1 #include <linux/bitops.h> 2 #include <linux/slab.h> 3 #include <linux/bio.h> 4 #include <linux/mm.h> 5 #include <linux/pagemap.h> 6 #include <linux/page-flags.h> 7 #include <linux/spinlock.h> 8 #include <linux/blkdev.h> 9 #include <linux/swap.h> 10 #include <linux/writeback.h> 11 #include <linux/pagevec.h> 12 #include <linux/prefetch.h> 13 #include <linux/cleancache.h> 14 #include "extent_io.h" 15 #include "extent_map.h" 16 #include "ctree.h" 17 #include "btrfs_inode.h" 18 #include "volumes.h" 19 #include "check-integrity.h" 20 #include "locking.h" 21 #include "rcu-string.h" 22 #include "backref.h" 23 24 static struct kmem_cache *extent_state_cache; 25 static struct kmem_cache *extent_buffer_cache; 26 static struct bio_set *btrfs_bioset; 27 28 static inline bool extent_state_in_tree(const struct extent_state *state) 29 { 30 return !RB_EMPTY_NODE(&state->rb_node); 31 } 32 33 #ifdef CONFIG_BTRFS_DEBUG 34 static LIST_HEAD(buffers); 35 static LIST_HEAD(states); 36 37 static DEFINE_SPINLOCK(leak_lock); 38 39 static inline 40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head) 41 { 42 unsigned long flags; 43 44 spin_lock_irqsave(&leak_lock, flags); 45 list_add(new, head); 46 spin_unlock_irqrestore(&leak_lock, flags); 47 } 48 49 static inline 50 void btrfs_leak_debug_del(struct list_head *entry) 51 { 52 unsigned long flags; 53 54 spin_lock_irqsave(&leak_lock, flags); 55 list_del(entry); 56 spin_unlock_irqrestore(&leak_lock, flags); 57 } 58 59 static inline 60 void btrfs_leak_debug_check(void) 61 { 62 struct extent_state *state; 63 struct extent_buffer *eb; 64 65 while (!list_empty(&states)) { 66 state = list_entry(states.next, struct extent_state, leak_list); 67 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n", 68 state->start, state->end, state->state, 69 extent_state_in_tree(state), 70 refcount_read(&state->refs)); 71 list_del(&state->leak_list); 72 kmem_cache_free(extent_state_cache, state); 73 } 74 75 while (!list_empty(&buffers)) { 76 eb = list_entry(buffers.next, struct extent_buffer, leak_list); 77 pr_err("BTRFS: buffer leak start %llu len %lu refs %d\n", 78 eb->start, eb->len, atomic_read(&eb->refs)); 79 list_del(&eb->leak_list); 80 kmem_cache_free(extent_buffer_cache, eb); 81 } 82 } 83 84 #define btrfs_debug_check_extent_io_range(tree, start, end) \ 85 __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end)) 86 static inline void __btrfs_debug_check_extent_io_range(const char *caller, 87 struct extent_io_tree *tree, u64 start, u64 end) 88 { 89 if (tree->ops && tree->ops->check_extent_io_range) 90 tree->ops->check_extent_io_range(tree->private_data, caller, 91 start, end); 92 } 93 #else 94 #define btrfs_leak_debug_add(new, head) do {} while (0) 95 #define btrfs_leak_debug_del(entry) do {} while (0) 96 #define btrfs_leak_debug_check() do {} while (0) 97 #define btrfs_debug_check_extent_io_range(c, s, e) do {} while (0) 98 #endif 99 100 #define BUFFER_LRU_MAX 64 101 102 struct tree_entry { 103 u64 start; 104 u64 end; 105 struct rb_node rb_node; 106 }; 107 108 struct extent_page_data { 109 struct bio *bio; 110 struct extent_io_tree *tree; 111 get_extent_t *get_extent; 112 unsigned long bio_flags; 113 114 /* tells writepage not to lock the state bits for this range 115 * it still does the unlocking 116 */ 117 unsigned int extent_locked:1; 118 119 /* tells the submit_bio code to use REQ_SYNC */ 120 unsigned int sync_io:1; 121 }; 122 123 static void add_extent_changeset(struct extent_state *state, unsigned bits, 124 struct extent_changeset *changeset, 125 int set) 126 { 127 int ret; 128 129 if (!changeset) 130 return; 131 if (set && (state->state & bits) == bits) 132 return; 133 if (!set && (state->state & bits) == 0) 134 return; 135 changeset->bytes_changed += state->end - state->start + 1; 136 ret = ulist_add(&changeset->range_changed, state->start, state->end, 137 GFP_ATOMIC); 138 /* ENOMEM */ 139 BUG_ON(ret < 0); 140 } 141 142 static noinline void flush_write_bio(void *data); 143 static inline struct btrfs_fs_info * 144 tree_fs_info(struct extent_io_tree *tree) 145 { 146 if (tree->ops) 147 return tree->ops->tree_fs_info(tree->private_data); 148 return NULL; 149 } 150 151 int __init extent_io_init(void) 152 { 153 extent_state_cache = kmem_cache_create("btrfs_extent_state", 154 sizeof(struct extent_state), 0, 155 SLAB_MEM_SPREAD, NULL); 156 if (!extent_state_cache) 157 return -ENOMEM; 158 159 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer", 160 sizeof(struct extent_buffer), 0, 161 SLAB_MEM_SPREAD, NULL); 162 if (!extent_buffer_cache) 163 goto free_state_cache; 164 165 btrfs_bioset = bioset_create(BIO_POOL_SIZE, 166 offsetof(struct btrfs_io_bio, bio), 167 BIOSET_NEED_BVECS); 168 if (!btrfs_bioset) 169 goto free_buffer_cache; 170 171 if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE)) 172 goto free_bioset; 173 174 return 0; 175 176 free_bioset: 177 bioset_free(btrfs_bioset); 178 btrfs_bioset = NULL; 179 180 free_buffer_cache: 181 kmem_cache_destroy(extent_buffer_cache); 182 extent_buffer_cache = NULL; 183 184 free_state_cache: 185 kmem_cache_destroy(extent_state_cache); 186 extent_state_cache = NULL; 187 return -ENOMEM; 188 } 189 190 void extent_io_exit(void) 191 { 192 btrfs_leak_debug_check(); 193 194 /* 195 * Make sure all delayed rcu free are flushed before we 196 * destroy caches. 197 */ 198 rcu_barrier(); 199 kmem_cache_destroy(extent_state_cache); 200 kmem_cache_destroy(extent_buffer_cache); 201 if (btrfs_bioset) 202 bioset_free(btrfs_bioset); 203 } 204 205 void extent_io_tree_init(struct extent_io_tree *tree, 206 void *private_data) 207 { 208 tree->state = RB_ROOT; 209 tree->ops = NULL; 210 tree->dirty_bytes = 0; 211 spin_lock_init(&tree->lock); 212 tree->private_data = private_data; 213 } 214 215 static struct extent_state *alloc_extent_state(gfp_t mask) 216 { 217 struct extent_state *state; 218 219 /* 220 * The given mask might be not appropriate for the slab allocator, 221 * drop the unsupported bits 222 */ 223 mask &= ~(__GFP_DMA32|__GFP_HIGHMEM); 224 state = kmem_cache_alloc(extent_state_cache, mask); 225 if (!state) 226 return state; 227 state->state = 0; 228 state->failrec = NULL; 229 RB_CLEAR_NODE(&state->rb_node); 230 btrfs_leak_debug_add(&state->leak_list, &states); 231 refcount_set(&state->refs, 1); 232 init_waitqueue_head(&state->wq); 233 trace_alloc_extent_state(state, mask, _RET_IP_); 234 return state; 235 } 236 237 void free_extent_state(struct extent_state *state) 238 { 239 if (!state) 240 return; 241 if (refcount_dec_and_test(&state->refs)) { 242 WARN_ON(extent_state_in_tree(state)); 243 btrfs_leak_debug_del(&state->leak_list); 244 trace_free_extent_state(state, _RET_IP_); 245 kmem_cache_free(extent_state_cache, state); 246 } 247 } 248 249 static struct rb_node *tree_insert(struct rb_root *root, 250 struct rb_node *search_start, 251 u64 offset, 252 struct rb_node *node, 253 struct rb_node ***p_in, 254 struct rb_node **parent_in) 255 { 256 struct rb_node **p; 257 struct rb_node *parent = NULL; 258 struct tree_entry *entry; 259 260 if (p_in && parent_in) { 261 p = *p_in; 262 parent = *parent_in; 263 goto do_insert; 264 } 265 266 p = search_start ? &search_start : &root->rb_node; 267 while (*p) { 268 parent = *p; 269 entry = rb_entry(parent, struct tree_entry, rb_node); 270 271 if (offset < entry->start) 272 p = &(*p)->rb_left; 273 else if (offset > entry->end) 274 p = &(*p)->rb_right; 275 else 276 return parent; 277 } 278 279 do_insert: 280 rb_link_node(node, parent, p); 281 rb_insert_color(node, root); 282 return NULL; 283 } 284 285 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset, 286 struct rb_node **prev_ret, 287 struct rb_node **next_ret, 288 struct rb_node ***p_ret, 289 struct rb_node **parent_ret) 290 { 291 struct rb_root *root = &tree->state; 292 struct rb_node **n = &root->rb_node; 293 struct rb_node *prev = NULL; 294 struct rb_node *orig_prev = NULL; 295 struct tree_entry *entry; 296 struct tree_entry *prev_entry = NULL; 297 298 while (*n) { 299 prev = *n; 300 entry = rb_entry(prev, struct tree_entry, rb_node); 301 prev_entry = entry; 302 303 if (offset < entry->start) 304 n = &(*n)->rb_left; 305 else if (offset > entry->end) 306 n = &(*n)->rb_right; 307 else 308 return *n; 309 } 310 311 if (p_ret) 312 *p_ret = n; 313 if (parent_ret) 314 *parent_ret = prev; 315 316 if (prev_ret) { 317 orig_prev = prev; 318 while (prev && offset > prev_entry->end) { 319 prev = rb_next(prev); 320 prev_entry = rb_entry(prev, struct tree_entry, rb_node); 321 } 322 *prev_ret = prev; 323 prev = orig_prev; 324 } 325 326 if (next_ret) { 327 prev_entry = rb_entry(prev, struct tree_entry, rb_node); 328 while (prev && offset < prev_entry->start) { 329 prev = rb_prev(prev); 330 prev_entry = rb_entry(prev, struct tree_entry, rb_node); 331 } 332 *next_ret = prev; 333 } 334 return NULL; 335 } 336 337 static inline struct rb_node * 338 tree_search_for_insert(struct extent_io_tree *tree, 339 u64 offset, 340 struct rb_node ***p_ret, 341 struct rb_node **parent_ret) 342 { 343 struct rb_node *prev = NULL; 344 struct rb_node *ret; 345 346 ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret); 347 if (!ret) 348 return prev; 349 return ret; 350 } 351 352 static inline struct rb_node *tree_search(struct extent_io_tree *tree, 353 u64 offset) 354 { 355 return tree_search_for_insert(tree, offset, NULL, NULL); 356 } 357 358 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new, 359 struct extent_state *other) 360 { 361 if (tree->ops && tree->ops->merge_extent_hook) 362 tree->ops->merge_extent_hook(tree->private_data, new, other); 363 } 364 365 /* 366 * utility function to look for merge candidates inside a given range. 367 * Any extents with matching state are merged together into a single 368 * extent in the tree. Extents with EXTENT_IO in their state field 369 * are not merged because the end_io handlers need to be able to do 370 * operations on them without sleeping (or doing allocations/splits). 371 * 372 * This should be called with the tree lock held. 373 */ 374 static void merge_state(struct extent_io_tree *tree, 375 struct extent_state *state) 376 { 377 struct extent_state *other; 378 struct rb_node *other_node; 379 380 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) 381 return; 382 383 other_node = rb_prev(&state->rb_node); 384 if (other_node) { 385 other = rb_entry(other_node, struct extent_state, rb_node); 386 if (other->end == state->start - 1 && 387 other->state == state->state) { 388 merge_cb(tree, state, other); 389 state->start = other->start; 390 rb_erase(&other->rb_node, &tree->state); 391 RB_CLEAR_NODE(&other->rb_node); 392 free_extent_state(other); 393 } 394 } 395 other_node = rb_next(&state->rb_node); 396 if (other_node) { 397 other = rb_entry(other_node, struct extent_state, rb_node); 398 if (other->start == state->end + 1 && 399 other->state == state->state) { 400 merge_cb(tree, state, other); 401 state->end = other->end; 402 rb_erase(&other->rb_node, &tree->state); 403 RB_CLEAR_NODE(&other->rb_node); 404 free_extent_state(other); 405 } 406 } 407 } 408 409 static void set_state_cb(struct extent_io_tree *tree, 410 struct extent_state *state, unsigned *bits) 411 { 412 if (tree->ops && tree->ops->set_bit_hook) 413 tree->ops->set_bit_hook(tree->private_data, state, bits); 414 } 415 416 static void clear_state_cb(struct extent_io_tree *tree, 417 struct extent_state *state, unsigned *bits) 418 { 419 if (tree->ops && tree->ops->clear_bit_hook) 420 tree->ops->clear_bit_hook(tree->private_data, state, bits); 421 } 422 423 static void set_state_bits(struct extent_io_tree *tree, 424 struct extent_state *state, unsigned *bits, 425 struct extent_changeset *changeset); 426 427 /* 428 * insert an extent_state struct into the tree. 'bits' are set on the 429 * struct before it is inserted. 430 * 431 * This may return -EEXIST if the extent is already there, in which case the 432 * state struct is freed. 433 * 434 * The tree lock is not taken internally. This is a utility function and 435 * probably isn't what you want to call (see set/clear_extent_bit). 436 */ 437 static int insert_state(struct extent_io_tree *tree, 438 struct extent_state *state, u64 start, u64 end, 439 struct rb_node ***p, 440 struct rb_node **parent, 441 unsigned *bits, struct extent_changeset *changeset) 442 { 443 struct rb_node *node; 444 445 if (end < start) 446 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n", 447 end, start); 448 state->start = start; 449 state->end = end; 450 451 set_state_bits(tree, state, bits, changeset); 452 453 node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent); 454 if (node) { 455 struct extent_state *found; 456 found = rb_entry(node, struct extent_state, rb_node); 457 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n", 458 found->start, found->end, start, end); 459 return -EEXIST; 460 } 461 merge_state(tree, state); 462 return 0; 463 } 464 465 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig, 466 u64 split) 467 { 468 if (tree->ops && tree->ops->split_extent_hook) 469 tree->ops->split_extent_hook(tree->private_data, orig, split); 470 } 471 472 /* 473 * split a given extent state struct in two, inserting the preallocated 474 * struct 'prealloc' as the newly created second half. 'split' indicates an 475 * offset inside 'orig' where it should be split. 476 * 477 * Before calling, 478 * the tree has 'orig' at [orig->start, orig->end]. After calling, there 479 * are two extent state structs in the tree: 480 * prealloc: [orig->start, split - 1] 481 * orig: [ split, orig->end ] 482 * 483 * The tree locks are not taken by this function. They need to be held 484 * by the caller. 485 */ 486 static int split_state(struct extent_io_tree *tree, struct extent_state *orig, 487 struct extent_state *prealloc, u64 split) 488 { 489 struct rb_node *node; 490 491 split_cb(tree, orig, split); 492 493 prealloc->start = orig->start; 494 prealloc->end = split - 1; 495 prealloc->state = orig->state; 496 orig->start = split; 497 498 node = tree_insert(&tree->state, &orig->rb_node, prealloc->end, 499 &prealloc->rb_node, NULL, NULL); 500 if (node) { 501 free_extent_state(prealloc); 502 return -EEXIST; 503 } 504 return 0; 505 } 506 507 static struct extent_state *next_state(struct extent_state *state) 508 { 509 struct rb_node *next = rb_next(&state->rb_node); 510 if (next) 511 return rb_entry(next, struct extent_state, rb_node); 512 else 513 return NULL; 514 } 515 516 /* 517 * utility function to clear some bits in an extent state struct. 518 * it will optionally wake up any one waiting on this state (wake == 1). 