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