fs/mpage.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * fs/mpage.c
   4  *
   5  * Copyright (C) 2002, Linus Torvalds.
   6  *
   7  * Contains functions related to preparing and submitting BIOs which contain
   8  * multiple pagecache pages.
   9  *
  10  * 15May2002    Andrew Morton
  11  *              Initial version
  12  * 27Jun2002    axboe@suse.de
  13  *              use bio_add_page() to build bio's just the right size
  14  */
  15
  16 #include <linux/kernel.h>
  17 #include <linux/export.h>
  18 #include <linux/mm.h>
  19 #include <linux/kdev_t.h>
  20 #include <linux/gfp.h>
  21 #include <linux/bio.h>
  22 #include <linux/fs.h>
  23 #include <linux/buffer_head.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/highmem.h>
  26 #include <linux/prefetch.h>
  27 #include <linux/mpage.h>
  28 #include <linux/mm_inline.h>
  29 #include <linux/writeback.h>
  30 #include <linux/backing-dev.h>
  31 #include <linux/pagevec.h>
  32 #include <linux/cleancache.h>
  33 #include "internal.h"
  34
  35 /*
  36  * I/O completion handler for multipage BIOs.
  37  *
  38  * The mpage code never puts partial pages into a BIO (except for end-of-file).
  39  * If a page does not map to a contiguous run of blocks then it simply falls
  40  * back to block_read_full_page().
  41  *
  42  * Why is this?  If a page's completion depends on a number of different BIOs
  43  * which can complete in any order (or at the same time) then determining the
  44  * status of that page is hard.  See end_buffer_async_read() for the details.
  45  * There is no point in duplicating all that complexity.
  46  */
  47 static void mpage_end_io(struct bio *bio)
  48 {
  49         struct bio_vec *bv;
  50         int i;
  51
  52         bio_for_each_segment_all(bv, bio, i) {
  53                 struct page *page = bv->bv_page;
  54                 page_endio(page, op_is_write(bio_op(bio)),
  55                                 blk_status_to_errno(bio->bi_status));
  56         }
  57
  58         bio_put(bio);
  59 }
  60
  61 static struct bio *mpage_bio_submit(int op, int op_flags, struct bio *bio)
  62 {
  63         bio->bi_end_io = mpage_end_io;
  64         bio_set_op_attrs(bio, op, op_flags);
  65         guard_bio_eod(op, bio);
  66         submit_bio(bio);
  67         return NULL;
  68 }
  69
  70 static struct bio *
  71 mpage_alloc(struct block_device *bdev,
  72                 sector_t first_sector, int nr_vecs,
  73                 gfp_t gfp_flags)
  74 {
  75         struct bio *bio;
  76
  77         /* Restrict the given (page cache) mask for slab allocations */
  78         gfp_flags &= GFP_KERNEL;
  79         bio = bio_alloc(gfp_flags, nr_vecs);
  80
  81         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
  82                 while (!bio && (nr_vecs /= 2))
  83                         bio = bio_alloc(gfp_flags, nr_vecs);
  84         }
  85
  86         if (bio) {
  87                 bio_set_dev(bio, bdev);
  88                 bio->bi_iter.bi_sector = first_sector;
  89         }
  90         return bio;
  91 }
  92
  93 /*
  94  * support function for mpage_readpages.  The fs supplied get_block might
  95  * return an up to date buffer.  This is used to map that buffer into
  96  * the page, which allows readpage to avoid triggering a duplicate call
  97  * to get_block.
  98  *
  99  * The idea is to avoid adding buffers to pages that don't already have
 100  * them.  So when the buffer is up to date and the page size == block size,
 101  * this marks the page up to date instead of adding new buffers.