519 * 520 * If no bits are set on the state struct after clearing things, the 521 * struct is freed and removed from the tree 522 */ 523 static struct extent_state *clear_state_bit(struct extent_io_tree *tree, 524 struct extent_state *state, 525 unsigned *bits, int wake, 526 struct extent_changeset *changeset) 527 { 528 struct extent_state *next; 529 unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS; 530 531 if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) { 532 u64 range = state->end - state->start + 1; 533 WARN_ON(range > tree->dirty_bytes); 534 tree->dirty_bytes -= range; 535 } 536 clear_state_cb(tree, state, bits); 537 add_extent_changeset(state, bits_to_clear, changeset, 0); 538 state->state &= ~bits_to_clear; 539 if (wake) 540 wake_up(&state->wq); 541 if (state->state == 0) { 542 next = next_state(state); 543 if (extent_state_in_tree(state)) { 544 rb_erase(&state->rb_node, &tree->state); 545 RB_CLEAR_NODE(&state->rb_node); 546 free_extent_state(state); 547 } else { 548 WARN_ON(1); 549 } 550 } else { 551 merge_state(tree, state); 552 next = next_state(state); 553 } 554 return next; 555 } 556 557 static struct extent_state * 558 alloc_extent_state_atomic(struct extent_state *prealloc) 559 { 560 if (!prealloc) 561 prealloc = alloc_extent_state(GFP_ATOMIC); 562 563 return prealloc; 564 } 565 566 static void extent_io_tree_panic(struct extent_io_tree *tree, int err) 567 { 568 btrfs_panic(tree_fs_info(tree), err, 569 "Locking error: Extent tree was modified by another thread while locked."); 570 } 571 572 /* 573 * clear some bits on a range in the tree. This may require splitting 574 * or inserting elements in the tree, so the gfp mask is used to 575 * indicate which allocations or sleeping are allowed. 576 * 577 * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove 578 * the given range from the tree regardless of state (ie for truncate). 579 * 580 * the range [start, end] is inclusive. 581 * 582 * This takes the tree lock, and returns 0 on success and < 0 on error. 583 */ 584 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, 585 unsigned bits, int wake, int delete, 586 struct extent_state **cached_state, 587 gfp_t mask, struct extent_changeset *changeset) 588 { 589 struct extent_state *state; 590 struct extent_state *cached; 591 struct extent_state *prealloc = NULL; 592 struct rb_node *node; 593 u64 last_end; 594 int err; 595 int clear = 0; 596 597 btrfs_debug_check_extent_io_range(tree, start, end); 598 599 if (bits & EXTENT_DELALLOC) 600 bits |= EXTENT_NORESERVE; 601 602 if (delete) 603 bits |= ~EXTENT_CTLBITS; 604 bits |= EXTENT_FIRST_DELALLOC; 605 606 if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY)) 607 clear = 1; 608 again: 609 if (!prealloc && gfpflags_allow_blocking(mask)) { 610 /* 611 * Don't care for allocation failure here because we might end 612 * up not needing the pre-allocated extent state at all, which 613 * is the case if we only have in the tree extent states that 614 * cover our input range and don't cover too any other range. 615 * If we end up needing a new extent state we allocate it later. 616 */ 617 prealloc = alloc_extent_state(mask); 618 } 619 620 spin_lock(&tree->lock); 621 if (cached_state) { 622 cached = *cached_state; 623 624 if (clear) { 625 *cached_state = NULL; 626 cached_state = NULL; 627 } 628 629 if (cached && extent_state_in_tree(cached) && 630 cached->start <= start && cached->end > start) { 631 if (clear) 632 refcount_dec(&cached->refs); 633 state = cached; 634 goto hit_next; 635 } 636 if (clear) 637 free_extent_state(cached); 638 } 639 /* 640 * this search will find the extents that end after 641 * our range starts 642 */ 643 node = tree_search(tree, start); 644 if (!node) 645 goto out; 646 state = rb_entry(node, struct extent_state, rb_node); 647 hit_next: 648 if (state->start > end) 649 goto out; 650 WARN_ON(state->end < start); 651 last_end = state->end; 652 653 /* the state doesn't have the wanted bits, go ahead */ 654 if (!(state->state & bits)) { 655 state = next_state(state); 656 goto next; 657 } 658 659 /* 660 * | ---- desired range ---- | 661 * | state | or 662 * | ------------- state -------------- | 663 * 664 * We need to split the extent we found, and may flip 665 * bits on second half. 666 * 667 * If the extent we found extends past our range, we 668 * just split and search again. It'll get split again 669 * the next time though. 670 * 671 * If the extent we found is inside our range, we clear 672 * the desired bit on it. 673 */ 674 675 if (state->start < start) { 676 prealloc = alloc_extent_state_atomic(prealloc); 677 BUG_ON(!prealloc); 678 err = split_state(tree, state, prealloc, start); 679 if (err) 680 extent_io_tree_panic(tree, err); 681 682 prealloc = NULL; 683 if (err) 684 goto out; 685 if (state->end <= end) { 686 state = clear_state_bit(tree, state, &bits, wake, 687 changeset); 688 goto next; 689 } 690 goto search_again; 691 } 692 /* 693 * | ---- desired range ---- | 694 * | state | 695 * We need to split the extent, and clear the bit 696 * on the first half 697 */ 698 if (state->start <= end && state->end > end) { 699 prealloc = alloc_extent_state_atomic(prealloc); 700 BUG_ON(!prealloc); 701 err = split_state(tree, state, prealloc, end + 1); 702 if (err) 703 extent_io_tree_panic(tree, err); 704 705 if (wake) 706 wake_up(&state->wq); 707 708 clear_state_bit(tree, prealloc, &bits, wake, changeset); 709 710 prealloc = NULL; 711 goto out; 712 } 713 714 state = clear_state_bit(tree, state, &bits, wake, changeset); 715 next: 716 if (last_end == (u64)-1) 717 goto out; 718 start = last_end + 1; 719 if (start <= end && state && !need_resched()) 720 goto hit_next; 721 722 search_again: 723 if (start > end) 724 goto out; 725 spin_unlock(&tree->lock); 726 if (gfpflags_allow_blocking(mask)) 727 cond_resched(); 728 goto again; 729 730 out: 731 spin_unlock(&tree->lock); 732 if (prealloc) 733 free_extent_state(prealloc); 734 735 return 0; 736 737 } 738 739 static void wait_on_state(struct extent_io_tree *tree, 740 struct extent_state *state) 741 __releases(tree->lock) 742 __acquires(tree->lock) 743 { 744 DEFINE_WAIT(wait); 745 prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE); 746 spin_unlock(&tree->lock); 747 schedule(); 748 spin_lock(&tree->lock); 749 finish_wait(&state->wq, &wait); 750 } 751 752 /* 753 * waits for one or more bits to clear on a range in the state tree. 754 * The range [start, end] is inclusive. 755 * The tree lock is taken by this function 756 */ 757 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, 758 unsigned long bits) 759 { 760 struct extent_state *state; 761 struct rb_node *node; 762 763 btrfs_debug_check_extent_io_range(tree, start, end); 764 765 spin_lock(&tree->lock); 766 again: 767 while (1) { 768 /* 769 * this search will find all the extents that end after 770 * our range starts 771 */ 772 node = tree_search(tree, start); 773 process_node: 774 if (!node) 775 break; 776 777 state = rb_entry(node, struct extent_state, rb_node); 778 779 if (state->start > end) 780 goto out; 781 782 if (state->state & bits) { 783 start = state->start; 784 refcount_inc(&state->refs); 785 wait_on_state(tree, state); 786 free_extent_state(state); 787 goto again; 788 } 789 start = state->end + 1; 790 791 if (start > end) 792 break; 793 794 if (!cond_resched_lock(&tree->lock)) { 795 node = rb_next(node); 796 goto process_node; 797 } 798 } 799 out: 800 spin_unlock(&tree->lock); 801 } 802 803 static void set_state_bits(struct extent_io_tree *tree, 804 struct extent_state *state, 805 unsigned *bits, struct extent_changeset *changeset) 806 { 807 unsigned bits_to_set = *bits & ~EXTENT_CTLBITS; 808 809 set_state_cb(tree, state, bits); 810 if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) { 811 u64 range = state->end - state->start + 1; 812 tree->dirty_bytes += range; 813 } 814 add_extent_changeset(state, bits_to_set, changeset, 1); 815 state->state |= bits_to_set; 816 } 817 818 static void cache_state_if_flags(struct extent_state *state, 819 struct extent_state **cached_ptr, 820 unsigned flags) 821 { 822 if (cached_ptr && !(*cached_ptr)) { 823 if (!flags || (state->state & flags)) { 824 *cached_ptr = state; 825 refcount_inc(&state->refs); 826 } 827 } 828 } 829 830 static void cache_state(struct extent_state *state, 831 struct extent_state **cached_ptr) 832 { 833 return cache_state_if_flags(state, cached_ptr, 834 EXTENT_IOBITS | EXTENT_BOUNDARY); 835 } 836 837 /* 838 * set some bits on a range in the tree. This may require allocations or 839 * sleeping, so the gfp mask is used to indicate what is allowed. 840 * 841 * If any of the exclusive bits are set, this will fail with -EEXIST if some 842 * part of the range already has the desired bits set. The start of the 843 * existing range is returned in failed_start in this case. 844 * 845 * [start, end] is inclusive This takes the tree lock. 846 */ 847 848 static int __must_check 849 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, 850 unsigned bits, unsigned exclusive_bits, 851 u64 *failed_start, struct extent_state **cached_state, 852 gfp_t mask, struct extent_changeset *changeset) 853 { 854 struct extent_state *state; 855 struct extent_state *prealloc = NULL; 856 struct rb_node *node; 857 struct rb_node **p; 858 struct rb_node *parent; 859 int err = 0; 860 u64 last_start; 861 u64 last_end; 862 863 btrfs_debug_check_extent_io_range(tree, start, end); 864 865 bits |= EXTENT_FIRST_DELALLOC; 866 again: 867 if (!prealloc && gfpflags_allow_blocking(mask)) { 868 /* 869 * Don't care for allocation failure here because we might end 870 * up not needing the pre-allocated extent state at all, which 871 * is the case if we only have in the tree extent states that 872 * cover our input range and don't cover too any other range. 873 * If we end up needing a new extent state we allocate it later. 874 */ 875 prealloc = alloc_extent_state(mask); 876 } 877 878 spin_lock(&tree->lock); 879 if (cached_state && *cached_state) { 880 state = *cached_state; 881 if (state->start <= start && state->end > start && 882 extent_state_in_tree(state)) { 883 node = &state->rb_node; 884 goto hit_next; 885 } 886 } 887 /* 888 * this search will find all the extents that end after 889 * our range starts. 890 */ 891 node = tree_search_for_insert(tree, start, &p, &parent); 892 if (!node) { 893 prealloc = alloc_extent_state_atomic(prealloc); 894 BUG_ON(!prealloc); 895 err = insert_state(tree, prealloc, start, end, 896 &p, &parent, &bits, changeset); 897 if (err) 898 extent_io_tree_panic(tree, err); 899 900 cache_state(prealloc, cached_state); 901 prealloc = NULL; 902 goto out; 903 } 904 state = rb_entry(node, struct extent_state, rb_node); 905 hit_next: 906 last_start = state->start; 907 last_end = state->end; 908 909 /* 910 * | ---- desired range ---- | 911 * | state | 912 * 913 * Just lock what we found and keep going 914 */ 915 if (state->start == start && state->end <= end) { 916 if (state->state & exclusive_bits) { 917 *failed_start = state->start; 918 err = -EEXIST; 919 goto out; 920 } 921 922 set_state_bits(tree, state, &bits, changeset); 923 cache_state(state, cached_state); 924 merge_state(tree, state); 925 if (last_end == (u64)-1) 926 goto out; 927 start = last_end + 1; 928 state = next_state(state); 929 if (start < end && state && state->start == start && 930 !need_resched()) 931 goto hit_next; 932 goto search_again; 933 } 934 935 /* 936 * | ---- desired range ---- | 937 * | state | 938 * or 939 * | ------------- state -------------- | 940 * 941 * We need to split the extent we found, and may flip bits on 942 * second half. 943 * 944 * If the extent we found extends past our 945 * range, we just split and search again. It'll get split 946 * again the next time though. 947 * 948 * If the extent we found is inside our range, we set the 949 * desired bit on it. 950 */ 951 if (state->start < start) { 952 if (state->state & exclusive_bits) { 953 *failed_start = start; 954 err = -EEXIST; 955 goto out; 956 } 957 958 prealloc = alloc_extent_state_atomic(prealloc); 959 BUG_ON(!prealloc); 960 err = split_state(tree, state, prealloc, start); 961 if (err) 962 extent_io_tree_panic(tree, err); 963 964 prealloc = NULL; 965 if (err) 966 goto out; 967 if (state->end <= end) { 968 set_state_bits(tree, state, &bits, changeset); 969 cache_state(state, cached_state); 970 merge_state(tree, state); 971 if (last_end == (u64)-1) 972 goto out; 973 start = last_end + 1; 974 state = next_state(state); 975 if (start < end && state && state->start == start && 976 !need_resched()) 977 goto hit_next; 978 } 979 goto search_again; 980 } 981 /* 982 * | ---- desired range ---- | 983 * | state | or | state | 984 * 985 * There's a hole, we need to insert something in it and 986 * ignore the extent we found. 987 */ 988 if (state->start > start) { 989 u64 this_end; 990 if (end < last_start) 991 this_end = end; 992 else 993 this_end = last_start - 1; 994 995 prealloc = alloc_extent_state_atomic(prealloc); 996 BUG_ON(!prealloc); 997 998 /* 999 * Avoid to free 'prealloc' if it can be merged with 1000 * the later extent. 1001 */ 1002 err = insert_state(tree, prealloc, start, this_end, 1003 NULL, NULL, &bits, changeset); 1004 if (err) 1005 extent_io_tree_panic(tree, err); 1006 1007 cache_state(prealloc, cached_state); 1008 prealloc = NULL; 1009 start = this_end + 1; 1010 goto search_again; 1011 } 1012 /* 1013 * | ---- desired range ---- | 1014 * | state | 1015 * We need to split the extent, and set the bit 1016 * on the first half 1017 */ 1018 if (state->start <= end && state->end > end) { 1019 if (state->state & exclusive_bits) { 1020 *failed_start = start; 1021 err = -EEXIST; 1022 goto out; 1023 } 1024 1025 prealloc = alloc_extent_state_atomic(prealloc); 1026 BUG_ON(!prealloc); 1027 err = split_state(tree, state, prealloc, end + 1); 1028 if (err) 1029 extent_io_tree_panic(tree, err); 1030 1031 set_state_bits(tree, prealloc, &bits, changeset); 1032 cache_state(prealloc, cached_state); 1033 merge_state(tree, prealloc); 1034 prealloc = NULL; 1035 goto out; 1036 } 1037 1038 search_again: 1039 if (start > end) 1040 goto out; 1041 spin_unlock(&tree->lock); 1042 if (gfpflags_allow_blocking(mask)) 1043 cond_resched(); 1044 goto again; 1045 1046 out: 1047 spin_unlock(&tree->lock); 1048 if (prealloc) 1049 free_extent_state(prealloc); 1050 1051 return err; 1052 1053 } 1054 1055 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, 1056 unsigned bits, u64 * failed_start, 1057 struct extent_state **cached_state, gfp_t mask) 1058 { 1059 return __set_extent_bit(tree, start, end, bits, 0, failed_start, 1060 cached_state, mask, NULL); 1061 } 1062 1063 1064 /** 1065 * convert_extent_bit - convert all bits in a given range from one bit to 1066 * another 1067 * @tree: the io tree to search 1068 * @start: the start offset in bytes 1069 * @end: the end offset in bytes (inclusive) 1070 * @bits: the bits to set in this range 1071 * @clear_bits: the bits to clear in this range 1072 * @cached_state: state that we're going to cache 1073 * 1074 * This will go through and set bits for the given range. If any states exist 1075 * already in this range they are set with the given bit and cleared of the 1076 * clear_bits. This is only meant to be used by things that are mergeable, ie 1077 * converting from say DELALLOC to DIRTY. This is not meant to be used with 1078 * boundary bits like LOCK. 1079 * 1080 * All allocations are done with GFP_NOFS. 1081 */ 1082 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, 1083 unsigned bits, unsigned clear_bits, 1084 struct extent_state **cached_state) 1085 { 1086 struct extent_state *state; 1087 struct extent_state *prealloc = NULL; 1088 struct rb_node *node; 1089 struct rb_node **p; 1090 struct rb_node *parent; 1091 int err = 0; 1092 u64 last_start; 1093 u64 last_end; 1094 bool first_iteration = true; 1095 1096 btrfs_debug_check_extent_io_range(tree, start, end); 1097 1098 again: 1099 if (!prealloc) { 1100 /* 1101 * Best effort, don't worry if extent state allocation fails 1102 * here for the first iteration. We might have a cached state 1103 * that matches exactly the target range, in which case no 1104 * extent state allocations are needed. We'll only know this 1105 * after locking the tree. 1106 */ 1107 prealloc = alloc_extent_state(GFP_NOFS); 1108 if (!prealloc && !first_iteration) 1109 return -ENOMEM; 1110 } 1111 1112 spin_lock(&tree->lock); 1113 if (cached_state && *cached_state) { 1114 state = *cached_state; 1115 if (state->start <= start && state->end > start && 1116 extent_state_in_tree(state)) { 1117 node = &state->rb_node; 1118 goto hit_next; 1119 } 1120 } 1121 1122 /* 1123 * this search will find all the extents that end after 1124 * our range starts. 