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, op_is_write(bio_op(bio)),
  54                                 blk_status_to_errno(bio->bi_status));
  55         }
  56
  57         bio_put(bio);
  58 }
  59
  60 static struct bio *mpage_bio_submit(int op, int op_flags, struct bio *bio)
  61 {
  62         bio->bi_end_io = mpage_end_io;
  63         bio_set_op_attrs(bio, op, op_flags);
  64         guard_bio_eod(op, bio);
  65         submit_bio(bio);
  66         return NULL;
  67 }
  68
  69 static struct bio *
  70 mpage_alloc(struct block_device *bdev,
  71                 sector_t first_sector, int nr_vecs,
  72                 gfp_t gfp_flags)
  73 {
  74         struct bio *bio;
  75
  76         /* Restrict the given (page cache) mask for slab allocations */
  77         gfp_flags &= GFP_KERNEL;
  78         bio = bio_alloc(gfp_flags, nr_vecs);
  79
  80         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
  81                 while (!bio && (nr_vecs /= 2))
  82                         bio = bio_alloc(gfp_flags, nr_vecs);
  83         }
  84
  85         if (bio) {
  86                 bio->bi_bdev = bdev;
  87                 bio->bi_iter.bi_sector = first_sector;
  88         }
  89         return bio;
  90 }
  91
  92 /*
  93  * support function for mpage_readpages.  The fs supplied get_block might
  94  * return an up to date buffer.  This is used to map that buffer into
  95  * the page, which allows readpage to avoid triggering a duplicate call
  96  * to get_block.
  97  *
  98  * The idea is to avoid adding buffers to pages that don't already have
  99  * them.  So when the buffer is up to date and the page size == block size,
 100  * this marks the page up to date instead of adding new buffers.
 101  */
 102 static void
 103 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
 104 {
 105         struct inode *inode = page->mapping->host;
 106         struct buffer_head *page_bh, *head;
 107         int block = 0;
 108
 109         if (!page_has_buffers(page)) {
 110                 /*
 111                  * don't make any buffers if there is only one buffer on
 112                  * the page and the page just needs to be set up to date
 113                  */
 114                 if (inode->i_blkbits == PAGE_SHIFT &&
 115                     buffer_uptodate(bh)) {
 116                         SetPageUptodate(page);
 117                         return;
 118                 }
 119                 create_empty_buffers(page, i_blocksize(inode), 0);
 120         }
 121         head = page_buffers(page);
 122         page_bh = head;
 123         do {
 124                 if (block == page_block) {
 125                         page_bh->b_state = bh->b_state;
 126                         page_bh->b_bdev = bh->b_bdev;
 127                         page_bh->b_blocknr = bh->b_blocknr;
 128                         break;
 129                 }
 130                 page_bh = page_bh->b_this_page;
 131                 block++;
 132         } while (page_bh != head);
 133 }
 134
 135 /*
 136  * This is the worker routine which does all the work of mapping the disk
 137  * blocks and constructs largest possible bios, submits them for IO if the
 138  * blocks are not contiguous on the disk.
 139  *
 140  * We pass a buffer_head back and forth and use its buffer_mapped() flag to
 141  * represent the validity of its disk mapping and to decide when to do the next
 142  * get_block() call.
 143  */
 144 static struct bio *
 145 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
 146                 sector_t *last_block_in_bio, struct buffer_head *map_bh,
 147                 unsigned long *first_logical_block, get_block_t get_block,
 148                 gfp_t gfp)
 149 {
 150         struct inode *inode = page->mapping->host;
 151         const unsigned blkbits = inode->i_blkbits;
 152         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 153         const unsigned blocksize = 1 << blkbits;
 154         sector_t block_in_file;
 155         sector_t last_block;
 156         sector_t last_block_in_file;
 157         sector_t blocks[MAX_BUF_PER_PAGE];
 158         unsigned page_block;
 159         unsigned first_hole = blocks_per_page;
 160         struct block_device *bdev = NULL;
 161         int length;
 162         int fully_mapped = 1;
 163         unsigned nblocks;
 164         unsigned relative_block;
 165
 166         if (page_has_buffers(page))
 167                 goto confused;
 168
 169         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
 170         last_block = block_in_file + nr_pages * blocks_per_page;
 171         last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
 172         if (last_block > last_block_in_file)
 173                 last_block = last_block_in_file;
 174         page_block = 0;
 175
 176         /*
 177          * Map blocks using the result from the previous get_blocks call first.