 102  */
 103 static void
 104 map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
 105 {
 106         struct inode *inode = page->mapping->host;
 107         struct buffer_head *page_bh, *head;
 108         int block = 0;
 109
 110         if (!page_has_buffers(page)) {
 111                 /*
 112                  * don't make any buffers if there is only one buffer on
 113                  * the page and the page just needs to be set up to date
 114                  */
 115                 if (inode->i_blkbits == PAGE_SHIFT &&
 116                     buffer_uptodate(bh)) {
 117                         SetPageUptodate(page);
 118                         return;
 119                 }
 120                 create_empty_buffers(page, i_blocksize(inode), 0);
 121         }
 122         head = page_buffers(page);
 123         page_bh = head;
 124         do {
 125                 if (block == page_block) {
 126                         page_bh->b_state = bh->b_state;
 127                         page_bh->b_bdev = bh->b_bdev;
 128                         page_bh->b_blocknr = bh->b_blocknr;
 129                         break;
 130                 }
 131                 page_bh = page_bh->b_this_page;
 132                 block++;
 133         } while (page_bh != head);
 134 }
 135
 136 /*
 137  * This is the worker routine which does all the work of mapping the disk
 138  * blocks and constructs largest possible bios, submits them for IO if the
 139  * blocks are not contiguous on the disk.
 140  *
 141  * We pass a buffer_head back and forth and use its buffer_mapped() flag to
 142  * represent the validity of its disk mapping and to decide when to do the next
 143  * get_block() call.
 144  */
 145 static struct bio *
 146 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
 147                 sector_t *last_block_in_bio, struct buffer_head *map_bh,
 148                 unsigned long *first_logical_block, get_block_t get_block,
 149                 gfp_t gfp)
 150 {
 151         struct inode *inode = page->mapping->host;
 152         const unsigned blkbits = inode->i_blkbits;
 153         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 154         const unsigned blocksize = 1 << blkbits;
 155         sector_t block_in_file;
 156         sector_t last_block;
 157         sector_t last_block_in_file;
 158         sector_t blocks[MAX_BUF_PER_PAGE];
 159         unsigned page_block;
 160         unsigned first_hole = blocks_per_page;
 161         struct block_device *bdev = NULL;
 162         int length;
 163         int fully_mapped = 1;
 164         unsigned nblocks;
 165         unsigned relative_block;
 166
 167         if (page_has_buffers(page))
 168                 goto confused;
 169
 170         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
 171         last_block = block_in_file + nr_pages * blocks_per_page;
 172         last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
 173         if (last_block > last_block_in_file)
 174                 last_block = last_block_in_file;
 175         page_block = 0;
 176
 177         /*
 178          * Map blocks using the result from the previous get_blocks call first.
 179          */
 180         nblocks = map_bh->b_size >> blkbits;
 181         if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
 182                         block_in_file < (*first_logical_block + nblocks)) {
 183                 unsigned map_offset = block_in_file - *first_logical_block;
 184                 unsigned last = nblocks - map_offset;
 185
 186                 for (relative_block = 0; ; relative_block++) {
 187                         if (relative_block == last) {
 188                                 clear_buffer_mapped(map_bh);
 189                                 break;
 190                         }
 191                         if (page_block == blocks_per_page)
 192                                 break;
 193                         blocks[page_block] = map_bh->b_blocknr + map_offset +
 194                                                 relative_block;
 195                         page_block++;
 196                         block_in_file++;
 197                 }
 198                 bdev = map_bh->b_bdev;
 199         }
 200
 201         /*
 202          * Then do more get_blocks calls until we are done with this page.
