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