 178          */
 179         nblocks = map_bh->b_size >> blkbits;
 180         if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
 181                         block_in_file < (*first_logical_block + nblocks)) {
 182                 unsigned map_offset = block_in_file - *first_logical_block;
 183                 unsigned last = nblocks - map_offset;
 184
 185                 for (relative_block = 0; ; relative_block++) {
 186                         if (relative_block == last) {
 187                                 clear_buffer_mapped(map_bh);
 188                                 break;
 189                         }
 190                         if (page_block == blocks_per_page)
 191                                 break;
 192                         blocks[page_block] = map_bh->b_blocknr + map_offset +
 193                                                 relative_block;
 194                         page_block++;
 195                         block_in_file++;
 196                 }
 197                 bdev = map_bh->b_bdev;
 198         }
 199
 200         /*
 201          * Then do more get_blocks calls until we are done with this page.
 202          */
 203         map_bh->b_page = page;
 204         while (page_block < blocks_per_page) {
 205                 map_bh->b_state = 0;
 206                 map_bh->b_size = 0;
 207
 208                 if (block_in_file < last_block) {
 209                         map_bh->b_size = (last_block-block_in_file) << blkbits;
 210                         if (get_block(inode, block_in_file, map_bh, 0))
 211                                 goto confused;
 212                         *first_logical_block = block_in_file;
 213                 }
 214
 215                 if (!buffer_mapped(map_bh)) {
 216                         fully_mapped = 0;
 217                         if (first_hole == blocks_per_page)
 218                                 first_hole = page_block;
 219                         page_block++;
 220                         block_in_file++;
 221                         continue;
 222                 }
 223
 224                 /* some filesystems will copy data into the page during
 225                  * the get_block call, in which case we don't want to
 226                  * read it again.  map_buffer_to_page copies the data
 227                  * we just collected from get_block into the page's buffers
 228                  * so readpage doesn't have to repeat the get_block call
 229                  */
 230                 if (buffer_uptodate(map_bh)) {
 231                         map_buffer_to_page(page, map_bh, page_block);
 232                         goto confused;
 233                 }
 234
 235                 if (first_hole != blocks_per_page)
 236                         goto confused;          /* hole -> non-hole */
 237
 238                 /* Contiguous blocks? */
 239                 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
 240                         goto confused;
 241                 nblocks = map_bh->b_size >> blkbits;
 242                 for (relative_block = 0; ; relative_block++) {
 243                         if (relative_block == nblocks) {
 244                                 clear_buffer_mapped(map_bh);
 245                                 break;
 246                         } else if (page_block == blocks_per_page)
 247                                 break;
 248                         blocks[page_block] = map_bh->b_blocknr+relative_block;
 249                         page_block++;
 250                         block_in_file++;
 251                 }
 252                 bdev = map_bh->b_bdev;
 253         }
 254
 255         if (first_hole != blocks_per_page) {
 256                 zero_user_segment(page, first_hole << blkbits, PAGE_SIZE);
 257                 if (first_hole == 0) {
 258                         SetPageUptodate(page);
 259                         unlock_page(page);
 260                         goto out;
 261                 }
 262         } else if (fully_mapped) {
 263                 SetPageMappedToDisk(page);
 264         }
 265
 266         if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
 267             cleancache_get_page(page) == 0) {
 268                 SetPageUptodate(page);
 269                 goto confused;
 270         }
 271
 272         /*
 273          * This page will go to BIO.  Do we need to send this BIO off first?
