244                         } else if (page_block == blocks_per_page)
 245                                 break;
 246                         blocks[page_block] = map_bh->b_blocknr+relative_block;
 247                         page_block++;
 248                         block_in_file++;
 249                 }
 250                 bdev = map_bh->b_bdev;
 251         }
 252
 253         if (first_hole != blocks_per_page) {
 254                 zero_user_segment(page, first_hole << blkbits, PAGE_SIZE);
 255                 if (first_hole == 0) {
 256                         SetPageUptodate(page);
 257                         unlock_page(page);
 258                         goto out;
 259                 }
 260         } else if (fully_mapped) {
 261                 SetPageMappedToDisk(page);
 262         }
 263
 264         if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
 265             cleancache_get_page(page) == 0) {
 266                 SetPageUptodate(page);
 267                 goto confused;
 268         }
 269
 270         /*
 271          * This page will go to BIO.  Do we need to send this BIO off first?
 272          */
 273         if (bio && (*last_block_in_bio != blocks[0] - 1))
 274                 bio = mpage_bio_submit(READ, bio);
 275
 276 alloc_new:
 277         if (bio == NULL) {
 278                 if (first_hole == blocks_per_page) {
 279                         if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
 280                                                                 page))
 281                                 goto out;
 282                 }
 283                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 284                                 min_t(int, nr_pages, BIO_MAX_PAGES), gfp);
 285                 if (bio == NULL)
 286                         goto confused;
 287         }
 288
 289         length = first_hole << blkbits;
 290         if (bio_add_page(bio, page, length, 0) < length) {
 291                 bio = mpage_bio_submit(READ, bio);
 292                 goto alloc_new;
 293         }
 294
 295         relative_block = block_in_file - *first_logical_block;
 296         nblocks = map_bh->b_size >> blkbits;
 297         if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
 298             (first_hole != blocks_per_page))
 299                 bio = mpage_bio_submit(READ, bio);
 300         else
 301                 *last_block_in_bio = blocks[blocks_per_page - 1];
 302 out:
 303         return bio;
 304
 305 confused:
 306         if (bio)
 307                 bio = mpage_bio_submit(READ, bio);
 308         if (!PageUptodate(page))
 309                 block_read_full_page(page, get_block);
 310         else
 311                 unlock_page(page);
 312         goto out;
 313 }
 314
 315 /**
 316  * mpage_readpages - populate an address space with some pages & start reads against them
 317  * @mapping: the address_space
 318  * @pages: The address of a list_head which contains the target pages.  These
 319  *   pages have their ->index populated and are otherwise uninitialised.
 320  *   The page at @pages->prev has the lowest file offset, and reads should be
 321  *   issued in @pages->prev to @pages->next order.
 322  * @nr_pages: The number of pages at *@pages
 323  * @get_block: The filesystem's block mapper function.
 324  *
 325  * This function walks the pages and the blocks within each page, building and
 326  * emitting large BIOs.
 327  *
 328  * If anything unusual happens, such as:
 329  *
 330  * - encountering a page which has buffers
 331  * - encountering a page which has a non-hole after a hole
 332  * - encountering a page with non-contiguous blocks
 333  *
 334  * then this code just gives up and calls the buffer_head-based read function.
 335  * It does handle a page which has holes at the end - that is a common case:
 336  * the end-of-file on blocksize < PAGE_SIZE setups.
 337  *
 338  * BH_Boundary explanation:
 339  *
 340  * There is a problem.  The mpage read code assembles several pages, gets all
 341  * their disk mappings, and then submits them all.  That's fine, but obtaining
 342  * the disk mappings may require I/O.  Reads of indirect blocks, for example.
 343  *
 344  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
 345  * submitted in the following order:
 346  *      12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
 347  *
 348  * because the indirect block has to be read to get the mappings of blocks
 349  * 13,14,15,16.  Obviously, this impacts performance.
 350  *
 351  * So what we do it to allow the filesystem's get_block() function to set
 352  * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
 353  * after this one will require I/O against a block which is probably close to
 354  * this one.  So you should push what I/O you have currently accumulated.
 355  *
 356  * This all causes the disk requests to be issued in the correct order.