 203          */
 204         map_bh->b_page = page;
 205         while (page_block < blocks_per_page) {
 206                 map_bh->b_state = 0;
 207                 map_bh->b_size = 0;
 208
 209                 if (block_in_file < last_block) {
 210                         map_bh->b_size = (last_block-block_in_file) << blkbits;
 211                         if (get_block(inode, block_in_file, map_bh, 0))
 212                                 goto confused;
 213                         *first_logical_block = block_in_file;
 214                 }
 215
 216                 if (!buffer_mapped(map_bh)) {
 217                         fully_mapped = 0;
 218                         if (first_hole == blocks_per_page)
 219                                 first_hole = page_block;
 220                         page_block++;
 221                         block_in_file++;
 222                         continue;
 223                 }
 224
 225                 /* some filesystems will copy data into the page during
 226                  * the get_block call, in which case we don't want to
 227                  * read it again.  map_buffer_to_page copies the data
 228                  * we just collected from get_block into the page's buffers
 229                  * so readpage doesn't have to repeat the get_block call
 230                  */
 231                 if (buffer_uptodate(map_bh)) {
 232                         map_buffer_to_page(page, map_bh, page_block);
 233                         goto confused;
 234                 }
 235
 236                 if (first_hole != blocks_per_page)
 237                         goto confused;          /* hole -> non-hole */
 238
 239                 /* Contiguous blocks? */
 240                 if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
 241                         goto confused;
 242                 nblocks = map_bh->b_size >> blkbits;
 243                 for (relative_block = 0; ; relative_block++) {
 244                         if (relative_block == nblocks) {
 245                                 clear_buffer_mapped(map_bh);
 246                                 break;
 247                         } else if (page_block == blocks_per_page)
 248                                 break;
 249                         blocks[page_block] = map_bh->b_blocknr+relative_block;
 250                         page_block++;
 251                         block_in_file++;
 252                 }
 253                 bdev = map_bh->b_bdev;
 254         }
 255
 256         if (first_hole != blocks_per_page) {
 257                 zero_user_segment(page, first_hole << blkbits, PAGE_SIZE);
 258                 if (first_hole == 0) {
 259                         SetPageUptodate(page);
 260                         unlock_page(page);
 261                         goto out;
 262                 }
 263         } else if (fully_mapped) {
 264                 SetPageMappedToDisk(page);
 265         }
 266
 267         if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
 268             cleancache_get_page(page) == 0) {
 269                 SetPageUptodate(page);
 270                 goto confused;
 271         }
 272
 273         /*
 274          * This page will go to BIO.  Do we need to send this BIO off first?
 275          */
 276         if (bio && (*last_block_in_bio != blocks[0] - 1))
 277                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
 278
 279 alloc_new:
 280         if (bio == NULL) {
 281                 if (first_hole == blocks_per_page) {
 282                         if (!bdev_read_page(bdev, blocks[0] << (blkbits - 9),
 283                                                                 page))
 284                                 goto out;
 285                 }
 286                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 287                                 min_t(int, nr_pages, BIO_MAX_PAGES), gfp);
 288                 if (bio == NULL)
 289                         goto confused;
 290         }
 291
 292         length = first_hole << blkbits;
 293         if (bio_add_page(bio, page, length, 0) < length) {
 294                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
 295                 goto alloc_new;
 296         }
 297
 298         relative_block = block_in_file - *first_logical_block;
 299         nblocks = map_bh->b_size >> blkbits;
 300         if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
 301             (first_hole != blocks_per_page))
 302                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
 303         else
 304                 *last_block_in_bio = blocks[blocks_per_page - 1];
 305 out:
 306         return bio;
 307
 308 confused:
 309         if (bio)
 310                 bio = mpage_bio_submit(REQ_OP_READ, 0, bio);
 311         if (!PageUptodate(page))
 312                 block_read_full_page(page, get_block);
 313         else
 314                 unlock_page(page);
 315         goto out;
 316 }
 317
 318 /**
 319  * mpage_readpages - populate an address space with some pages & start reads against them
 320  * @mapping: the address_space
 321  * @pages: The address of a list_head which contains the target pages.  These
 322  *   pages have their ->index populated and are otherwise uninitialised.
 323  *   The page at @pages->prev has the lowest file offset, and reads should be
 324  *   issued in @pages->prev to @pages->next order.
 325  * @nr_pages: The number of pages at *@pages
 326  * @get_block: The filesystem's block mapper function.