 274          */
 275         if (bio && (*last_block_in_bio != blocks[0] - 1))
 276                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
 277
 278 alloc_new:
 279         if (bio == NULL) {
 280                 if (first_hole == blocks_per_page) {
 281                         if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
 282                                                                 page))
 283                                 goto out;
 284                 }
 285                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 286                                 min_t(int, nr_pages, BIO_MAX_PAGES), gfp);
 287                 if (bio == NULL)
 288                         goto confused;
 289         }
 290
 291         length = first_hole << blkbits;
 292         if (bio_add_page(bio, page, length, 0) < length) {
 293                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
 294                 goto alloc_new;
 295         }
 296
 297         relative_block = block_in_file - *first_logical_block;
 298         nblocks = map_bh->b_size >> blkbits;
 299         if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
 300             (first_hole != blocks_per_page))
 301                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
 302         else
 303                 *last_block_in_bio = blocks[blocks_per_page - 1];
 304 out:
 305         return bio;
 306
 307 confused:
 308         if (bio)
 309                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
 310         if (!PageUptodate(page))
 311                 block_read_full_page(page, get_block);
 312         else
 313                 unlock_page(page);
 314         goto out;
 315 }
 316
 317 /**
 318  * mpage_readpages - populate an address space with some pages & start reads against them
 319  * @mapping: the address_space
 320  * @pages: The address of a list_head which contains the target pages.  These
 321  *   pages have their ->index populated and are otherwise uninitialised.
 322  *   The page at @pages->prev has the lowest file offset, and reads should be
 323  *   issued in @pages->prev to @pages->next order.
 324  * @nr_pages: The number of pages at *@pages
 325  * @get_block: The filesystem's block mapper function.
 326  *
 327  * This function walks the pages and the blocks within each page, building and
 328  * emitting large BIOs.
 329  *
 330  * If anything unusual happens, such as:
 331  *
 332  * - encountering a page which has buffers
 333  * - encountering a page which has a non-hole after a hole
 334  * - encountering a page with non-contiguous blocks
 335  *
 336  * then this code just gives up and calls the buffer_head-based read function.
 337  * It does handle a page which has holes at the end - that is a common case:
 338  * the end-of-file on blocksize < PAGE_SIZE setups.
 339  *
 340  * BH_Boundary explanation:
 341  *
 342  * There is a problem.  The mpage read code assembles several pages, gets all
 343  * their disk mappings, and then submits them all.  That's fine, but obtaining
 344  * the disk mappings may require I/O.  Reads of indirect blocks, for example.
 345  *
 346  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
 347  * submitted in the following order:
 348  *      12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
 349  *
 350  * because the indirect block has to be read to get the mappings of blocks
 351  * 13,14,15,16.  Obviously, this impacts performance.
 352  *
 353  * So what we do it to allow the filesystem's get_block() function to set
 354  * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
 355  * after this one will require I/O against a block which is probably close to
 356  * this one.  So you should push what I/O you have currently accumulated.
 357  *
 358  * This all causes the disk requests to be issued in the correct order.
 359  */
 360 int
 361 mpage_readpages(struct address_space *mapping, struct list_head *pages,
 362                                 unsigned nr_pages, get_block_t get_block)
 363 {
 364         struct bio *bio = NULL;
 365         unsigned page_idx;
 366         sector_t last_block_in_bio = 0;
 367         struct buffer_head map_bh;
 368         unsigned long first_logical_block = 0;
 369         gfp_t gfp = readahead_gfp_mask(mapping);
 370
 371         map_bh.b_state = 0;
 372         map_bh.b_size = 0;
 373         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 374                 struct page *page = lru_to_page(pages);
 375
 376                 prefetchw(&page->flags);
 377                 list_del(&page->lru);
 378                 if (!add_to_page_cache_lru(page, mapping,
 379                                         page->index,
 380                                         gfp)) {
 381                         bio = do_mpage_readpage(bio, page,
 382                                         nr_pages - page_idx,
 383                                         &last_block_in_bio, &map_bh,
 384                                         &first_logical_block,
 385                                         get_block, gfp);
 386                 }
 387                 put_page(page);
 388         }
 389         BUG_ON(!list_empty(pages));
 390         if (bio)
 391                 mpage_bio_submit(REQ_OP_READ, 0, bio);
 392         return 0;
 393 }
 394 EXPORT_SYMBOL(mpage_readpages);
 395
 396 /*
 397  * This isn't called much at all
 398  */
 399 int mpage_readpage(struct page *page, get_block_t get_block)
 400 {
 401         struct bio *bio = NULL;
 402         sector_t last_block_in_bio = 0;
 403         struct buffer_head map_bh;
 404         unsigned long first_logical_block = 0;
 405         gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 406
 407         map_bh.b_state = 0;
 408         map_bh.b_size = 0;
 409         bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
 410                         &map_bh, &first_logical_block, get_block, gfp);
 411         if (bio)
 412                 mpage_bio_submit(REQ_OP_READ, 0, bio);
 413         return 0;
 414 }
 415 EXPORT_SYMBOL(mpage_readpage);
 416
 417 /*
 418  * Writing is not so simple.