 357  */
 358 int
 359 mpage_readpages(struct address_space *mapping, struct list_head *pages,
 360                                 unsigned nr_pages, get_block_t get_block)
 361 {
 362         struct bio *bio = NULL;
 363         unsigned page_idx;
 364         sector_t last_block_in_bio = 0;
 365         struct buffer_head map_bh;
 366         unsigned long first_logical_block = 0;
 367         gfp_t gfp = readahead_gfp_mask(mapping);
 368
 369         map_bh.b_state = 0;
 370         map_bh.b_size = 0;
 371         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 372                 struct page *page = lru_to_page(pages);
 373
 374                 prefetchw(&page->flags);
 375                 list_del(&page->lru);
 376                 if (!add_to_page_cache_lru(page, mapping,
 377                                         page->index,
 378                                         gfp)) {
 379                         bio = do_mpage_readpage(bio, page,
 380                                         nr_pages - page_idx,
 381                                         &last_block_in_bio, &map_bh,
 382                                         &first_logical_block,
 383                                         get_block, gfp);
 384                 }
 385                 put_page(page);
 386         }
 387         BUG_ON(!list_empty(pages));
 388         if (bio)
 389                 mpage_bio_submit(READ, bio);
 390         return 0;
 391 }
 392 EXPORT_SYMBOL(mpage_readpages);
 393
 394 /*
 395  * This isn't called much at all
 396  */
 397 int mpage_readpage(struct page *page, get_block_t get_block)
 398 {
 399         struct bio *bio = NULL;
 400         sector_t last_block_in_bio = 0;
 401         struct buffer_head map_bh;
 402         unsigned long first_logical_block = 0;
 403         gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 404
 405         map_bh.b_state = 0;
 406         map_bh.b_size = 0;
 407         bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
 408                         &map_bh, &first_logical_block, get_block, gfp);
 409         if (bio)
 410                 mpage_bio_submit(READ, bio);
 411         return 0;
 412 }
 413 EXPORT_SYMBOL(mpage_readpage);
 414
 415 /*
 416  * Writing is not so simple.
 417  *
 418  * If the page has buffers then they will be used for obtaining the disk
 419  * mapping.  We only support pages which are fully mapped-and-dirty, with a
 420  * special case for pages which are unmapped at the end: end-of-file.
 421  *
 422  * If the page has no buffers (preferred) then the page is mapped here.
 423  *
 424  * If all blocks are found to be contiguous then the page can go into the
 425  * BIO.  Otherwise fall back to the mapping's writepage().
 426  *
 427  * FIXME: This code wants an estimate of how many pages are still to be
 428  * written, so it can intelligently allocate a suitably-sized BIO.  For now,
 429  * just allocate full-size (16-page) BIOs.
 430  */
 431
 432 struct mpage_data {
 433         struct bio *bio;
 434         sector_t last_block_in_bio;
 435         get_block_t *get_block;
 436         unsigned use_writepage;
 437 };
 438
 439 /*
 440  * We have our BIO, so we can now mark the buffers clean.  Make
 441  * sure to only clean buffers which we know we'll be writing.
 442  */
 443 static void clean_buffers(struct page *page, unsigned first_unmapped)
 444 {
 445         unsigned buffer_counter = 0;
 446         struct buffer_head *bh, *head;
 447         if (!page_has_buffers(page))
 448                 return;
 449         head = page_buffers(page);
 450         bh = head;
 451
 452         do {
 453                 if (buffer_counter++ == first_unmapped)
 454                         break;
 455                 clear_buffer_dirty(bh);
 456                 bh = bh->b_this_page;
 457         } while (bh != head);
 458
 459         /*
 460          * we cannot drop the bh if the page is not uptodate or a concurrent
 461          * readpage would fail to serialize with the bh and it would read from
 462          * disk before we reach the platter.