 327  *
 328  * This function walks the pages and the blocks within each page, building and
 329  * emitting large BIOs.
 330  *
 331  * If anything unusual happens, such as:
 332  *
 333  * - encountering a page which has buffers
 334  * - encountering a page which has a non-hole after a hole
 335  * - encountering a page with non-contiguous blocks
 336  *
 337  * then this code just gives up and calls the buffer_head-based read function.
 338  * It does handle a page which has holes at the end - that is a common case:
 339  * the end-of-file on blocksize < PAGE_SIZE setups.
 340  *
 341  * BH_Boundary explanation:
 342  *
 343  * There is a problem.  The mpage read code assembles several pages, gets all
 344  * their disk mappings, and then submits them all.  That's fine, but obtaining
 345  * the disk mappings may require I/O.  Reads of indirect blocks, for example.
 346  *
 347  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
 348  * submitted in the following order:
 349  *
 350  *      12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
 351  *
 352  * because the indirect block has to be read to get the mappings of blocks
 353  * 13,14,15,16.  Obviously, this impacts performance.
 354  *
 355  * So what we do it to allow the filesystem's get_block() function to set
 356  * BH_Boundary when it maps block 11.  BH_Boundary says: mapping of the block
 357  * after this one will require I/O against a block which is probably close to
 358  * this one.  So you should push what I/O you have currently accumulated.
 359  *
 360  * This all causes the disk requests to be issued in the correct order.
 361  */
 362 int
 363 mpage_readpages(struct address_space *mapping, struct list_head *pages,
 364                                 unsigned nr_pages, get_block_t get_block)
 365 {
 366         struct bio *bio = NULL;
 367         unsigned page_idx;
 368         sector_t last_block_in_bio = 0;
 369         struct buffer_head map_bh;
 370         unsigned long first_logical_block = 0;
 371         gfp_t gfp = readahead_gfp_mask(mapping);
 372
 373         map_bh.b_state = 0;
 374         map_bh.b_size = 0;
 375         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
 376                 struct page *page = lru_to_page(pages);
 377
 378                 prefetchw(&page->flags);
 379                 list_del(&page->lru);
 380                 if (!add_to_page_cache_lru(page, mapping,
 381                                         page->index,
 382                                         gfp)) {
 383                         bio = do_mpage_readpage(bio, page,
 384                                         nr_pages - page_idx,
 385                                         &last_block_in_bio, &map_bh,
 386                                         &first_logical_block,
 387                                         get_block, gfp);
 388                 }
 389                 put_page(page);
 390         }
 391         BUG_ON(!list_empty(pages));
 392         if (bio)
 393                 mpage_bio_submit(REQ_OP_READ, 0, bio);
 394         return 0;
 395 }
 396 EXPORT_SYMBOL(mpage_readpages);
 397
 398 /*
 399  * This isn't called much at all
 400  */
 401 int mpage_readpage(struct page *page, get_block_t get_block)
 402 {
 403         struct bio *bio = NULL;
 404         sector_t last_block_in_bio = 0;
 405         struct buffer_head map_bh;
 406         unsigned long first_logical_block = 0;
 407         gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
 408
 409         map_bh.b_state = 0;
 410         map_bh.b_size = 0;
 411         bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
 412                         &map_bh, &first_logical_block, get_block, gfp);
 413         if (bio)
 414                 mpage_bio_submit(REQ_OP_READ, 0, bio);
 415         return 0;
 416 }
 417 EXPORT_SYMBOL(mpage_readpage);
 418
 419 /*
 420  * Writing is not so simple.
 421  *
 422  * If the page has buffers then they will be used for obtaining the disk
 423  * mapping.  We only support pages which are fully mapped-and-dirty, with a
 424  * special case for pages which are unmapped at the end: end-of-file.
 425  *
 426  * If the page has no buffers (preferred) then the page is mapped here.