 419  *
 420  * If the page has buffers then they will be used for obtaining the disk
 421  * mapping.  We only support pages which are fully mapped-and-dirty, with a
 422  * special case for pages which are unmapped at the end: end-of-file.
 423  *
 424  * If the page has no buffers (preferred) then the page is mapped here.
 425  *
 426  * If all blocks are found to be contiguous then the page can go into the
 427  * BIO.  Otherwise fall back to the mapping's writepage().
 428  *
 429  * FIXME: This code wants an estimate of how many pages are still to be
 430  * written, so it can intelligently allocate a suitably-sized BIO.  For now,
 431  * just allocate full-size (16-page) BIOs.
 432  */
 433
 434 struct mpage_data {
 435         struct bio *bio;
 436         sector_t last_block_in_bio;
 437         get_block_t *get_block;
 438         unsigned use_writepage;
 439 };
 440
 441 /*
 442  * We have our BIO, so we can now mark the buffers clean.  Make
 443  * sure to only clean buffers which we know we'll be writing.
 444  */
 445 static void clean_buffers(struct page *page, unsigned first_unmapped)
 446 {
 447         unsigned buffer_counter = 0;
 448         struct buffer_head *bh, *head;
 449         if (!page_has_buffers(page))
 450                 return;
 451         head = page_buffers(page);
 452         bh = head;
 453
 454         do {
 455                 if (buffer_counter++ == first_unmapped)
 456                         break;
 457                 clear_buffer_dirty(bh);
 458                 bh = bh->b_this_page;
 459         } while (bh != head);
 460
 461         /*
 462          * we cannot drop the bh if the page is not uptodate or a concurrent
 463          * readpage would fail to serialize with the bh and it would read from
 464          * disk before we reach the platter.
 465          */
 466         if (buffer_heads_over_limit && PageUptodate(page))
 467                 try_to_free_buffers(page);
 468 }
 469
 470 static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
 471                       void *data)
 472 {
 473         struct mpage_data *mpd = data;
 474         struct bio *bio = mpd->bio;
 475         struct address_space *mapping = page->mapping;
 476         struct inode *inode = page->mapping->host;
 477         const unsigned blkbits = inode->i_blkbits;
 478         unsigned long end_index;
 479         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 480         sector_t last_block;
 481         sector_t block_in_file;
 482         sector_t blocks[MAX_BUF_PER_PAGE];
 483         unsigned page_block;
 484         unsigned first_unmapped = blocks_per_page;
 485         struct block_device *bdev = NULL;
 486         int boundary = 0;
 487         sector_t boundary_block = 0;
 488         struct block_device *boundary_bdev = NULL;
 489         int length;
 490         struct buffer_head map_bh;
 491         loff_t i_size = i_size_read(inode);
 492         int ret = 0;
 493         int op_flags = wbc_to_write_flags(wbc);
 494
 495         if (page_has_buffers(page)) {
 496                 struct buffer_head *head = page_buffers(page);
 497                 struct buffer_head *bh = head;
 498
 499                 /* If they're all mapped and dirty, do it */
 500                 page_block = 0;
 501                 do {
 502                         BUG_ON(buffer_locked(bh));
 503                         if (!buffer_mapped(bh)) {
 504                                 /*
 505                                  * unmapped dirty buffers are created by
 506                                  * __set_page_dirty_buffers -> mmapped data
 507                                  */
 508                                 if (buffer_dirty(bh))
 509                                         goto confused;
 510                                 if (first_unmapped == blocks_per_page)
 511                                         first_unmapped = page_block;
 512                                 continue;