 463          */
 464         if (buffer_heads_over_limit && PageUptodate(page))
 465                 try_to_free_buffers(page);
 466 }
 467
 468 static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
 469                       void *data)
 470 {
 471         struct mpage_data *mpd = data;
 472         struct bio *bio = mpd->bio;
 473         struct address_space *mapping = page->mapping;
 474         struct inode *inode = page->mapping->host;
 475         const unsigned blkbits = inode->i_blkbits;
 476         unsigned long end_index;
 477         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 478         sector_t last_block;
 479         sector_t block_in_file;
 480         sector_t blocks[MAX_BUF_PER_PAGE];
 481         unsigned page_block;
 482         unsigned first_unmapped = blocks_per_page;
 483         struct block_device *bdev = NULL;
 484         int boundary = 0;
 485         sector_t boundary_block = 0;
 486         struct block_device *boundary_bdev = NULL;
 487         int length;
 488         struct buffer_head map_bh;
 489         loff_t i_size = i_size_read(inode);
 490         int ret = 0;
 491         int wr = (wbc->sync_mode == WB_SYNC_ALL ?  WRITE_SYNC : WRITE);
 492
 493         if (page_has_buffers(page)) {
 494                 struct buffer_head *head = page_buffers(page);
 495                 struct buffer_head *bh = head;
 496
 497                 /* If they're all mapped and dirty, do it */
 498                 page_block = 0;
 499                 do {
 500                         BUG_ON(buffer_locked(bh));
 501                         if (!buffer_mapped(bh)) {
 502                                 /*
 503                                  * unmapped dirty buffers are created by
 504                                  * __set_page_dirty_buffers -> mmapped data
 505                                  */
 506                                 if (buffer_dirty(bh))
 507                                         goto confused;
 508                                 if (first_unmapped == blocks_per_page)
 509                                         first_unmapped = page_block;
 510                                 continue;
 511                         }
 512
 513                         if (first_unmapped != blocks_per_page)
 514                                 goto confused;  /* hole -> non-hole */
 515
 516                         if (!buffer_dirty(bh) || !buffer_uptodate(bh))
 517                                 goto confused;
 518                         if (page_block) {
 519                                 if (bh->b_blocknr != blocks[page_block-1] + 1)
 520                                         goto confused;
 521                         }
 522                         blocks[page_block++] = bh->b_blocknr;
 523                         boundary = buffer_boundary(bh);
 524                         if (boundary) {
 525                                 boundary_block = bh->b_blocknr;
 526                                 boundary_bdev = bh->b_bdev;
 527                         }
 528                         bdev = bh->b_bdev;
 529                 } while ((bh = bh->b_this_page) != head);
 530
 531                 if (first_unmapped)
 532                         goto page_is_mapped;
 533
 534                 /*
 535                  * Page has buffers, but they are all unmapped. The page was
 536                  * created by pagein or read over a hole which was handled by
 537                  * block_read_full_page().  If this address_space is also
 538                  * using mpage_readpages then this can rarely happen.
 539                  */
 540                 goto confused;
 541         }
 542
 543         /*
 544          * The page has no buffers: map it to disk
 545          */
 546         BUG_ON(!PageUptodate(page));
 547         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
 548         last_block = (i_size - 1) >> blkbits;
 549         map_bh.b_page = page;
 550         for (page_block = 0; page_block < blocks_per_page; ) {
 551
 552                 map_bh.b_state = 0;
 553                 map_bh.b_size = 1 << blkbits;
 554                 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
 555                         goto confused;
 556                 if (buffer_new(&map_bh))
 557                         unmap_underlying_metadata(map_bh.b_bdev,
 558                                                 map_bh.b_blocknr);
 559                 if (buffer_boundary(&map_bh)) {
 560                         boundary_block = map_bh.b_blocknr;
 561                         boundary_bdev = map_bh.b_bdev;
 562                 }
 563                 if (page_block) {
 564                         if (map_bh.b_blocknr != blocks[page_block-1] + 1)
 565                                 goto confused;
 566                 }
 567                 blocks[page_block++] = map_bh.b_blocknr;
 568                 boundary = buffer_boundary(&map_bh);
 569                 bdev = map_bh.b_bdev;
 570                 if (block_in_file == last_block)
 571                         break;
 572                 block_in_file++;
 573         }
 574         BUG_ON(page_block == 0);
 575
 576         first_unmapped = page_block;
 577
 578 page_is_mapped:
 579         end_index = i_size >> PAGE_SHIFT;
 580         if (page->index >= end_index) {
 581                 /*
 582                  * The page straddles i_size.  It must be zeroed out on each
 583                  * and every writepage invocation because it may be mmapped.
 584                  * "A file is mapped in multiples of the page size.  For a file
 585                  * that is not a multiple of the page size, the remaining memory
 586                  * is zeroed when mapped, and writes to that region are not
 587                  * written out to the file."
 588                  */
 589                 unsigned offset = i_size & (PAGE_SIZE - 1);
 590
 591                 if (page->index > end_index || !offset)
 592                         goto confused;
 593                 zero_user_segment(page, offset, PAGE_SIZE);
 594         }
 595
 596         /*
 597          * This page will go to BIO.  Do we need to send this BIO off first?