1125 */ 1126 node = tree_search_for_insert(tree, start, &p, &parent); 1127 if (!node) { 1128 prealloc = alloc_extent_state_atomic(prealloc); 1129 if (!prealloc) { 1130 err = -ENOMEM; 1131 goto out; 1132 } 1133 err = insert_state(tree, prealloc, start, end, 1134 &p, &parent, &bits, NULL); 1135 if (err) 1136 extent_io_tree_panic(tree, err); 1137 cache_state(prealloc, cached_state); 1138 prealloc = NULL; 1139 goto out; 1140 } 1141 state = rb_entry(node, struct extent_state, rb_node); 1142 hit_next: 1143 last_start = state->start; 1144 last_end = state->end; 1145 1146 /* 1147 * | ---- desired range ---- | 1148 * | state | 1149 * 1150 * Just lock what we found and keep going 1151 */ 1152 if (state->start == start && state->end <= end) { 1153 set_state_bits(tree, state, &bits, NULL); 1154 cache_state(state, cached_state); 1155 state = clear_state_bit(tree, state, &clear_bits, 0, NULL); 1156 if (last_end == (u64)-1) 1157 goto out; 1158 start = last_end + 1; 1159 if (start < end && state && state->start == start && 1160 !need_resched()) 1161 goto hit_next; 1162 goto search_again; 1163 } 1164 1165 /* 1166 * | ---- desired range ---- | 1167 * | state | 1168 * or 1169 * | ------------- state -------------- | 1170 * 1171 * We need to split the extent we found, and may flip bits on 1172 * second half. 1173 * 1174 * If the extent we found extends past our 1175 * range, we just split and search again. It'll get split 1176 * again the next time though. 1177 * 1178 * If the extent we found is inside our range, we set the 1179 * desired bit on it. 1180 */ 1181 if (state->start < start) { 1182 prealloc = alloc_extent_state_atomic(prealloc); 1183 if (!prealloc) { 1184 err = -ENOMEM; 1185 goto out; 1186 } 1187 err = split_state(tree, state, prealloc, start); 1188 if (err) 1189 extent_io_tree_panic(tree, err); 1190 prealloc = NULL; 1191 if (err) 1192 goto out; 1193 if (state->end <= end) { 1194 set_state_bits(tree, state, &bits, NULL); 1195 cache_state(state, cached_state); 1196 state = clear_state_bit(tree, state, &clear_bits, 0, 1197 NULL); 1198 if (last_end == (u64)-1) 1199 goto out; 1200 start = last_end + 1; 1201 if (start < end && state && state->start == start && 1202 !need_resched()) 1203 goto hit_next; 1204 } 1205 goto search_again; 1206 } 1207 /* 1208 * | ---- desired range ---- | 1209 * | state | or | state | 1210 * 1211 * There's a hole, we need to insert something in it and 1212 * ignore the extent we found. 1213 */ 1214 if (state->start > start) { 1215 u64 this_end; 1216 if (end < last_start) 1217 this_end = end; 1218 else 1219 this_end = last_start - 1; 1220 1221 prealloc = alloc_extent_state_atomic(prealloc); 1222 if (!prealloc) { 1223 err = -ENOMEM; 1224 goto out; 1225 } 1226 1227 /* 1228 * Avoid to free 'prealloc' if it can be merged with 1229 * the later extent. 1230 */ 1231 err = insert_state(tree, prealloc, start, this_end, 1232 NULL, NULL, &bits, NULL); 1233 if (err) 1234 extent_io_tree_panic(tree, err); 1235 cache_state(prealloc, cached_state); 1236 prealloc = NULL; 1237 start = this_end + 1; 1238 goto search_again; 1239 } 1240 /* 1241 * | ---- desired range ---- | 1242 * | state | 1243 * We need to split the extent, and set the bit 1244 * on the first half 1245 */ 1246 if (state->start <= end && state->end > end) { 1247 prealloc = alloc_extent_state_atomic(prealloc); 1248 if (!prealloc) { 1249 err = -ENOMEM; 1250 goto out; 1251 } 1252 1253 err = split_state(tree, state, prealloc, end + 1); 1254 if (err) 1255 extent_io_tree_panic(tree, err); 1256 1257 set_state_bits(tree, prealloc, &bits, NULL); 1258 cache_state(prealloc, cached_state); 1259 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL); 1260 prealloc = NULL; 1261 goto out; 1262 } 1263 1264 search_again: 1265 if (start > end) 1266 goto out; 1267 spin_unlock(&tree->lock); 1268 cond_resched(); 1269 first_iteration = false; 1270 goto again; 1271 1272 out: 1273 spin_unlock(&tree->lock); 1274 if (prealloc) 1275 free_extent_state(prealloc); 1276 1277 return err; 1278 } 1279 1280 /* wrappers around set/clear extent bit */ 1281 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end, 1282 unsigned bits, struct extent_changeset *changeset) 1283 { 1284 /* 1285 * We don't support EXTENT_LOCKED yet, as current changeset will 1286 * record any bits changed, so for EXTENT_LOCKED case, it will 1287 * either fail with -EEXIST or changeset will record the whole 1288 * range. 1289 */ 1290 BUG_ON(bits & EXTENT_LOCKED); 1291 1292 return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS, 1293 changeset); 1294 } 1295 1296 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, 1297 unsigned bits, int wake, int delete, 1298 struct extent_state **cached, gfp_t mask) 1299 { 1300 return __clear_extent_bit(tree, start, end, bits, wake, delete, 1301 cached, mask, NULL); 1302 } 1303 1304 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end, 1305 unsigned bits, struct extent_changeset *changeset) 1306 { 1307 /* 1308 * Don't support EXTENT_LOCKED case, same reason as 1309 * set_record_extent_bits(). 1310 */ 1311 BUG_ON(bits & EXTENT_LOCKED); 1312 1313 return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS, 1314 changeset); 1315 } 1316 1317 /* 1318 * either insert or lock state struct between start and end use mask to tell 1319 * us if waiting is desired. 1320 */ 1321 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end, 1322 struct extent_state **cached_state) 1323 { 1324 int err; 1325 u64 failed_start; 1326 1327 while (1) { 1328 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, 1329 EXTENT_LOCKED, &failed_start, 1330 cached_state, GFP_NOFS, NULL); 1331 if (err == -EEXIST) { 1332 wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED); 1333 start = failed_start; 1334 } else 1335 break; 1336 WARN_ON(start > end); 1337 } 1338 return err; 1339 } 1340 1341 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end) 1342 { 1343 int err; 1344 u64 failed_start; 1345 1346 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED, 1347 &failed_start, NULL, GFP_NOFS, NULL); 1348 if (err == -EEXIST) { 1349 if (failed_start > start) 1350 clear_extent_bit(tree, start, failed_start - 1, 1351 EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS); 1352 return 0; 1353 } 1354 return 1; 1355 } 1356 1357 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end) 1358 { 1359 unsigned long index = start >> PAGE_SHIFT; 1360 unsigned long end_index = end >> PAGE_SHIFT; 1361 struct page *page; 1362 1363 while (index <= end_index) { 1364 page = find_get_page(inode->i_mapping, index); 1365 BUG_ON(!page); /* Pages should be in the extent_io_tree */ 1366 clear_page_dirty_for_io(page); 1367 put_page(page); 1368 index++; 1369 } 1370 } 1371 1372 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end) 1373 { 1374 unsigned long index = start >> PAGE_SHIFT; 1375 unsigned long end_index = end >> PAGE_SHIFT; 1376 struct page *page; 1377 1378 while (index <= end_index) { 1379 page = find_get_page(inode->i_mapping, index); 1380 BUG_ON(!page); /* Pages should be in the extent_io_tree */ 1381 __set_page_dirty_nobuffers(page); 1382 account_page_redirty(page); 1383 put_page(page); 1384 index++; 1385 } 1386 } 1387 1388 /* 1389 * helper function to set both pages and extents in the tree writeback 1390 */ 1391 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end) 1392 { 1393 tree->ops->set_range_writeback(tree->private_data, start, end); 1394 } 1395 1396 /* find the first state struct with 'bits' set after 'start', and 1397 * return it. tree->lock must be held. NULL will returned if 1398 * nothing was found after 'start' 1399 */ 1400 static struct extent_state * 1401 find_first_extent_bit_state(struct extent_io_tree *tree, 1402 u64 start, unsigned bits) 1403 { 1404 struct rb_node *node; 1405 struct extent_state *state; 1406 1407 /* 1408 * this search will find all the extents that end after 1409 * our range starts. 1410 */ 1411 node = tree_search(tree, start); 1412 if (!node) 1413 goto out; 1414 1415 while (1) { 1416 state = rb_entry(node, struct extent_state, rb_node); 1417 if (state->end >= start && (state->state & bits)) 1418 return state; 1419 1420 node = rb_next(node); 1421 if (!node) 1422 break; 1423 } 1424 out: 1425 return NULL; 1426 } 1427 1428 /* 1429 * find the first offset in the io tree with 'bits' set. zero is 1430 * returned if we find something, and *start_ret and *end_ret are 1431 * set to reflect the state struct that was found. 1432 * 1433 * If nothing was found, 1 is returned. If found something, return 0. 1434 */ 1435 int find_first_extent_bit(struct extent_io_tree *tree, u64 start, 1436 u64 *start_ret, u64 *end_ret, unsigned bits, 1437 struct extent_state **cached_state) 1438 { 1439 struct extent_state *state; 1440 struct rb_node *n; 1441 int ret = 1; 1442 1443 spin_lock(&tree->lock); 1444 if (cached_state && *cached_state) { 1445 state = *cached_state; 1446 if (state->end == start - 1 && extent_state_in_tree(state)) { 1447 n = rb_next(&state->rb_node); 1448 while (n) { 1449 state = rb_entry(n, struct extent_state, 1450 rb_node); 1451 if (state->state & bits) 1452 goto got_it; 1453 n = rb_next(n); 1454 } 1455 free_extent_state(*cached_state); 1456 *cached_state = NULL; 1457 goto out; 1458 } 1459 free_extent_state(*cached_state); 1460 *cached_state = NULL; 1461 } 1462 1463 state = find_first_extent_bit_state(tree, start, bits); 1464 got_it: 1465 if (state) { 1466 cache_state_if_flags(state, cached_state, 0); 1467 *start_ret = state->start; 1468 *end_ret = state->end; 1469 ret = 0; 1470 } 1471 out: 1472 spin_unlock(&tree->lock); 1473 return ret; 1474 } 1475 1476 /* 1477 * find a contiguous range of bytes in the file marked as delalloc, not 1478 * more than 'max_bytes'. start and end are used to return the range, 1479 * 1480 * 1 is returned if we find something, 0 if nothing was in the tree 1481 */ 1482 static noinline u64 find_delalloc_range(struct extent_io_tree *tree, 1483 u64 *start, u64 *end, u64 max_bytes, 1484 struct extent_state **cached_state) 1485 { 1486 struct rb_node *node; 1487 struct extent_state *state; 1488 u64 cur_start = *start; 1489 u64 found = 0; 1490 u64 total_bytes = 0; 1491 1492 spin_lock(&tree->lock); 1493 1494 /* 1495 * this search will find all the extents that end after 1496 * our range starts. 1497 */ 1498 node = tree_search(tree, cur_start); 1499 if (!node) { 1500 if (!found) 1501 *end = (u64)-1; 1502 goto out; 1503 } 1504 1505 while (1) { 1506 state = rb_entry(node, struct extent_state, rb_node); 1507 if (found && (state->start != cur_start || 1508 (state->state & EXTENT_BOUNDARY))) { 1509 goto out; 1510 } 1511 if (!(state->state & EXTENT_DELALLOC)) { 1512 if (!found) 1513 *end = state->end; 1514 goto out; 1515 } 1516 if (!found) { 1517 *start = state->start; 1518 *cached_state = state; 1519 refcount_inc(&state->refs); 1520 } 1521 found++; 1522 *end = state->end; 1523 cur_start = state->end + 1; 1524 node = rb_next(node); 1525 total_bytes += state->end - state->start + 1; 1526 if (total_bytes >= max_bytes) 1527 break; 1528 if (!node) 1529 break; 1530 } 1531 out: 1532 spin_unlock(&tree->lock); 1533 return found; 1534 } 1535 1536 static int __process_pages_contig(struct address_space *mapping, 1537 struct page *locked_page, 1538 pgoff_t start_index, pgoff_t end_index, 1539 unsigned long page_ops, pgoff_t *index_ret); 1540 1541 static noinline void __unlock_for_delalloc(struct inode *inode, 1542 struct page *locked_page, 1543 u64 start, u64 end) 1544 { 1545 unsigned long index = start >> PAGE_SHIFT; 1546 unsigned long end_index = end >> PAGE_SHIFT; 1547 1548 ASSERT(locked_page); 1549 if (index == locked_page->index && end_index == index) 1550 return; 1551 1552 __process_pages_contig(inode->i_mapping, locked_page, index, end_index, 1553 PAGE_UNLOCK, NULL); 1554 } 1555 1556 static noinline int lock_delalloc_pages(struct inode *inode, 1557 struct page *locked_page, 1558 u64 delalloc_start, 1559 u64 delalloc_end) 1560 { 1561 unsigned long index = delalloc_start >> PAGE_SHIFT; 1562 unsigned long index_ret = index; 1563 unsigned long end_index = delalloc_end >> PAGE_SHIFT; 1564 int ret; 1565 1566 ASSERT(locked_page); 1567 if (index == locked_page->index && index == end_index) 1568 return 0; 1569 1570 ret = __process_pages_contig(inode->i_mapping, locked_page, index, 1571 end_index, PAGE_LOCK, &index_ret); 1572 if (ret == -EAGAIN) 1573 __unlock_for_delalloc(inode, locked_page, delalloc_start, 1574 (u64)index_ret << PAGE_SHIFT); 1575 return ret; 1576 } 1577 1578 /* 1579 * find a contiguous range of bytes in the file marked as delalloc, not 1580 * more than 'max_bytes'. start and end are used to return the range, 1581 * 1582 * 1 is returned if we find something, 0 if nothing was in the tree 1583 */ 1584 STATIC u64 find_lock_delalloc_range(struct inode *inode, 1585 struct extent_io_tree *tree, 1586 struct page *locked_page, u64 *start, 1587 u64 *end, u64 max_bytes) 1588 { 1589 u64 delalloc_start; 1590 u64 delalloc_end; 1591 u64 found; 1592 struct extent_state *cached_state = NULL; 1593 int ret; 1594 int loops = 0; 1595 1596 again: 1597 /* step one, find a bunch of delalloc bytes starting at start */ 1598 delalloc_start = *start; 1599 delalloc_end = 0; 1600 found = find_delalloc_range(tree, &delalloc_start, &delalloc_end, 1601 max_bytes, &cached_state); 1602 if (!found || delalloc_end <= *start) { 1603 *start = delalloc_start; 1604 *end = delalloc_end; 1605 free_extent_state(cached_state); 1606 return 0; 1607 } 1608 1609 /* 1610 * start comes from the offset of locked_page. We have to lock 1611 * pages in order, so we can't process delalloc bytes before 1612 * locked_page 1613 */ 1614 if (delalloc_start < *start) 1615 delalloc_start = *start; 1616 1617 /* 1618 * make sure to limit the number of pages we try to lock down 1619 */ 1620 if (delalloc_end + 1 - delalloc_start > max_bytes) 1621 delalloc_end = delalloc_start + max_bytes - 1; 1622 1623 /* step two, lock all the pages after the page that has start */ 1624 ret = lock_delalloc_pages(inode, locked_page, 1625 delalloc_start, delalloc_end); 1626 if (ret == -EAGAIN) { 1627 /* some of the pages are gone, lets avoid looping by 1628 * shortening the size of the delalloc range we're searching 1629 */ 1630 free_extent_state(cached_state); 1631 cached_state = NULL; 1632 if (!loops) { 1633 max_bytes = PAGE_SIZE; 1634 loops = 1; 1635 goto again; 1636 } else { 1637 found = 0; 1638 goto out_failed; 1639 } 1640 } 1641 BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */ 1642 1643 /* step three, lock the state bits for the whole range */ 1644 lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state); 1645 1646 /* then test to make sure it is all still delalloc */ 1647 ret = test_range_bit(tree, delalloc_start, delalloc_end, 1648 EXTENT_DELALLOC, 1, cached_state); 1649 if (!ret) { 1650 unlock_extent_cached(tree, delalloc_start, delalloc_end, 1651 &cached_state, GFP_NOFS); 1652 __unlock_for_delalloc(inode, locked_page, 1653 delalloc_start, delalloc_end); 1654 cond_resched(); 1655 goto again; 1656 } 1657 free_extent_state(cached_state); 1658 *start = delalloc_start; 1659 *end = delalloc_end; 1660 out_failed: 1661 return found; 1662 } 1663 1664 static int __process_pages_contig(struct address_space *mapping, 1665 struct page *locked_page, 1666 pgoff_t start_index, pgoff_t end_index, 1667 unsigned long page_ops, pgoff_t *index_ret) 1668 { 1669 unsigned long nr_pages = end_index - start_index + 1; 1670 unsigned long pages_locked = 0; 1671 pgoff_t index = start_index; 1672 struct page *pages[16]; 1673 unsigned ret; 1674 int err = 0; 1675 int i; 1676 1677 if (page_ops & PAGE_LOCK) { 1678 ASSERT(page_ops == PAGE_LOCK); 1679 ASSERT(index_ret && *index_ret == start_index); 1680 } 1681 1682 if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0) 1683 mapping_set_error(mapping, -EIO); 1684 1685 while (nr_pages > 0) { 1686 ret = find_get_pages_contig(mapping, index, 1687 min_t(unsigned long, 1688 nr_pages, ARRAY_SIZE(pages)), pages); 1689 if (ret == 0) { 1690 /* 1691 * Only if we're going to lock these pages, 1692 * can we find nothing at @index. 