 513                         }
 514
 515                         if (first_unmapped != blocks_per_page)
 516                                 goto confused;  /* hole -> non-hole */
 517
 518                         if (!buffer_dirty(bh) || !buffer_uptodate(bh))
 519                                 goto confused;
 520                         if (page_block) {
 521                                 if (bh->b_blocknr != blocks[page_block-1] + 1)
 522                                         goto confused;
 523                         }
 524                         blocks[page_block++] = bh->b_blocknr;
 525                         boundary = buffer_boundary(bh);
 526                         if (boundary) {
 527                                 boundary_block = bh->b_blocknr;
 528                                 boundary_bdev = bh->b_bdev;
 529                         }
 530                         bdev = bh->b_bdev;
 531                 } while ((bh = bh->b_this_page) != head);
 532
 533                 if (first_unmapped)
 534                         goto page_is_mapped;
 535
 536                 /*
 537                  * Page has buffers, but they are all unmapped. The page was
 538                  * created by pagein or read over a hole which was handled by
 539                  * block_read_full_page().  If this address_space is also
 540                  * using mpage_readpages then this can rarely happen.
 541                  */
 542                 goto confused;
 543         }
 544
 545         /*
 546          * The page has no buffers: map it to disk
 547          */
 548         BUG_ON(!PageUptodate(page));
 549         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
 550         last_block = (i_size - 1) >> blkbits;
 551         map_bh.b_page = page;
 552         for (page_block = 0; page_block < blocks_per_page; ) {
 553
 554                 map_bh.b_state = 0;
 555                 map_bh.b_size = 1 << blkbits;
 556                 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
 557                         goto confused;
 558                 if (buffer_new(&map_bh))
 559                         clean_bdev_bh_alias(&map_bh);
 560                 if (buffer_boundary(&map_bh)) {
 561                         boundary_block = map_bh.b_blocknr;
 562                         boundary_bdev = map_bh.b_bdev;
 563                 }
 564                 if (page_block) {
 565                         if (map_bh.b_blocknr != blocks[page_block-1] + 1)
 566                                 goto confused;
 567                 }
 568                 blocks[page_block++] = map_bh.b_blocknr;
 569                 boundary = buffer_boundary(&map_bh);
 570                 bdev = map_bh.b_bdev;
 571                 if (block_in_file == last_block)
 572                         break;
 573                 block_in_file++;
 574         }
 575         BUG_ON(page_block == 0);
 576
 577         first_unmapped = page_block;
 578
 579 page_is_mapped:
 580         end_index = i_size >> PAGE_SHIFT;
 581         if (page->index >= end_index) {
 582                 /*
 583                  * The page straddles i_size.  It must be zeroed out on each
 584                  * and every writepage invocation because it may be mmapped.
 585                  * "A file is mapped in multiples of the page size.  For a file
 586                  * that is not a multiple of the page size, the remaining memory
 587                  * is zeroed when mapped, and writes to that region are not
 588                  * written out to the file."
 589                  */
 590                 unsigned offset = i_size & (PAGE_SIZE - 1);
 591
 592                 if (page->index > end_index || !offset)
 593                         goto confused;
 594                 zero_user_segment(page, offset, PAGE_SIZE);
 595         }
 596
 597         /*
 598          * This page will go to BIO.  Do we need to send this BIO off first?