 427  *
 428  * If all blocks are found to be contiguous then the page can go into the
 429  * BIO.  Otherwise fall back to the mapping's writepage().
 430  *
 431  * FIXME: This code wants an estimate of how many pages are still to be
 432  * written, so it can intelligently allocate a suitably-sized BIO.  For now,
 433  * just allocate full-size (16-page) BIOs.
 434  */
 435
 436 struct mpage_data {
 437         struct bio *bio;
 438         sector_t last_block_in_bio;
 439         get_block_t *get_block;
 440         unsigned use_writepage;
 441 };
 442
 443 /*
 444  * We have our BIO, so we can now mark the buffers clean.  Make
 445  * sure to only clean buffers which we know we'll be writing.
 446  */
 447 static void clean_buffers(struct page *page, unsigned first_unmapped)
 448 {
 449         unsigned buffer_counter = 0;
 450         struct buffer_head *bh, *head;
 451         if (!page_has_buffers(page))
 452                 return;
 453         head = page_buffers(page);
 454         bh = head;
 455
 456         do {
 457                 if (buffer_counter++ == first_unmapped)
 458                         break;
 459                 clear_buffer_dirty(bh);
 460                 bh = bh->b_this_page;
 461         } while (bh != head);
 462
 463         /*
 464          * we cannot drop the bh if the page is not uptodate or a concurrent
 465          * readpage would fail to serialize with the bh and it would read from
 466          * disk before we reach the platter.
 467          */
 468         if (buffer_heads_over_limit && PageUptodate(page))
 469                 try_to_free_buffers(page);
 470 }
 471
 472 /*
 473  * For situations where we want to clean all buffers attached to a page.
 474  * We don't need to calculate how many buffers are attached to the page,
 475  * we just need to specify a number larger than the maximum number of buffers.
 476  */
 477 void clean_page_buffers(struct page *page)
 478 {
 479         clean_buffers(page, ~0U);
 480 }
 481
 482 static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
 483                       void *data)
 484 {
 485         struct mpage_data *mpd = data;
 486         struct bio *bio = mpd->bio;
 487         struct address_space *mapping = page->mapping;
 488         struct inode *inode = page->mapping->host;
 489         const unsigned blkbits = inode->i_blkbits;
 490         unsigned long end_index;
 491         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 492         sector_t last_block;
 493         sector_t block_in_file;
 494         sector_t blocks[MAX_BUF_PER_PAGE];
 495         unsigned page_block;
 496         unsigned first_unmapped = blocks_per_page;
 497         struct block_device *bdev = NULL;
 498         int boundary = 0;
 499         sector_t boundary_block = 0;
 500         struct block_device *boundary_bdev = NULL;
 501         int length;
 502         struct buffer_head map_bh;
 503         loff_t i_size = i_size_read(inode);
 504         int ret = 0;
 505         int op_flags = wbc_to_write_flags(wbc);
 506
 507         if (page_has_buffers(page)) {
 508                 struct buffer_head *head = page_buffers(page);
 509                 struct buffer_head *bh = head;
 510
 511                 /* If they're all mapped and dirty, do it */
 512                 page_block = 0;
 513                 do {
 514                         BUG_ON(buffer_locked(bh));
 515                         if (!buffer_mapped(bh)) {
 516                                 /*
 517                                  * unmapped dirty buffers are created by
 518                                  * __set_page_dirty_buffers -> mmapped data
 519                                  */
 520                                 if (buffer_dirty(bh))
 521                                         goto confused;
 522                                 if (first_unmapped == blocks_per_page)
 523                                         first_unmapped = page_block;
 524                                 continue;
 525                         }
 526
 527                         if (first_unmapped != blocks_per_page)
 528                                 goto confused;  /* hole -> non-hole */
 529
 530                         if (!buffer_dirty(bh) || !buffer_uptodate(bh))
 531                                 goto confused;
 532                         if (page_block) {
 533                                 if (bh->b_blocknr != blocks[page_block-1] + 1)
 534                                         goto confused;
 535                         }
 536                         blocks[page_block++] = bh->b_blocknr;
 537                         boundary = buffer_boundary(bh);
 538                         if (boundary) {
 539                                 boundary_block = bh->b_blocknr;
 540                                 boundary_bdev = bh->b_bdev;
 541                         }
 542                         bdev = bh->b_bdev;
 543                 } while ((bh = bh->b_this_page) != head);
 544
 545                 if (first_unmapped)
 546                         goto page_is_mapped;
 547
 548                 /*
 549                  * Page has buffers, but they are all unmapped. The page was
 550                  * created by pagein or read over a hole which was handled by
 551                  * block_read_full_page().  If this address_space is also
 552                  * using mpage_readpages then this can rarely happen.