1693 */ 1694 ASSERT(page_ops & PAGE_LOCK); 1695 err = -EAGAIN; 1696 goto out; 1697 } 1698 1699 for (i = 0; i < ret; i++) { 1700 if (page_ops & PAGE_SET_PRIVATE2) 1701 SetPagePrivate2(pages[i]); 1702 1703 if (pages[i] == locked_page) { 1704 put_page(pages[i]); 1705 pages_locked++; 1706 continue; 1707 } 1708 if (page_ops & PAGE_CLEAR_DIRTY) 1709 clear_page_dirty_for_io(pages[i]); 1710 if (page_ops & PAGE_SET_WRITEBACK) 1711 set_page_writeback(pages[i]); 1712 if (page_ops & PAGE_SET_ERROR) 1713 SetPageError(pages[i]); 1714 if (page_ops & PAGE_END_WRITEBACK) 1715 end_page_writeback(pages[i]); 1716 if (page_ops & PAGE_UNLOCK) 1717 unlock_page(pages[i]); 1718 if (page_ops & PAGE_LOCK) { 1719 lock_page(pages[i]); 1720 if (!PageDirty(pages[i]) || 1721 pages[i]->mapping != mapping) { 1722 unlock_page(pages[i]); 1723 put_page(pages[i]); 1724 err = -EAGAIN; 1725 goto out; 1726 } 1727 } 1728 put_page(pages[i]); 1729 pages_locked++; 1730 } 1731 nr_pages -= ret; 1732 index += ret; 1733 cond_resched(); 1734 } 1735 out: 1736 if (err && index_ret) 1737 *index_ret = start_index + pages_locked - 1; 1738 return err; 1739 } 1740 1741 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end, 1742 u64 delalloc_end, struct page *locked_page, 1743 unsigned clear_bits, 1744 unsigned long page_ops) 1745 { 1746 clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0, 1747 NULL, GFP_NOFS); 1748 1749 __process_pages_contig(inode->i_mapping, locked_page, 1750 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 1751 page_ops, NULL); 1752 } 1753 1754 /* 1755 * count the number of bytes in the tree that have a given bit(s) 1756 * set. This can be fairly slow, except for EXTENT_DIRTY which is 1757 * cached. The total number found is returned. 1758 */ 1759 u64 count_range_bits(struct extent_io_tree *tree, 1760 u64 *start, u64 search_end, u64 max_bytes, 1761 unsigned bits, int contig) 1762 { 1763 struct rb_node *node; 1764 struct extent_state *state; 1765 u64 cur_start = *start; 1766 u64 total_bytes = 0; 1767 u64 last = 0; 1768 int found = 0; 1769 1770 if (WARN_ON(search_end <= cur_start)) 1771 return 0; 1772 1773 spin_lock(&tree->lock); 1774 if (cur_start == 0 && bits == EXTENT_DIRTY) { 1775 total_bytes = tree->dirty_bytes; 1776 goto out; 1777 } 1778 /* 1779 * this search will find all the extents that end after 1780 * our range starts. 1781 */ 1782 node = tree_search(tree, cur_start); 1783 if (!node) 1784 goto out; 1785 1786 while (1) { 1787 state = rb_entry(node, struct extent_state, rb_node); 1788 if (state->start > search_end) 1789 break; 1790 if (contig && found && state->start > last + 1) 1791 break; 1792 if (state->end >= cur_start && (state->state & bits) == bits) { 1793 total_bytes += min(search_end, state->end) + 1 - 1794 max(cur_start, state->start); 1795 if (total_bytes >= max_bytes) 1796 break; 1797 if (!found) { 1798 *start = max(cur_start, state->start); 1799 found = 1; 1800 } 1801 last = state->end; 1802 } else if (contig && found) { 1803 break; 1804 } 1805 node = rb_next(node); 1806 if (!node) 1807 break; 1808 } 1809 out: 1810 spin_unlock(&tree->lock); 1811 return total_bytes; 1812 } 1813 1814 /* 1815 * set the private field for a given byte offset in the tree. If there isn't 1816 * an extent_state there already, this does nothing. 1817 */ 1818 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start, 1819 struct io_failure_record *failrec) 1820 { 1821 struct rb_node *node; 1822 struct extent_state *state; 1823 int ret = 0; 1824 1825 spin_lock(&tree->lock); 1826 /* 1827 * this search will find all the extents that end after 1828 * our range starts. 1829 */ 1830 node = tree_search(tree, start); 1831 if (!node) { 1832 ret = -ENOENT; 1833 goto out; 1834 } 1835 state = rb_entry(node, struct extent_state, rb_node); 1836 if (state->start != start) { 1837 ret = -ENOENT; 1838 goto out; 1839 } 1840 state->failrec = failrec; 1841 out: 1842 spin_unlock(&tree->lock); 1843 return ret; 1844 } 1845 1846 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start, 1847 struct io_failure_record **failrec) 1848 { 1849 struct rb_node *node; 1850 struct extent_state *state; 1851 int ret = 0; 1852 1853 spin_lock(&tree->lock); 1854 /* 1855 * this search will find all the extents that end after 1856 * our range starts. 1857 */ 1858 node = tree_search(tree, start); 1859 if (!node) { 1860 ret = -ENOENT; 1861 goto out; 1862 } 1863 state = rb_entry(node, struct extent_state, rb_node); 1864 if (state->start != start) { 1865 ret = -ENOENT; 1866 goto out; 1867 } 1868 *failrec = state->failrec; 1869 out: 1870 spin_unlock(&tree->lock); 1871 return ret; 1872 } 1873 1874 /* 1875 * searches a range in the state tree for a given mask. 1876 * If 'filled' == 1, this returns 1 only if every extent in the tree 1877 * has the bits set. Otherwise, 1 is returned if any bit in the 1878 * range is found set. 1879 */ 1880 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end, 1881 unsigned bits, int filled, struct extent_state *cached) 1882 { 1883 struct extent_state *state = NULL; 1884 struct rb_node *node; 1885 int bitset = 0; 1886 1887 spin_lock(&tree->lock); 1888 if (cached && extent_state_in_tree(cached) && cached->start <= start && 1889 cached->end > start) 1890 node = &cached->rb_node; 1891 else 1892 node = tree_search(tree, start); 1893 while (node && start <= end) { 1894 state = rb_entry(node, struct extent_state, rb_node); 1895 1896 if (filled && state->start > start) { 1897 bitset = 0; 1898 break; 1899 } 1900 1901 if (state->start > end) 1902 break; 1903 1904 if (state->state & bits) { 1905 bitset = 1; 1906 if (!filled) 1907 break; 1908 } else if (filled) { 1909 bitset = 0; 1910 break; 1911 } 1912 1913 if (state->end == (u64)-1) 1914 break; 1915 1916 start = state->end + 1; 1917 if (start > end) 1918 break; 1919 node = rb_next(node); 1920 if (!node) { 1921 if (filled) 1922 bitset = 0; 1923 break; 1924 } 1925 } 1926 spin_unlock(&tree->lock); 1927 return bitset; 1928 } 1929 1930 /* 1931 * helper function to set a given page up to date if all the 1932 * extents in the tree for that page are up to date 1933 */ 1934 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page) 1935 { 1936 u64 start = page_offset(page); 1937 u64 end = start + PAGE_SIZE - 1; 1938 if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL)) 1939 SetPageUptodate(page); 1940 } 1941 1942 int free_io_failure(struct extent_io_tree *failure_tree, 1943 struct extent_io_tree *io_tree, 1944 struct io_failure_record *rec) 1945 { 1946 int ret; 1947 int err = 0; 1948 1949 set_state_failrec(failure_tree, rec->start, NULL); 1950 ret = clear_extent_bits(failure_tree, rec->start, 1951 rec->start + rec->len - 1, 1952 EXTENT_LOCKED | EXTENT_DIRTY); 1953 if (ret) 1954 err = ret; 1955 1956 ret = clear_extent_bits(io_tree, rec->start, 1957 rec->start + rec->len - 1, 1958 EXTENT_DAMAGED); 1959 if (ret && !err) 1960 err = ret; 1961 1962 kfree(rec); 1963 return err; 1964 } 1965 1966 /* 1967 * this bypasses the standard btrfs submit functions deliberately, as 1968 * the standard behavior is to write all copies in a raid setup. here we only 1969 * want to write the one bad copy. so we do the mapping for ourselves and issue 1970 * submit_bio directly. 1971 * to avoid any synchronization issues, wait for the data after writing, which 1972 * actually prevents the read that triggered the error from finishing. 1973 * currently, there can be no more than two copies of every data bit. thus, 1974 * exactly one rewrite is required. 1975 */ 1976 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start, 1977 u64 length, u64 logical, struct page *page, 1978 unsigned int pg_offset, int mirror_num) 1979 { 1980 struct bio *bio; 1981 struct btrfs_device *dev; 1982 u64 map_length = 0; 1983 u64 sector; 1984 struct btrfs_bio *bbio = NULL; 1985 int ret; 1986 1987 ASSERT(!(fs_info->sb->s_flags & MS_RDONLY)); 1988 BUG_ON(!mirror_num); 1989 1990 bio = btrfs_io_bio_alloc(1); 1991 bio->bi_iter.bi_size = 0; 1992 map_length = length; 1993 1994 /* 1995 * Avoid races with device replace and make sure our bbio has devices 1996 * associated to its stripes that don't go away while we are doing the 1997 * read repair operation. 1998 */ 1999 btrfs_bio_counter_inc_blocked(fs_info); 2000 if (btrfs_is_parity_mirror(fs_info, logical, length)) { 2001 /* 2002 * Note that we don't use BTRFS_MAP_WRITE because it's supposed 2003 * to update all raid stripes, but here we just want to correct 2004 * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad 2005 * stripe's dev and sector. 2006 */ 2007 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical, 2008 &map_length, &bbio, 0); 2009 if (ret) { 2010 btrfs_bio_counter_dec(fs_info); 2011 bio_put(bio); 2012 return -EIO; 2013 } 2014 ASSERT(bbio->mirror_num == 1); 2015 } else { 2016 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical, 2017 &map_length, &bbio, mirror_num); 2018 if (ret) { 2019 btrfs_bio_counter_dec(fs_info); 2020 bio_put(bio); 2021 return -EIO; 2022 } 2023 BUG_ON(mirror_num != bbio->mirror_num); 2024 } 2025 2026 sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9; 2027 bio->bi_iter.bi_sector = sector; 2028 dev = bbio->stripes[bbio->mirror_num - 1].dev; 2029 btrfs_put_bbio(bbio); 2030 if (!dev || !dev->bdev || !dev->writeable) { 2031 btrfs_bio_counter_dec(fs_info); 2032 bio_put(bio); 2033 return -EIO; 2034 } 2035 bio_set_dev(bio, dev->bdev); 2036 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC; 2037 bio_add_page(bio, page, length, pg_offset); 2038 2039 if (btrfsic_submit_bio_wait(bio)) { 2040 /* try to remap that extent elsewhere? */ 2041 btrfs_bio_counter_dec(fs_info); 2042 bio_put(bio); 2043 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS); 2044 return -EIO; 2045 } 2046 2047 btrfs_info_rl_in_rcu(fs_info, 2048 "read error corrected: ino %llu off %llu (dev %s sector %llu)", 2049 ino, start, 2050 rcu_str_deref(dev->name), sector); 2051 btrfs_bio_counter_dec(fs_info); 2052 bio_put(bio); 2053 return 0; 2054 } 2055 2056 int repair_eb_io_failure(struct btrfs_fs_info *fs_info, 2057 struct extent_buffer *eb, int mirror_num) 2058 { 2059 u64 start = eb->start; 2060 unsigned long i, num_pages = num_extent_pages(eb->start, eb->len); 2061 int ret = 0; 2062 2063 if (sb_rdonly(fs_info->sb)) 2064 return -EROFS; 2065 2066 for (i = 0; i < num_pages; i++) { 2067 struct page *p = eb->pages[i]; 2068 2069 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p, 2070 start - page_offset(p), mirror_num); 2071 if (ret) 2072 break; 2073 start += PAGE_SIZE; 2074 } 2075 2076 return ret; 2077 } 2078 2079 /* 2080 * each time an IO finishes, we do a fast check in the IO failure tree 2081 * to see if we need to process or clean up an io_failure_record 2082 */ 2083 int clean_io_failure(struct btrfs_fs_info *fs_info, 2084 struct extent_io_tree *failure_tree, 2085 struct extent_io_tree *io_tree, u64 start, 2086 struct page *page, u64 ino, unsigned int pg_offset) 2087 { 2088 u64 private; 2089 struct io_failure_record *failrec; 2090 struct extent_state *state; 2091 int num_copies; 2092 int ret; 2093 2094 private = 0; 2095 ret = count_range_bits(failure_tree, &private, (u64)-1, 1, 2096 EXTENT_DIRTY, 0); 2097 if (!ret) 2098 return 0; 2099 2100 ret = get_state_failrec(failure_tree, start, &failrec); 2101 if (ret) 2102 return 0; 2103 2104 BUG_ON(!failrec->this_mirror); 2105 2106 if (failrec->in_validation) { 2107 /* there was no real error, just free the record */ 2108 btrfs_debug(fs_info, 2109 "clean_io_failure: freeing dummy error at %llu", 2110 failrec->start); 2111 goto out; 2112 } 2113 if (sb_rdonly(fs_info->sb)) 2114 goto out; 2115 2116 spin_lock(&io_tree->lock); 2117 state = find_first_extent_bit_state(io_tree, 2118 failrec->start, 2119 EXTENT_LOCKED); 2120 spin_unlock(&io_tree->lock); 2121 2122 if (state && state->start <= failrec->start && 2123 state->end >= failrec->start + failrec->len - 1) { 2124 num_copies = btrfs_num_copies(fs_info, failrec->logical, 2125 failrec->len); 2126 if (num_copies > 1) { 2127 repair_io_failure(fs_info, ino, start, failrec->len, 2128 failrec->logical, page, pg_offset, 2129 failrec->failed_mirror); 2130 } 2131 } 2132 2133 out: 2134 free_io_failure(failure_tree, io_tree, failrec); 2135 2136 return 0; 2137 } 2138 2139 /* 2140 * Can be called when 2141 * - hold extent lock 2142 * - under ordered extent 2143 * - the inode is freeing 2144 */ 2145 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end) 2146 { 2147 struct extent_io_tree *failure_tree = &inode->io_failure_tree; 2148 struct io_failure_record *failrec; 2149 struct extent_state *state, *next; 2150 2151 if (RB_EMPTY_ROOT(&failure_tree->state)) 2152 return; 2153 2154 spin_lock(&failure_tree->lock); 2155 state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY); 2156 while (state) { 2157 if (state->start > end) 2158 break; 2159 2160 ASSERT(state->end <= end); 2161 2162 next = next_state(state); 2163 2164 failrec = state->failrec; 2165 free_extent_state(state); 2166 kfree(failrec); 2167 2168 state = next; 2169 } 2170 spin_unlock(&failure_tree->lock); 2171 } 2172 2173 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end, 2174 struct io_failure_record **failrec_ret) 2175 { 2176 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 2177 struct io_failure_record *failrec; 2178 struct extent_map *em; 2179 struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree; 2180 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree; 2181 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 2182 int ret; 2183 u64 logical; 2184 