 599          */
 600         if (bio && mpd->last_block_in_bio != blocks[0] - 1)
 601                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
 602
 603 alloc_new:
 604         if (bio == NULL) {
 605                 if (first_unmapped == blocks_per_page) {
 606                         if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
 607                                                                 page, wbc)) {
 608                                 clean_buffers(page, first_unmapped);
 609                                 goto out;
 610                         }
 611                 }
 612                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 613                                 BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
 614                 if (bio == NULL)
 615                         goto confused;
 616
 617                 wbc_init_bio(wbc, bio);
 618                 bio->bi_write_hint = inode->i_write_hint;
 619         }
 620
 621         /*
 622          * Must try to add the page before marking the buffer clean or
 623          * the confused fail path above (OOM) will be very confused when
 624          * it finds all bh marked clean (i.e. it will not write anything)
 625          */
 626         wbc_account_io(wbc, page, PAGE_SIZE);
 627         length = first_unmapped << blkbits;
 628         if (bio_add_page(bio, page, length, 0) < length) {
 629                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
 630                 goto alloc_new;
 631         }
 632
 633         clean_buffers(page, first_unmapped);
 634
 635         BUG_ON(PageWriteback(page));
 636         set_page_writeback(page);
 637         unlock_page(page);
 638         if (boundary || (first_unmapped != blocks_per_page)) {
 639                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
 640                 if (boundary_block) {
 641                         write_boundary_block(boundary_bdev,
 642                                         boundary_block, 1 << blkbits);
 643                 }
 644         } else {
 645                 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
 646         }
 647         goto out;
 648
 649 confused:
 650         if (bio)
 651                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
 652
 653         if (mpd->use_writepage) {
 654                 ret = mapping->a_ops->writepage(page, wbc);
 655         } else {
 656                 ret = -EAGAIN;
 657                 goto out;
 658         }
 659         /*
 660          * The caller has a ref on the inode, so *mapping is stable
 661          */
 662         mapping_set_error(mapping, ret);
 663 out:
 664         mpd->bio = bio;
 665         return ret;
 666 }
 667
 668 /**
 669  * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
 670  * @mapping: address space structure to write
 671  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
 672  * @get_block: the filesystem's block mapper function.
 673  *             If this is NULL then use a_ops->writepage.  Otherwise, go
 674  *             direct-to-BIO.
 675  *
 676  * This is a library function, which implements the writepages()
 677  * address_space_operation.
 678  *
 679  * If a page is already under I/O, generic_writepages() skips it, even
 680  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
 681  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
 682  * and msync() need to guarantee that all the data which was dirty at the time
 683  * the call was made get new I/O started against them.  If wbc->sync_mode is
 684  * WB_SYNC_ALL then we were called for data integrity and we must wait for
 685  * existing IO to complete.
 686  */
 687 int
 688 mpage_writepages(struct address_space *mapping,
 689                 struct writeback_control *wbc, get_block_t get_block)
 690 {
 691         struct blk_plug plug;
 692         int ret;
 693
 694         blk_start_plug(&plug);
 695
 696         if (!get_block)
 697                 ret = generic_writepages(mapping, wbc);
 698         else {
 699                 struct mpage_data mpd = {
 700                         .bio = NULL,
 701                         .last_block_in_bio = 0,
 702                         .get_block = get_block,
 703                         .use_writepage = 1,
 704                 };
 705
 706                 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
 707                 if (mpd.bio) {
 708                         int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
 709                                   REQ_SYNC : 0);
 710                         mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
 711                 }
 712         }
 713         blk_finish_plug(&plug);
 714         return ret;
 715 }
 716 EXPORT_SYMBOL(mpage_writepages);
 717
 718 int mpage_writepage(struct page *page, get_block_t get_block,
 719         struct writeback_control *wbc)
 720 {
 721         struct mpage_data mpd = {
 722                 .bio = NULL,
 723                 .last_block_in_bio = 0,
 724                 .get_block = get_block,
 725                 .use_writepage = 0,
 726         };
 727         int ret = __mpage_writepage(page, wbc, &mpd);
 728         if (mpd.bio) {
 729                 int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
 730                           REQ_SYNC : 0);
 731                 mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
 732         }
 733         return ret;
 734 }
 735 EXPORT_SYMBOL(mpage_writepage);