 598          */
 599         if (bio && mpd->last_block_in_bio != blocks[0] - 1)
 600                 bio = mpage_bio_submit(wr, bio);
 601
 602 alloc_new:
 603         if (bio == NULL) {
 604                 if (first_unmapped == blocks_per_page) {
 605                         if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
 606                                                                 page, wbc)) {
 607                                 clean_buffers(page, first_unmapped);
 608                                 goto out;
 609                         }
 610                 }
 611                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 612                                 BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
 613                 if (bio == NULL)
 614                         goto confused;
 615
 616                 wbc_init_bio(wbc, bio);
 617         }
 618
 619         /*
 620          * Must try to add the page before marking the buffer clean or
 621          * the confused fail path above (OOM) will be very confused when
 622          * it finds all bh marked clean (i.e. it will not write anything)
 623          */
 624         wbc_account_io(wbc, page, PAGE_SIZE);
 625         length = first_unmapped << blkbits;
 626         if (bio_add_page(bio, page, length, 0) < length) {
 627                 bio = mpage_bio_submit(wr, bio);
 628                 goto alloc_new;
 629         }
 630
 631         clean_buffers(page, first_unmapped);
 632
 633         BUG_ON(PageWriteback(page));
 634         set_page_writeback(page);
 635         unlock_page(page);
 636         if (boundary || (first_unmapped != blocks_per_page)) {
 637                 bio = mpage_bio_submit(wr, bio);
 638                 if (boundary_block) {
 639                         write_boundary_block(boundary_bdev,
 640                                         boundary_block, 1 << blkbits);
 641                 }
 642         } else {
 643                 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
 644         }
 645         goto out;
 646
 647 confused:
 648         if (bio)
 649                 bio = mpage_bio_submit(wr, bio);
 650
 651         if (mpd->use_writepage) {
 652                 ret = mapping->a_ops->writepage(page, wbc);
 653         } else {
 654                 ret = -EAGAIN;
 655                 goto out;
 656         }
 657         /*
 658          * The caller has a ref on the inode, so *mapping is stable
 659          */
 660         mapping_set_error(mapping, ret);
 661 out:
 662         mpd->bio = bio;
 663         return ret;
 664 }
 665
 666 /**
 667  * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
 668  * @mapping: address space structure to write
 669  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
 670  * @get_block: the filesystem's block mapper function.
 671  *             If this is NULL then use a_ops->writepage.  Otherwise, go
 672  *             direct-to-BIO.
 673  *
 674  * This is a library function, which implements the writepages()
 675  * address_space_operation.
 676  *
 677  * If a page is already under I/O, generic_writepages() skips it, even
 678  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
 679  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
 680  * and msync() need to guarantee that all the data which was dirty at the time
 681  * the call was made get new I/O started against them.  If wbc->sync_mode is
 682  * WB_SYNC_ALL then we were called for data integrity and we must wait for
 683  * existing IO to complete.
 684  */
 685 int
 686 mpage_writepages(struct address_space *mapping,
 687                 struct writeback_control *wbc, get_block_t get_block)
 688 {
 689         struct blk_plug plug;
 690         int ret;
 691
 692         blk_start_plug(&plug);
 693
 694         if (!get_block)
 695                 ret = generic_writepages(mapping, wbc);
 696         else {
 697                 struct mpage_data mpd = {
 698                         .bio = NULL,
 699                         .last_block_in_bio = 0,
 700                         .get_block = get_block,
 701                         .use_writepage = 1,
 702                 };
 703
 704                 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
 705                 if (mpd.bio) {
 706                         int wr = (wbc->sync_mode == WB_SYNC_ALL ?
 707                                   WRITE_SYNC : WRITE);
 708                         mpage_bio_submit(wr, mpd.bio);
 709                 }
 710         }
 711         blk_finish_plug(&plug);
 712         return ret;
 713 }
 714 EXPORT_SYMBOL(mpage_writepages);
 715
 716 int mpage_writepage(struct page *page, get_block_t get_block,
 717         struct writeback_control *wbc)
 718 {
 719         struct mpage_data mpd = {
 720                 .bio = NULL,
 721                 .last_block_in_bio = 0,
 722                 .get_block = get_block,
 723                 .use_writepage = 0,
 724         };
 725         int ret = __mpage_writepage(page, wbc, &mpd);
 726         if (mpd.bio) {
 727                 int wr = (wbc->sync_mode == WB_SYNC_ALL ?
 728                           WRITE_SYNC : WRITE);
 729                 mpage_bio_submit(wr, mpd.bio);
 730         }
 731         return ret;
 732 }
 733 EXPORT_SYMBOL(mpage_writepage);