2185 ret = get_state_failrec(failure_tree, start, &failrec); 2186 if (ret) { 2187 failrec = kzalloc(sizeof(*failrec), GFP_NOFS); 2188 if (!failrec) 2189 return -ENOMEM; 2190 2191 failrec->start = start; 2192 failrec->len = end - start + 1; 2193 failrec->this_mirror = 0; 2194 failrec->bio_flags = 0; 2195 failrec->in_validation = 0; 2196 2197 read_lock(&em_tree->lock); 2198 em = lookup_extent_mapping(em_tree, start, failrec->len); 2199 if (!em) { 2200 read_unlock(&em_tree->lock); 2201 kfree(failrec); 2202 return -EIO; 2203 } 2204 2205 if (em->start > start || em->start + em->len <= start) { 2206 free_extent_map(em); 2207 em = NULL; 2208 } 2209 read_unlock(&em_tree->lock); 2210 if (!em) { 2211 kfree(failrec); 2212 return -EIO; 2213 } 2214 2215 logical = start - em->start; 2216 logical = em->block_start + logical; 2217 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) { 2218 logical = em->block_start; 2219 failrec->bio_flags = EXTENT_BIO_COMPRESSED; 2220 extent_set_compress_type(&failrec->bio_flags, 2221 em->compress_type); 2222 } 2223 2224 btrfs_debug(fs_info, 2225 "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu", 2226 logical, start, failrec->len); 2227 2228 failrec->logical = logical; 2229 free_extent_map(em); 2230 2231 /* set the bits in the private failure tree */ 2232 ret = set_extent_bits(failure_tree, start, end, 2233 EXTENT_LOCKED | EXTENT_DIRTY); 2234 if (ret >= 0) 2235 ret = set_state_failrec(failure_tree, start, failrec); 2236 /* set the bits in the inode's tree */ 2237 if (ret >= 0) 2238 ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED); 2239 if (ret < 0) { 2240 kfree(failrec); 2241 return ret; 2242 } 2243 } else { 2244 btrfs_debug(fs_info, 2245 "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d", 2246 failrec->logical, failrec->start, failrec->len, 2247 failrec->in_validation); 2248 /* 2249 * when data can be on disk more than twice, add to failrec here 2250 * (e.g. with a list for failed_mirror) to make 2251 * clean_io_failure() clean all those errors at once. 2252 */ 2253 } 2254 2255 *failrec_ret = failrec; 2256 2257 return 0; 2258 } 2259 2260 bool btrfs_check_repairable(struct inode *inode, struct bio *failed_bio, 2261 struct io_failure_record *failrec, int failed_mirror) 2262 { 2263 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 2264 int num_copies; 2265 2266 num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len); 2267 if (num_copies == 1) { 2268 /* 2269 * we only have a single copy of the data, so don't bother with 2270 * all the retry and error correction code that follows. no 2271 * matter what the error is, it is very likely to persist. 2272 */ 2273 btrfs_debug(fs_info, 2274 "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d", 2275 num_copies, failrec->this_mirror, failed_mirror); 2276 return false; 2277 } 2278 2279 /* 2280 * there are two premises: 2281 * a) deliver good data to the caller 2282 * b) correct the bad sectors on disk 2283 */ 2284 if (failed_bio->bi_vcnt > 1) { 2285 /* 2286 * to fulfill b), we need to know the exact failing sectors, as 2287 * we don't want to rewrite any more than the failed ones. thus, 2288 * we need separate read requests for the failed bio 2289 * 2290 * if the following BUG_ON triggers, our validation request got 2291 * merged. we need separate requests for our algorithm to work. 2292 */ 2293 BUG_ON(failrec->in_validation); 2294 failrec->in_validation = 1; 2295 failrec->this_mirror = failed_mirror; 2296 } else { 2297 /* 2298 * we're ready to fulfill a) and b) alongside. get a good copy 2299 * of the failed sector and if we succeed, we have setup 2300 * everything for repair_io_failure to do the rest for us. 2301 */ 2302 if (failrec->in_validation) { 2303 BUG_ON(failrec->this_mirror != failed_mirror); 2304 failrec->in_validation = 0; 2305 failrec->this_mirror = 0; 2306 } 2307 failrec->failed_mirror = failed_mirror; 2308 failrec->this_mirror++; 2309 if (failrec->this_mirror == failed_mirror) 2310 failrec->this_mirror++; 2311 } 2312 2313 if (failrec->this_mirror > num_copies) { 2314 btrfs_debug(fs_info, 2315 "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d", 2316 num_copies, failrec->this_mirror, failed_mirror); 2317 return false; 2318 } 2319 2320 return true; 2321 } 2322 2323 2324 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio, 2325 struct io_failure_record *failrec, 2326 struct page *page, int pg_offset, int icsum, 2327 bio_end_io_t *endio_func, void *data) 2328 { 2329 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 2330 struct bio *bio; 2331 struct btrfs_io_bio *btrfs_failed_bio; 2332 struct btrfs_io_bio *btrfs_bio; 2333 2334 bio = btrfs_io_bio_alloc(1); 2335 bio->bi_end_io = endio_func; 2336 bio->bi_iter.bi_sector = failrec->logical >> 9; 2337 bio_set_dev(bio, fs_info->fs_devices->latest_bdev); 2338 bio->bi_iter.bi_size = 0; 2339 bio->bi_private = data; 2340 2341 btrfs_failed_bio = btrfs_io_bio(failed_bio); 2342 if (btrfs_failed_bio->csum) { 2343 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy); 2344 2345 btrfs_bio = btrfs_io_bio(bio); 2346 btrfs_bio->csum = btrfs_bio->csum_inline; 2347 icsum *= csum_size; 2348 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum, 2349 csum_size); 2350 } 2351 2352 bio_add_page(bio, page, failrec->len, pg_offset); 2353 2354 return bio; 2355 } 2356 2357 /* 2358 * this is a generic handler for readpage errors (default 2359 * readpage_io_failed_hook). if other copies exist, read those and write back 2360 * good data to the failed position. does not investigate in remapping the 2361 * failed extent elsewhere, hoping the device will be smart enough to do this as 2362 * needed 2363 */ 2364 2365 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset, 2366 struct page *page, u64 start, u64 end, 2367 int failed_mirror) 2368 { 2369 struct io_failure_record *failrec; 2370 struct inode *inode = page->mapping->host; 2371 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree; 2372 struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree; 2373 struct bio *bio; 2374 int read_mode = 0; 2375 blk_status_t status; 2376 int ret; 2377 2378 BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE); 2379 2380 ret = btrfs_get_io_failure_record(inode, start, end, &failrec); 2381 if (ret) 2382 return ret; 2383 2384 if (!btrfs_check_repairable(inode, failed_bio, failrec, 2385 failed_mirror)) { 2386 free_io_failure(failure_tree, tree, failrec); 2387 return -EIO; 2388 } 2389 2390 if (failed_bio->bi_vcnt > 1) 2391 read_mode |= REQ_FAILFAST_DEV; 2392 2393 phy_offset >>= inode->i_sb->s_blocksize_bits; 2394 bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page, 2395 start - page_offset(page), 2396 (int)phy_offset, failed_bio->bi_end_io, 2397 NULL); 2398 bio_set_op_attrs(bio, REQ_OP_READ, read_mode); 2399 2400 btrfs_debug(btrfs_sb(inode->i_sb), 2401 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d", 2402 read_mode, failrec->this_mirror, failrec->in_validation); 2403 2404 status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror, 2405 failrec->bio_flags, 0); 2406 if (status) { 2407 free_io_failure(failure_tree, tree, failrec); 2408 bio_put(bio); 2409 ret = blk_status_to_errno(status); 2410 } 2411 2412 return ret; 2413 } 2414 2415 /* lots and lots of room for performance fixes in the end_bio funcs */ 2416 2417 void end_extent_writepage(struct page *page, int err, u64 start, u64 end) 2418 { 2419 int uptodate = (err == 0); 2420 struct extent_io_tree *tree; 2421 int ret = 0; 2422 2423 tree = &BTRFS_I(page->mapping->host)->io_tree; 2424 2425 if (tree->ops && tree->ops->writepage_end_io_hook) 2426 tree->ops->writepage_end_io_hook(page, start, end, NULL, 2427 uptodate); 2428 2429 if (!uptodate) { 2430 ClearPageUptodate(page); 2431 SetPageError(page); 2432 ret = err < 0 ? err : -EIO; 2433 mapping_set_error(page->mapping, ret); 2434 } 2435 } 2436 2437 /* 2438 * after a writepage IO is done, we need to: 2439 * clear the uptodate bits on error 2440 * clear the writeback bits in the extent tree for this IO 2441 * end_page_writeback if the page has no more pending IO 2442 * 2443 * Scheduling is not allowed, so the extent state tree is expected 2444 * to have one and only one object corresponding to this IO. 2445 */ 2446 static void end_bio_extent_writepage(struct bio *bio) 2447 { 2448 int error = blk_status_to_errno(bio->bi_status); 2449 struct bio_vec *bvec; 2450 u64 start; 2451 u64 end; 2452 int i; 2453 2454 ASSERT(!bio_flagged(bio, BIO_CLONED)); 2455 bio_for_each_segment_all(bvec, bio, i) { 2456 struct page *page = bvec->bv_page; 2457 struct inode *inode = page->mapping->host; 2458 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 2459 2460 /* We always issue full-page reads, but if some block 2461 * in a page fails to read, blk_update_request() will 2462 * advance bv_offset and adjust bv_len to compensate. 2463 * Print a warning for nonzero offsets, and an error 2464 * if they don't add up to a full page. */ 2465 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) { 2466 if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE) 2467 btrfs_err(fs_info, 2468 "partial page write in btrfs with offset %u and length %u", 2469 bvec->bv_offset, bvec->bv_len); 2470 else 2471 btrfs_info(fs_info, 2472 "incomplete page write in btrfs with offset %u and length %u", 2473 bvec->bv_offset, bvec->bv_len); 2474 } 2475 2476 start = page_offset(page); 2477 end = start + bvec->bv_offset + bvec->bv_len - 1; 2478 2479 end_extent_writepage(page, error, start, end); 2480 end_page_writeback(page); 2481 } 2482 2483 bio_put(bio); 2484 } 2485 2486 static void 2487 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len, 2488 int uptodate) 2489 { 2490 struct extent_state *cached = NULL; 2491 u64 end = start + len - 1; 2492 2493 if (uptodate && tree->track_uptodate) 2494 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC); 2495 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC); 2496 } 2497 2498 /* 2499 * after a readpage IO is done, we need to: 2500 * clear the uptodate bits on error 2501 * set the uptodate bits if things worked 2502 * set the page up to date if all extents in the tree are uptodate 2503 * clear the lock bit in the extent tree 2504 * unlock the page if there are no other extents locked for it 2505 * 2506 * Scheduling is not allowed, so the extent state tree is expected 2507 * to have one and only one object corresponding to this IO. 2508 */ 2509 static void end_bio_extent_readpage(struct bio *bio) 2510 { 2511 struct bio_vec *bvec; 2512 int uptodate = !bio->bi_status; 2513 struct btrfs_io_bio *io_bio = btrfs_io_bio(bio); 2514 struct extent_io_tree *tree, *failure_tree; 2515 u64 offset = 0; 2516 u64 start; 2517 u64 end; 2518 u64 len; 2519 u64 extent_start = 0; 2520 u64 extent_len = 0; 2521 int mirror; 2522 int ret; 2523 int i; 2524 2525 ASSERT(!bio_flagged(bio, BIO_CLONED)); 2526 bio_for_each_segment_all(bvec, bio, i) { 2527 struct page *page = bvec->bv_page; 2528 struct inode *inode = page->mapping->host; 2529 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 2530 2531 btrfs_debug(fs_info, 2532 "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u", 2533 (u64)bio->bi_iter.bi_sector, bio->bi_status, 2534 io_bio->mirror_num); 2535 tree = &BTRFS_I(inode)->io_tree; 2536 failure_tree = &BTRFS_I(inode)->io_failure_tree; 2537 2538 /* We always issue full-page reads, but if some block 2539 * in a page fails to read, blk_update_request() will 2540 * advance bv_offset and adjust bv_len to compensate. 2541 * Print a warning for nonzero offsets, and an error 2542 * if they don't add up to a full page. */ 2543 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) { 2544 if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE) 2545 btrfs_err(fs_info, 2546 "partial page read in btrfs with offset %u and length %u", 2547 bvec->bv_offset, bvec->bv_len); 2548 else 2549 btrfs_info(fs_info, 2550 "incomplete page read in btrfs with offset %u and length %u", 2551 bvec->bv_offset, bvec->bv_len); 2552 } 2553 2554 start = page_offset(page); 2555 end = start + bvec->bv_offset + bvec->bv_len - 1; 2556 len = bvec->bv_len; 2557 2558 mirror = io_bio->mirror_num; 2559 if (likely(uptodate && tree->ops)) { 2560 ret = tree->ops->readpage_end_io_hook(io_bio, offset, 2561 page, start, end, 2562 mirror); 2563 if (ret) 2564 uptodate = 0; 2565 else 2566 clean_io_failure(BTRFS_I(inode)->root->fs_info, 2567 failure_tree, tree, start, 2568 page, 2569 btrfs_ino(BTRFS_I(inode)), 0); 2570 } 2571 2572 if (likely(uptodate)) 2573 goto readpage_ok; 2574 2575 if (tree->ops) { 2576 ret = tree->ops->readpage_io_failed_hook(page, mirror); 2577 if (ret == -EAGAIN) { 2578 /* 2579 * Data inode's readpage_io_failed_hook() always 2580 * returns -EAGAIN. 2581 * 2582 * The generic bio_readpage_error handles errors 2583 * the following way: If possible, new read 2584 * requests are created and submitted and will 2585 * end up in end_bio_extent_readpage as well (if 2586 * we're lucky, not in the !uptodate case). In 2587 * that case it returns 0 and we just go on with 2588 * the next page in our bio. If it can't handle 2589 * the error it will return -EIO and we remain 2590 * responsible for that page. 2591 */ 2592 ret = bio_readpage_error(bio, offset, page, 2593 start, end, mirror); 2594 if (ret == 0) { 2595 uptodate = !bio->bi_status; 2596 offset += len; 2597 continue; 2598 } 2599 } 2600 2601 /* 2602 * metadata's readpage_io_failed_hook() always returns 2603 * -EIO and fixes nothing. -EIO is also returned if 2604 * data inode error could not be fixed. 2605 */ 2606 ASSERT(ret == -EIO); 2607 } 2608 readpage_ok: 2609 if (likely(uptodate)) { 2610 loff_t i_size = i_size_read(inode); 2611 pgoff_t end_index = i_size >> PAGE_SHIFT; 2612 unsigned off; 2613 2614 /* Zero out the end if this page straddles i_size */ 2615 off = i_size & (PAGE_SIZE-1); 2616 if (page->index == end_index && off) 2617 zero_user_segment(page, off, PAGE_SIZE); 2618 SetPageUptodate(page); 2619 } else { 2620 ClearPageUptodate(page); 2621 SetPageError(page); 2622 } 2623 unlock_page(page); 2624 offset += len; 2625 2626 if (unlikely(!uptodate)) { 2627 if (extent_len) { 2628 endio_readpage_release_extent(tree, 2629 extent_start, 2630 extent_len, 1); 2631 extent_start = 0; 2632 extent_len = 0; 2633 } 2634 endio_readpage_release_extent(tree, start, 2635 end - start + 1, 0); 2636 } else if (!extent_len) { 2637 extent_start = start; 2638 extent_len = end + 1 - start; 2639 } else if (extent_start + extent_len == start) { 2640 extent_len += end + 1 - start; 2641 } else { 2642 endio_readpage_release_extent(tree, extent_start, 2643 extent_len, uptodate); 2644 extent_start = start; 2645 extent_len = end + 1 - start; 2646 } 2647 } 2648 2649 if (extent_len) 2650 endio_readpage_release_extent(tree, extent_start, extent_len, 2651 uptodate); 2652 if (io_bio->end_io) 2653 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status)); 2654 bio_put(bio); 2655 } 2656 2657 /* 2658 * Initialize the members up to but not including 'bio'. Use after allocating a 2659 * new bio by bio_alloc_bioset as it does not initialize the bytes outside of 2660 * 'bio' because use of __GFP_ZERO is not supported. 