 553                  */
 554                 goto confused;
 555         }
 556
 557         /*
 558          * The page has no buffers: map it to disk
 559          */
 560         BUG_ON(!PageUptodate(page));
 561         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
 562         last_block = (i_size - 1) >> blkbits;
 563         map_bh.b_page = page;
 564         for (page_block = 0; page_block < blocks_per_page; ) {
 565
 566                 map_bh.b_state = 0;
 567                 map_bh.b_size = 1 << blkbits;
 568                 if (mpd->get_block(inode, block_in_file, &map_bh, 1))
 569                         goto confused;
 570                 if (buffer_new(&map_bh))
 571                         clean_bdev_bh_alias(&map_bh);
 572                 if (buffer_boundary(&map_bh)) {
 573                         boundary_block = map_bh.b_blocknr;
 574                         boundary_bdev = map_bh.b_bdev;
 575                 }
 576                 if (page_block) {
 577                         if (map_bh.b_blocknr != blocks[page_block-1] + 1)
 578                                 goto confused;
 579                 }
 580                 blocks[page_block++] = map_bh.b_blocknr;
 581                 boundary = buffer_boundary(&map_bh);
 582                 bdev = map_bh.b_bdev;
 583                 if (block_in_file == last_block)
 584                         break;
 585                 block_in_file++;
 586         }
 587         BUG_ON(page_block == 0);
 588
 589         first_unmapped = page_block;
 590
 591 page_is_mapped:
 592         end_index = i_size >> PAGE_SHIFT;
 593         if (page->index >= end_index) {
 594                 /*
 595                  * The page straddles i_size.  It must be zeroed out on each
 596                  * and every writepage invocation because it may be mmapped.
 597                  * "A file is mapped in multiples of the page size.  For a file
 598                  * that is not a multiple of the page size, the remaining memory
 599                  * is zeroed when mapped, and writes to that region are not
 600                  * written out to the file."
 601                  */
 602                 unsigned offset = i_size & (PAGE_SIZE - 1);
 603
 604                 if (page->index > end_index || !offset)
 605                         goto confused;
 606                 zero_user_segment(page, offset, PAGE_SIZE);
 607         }
 608
 609         /*
 610          * This page will go to BIO.  Do we need to send this BIO off first?