2661 */ 2662 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio) 2663 { 2664 memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio)); 2665 } 2666 2667 /* 2668 * The following helpers allocate a bio. As it's backed by a bioset, it'll 2669 * never fail. We're returning a bio right now but you can call btrfs_io_bio 2670 * for the appropriate container_of magic 2671 */ 2672 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte) 2673 { 2674 struct bio *bio; 2675 2676 bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset); 2677 bio_set_dev(bio, bdev); 2678 bio->bi_iter.bi_sector = first_byte >> 9; 2679 btrfs_io_bio_init(btrfs_io_bio(bio)); 2680 return bio; 2681 } 2682 2683 struct bio *btrfs_bio_clone(struct bio *bio) 2684 { 2685 struct btrfs_io_bio *btrfs_bio; 2686 struct bio *new; 2687 2688 /* Bio allocation backed by a bioset does not fail */ 2689 new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset); 2690 btrfs_bio = btrfs_io_bio(new); 2691 btrfs_io_bio_init(btrfs_bio); 2692 btrfs_bio->iter = bio->bi_iter; 2693 return new; 2694 } 2695 2696 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs) 2697 { 2698 struct bio *bio; 2699 2700 /* Bio allocation backed by a bioset does not fail */ 2701 bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset); 2702 btrfs_io_bio_init(btrfs_io_bio(bio)); 2703 return bio; 2704 } 2705 2706 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size) 2707 { 2708 struct bio *bio; 2709 struct btrfs_io_bio *btrfs_bio; 2710 2711 /* this will never fail when it's backed by a bioset */ 2712 bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset); 2713 ASSERT(bio); 2714 2715 btrfs_bio = btrfs_io_bio(bio); 2716 btrfs_io_bio_init(btrfs_bio); 2717 2718 bio_trim(bio, offset >> 9, size >> 9); 2719 btrfs_bio->iter = bio->bi_iter; 2720 return bio; 2721 } 2722 2723 static int __must_check submit_one_bio(struct bio *bio, int mirror_num, 2724 unsigned long bio_flags) 2725 { 2726 blk_status_t ret = 0; 2727 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1; 2728 struct page *page = bvec->bv_page; 2729 struct extent_io_tree *tree = bio->bi_private; 2730 u64 start; 2731 2732 start = page_offset(page) + bvec->bv_offset; 2733 2734 bio->bi_private = NULL; 2735 bio_get(bio); 2736 2737 if (tree->ops) 2738 ret = tree->ops->submit_bio_hook(tree->private_data, bio, 2739 mirror_num, bio_flags, start); 2740 else 2741 btrfsic_submit_bio(bio); 2742 2743 bio_put(bio); 2744 return blk_status_to_errno(ret); 2745 } 2746 2747 static int merge_bio(struct extent_io_tree *tree, struct page *page, 2748 unsigned long offset, size_t size, struct bio *bio, 2749 unsigned long bio_flags) 2750 { 2751 int ret = 0; 2752 if (tree->ops) 2753 ret = tree->ops->merge_bio_hook(page, offset, size, bio, 2754 bio_flags); 2755 return ret; 2756 2757 } 2758 2759 /* 2760 * @opf: bio REQ_OP_* and REQ_* flags as one value 2761 */ 2762 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree, 2763 struct writeback_control *wbc, 2764 struct page *page, sector_t sector, 2765 size_t size, unsigned long offset, 2766 struct block_device *bdev, 2767 struct bio **bio_ret, 2768 bio_end_io_t end_io_func, 2769 int mirror_num, 2770 unsigned long prev_bio_flags, 2771 unsigned long bio_flags, 2772 bool force_bio_submit) 2773 { 2774 int ret = 0; 2775 struct bio *bio; 2776 int contig = 0; 2777 int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED; 2778 size_t page_size = min_t(size_t, size, PAGE_SIZE); 2779 2780 if (bio_ret && *bio_ret) { 2781 bio = *bio_ret; 2782 if (old_compressed) 2783 contig = bio->bi_iter.bi_sector == sector; 2784 else 2785 contig = bio_end_sector(bio) == sector; 2786 2787 if (prev_bio_flags != bio_flags || !contig || 2788 force_bio_submit || 2789 merge_bio(tree, page, offset, page_size, bio, bio_flags) || 2790 bio_add_page(bio, page, page_size, offset) < page_size) { 2791 ret = submit_one_bio(bio, mirror_num, prev_bio_flags); 2792 if (ret < 0) { 2793 *bio_ret = NULL; 2794 return ret; 2795 } 2796 bio = NULL; 2797 } else { 2798 if (wbc) 2799 wbc_account_io(wbc, page, page_size); 2800 return 0; 2801 } 2802 } 2803 2804 bio = btrfs_bio_alloc(bdev, sector << 9); 2805 bio_add_page(bio, page, page_size, offset); 2806 bio->bi_end_io = end_io_func; 2807 bio->bi_private = tree; 2808 bio->bi_write_hint = page->mapping->host->i_write_hint; 2809 bio->bi_opf = opf; 2810 if (wbc) { 2811 wbc_init_bio(wbc, bio); 2812 wbc_account_io(wbc, page, page_size); 2813 } 2814 2815 if (bio_ret) 2816 *bio_ret = bio; 2817 else 2818 ret = submit_one_bio(bio, mirror_num, bio_flags); 2819 2820 return ret; 2821 } 2822 2823 static void attach_extent_buffer_page(struct extent_buffer *eb, 2824 struct page *page) 2825 { 2826 if (!PagePrivate(page)) { 2827 SetPagePrivate(page); 2828 get_page(page); 2829 set_page_private(page, (unsigned long)eb); 2830 } else { 2831 WARN_ON(page->private != (unsigned long)eb); 2832 } 2833 } 2834 2835 void set_page_extent_mapped(struct page *page) 2836 { 2837 if (!PagePrivate(page)) { 2838 SetPagePrivate(page); 2839 get_page(page); 2840 set_page_private(page, EXTENT_PAGE_PRIVATE); 2841 } 2842 } 2843 2844 static struct extent_map * 2845 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset, 2846 u64 start, u64 len, get_extent_t *get_extent, 2847 struct extent_map **em_cached) 2848 { 2849 struct extent_map *em; 2850 2851 if (em_cached && *em_cached) { 2852 em = *em_cached; 2853 if (extent_map_in_tree(em) && start >= em->start && 2854 start < extent_map_end(em)) { 2855 refcount_inc(&em->refs); 2856 return em; 2857 } 2858 2859 free_extent_map(em); 2860 *em_cached = NULL; 2861 } 2862 2863 em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0); 2864 if (em_cached && !IS_ERR_OR_NULL(em)) { 2865 BUG_ON(*em_cached); 2866 refcount_inc(&em->refs); 2867 *em_cached = em; 2868 } 2869 return em; 2870 } 2871 /* 2872 * basic readpage implementation. Locked extent state structs are inserted 2873 * into the tree that are removed when the IO is done (by the end_io 2874 * handlers) 2875 * XXX JDM: This needs looking at to ensure proper page locking 2876 * return 0 on success, otherwise return error 2877 */ 2878 static int __do_readpage(struct extent_io_tree *tree, 2879 struct page *page, 2880 get_extent_t *get_extent, 2881 struct extent_map **em_cached, 2882 struct bio **bio, int mirror_num, 2883 unsigned long *bio_flags, unsigned int read_flags, 2884 u64 *prev_em_start) 2885 { 2886 struct inode *inode = page->mapping->host; 2887 u64 start = page_offset(page); 2888 u64 page_end = start + PAGE_SIZE - 1; 2889 u64 end; 2890 u64 cur = start; 2891 u64 extent_offset; 2892 u64 last_byte = i_size_read(inode); 2893 u64 block_start; 2894 u64 cur_end; 2895 sector_t sector; 2896 struct extent_map *em; 2897 struct block_device *bdev; 2898 int ret = 0; 2899 int nr = 0; 2900 size_t pg_offset = 0; 2901 size_t iosize; 2902 size_t disk_io_size; 2903 size_t blocksize = inode->i_sb->s_blocksize; 2904 unsigned long this_bio_flag = 0; 2905 2906 set_page_extent_mapped(page); 2907 2908 end = page_end; 2909 if (!PageUptodate(page)) { 2910 if (cleancache_get_page(page) == 0) { 2911 BUG_ON(blocksize != PAGE_SIZE); 2912 unlock_extent(tree, start, end); 2913 goto out; 2914 } 2915 } 2916 2917 if (page->index == last_byte >> PAGE_SHIFT) { 2918 char *userpage; 2919 size_t zero_offset = last_byte & (PAGE_SIZE - 1); 2920 2921 if (zero_offset) { 2922 iosize = PAGE_SIZE - zero_offset; 2923 userpage = kmap_atomic(page); 2924 memset(userpage + zero_offset, 0, iosize); 2925 flush_dcache_page(page); 2926 kunmap_atomic(userpage); 2927 } 2928 } 2929 while (cur <= end) { 2930 bool force_bio_submit = false; 2931 2932 if (cur >= last_byte) { 2933 char *userpage; 2934 struct extent_state *cached = NULL; 2935 2936 iosize = PAGE_SIZE - pg_offset; 2937 userpage = kmap_atomic(page); 2938 memset(userpage + pg_offset, 0, iosize); 2939 flush_dcache_page(page); 2940 kunmap_atomic(userpage); 2941 set_extent_uptodate(tree, cur, cur + iosize - 1, 2942 &cached, GFP_NOFS); 2943 unlock_extent_cached(tree, cur, 2944 cur + iosize - 1, 2945 &cached, GFP_NOFS); 2946 break; 2947 } 2948 em = __get_extent_map(inode, page, pg_offset, cur, 2949 end - cur + 1, get_extent, em_cached); 2950 if (IS_ERR_OR_NULL(em)) { 2951 SetPageError(page); 2952 unlock_extent(tree, cur, end); 2953 break; 2954 } 2955 extent_offset = cur - em->start; 2956 BUG_ON(extent_map_end(em) <= cur); 2957 BUG_ON(end < cur); 2958 2959 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) { 2960 this_bio_flag |= EXTENT_BIO_COMPRESSED; 2961 extent_set_compress_type(&this_bio_flag, 2962 em->compress_type); 2963 } 2964 2965 iosize = min(extent_map_end(em) - cur, end - cur + 1); 2966 cur_end = min(extent_map_end(em) - 1, end); 2967 iosize = ALIGN(iosize, blocksize); 2968 if (this_bio_flag & EXTENT_BIO_COMPRESSED) { 2969 disk_io_size = em->block_len; 2970 sector = em->block_start >> 9; 2971 } else { 2972 sector = (em->block_start + extent_offset) >> 9; 2973 disk_io_size = iosize; 2974 } 2975 bdev = em->bdev; 2976 block_start = em->block_start; 2977 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) 2978 block_start = EXTENT_MAP_HOLE; 2979 2980 /* 2981 * If we have a file range that points to a compressed extent 2982 * and it's followed by a consecutive file range that points to 2983 * to the same compressed extent (possibly with a different 2984 * offset and/or length, so it either points to the whole extent 2985 * or only part of it), we must make sure we do not submit a 2986 * single bio to populate the pages for the 2 ranges because 2987 * this makes the compressed extent read zero out the pages 2988 * belonging to the 2nd range. Imagine the following scenario: 2989 * 2990 * File layout 2991 * [0 - 8K] [8K - 24K] 2992 * | | 2993 * | | 2994 * points to extent X, points to extent X, 2995 * offset 4K, length of 8K offset 0, length 16K 2996 * 2997 * [extent X, compressed length = 4K uncompressed length = 16K] 2998 * 2999 * If the bio to read the compressed extent covers both ranges, 3000 * it will decompress extent X into the pages belonging to the 3001 * first range and then it will stop, zeroing out the remaining 3002 * pages that belong to the other range that points to extent X. 3003 * So here we make sure we submit 2 bios, one for the first 3004 * range and another one for the third range. Both will target 3005 * the same physical extent from disk, but we can't currently 3006 * make the compressed bio endio callback populate the pages 3007 * for both ranges because each compressed bio is tightly 3008 * coupled with a single extent map, and each range can have 3009 * an extent map with a different offset value relative to the 3010 * uncompressed data of our extent and different lengths. This 3011 * is a corner case so we prioritize correctness over 3012 * non-optimal behavior (submitting 2 bios for the same extent). 3013 */ 3014 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) && 3015 prev_em_start && *prev_em_start != (u64)-1 && 3016 *prev_em_start != em->orig_start) 3017 force_bio_submit = true; 3018 3019 if (prev_em_start) 3020 *prev_em_start = em->orig_start; 3021 3022 free_extent_map(em); 3023 em = NULL; 3024 3025 /* we've found a hole, just zero and go on */ 3026 if (block_start == EXTENT_MAP_HOLE) { 3027 char *userpage; 3028 struct extent_state *cached = NULL; 3029 3030 userpage = kmap_atomic(page); 3031 memset(userpage + pg_offset, 0, iosize); 3032 flush_dcache_page(page); 3033 kunmap_atomic(userpage); 3034 3035 set_extent_uptodate(tree, cur, cur + iosize - 1, 3036 &cached, GFP_NOFS); 3037 unlock_extent_cached(tree, cur, 3038 cur + iosize - 1, 3039 &cached, GFP_NOFS); 3040 cur = cur + iosize; 3041 pg_offset += iosize; 3042 continue; 3043 } 3044 /* the get_extent function already copied into the page */ 3045 if (test_range_bit(tree, cur, cur_end, 3046 EXTENT_UPTODATE, 1, NULL)) { 3047 check_page_uptodate(tree, page); 3048 unlock_extent(tree, cur, cur + iosize - 1); 3049 cur = cur + iosize; 3050 pg_offset += iosize; 3051 continue; 3052 } 3053 /* we have an inline extent but it didn't get marked up 3054 * to date. Error out 3055 */ 3056 if (block_start == EXTENT_MAP_INLINE) { 3057 SetPageError(page); 3058 unlock_extent(tree, cur, cur + iosize - 1); 3059 cur = cur + iosize; 3060 pg_offset += iosize; 3061 continue; 3062 } 3063 3064 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL, 3065 page, sector, disk_io_size, pg_offset, 3066 bdev, bio, 3067 end_bio_extent_readpage, mirror_num, 3068 *bio_flags, 3069 this_bio_flag, 3070 force_bio_submit); 3071 if (!ret) { 3072 nr++; 3073 *bio_flags = this_bio_flag; 3074 } else { 3075 SetPageError(page); 3076 unlock_extent(tree, cur, cur + iosize - 1); 3077 goto out; 3078 } 3079 cur = cur + iosize; 3080 pg_offset += iosize; 3081 } 3082 out: 3083 if (!nr) { 3084 if (!PageError(page)) 3085 SetPageUptodate(page); 3086 unlock_page(page); 3087 } 3088 return ret; 3089 } 3090 3091 static inline void __do_contiguous_readpages(struct extent_io_tree *tree, 3092 struct page *pages[], int nr_pages, 3093 u64 start, u64 end, 3094 get_extent_t *get_extent, 3095 struct extent_map **em_cached, 3096 struct bio **bio, int mirror_num, 3097 unsigned long *bio_flags, 3098 u64 *prev_em_start) 3099 { 3100 struct inode *inode; 3101 struct btrfs_ordered_extent *ordered; 3102 int index; 3103 3104 inode = pages[0]->mapping->host; 3105 while (1) { 3106 lock_extent(tree, start, end); 3107 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start, 3108 end - start + 1); 3109 if (!ordered) 3110 break; 3111 unlock_extent(tree, start, end); 3112 btrfs_start_ordered_extent(inode, ordered, 1); 3113 btrfs_put_ordered_extent(ordered); 3114 } 3115 3116 for (index = 0; index < nr_pages; index++) { 3117 __do_readpage(tree, pages[index], get_extent, em_cached, bio, 3118 mirror_num, bio_flags, 0, prev_em_start); 3119 put_page(pages[index]); 3120 } 3121 } 3122 3123 static void __extent_readpages(struct extent_io_tree *tree, 3124 struct page *pages[], 3125 int nr_pages, get_extent_t *get_extent, 3126 struct extent_map **em_cached, 3127 struct bio **bio, int mirror_num, 3128 unsigned long *bio_flags, 3129 u64 *prev_em_start) 3130 { 3131 u64 start = 0; 3132 u64 end = 0; 3133 u64 page_start; 3134 int index; 3135 int first_index = 0; 3136 3137 for (index = 0; index < nr_pages; index++) { 3138 page_start = page_offset(pages[index]); 3139 if (!end) { 3140 start = page_start; 3141 end = start + PAGE_SIZE - 1; 3142 first_index = index; 3143 } else if (end + 1 == page_start) { 3144 end += PAGE_SIZE; 3145 } else { 3146 __do_contiguous_readpages(tree, &pages[first_index], 3147 index - first_index, start, 3148 end, get_extent, em_cached, 3149 bio, mirror_num, bio_flags, 3150 prev_em_start); 3151 start = page_start; 3152 end = start + PAGE_SIZE - 1; 3153 first_index = index; 3154 } 3155 } 3156 3157 if (end) 3158 __do_contiguous_readpages(tree, &pages[first_index], 3159 index - first_index, start, 3160 end, get_extent, em_cached, bio, 3161 mirror_num, bio_flags, 3162 prev_em_start); 3163 } 3164 3165 static int __extent_read_full_page(struct extent_io_tree *tree, 3166 struct page *page, 3167 get_extent_t *get_extent, 3168 struct bio **bio, int mirror_num, 3169 unsigned long *bio_flags, 3170 unsigned int read_flags) 3171 { 3172 struct inode *inode = page->mapping->host; 3173 struct btrfs_ordered_extent *ordered; 3174 u64 start = page_offset(page); 3175 u64 end = start + PAGE_SIZE - 1; 3176 int ret; 3177 3178 while (1) { 3179 lock_extent(tree, start, end); 3180 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start, 3181 PAGE_SIZE); 3182 if (!ordered) 3183 break; 3184 unlock_extent(tree, start, end); 3185 btrfs_start_ordered_extent(inode, ordered, 1); 3186 btrfs_put_ordered_extent(ordered); 3187 } 3188 3189 ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num, 3190 bio_flags, read_flags, NULL); 3191 return ret; 3192 } 3193 3194 int extent_read_full_page(struct extent_io_tree *tree, struct page *page, 3195 get_extent_t *get_extent, int mirror_num) 3196 { 3197 struct bio *bio = NULL; 3198 unsigned long bio_flags = 0; 3199 int ret; 3200 3201 ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num, 3202 &bio_flags, 0); 3203 if (bio) 3204 ret = submit_one_bio(bio, mirror_num, bio_flags); 3205 return ret; 3206 } 3207 3208 static void update_nr_written(struct writeback_control *wbc, 3209 unsigned long nr_written) 3210 { 3211 wbc->nr_to_write -= nr_written; 3212 } 3213 3214 /* 3215 * helper for __extent_writepage, doing all of the delayed allocation setup. 3216 * 3217 * This returns 1 if our fill_delalloc function did all the work required 3218 * to write the page (copy into inline extent). In this case the IO has 3219 * been started and the page is already unlocked. 3220 * 3221 * This returns 0 if all went well (page still locked) 3222 * This returns < 0 if there were errors (page still locked) 3223 */ 3224 static noinline_for_stack int writepage_delalloc(struct inode *inode, 3225 struct page *page, struct writeback_control *wbc, 3226 struct extent_page_data *epd, 3227 u64 delalloc_start, 3228 unsigned long *nr_written) 3229 { 3230 struct extent_io_tree *tree = epd->tree; 3231 u64 page_end = delalloc_start + PAGE_SIZE - 1; 3232 u64 nr_delalloc; 3233 u64 delalloc_to_write = 0; 3234 u64 delalloc_end = 0; 3235 int ret; 3236 int page_started = 0; 3237 3238 if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc) 3239 return 0; 3240 3241 while (delalloc_end < page_end) { 3242 nr_delalloc = find_lock_delalloc_range(inode, tree, 3243 page, 3244 &delalloc_start, 3245 &delalloc_end, 3246 BTRFS_MAX_EXTENT_SIZE); 3247 if (nr_delalloc == 0) { 3248 delalloc_start = delalloc_end + 1; 3249 continue; 3250 } 3251 ret = tree->ops->fill_delalloc(inode, page, 3252 delalloc_start, 3253 delalloc_end, 3254 &page_started, 3255 nr_written); 3256 /* File system has been set read-only */ 3257 if (ret) { 3258 SetPageError(page); 3259 /* fill_delalloc should be return < 0 for error 3260 * but just in case, we use > 0 here meaning the 3261 * IO is started, so we don't want to return > 0 3262 * unless things are going well. 3263 */ 3264 ret = ret < 0 ? ret : -EIO; 3265 goto done; 3266 } 3267 /* 3268 * delalloc_end is already one less than the total length, so 3269 * we don't subtract one from PAGE_SIZE 3270 */ 3271 delalloc_to_write += (delalloc_end - delalloc_start + 3272 PAGE_SIZE) >> PAGE_SHIFT; 3273 delalloc_start = delalloc_end + 1; 3274 } 3275 if (wbc->nr_to_write < delalloc_to_write) { 3276 int thresh = 8192; 3277 3278 if (delalloc_to_write < thresh * 2) 3279 thresh = delalloc_to_write; 3280 wbc->nr_to_write = min_t(u64, delalloc_to_write, 3281 thresh); 3282 } 3283 3284 /* did the fill delalloc function already unlock and start 3285 * the IO? 3286 */ 3287 if (page_started) { 3288 /* 3289 * we've unlocked the page, so we can't update 3290 * the mapping's writeback index, just update 3291 * nr_to_write. 3292 */ 3293 wbc->nr_to_write -= *nr_written; 3294 return 1; 3295 } 3296 3297 ret = 0; 3298 3299 done: 3300 return ret; 3301 } 3302 3303 /* 3304 * helper for __extent_writepage. This calls the writepage start hooks, 3305 * and does the loop to map the page into extents and bios. 3306 * 3307 * We return 1 if the IO is started and the page is unlocked, 3308 * 0 if all went well (page still locked) 3309 * < 0 if there were errors (page still locked) 3310 */ 3311 static noinline_for_stack int __extent_writepage_io(struct inode *inode, 3312 struct page *page, 3313 struct writeback_control *wbc, 3314 struct extent_page_data *epd, 3315 loff_t i_size, 3316 unsigned long nr_written, 3317 unsigned int write_flags, int *nr_ret) 3318 { 3319 struct extent_io_tree *tree = epd->tree; 3320 u64 start = page_offset(page); 3321 u64 page_end = start + PAGE_SIZE - 1; 3322 u64 end; 3323 u64 cur = start; 3324 u64 extent_offset; 3325 u64 block_start; 3326 u64 iosize; 3327 sector_t sector; 3328 struct extent_map *em; 3329 struct block_device *bdev; 3330 size_t pg_offset = 0; 3331 size_t blocksize; 3332 int ret = 0; 3333 int nr = 0; 3334 bool compressed; 3335 3336 if (tree->ops && tree->ops->writepage_start_hook) { 3337 ret = tree->ops->writepage_start_hook(page, start, 3338 page_end); 3339 if (ret) { 3340 /* Fixup worker will requeue */ 3341 if (ret == -EBUSY) 3342 wbc->pages_skipped++; 3343 else 3344 redirty_page_for_writepage(wbc, page); 3345 3346 update_nr_written(wbc, nr_written); 3347 unlock_page(page); 3348 return 1; 3349 } 3350 } 3351 3352 /* 3353 * we don't want to touch the inode after unlocking the page, 3354 * so we update the mapping writeback index now 3355 */ 3356 update_nr_written(wbc, nr_written + 1); 3357 3358 end = page_end; 3359 if (i_size <= start) { 3360 if (tree->ops && tree->ops->writepage_end_io_hook) 3361 tree->ops->writepage_end_io_hook(page, start, 3362 page_end, NULL, 1); 3363 goto done; 3364 } 3365 3366 blocksize = inode->i_sb->s_blocksize; 3367 3368 while (cur <= end) { 3369 u64 em_end; 3370 3371 if (cur >= i_size) { 3372 if (tree->ops && tree->ops->writepage_end_io_hook) 3373 tree->ops->writepage_end_io_hook(page, cur, 3374 page_end, NULL, 1); 3375 break; 3376 } 3377 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur, 3378 end - cur + 1, 1); 3379 if (IS_ERR_OR_NULL(em)) { 3380 SetPageError(page); 3381 ret = PTR_ERR_OR_ZERO(em); 3382 break; 3383 } 3384 3385 extent_offset = cur - em->start; 3386 em_end = extent_map_end(em); 3387 BUG_ON(em_end <= cur); 3388 BUG_ON(end < cur); 3389 iosize = min(em_end - cur, end - cur + 1); 3390 iosize = ALIGN(iosize, blocksize); 3391 sector = (em->block_start + extent_offset) >> 9; 3392 bdev = em->bdev; 3393 block_start = em->block_start; 3394 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags); 3395 free_extent_map(em); 3396 em = NULL; 3397 3398 /* 3399 * compressed and inline extents are written through other 3400 * paths in the FS 3401 */ 3402 if (compressed || block_start == EXTENT_MAP_HOLE || 3403 block_start == EXTENT_MAP_INLINE) { 3404 /* 3405 * end_io notification does not happen here for 3406 * compressed extents 3407 */ 3408 if (!compressed && tree->ops && 3409 tree->ops->writepage_end_io_hook) 3410 tree->ops->writepage_end_io_hook(page, cur, 3411 cur + iosize - 1, 3412 NULL, 1); 3413 else if (compressed) { 3414 /* we don't want to end_page_writeback on 3415 * a compressed extent. this happens 3416 * elsewhere 3417 */ 3418 nr++; 3419 } 3420 3421 cur += iosize; 3422 pg_offset += iosize; 3423 continue; 3424 } 3425 3426 set_range_writeback(tree, cur, cur + iosize - 1); 3427 if (!PageWriteback(page)) { 3428 btrfs_err(BTRFS_I(inode)->root->fs_info, 3429 "page %lu not writeback, cur %llu end %llu", 3430 page->index, cur, end); 3431 } 3432 3433 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc, 3434 page, sector, iosize, pg_offset, 3435 bdev, &epd->bio, 3436 end_bio_extent_writepage, 3437 0, 0, 0, false); 3438 if (ret) { 3439 SetPageError(page); 3440 if (PageWriteback(page)) 3441 end_page_writeback(page); 3442 } 3443 3444 cur = cur + iosize; 3445 pg_offset += iosize; 3446 nr++; 3447 } 3448 done: 3449 *nr_ret = nr; 3450 return ret; 3451 } 3452 3453 /* 3454 * the writepage semantics are similar to regular writepage. extent 3455 * records are inserted to lock ranges in the tree, and as dirty areas 3456 * are found, they are marked writeback. Then the lock bits are removed 3457 * and the end_io handler clears the writeback ranges 3458 */ 3459 static int __extent_writepage(struct page *page, struct writeback_control *wbc, 3460 void *data) 3461 { 3462 struct inode *inode = page->mapping->host; 3463 struct extent_page_data *epd = data; 3464 u64 start = page_offset(page); 3465 u64 page_end = start + PAGE_SIZE - 1; 3466 int ret; 3467 int nr = 0; 3468 size_t pg_offset = 0; 3469 loff_t i_size = i_size_read(inode); 3470 unsigned long end_index = i_size >> PAGE_SHIFT; 3471 unsigned int write_flags = 0; 3472 unsigned long nr_written = 0; 3473 3474 if (wbc->sync_mode == WB_SYNC_ALL) 3475 write_flags = REQ_SYNC; 3476 3477 trace___extent_writepage(page, inode, wbc); 3478 3479 WARN_ON(!PageLocked(page)); 3480 3481 ClearPageError(page); 3482 3483 pg_offset = i_size & (PAGE_SIZE - 1); 3484 if (page->index > end_index || 3485 (page->index == end_index && !pg_offset)) { 3486 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE); 3487 unlock_page(page); 3488 return 0; 3489 } 3490 3491 if (page->index == end_index) { 3492 char *userpage; 3493 3494 userpage = kmap_atomic(page); 3495 memset(userpage + pg_offset, 0, 3496 PAGE_SIZE - pg_offset); 3497 kunmap_atomic(userpage); 3498 flush_dcache_page(page); 3499 } 3500 3501 pg_offset = 0; 3502 3503 set_page_extent_mapped(page); 3504 3505 ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written); 3506 if (ret == 1) 3507 goto done_unlocked; 3508 if (ret) 3509 goto done; 3510 3511 ret = __extent_writepage_io(inode, page, wbc, epd, 3512 i_size, nr_written, write_flags, &nr); 3513 if (ret == 1) 3514 goto done_unlocked; 3515 3516 done: 3517 if (nr == 0) { 3518 /* make sure the mapping tag for page dirty gets cleared */ 3519 set_page_writeback(page); 3520 end_page_writeback(page); 3521 } 3522 if (PageError(page)) { 3523 ret = ret < 0 ? ret : -EIO; 3524 end_extent_writepage(page, ret, start, page_end); 3525 } 3526 unlock_page(page); 3527 return ret; 3528 3529 done_unlocked: 3530 return 0; 3531 } 3532 3533 void wait_on_extent_buffer_writeback(struct extent_buffer *eb) 3534 { 3535 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK, 3536 TASK_UNINTERRUPTIBLE); 3537 } 3538 3539 static noinline_for_stack int 3540 lock_extent_buffer_for_io(struct extent_buffer *eb, 3541 struct btrfs_fs_info *fs_info, 3542 struct extent_page_data *epd) 3543 { 3544 unsigned long i, num_pages; 3545 int flush = 0; 3546 int ret = 0; 3547 3548 if (!btrfs_try_tree_write_lock(eb)) { 3549 flush = 1; 3550 flush_write_bio(epd); 3551 btrfs_tree_lock(eb); 3552 } 3553 3554 if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) { 3555 btrfs_tree_unlock(eb); 3556 if (!epd->sync_io) 3557 return 0; 3558 if (!flush) { 3559 flush_write_bio(epd); 3560 flush = 1; 3561 } 3562 while (1) { 3563 wait_on_extent_buffer_writeback(eb); 3564 btrfs_tree_lock(eb); 3565 if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) 3566 break; 3567 btrfs_tree_unlock(eb); 3568 } 3569 } 3570 3571 /* 3572 * We need to do this to prevent races in people who check if the eb is 3573 * under IO since we can end up having no IO bits set for a short period 3574 * of time. 3575 */ 3576 spin_lock(&eb->refs_lock); 3577 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) { 3578 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags); 3579 spin_unlock(&eb->refs_lock); 3580 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN); 3581 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, 3582 -eb->len, 3583 fs_info->dirty_metadata_batch); 3584 ret = 1; 3585 } else { 3586 spin_unlock(&eb->refs_lock); 3587 } 3588 3589 btrfs_tree_unlock(eb); 3590 3591 if (!ret) 3592 return ret; 3593 3594 num_pages = num_extent_pages(eb->start, eb->len); 3595 for (i = 0; i < num_pages; i++) { 3596 struct page *p = eb->pages[i]; 3597 3598 if (!trylock_page(p)) { 3599 if (!flush) { 3600 flush_write_bio(epd); 3601 flush = 1; 3602 } 3603 lock_page(p); 3604 } 3605 } 3606 3607 return ret; 3608 } 3609 3610 static void end_extent_buffer_writeback(struct extent_buffer *eb) 3611 { 3612 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags); 3613 smp_mb__after_atomic(); 3614 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK); 3615 } 3616 3617 static void set_btree_ioerr(struct page *page) 3618 { 3619 struct extent_buffer *eb = (struct extent_buffer *)page->private; 3620 3621 SetPageError(page); 3622 if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) 3623 return; 3624 3625 /* 3626 * If writeback for a btree extent that doesn't belong to a log tree 3627 * failed, increment the counter transaction->eb_write_errors. 3628 * We do this because while the transaction is running and before it's 3629 * committing (when we call filemap_fdata[write|wait]_range against 3630 * the btree inode), we might have 3631 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it 3632 * returns an error or an error happens during writeback, when we're 3633 * committing the transaction we wouldn't know about it, since the pages 3634 * can be no longer dirty nor marked anymore for writeback (if a 3635 * subsequent modification to the extent buffer didn't happen before the 3636 * transaction commit), which makes filemap_fdata[write|wait]_range not 3637 * able to find the pages tagged with SetPageError at transaction 3638 * commit time. So if this happens we must abort the transaction, 3639 * otherwise we commit a super block with btree roots that point to 3640 * btree nodes/leafs whose content on disk is invalid - either garbage 3641 * or the content of some node/leaf from a past generation that got 3642 * cowed or deleted and is no longer valid. 3643 * 3644 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would 3645 * not be enough - we need to distinguish between log tree extents vs 3646 * non-log tree extents, and the next filemap_fdatawait_range() call 3647 * will catch and clear such errors in the mapping - and that call might 3648 * be from a log sync and not from a transaction commit. Also, checking 3649 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is 3650 * not done and would not be reliable - the eb might have been released 3651 * from memory and reading it back again means that flag would not be 3652 * set (since it's a runtime flag, not persisted on disk). 3653 * 3654 * Using the flags below in the btree inode also makes us achieve the 3655 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started 3656 * writeback for all dirty pages and before filemap_fdatawait_range() 3657 * is called, the writeback for all dirty pages had already finished 3658 * with errors - because we were not using AS_EIO/AS_ENOSPC, 3659 * filemap_fdatawait_range() would return success, as it could not know 3660 * that writeback errors happened (the pages were no longer tagged for 3661 * writeback). 