 611          */
 612         if (bio && mpd->last_block_in_bio != blocks[0] - 1)
 613                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
 614
 615 alloc_new:
 616         if (bio == NULL) {
 617                 if (first_unmapped == blocks_per_page) {
 618                         if (!bdev_write_page(bdev, blocks[0] << (blkbits - 9),
 619                                                                 page, wbc))
 620                                 goto out;
 621                 }
 622                 bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
 623                                 BIO_MAX_PAGES, GFP_NOFS|__GFP_HIGH);
 624                 if (bio == NULL)
 625                         goto confused;
 626
 627                 wbc_init_bio(wbc, bio);
 628                 bio->bi_write_hint = inode->i_write_hint;
 629         }
 630
 631         /*
 632          * Must try to add the page before marking the buffer clean or
 633          * the confused fail path above (OOM) will be very confused when
 634          * it finds all bh marked clean (i.e. it will not write anything)
 635          */
 636         wbc_account_io(wbc, page, PAGE_SIZE);
 637         length = first_unmapped << blkbits;
 638         if (bio_add_page(bio, page, length, 0) < length) {
 639                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
 640                 goto alloc_new;
 641         }
 642
 643         clean_buffers(page, first_unmapped);
 644
 645         BUG_ON(PageWriteback(page));
 646         set_page_writeback(page);
 647         unlock_page(page);
 648         if (boundary || (first_unmapped != blocks_per_page)) {
 649                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
 650                 if (boundary_block) {
 651                         write_boundary_block(boundary_bdev,
 652                                         boundary_block, 1 << blkbits);
 653                 }
 654         } else {
 655                 mpd->last_block_in_bio = blocks[blocks_per_page - 1];
 656         }
 657         goto out;
 658
 659 confused:
 660         if (bio)
 661                 bio = mpage_bio_submit(REQ_OP_WRITE, op_flags, bio);
 662
 663         if (mpd->use_writepage) {
 664                 ret = mapping->a_ops->writepage(page, wbc);
 665         } else {
 666                 ret = -EAGAIN;
 667                 goto out;
 668         }
 669         /*
 670          * The caller has a ref on the inode, so *mapping is stable
 671          */
 672         mapping_set_error(mapping, ret);
 673 out:
 674         mpd->bio = bio;
 675         return ret;
 676 }
 677
 678 /**
 679  * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
 680  * @mapping: address space structure to write
 681  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
 682  * @get_block: the filesystem's block mapper function.
 683  *             If this is NULL then use a_ops->writepage.  Otherwise, go
 684  *             direct-to-BIO.
 685  *
 686  * This is a library function, which implements the writepages()
 687  * address_space_operation.
 688  *
 689  * If a page is already under I/O, generic_writepages() skips it, even
 690  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
 691  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
 692  * and msync() need to guarantee that all the data which was dirty at the time
 693  * the call was made get new I/O started against them.  If wbc->sync_mode is
 694  * WB_SYNC_ALL then we were called for data integrity and we must wait for
 695  * existing IO to complete.
 696  */
 697 int
 698 mpage_writepages(struct address_space *mapping,
 699                 struct writeback_control *wbc, get_block_t get_block)
 700 {
 701         struct blk_plug plug;
 702         int ret;
 703
 704         blk_start_plug(&plug);
 705
 706         if (!get_block)
 707                 ret = generic_writepages(mapping, wbc);
 708         else {
 709                 struct mpage_data mpd = {
 710                         .bio = NULL,
 711                         .last_block_in_bio = 0,
 712                         .get_block = get_block,
 713                         .use_writepage = 1,
 714                 };
 715
 716                 ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
 717                 if (mpd.bio) {
 718                         int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
 719                                   REQ_SYNC : 0);
 720                         mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
 721                 }
 722         }
 723         blk_finish_plug(&plug);
 724         return ret;
 725 }
 726 EXPORT_SYMBOL(mpage_writepages);
 727
 728 int mpage_writepage(struct page *page, get_block_t get_block,
 729         struct writeback_control *wbc)
 730 {
 731         struct mpage_data mpd = {
 732                 .bio = NULL,
 733                 .last_block_in_bio = 0,
 734                 .get_block = get_block,
 735                 .use_writepage = 0,
 736         };
 737         int ret = __mpage_writepage(page, wbc, &mpd);
 738         if (mpd.bio) {
 739                 int op_flags = (wbc->sync_mode == WB_SYNC_ALL ?
 740                           REQ_SYNC : 0);
 741                 mpage_bio_submit(REQ_OP_WRITE, op_flags, mpd.bio);
 742         }
 743         return ret;
 744 }
 745 EXPORT_SYMBOL(mpage_writepage);