3662 */ 3663 switch (eb->log_index) { 3664 case -1: 3665 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags); 3666 break; 3667 case 0: 3668 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags); 3669 break; 3670 case 1: 3671 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags); 3672 break; 3673 default: 3674 BUG(); /* unexpected, logic error */ 3675 } 3676 } 3677 3678 static void end_bio_extent_buffer_writepage(struct bio *bio) 3679 { 3680 struct bio_vec *bvec; 3681 struct extent_buffer *eb; 3682 int i, done; 3683 3684 ASSERT(!bio_flagged(bio, BIO_CLONED)); 3685 bio_for_each_segment_all(bvec, bio, i) { 3686 struct page *page = bvec->bv_page; 3687 3688 eb = (struct extent_buffer *)page->private; 3689 BUG_ON(!eb); 3690 done = atomic_dec_and_test(&eb->io_pages); 3691 3692 if (bio->bi_status || 3693 test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) { 3694 ClearPageUptodate(page); 3695 set_btree_ioerr(page); 3696 } 3697 3698 end_page_writeback(page); 3699 3700 if (!done) 3701 continue; 3702 3703 end_extent_buffer_writeback(eb); 3704 } 3705 3706 bio_put(bio); 3707 } 3708 3709 static noinline_for_stack int write_one_eb(struct extent_buffer *eb, 3710 struct btrfs_fs_info *fs_info, 3711 struct writeback_control *wbc, 3712 struct extent_page_data *epd) 3713 { 3714 struct block_device *bdev = fs_info->fs_devices->latest_bdev; 3715 struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree; 3716 u64 offset = eb->start; 3717 u32 nritems; 3718 unsigned long i, num_pages; 3719 unsigned long bio_flags = 0; 3720 unsigned long start, end; 3721 unsigned int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META; 3722 int ret = 0; 3723 3724 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags); 3725 num_pages = num_extent_pages(eb->start, eb->len); 3726 atomic_set(&eb->io_pages, num_pages); 3727 if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID) 3728 bio_flags = EXTENT_BIO_TREE_LOG; 3729 3730 /* set btree blocks beyond nritems with 0 to avoid stale content. */ 3731 nritems = btrfs_header_nritems(eb); 3732 if (btrfs_header_level(eb) > 0) { 3733 end = btrfs_node_key_ptr_offset(nritems); 3734 3735 memzero_extent_buffer(eb, end, eb->len - end); 3736 } else { 3737 /* 3738 * leaf: 3739 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0 3740 */ 3741 start = btrfs_item_nr_offset(nritems); 3742 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, eb); 3743 memzero_extent_buffer(eb, start, end - start); 3744 } 3745 3746 for (i = 0; i < num_pages; i++) { 3747 struct page *p = eb->pages[i]; 3748 3749 clear_page_dirty_for_io(p); 3750 set_page_writeback(p); 3751 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc, 3752 p, offset >> 9, PAGE_SIZE, 0, bdev, 3753 &epd->bio, 3754 end_bio_extent_buffer_writepage, 3755 0, epd->bio_flags, bio_flags, false); 3756 epd->bio_flags = bio_flags; 3757 if (ret) { 3758 set_btree_ioerr(p); 3759 if (PageWriteback(p)) 3760 end_page_writeback(p); 3761 if (atomic_sub_and_test(num_pages - i, &eb->io_pages)) 3762 end_extent_buffer_writeback(eb); 3763 ret = -EIO; 3764 break; 3765 } 3766 offset += PAGE_SIZE; 3767 update_nr_written(wbc, 1); 3768 unlock_page(p); 3769 } 3770 3771 if (unlikely(ret)) { 3772 for (; i < num_pages; i++) { 3773 struct page *p = eb->pages[i]; 3774 clear_page_dirty_for_io(p); 3775 unlock_page(p); 3776 } 3777 } 3778 3779 return ret; 3780 } 3781 3782 int btree_write_cache_pages(struct address_space *mapping, 3783 struct writeback_control *wbc) 3784 { 3785 struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree; 3786 struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info; 3787 struct extent_buffer *eb, *prev_eb = NULL; 3788 struct extent_page_data epd = { 3789 .bio = NULL, 3790 .tree = tree, 3791 .extent_locked = 0, 3792 .sync_io = wbc->sync_mode == WB_SYNC_ALL, 3793 .bio_flags = 0, 3794 }; 3795 int ret = 0; 3796 int done = 0; 3797 int nr_to_write_done = 0; 3798 struct pagevec pvec; 3799 int nr_pages; 3800 pgoff_t index; 3801 pgoff_t end; /* Inclusive */ 3802 int scanned = 0; 3803 int tag; 3804 3805 pagevec_init(&pvec, 0); 3806 if (wbc->range_cyclic) { 3807 index = mapping->writeback_index; /* Start from prev offset */ 3808 end = -1; 3809 } else { 3810 index = wbc->range_start >> PAGE_SHIFT; 3811 end = wbc->range_end >> PAGE_SHIFT; 3812 scanned = 1; 3813 } 3814 if (wbc->sync_mode == WB_SYNC_ALL) 3815 tag = PAGECACHE_TAG_TOWRITE; 3816 else 3817 tag = PAGECACHE_TAG_DIRTY; 3818 retry: 3819 if (wbc->sync_mode == WB_SYNC_ALL) 3820 tag_pages_for_writeback(mapping, index, end); 3821 while (!done && !nr_to_write_done && (index <= end) && 3822 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, 3823 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) { 3824 unsigned i; 3825 3826 scanned = 1; 3827 for (i = 0; i < nr_pages; i++) { 3828 struct page *page = pvec.pages[i]; 3829 3830 if (!PagePrivate(page)) 3831 continue; 3832 3833 if (!wbc->range_cyclic && page->index > end) { 3834 done = 1; 3835 break; 3836 } 3837 3838 spin_lock(&mapping->private_lock); 3839 if (!PagePrivate(page)) { 3840 spin_unlock(&mapping->private_lock); 3841 continue; 3842 } 3843 3844 eb = (struct extent_buffer *)page->private; 3845 3846 /* 3847 * Shouldn't happen and normally this would be a BUG_ON 3848 * but no sense in crashing the users box for something 3849 * we can survive anyway. 3850 */ 3851 if (WARN_ON(!eb)) { 3852 spin_unlock(&mapping->private_lock); 3853 continue; 3854 } 3855 3856 if (eb == prev_eb) { 3857 spin_unlock(&mapping->private_lock); 3858 continue; 3859 } 3860 3861 ret = atomic_inc_not_zero(&eb->refs); 3862 spin_unlock(&mapping->private_lock); 3863 if (!ret) 3864 continue; 3865 3866 prev_eb = eb; 3867 ret = lock_extent_buffer_for_io(eb, fs_info, &epd); 3868 if (!ret) { 3869 free_extent_buffer(eb); 3870 continue; 3871 } 3872 3873 ret = write_one_eb(eb, fs_info, wbc, &epd); 3874 if (ret) { 3875 done = 1; 3876 free_extent_buffer(eb); 3877 break; 3878 } 3879 free_extent_buffer(eb); 3880 3881 /* 3882 * the filesystem may choose to bump up nr_to_write. 3883 * We have to make sure to honor the new nr_to_write 3884 * at any time 3885 */ 3886 nr_to_write_done = wbc->nr_to_write <= 0; 3887 } 3888 pagevec_release(&pvec); 3889 cond_resched(); 3890 } 3891 if (!scanned && !done) { 3892 /* 3893 * We hit the last page and there is more work to be done: wrap 3894 * back to the start of the file 3895 */ 3896 scanned = 1; 3897 index = 0; 3898 goto retry; 3899 } 3900 flush_write_bio(&epd); 3901 return ret; 3902 } 3903 3904 /** 3905 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. 3906 * @mapping: address space structure to write 3907 * @wbc: subtract the number of written pages from *@wbc->nr_to_write 3908 * @writepage: function called for each page 3909 * @data: data passed to writepage function 3910 * 3911 * If a page is already under I/O, write_cache_pages() skips it, even 3912 * if it's dirty. This is desirable behaviour for memory-cleaning writeback, 3913 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() 3914 * and msync() need to guarantee that all the data which was dirty at the time 3915 * the call was made get new I/O started against them. If wbc->sync_mode is 3916 * WB_SYNC_ALL then we were called for data integrity and we must wait for 3917 * existing IO to complete. 3918 */ 3919 static int extent_write_cache_pages(struct address_space *mapping, 3920 struct writeback_control *wbc, 3921 writepage_t writepage, void *data, 3922 void (*flush_fn)(void *)) 3923 { 3924 struct inode *inode = mapping->host; 3925 int ret = 0; 3926 int done = 0; 3927 int nr_to_write_done = 0; 3928 struct pagevec pvec; 3929 int nr_pages; 3930 pgoff_t index; 3931 pgoff_t end; /* Inclusive */ 3932 pgoff_t done_index; 3933 int range_whole = 0; 3934 int scanned = 0; 3935 int tag; 3936 3937 /* 3938 * We have to hold onto the inode so that ordered extents can do their 3939 * work when the IO finishes. The alternative to this is failing to add 3940 * an ordered extent if the igrab() fails there and that is a huge pain 3941 * to deal with, so instead just hold onto the inode throughout the 3942 * writepages operation. If it fails here we are freeing up the inode 3943 * anyway and we'd rather not waste our time writing out stuff that is 3944 * going to be truncated anyway. 3945 */ 3946 if (!igrab(inode)) 3947 return 0; 3948 3949 pagevec_init(&pvec, 0); 3950 if (wbc->range_cyclic) { 3951 index = mapping->writeback_index; /* Start from prev offset */ 3952 end = -1; 3953 } else { 3954 index = wbc->range_start >> PAGE_SHIFT; 3955 end = wbc->range_end >> PAGE_SHIFT; 3956 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) 3957 range_whole = 1; 3958 scanned = 1; 3959 } 3960 if (wbc->sync_mode == WB_SYNC_ALL) 3961 tag = PAGECACHE_TAG_TOWRITE; 3962 else 3963 tag = PAGECACHE_TAG_DIRTY; 3964 retry: 3965 if (wbc->sync_mode == WB_SYNC_ALL) 3966 tag_pages_for_writeback(mapping, index, end); 3967 done_index = index; 3968 while (!done && !nr_to_write_done && (index <= end) && 3969 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, 3970 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) { 3971 unsigned i; 3972 3973 scanned = 1; 3974 for (i = 0; i < nr_pages; i++) { 3975 struct page *page = pvec.pages[i]; 3976 3977 done_index = page->index; 3978 /* 3979 * At this point we hold neither mapping->tree_lock nor 3980 * lock on the page itself: the page may be truncated or 3981 * invalidated (changing page->mapping to NULL), or even 3982 * swizzled back from swapper_space to tmpfs file 3983 * mapping 3984 */ 3985 if (!trylock_page(page)) { 3986 flush_fn(data); 3987 lock_page(page); 3988 } 3989 3990 if (unlikely(page->mapping != mapping)) { 3991 unlock_page(page); 3992 continue; 3993 } 3994 3995 if (!wbc->range_cyclic && page->index > end) { 3996 done = 1; 3997 unlock_page(page); 3998 continue; 3999 } 4000 4001 if (wbc->sync_mode != WB_SYNC_NONE) { 4002 if (PageWriteback(page)) 4003 flush_fn(data); 4004 wait_on_page_writeback(page); 4005 } 4006 4007 if (PageWriteback(page) || 4008 !clear_page_dirty_for_io(page)) { 4009 unlock_page(page); 4010 continue; 4011 } 4012 4013 ret = (*writepage)(page, wbc, data); 4014 4015 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) { 4016 unlock_page(page); 4017 ret = 0; 4018 } 4019 if (ret < 0) { 4020 /* 4021 * done_index is set past this page, 4022 * so media errors will not choke 4023 * background writeout for the entire 4024 * file. This has consequences for 4025 * range_cyclic semantics (ie. it may 4026 * not be suitable for data integrity 4027 * writeout). 4028 */ 4029 done_index = page->index + 1; 4030 done = 1; 4031 break; 4032 } 4033 4034 /* 4035 * the filesystem may choose to bump up nr_to_write. 4036 * We have to make sure to honor the new nr_to_write 4037 * at any time 4038 */ 4039 nr_to_write_done = wbc->nr_to_write <= 0; 4040 } 4041 pagevec_release(&pvec); 4042 cond_resched(); 4043 } 4044 if (!scanned && !done) { 4045 /* 4046 * We hit the last page and there is more work to be done: wrap 4047 * back to the start of the file 4048 */ 4049 scanned = 1; 4050 index = 0; 4051 goto retry; 4052 } 4053 4054 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole)) 4055 mapping->writeback_index = done_index; 4056 4057 btrfs_add_delayed_iput(inode); 4058 return ret; 4059 } 4060 4061 static void flush_epd_write_bio(struct extent_page_data *epd) 4062 { 4063 if (epd->bio) { 4064 int ret; 4065 4066 bio_set_op_attrs(epd->bio, REQ_OP_WRITE, 4067 epd->sync_io ? REQ_SYNC : 0); 4068 4069 ret = submit_one_bio(epd->bio, 0, epd->bio_flags); 4070 BUG_ON(ret < 0); /* -ENOMEM */ 4071 epd->bio = NULL; 4072 } 4073 } 4074 4075 static noinline void flush_write_bio(void *data) 4076 { 4077 struct extent_page_data *epd = data; 4078 flush_epd_write_bio(epd); 4079 } 4080 4081 int extent_write_full_page(struct extent_io_tree *tree, struct page *page, 4082 get_extent_t *get_extent, 4083 struct writeback_control *wbc) 4084 { 4085 int ret; 4086 struct extent_page_data epd = { 4087 .bio = NULL, 4088 .tree = tree, 4089 .get_extent = get_extent, 4090 .extent_locked = 0, 4091 .sync_io = wbc->sync_mode == WB_SYNC_ALL, 4092 .bio_flags = 0, 4093 }; 4094 4095 ret = __extent_writepage(page, wbc, &epd); 4096 4097 flush_epd_write_bio(&epd); 4098 return ret; 4099 } 4100 4101 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode, 4102 u64 start, u64 end, get_extent_t *get_extent, 4103 int mode) 4104 { 4105 int ret = 0; 4106 struct address_space *mapping = inode->i_mapping; 4107 struct page *page; 4108 unsigned long nr_pages = (end - start + PAGE_SIZE) >> 4109 PAGE_SHIFT; 4110 4111 struct extent_page_data epd = { 4112 .bio = NULL, 4113 .tree = tree, 4114 .get_extent = get_extent, 4115 .extent_locked = 1, 4116 .sync_io = mode == WB_SYNC_ALL, 4117 .bio_flags = 0, 4118 }; 4119 struct writeback_control wbc_writepages = { 4120 .sync_mode = mode, 4121 .nr_to_write = nr_pages * 2, 4122 .range_start = start, 4123 .range_end = end + 1, 4124 }; 4125 4126 while (start <= end) { 4127 page = find_get_page(mapping, start >> PAGE_SHIFT); 4128 if (clear_page_dirty_for_io(page)) 4129 ret = __extent_writepage(page, &wbc_writepages, &epd); 4130 else { 4131 if (tree->ops && tree->ops->writepage_end_io_hook) 4132 tree->ops->writepage_end_io_hook(page, start, 4133 start + PAGE_SIZE - 1, 4134 NULL, 1); 4135 unlock_page(page); 4136 } 4137 put_page(page); 4138 start += PAGE_SIZE; 4139 } 4140 4141 flush_epd_write_bio(&epd); 4142 return ret; 4143 } 4144 4145 int extent_writepages(struct extent_io_tree *tree, 4146 struct address_space *mapping, 4147 get_extent_t *get_extent, 4148 struct writeback_control *wbc) 4149 { 4150 int ret = 0; 4151 struct extent_page_data epd = { 4152 .bio = NULL, 4153 .tree = tree, 4154 .get_extent = get_extent, 4155 .extent_locked = 0, 4156 .sync_io = wbc->sync_mode == WB_SYNC_ALL, 4157 .bio_flags = 0, 4158 }; 4159 4160 ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd, 4161 flush_write_bio); 4162 flush_epd_write_bio(&epd); 4163 return ret; 4164 } 4165 4166 int extent_readpages(struct extent_io_tree *tree, 4167 struct address_space *mapping, 4168 struct list_head *pages, unsigned nr_pages, 4169 get_extent_t get_extent) 4170 { 4171 struct bio *bio = NULL; 4172 unsigned page_idx; 4173 unsigned long bio_flags = 0; 4174 struct page *pagepool[16]; 4175 struct page *page; 4176 struct extent_map *em_cached = NULL; 4177 int nr = 0; 4178 u64 prev_em_start = (u64)-1; 4179 4180 for (page_idx = 0; page_idx < nr_pages; page_idx++) { 4181 page = list_entry(pages->prev, struct page, lru); 4182 4183 prefetchw(&page->flags); 4184 list_del(&page->lru); 4185 if (add_to_page_cache_lru(page, mapping, 4186 page->index, 4187 readahead_gfp_mask(mapping))) { 4188 put_page(page); 4189 continue; 4190 } 4191 4192 pagepool[nr++] = page; 4193 if (nr < ARRAY_SIZE(pagepool)) 4194 continue; 4195 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached, 4196 &bio, 0, &bio_flags, &prev_em_start); 4197 nr = 0; 4198 } 4199 if (nr) 4200 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached, 4201 &bio, 0, &bio_flags, &prev_em_start); 4202 4203 if (em_cached) 4204 free_extent_map(em_cached); 4205 4206 BUG_ON(!list_empty(pages)); 4207 if (bio) 4208 return submit_one_bio(bio, 0, bio_flags); 4209 return 0; 4210 } 4211 4212 /* 4213 * basic invalidatepage code, this waits on any locked or writeback 4214 * ranges corresponding to the page, and then deletes any extent state 4215 * records from the tree 4216 */ 4217 int extent_invalidatepage(struct extent_io_tree *tree, 4218 struct page *page, unsigned long offset) 4219 { 4220 struct extent_state *cached_state = NULL; 4221 u64 start = page_offset(page); 4222 u64 end = start + PAGE_SIZE - 1; 4223 size_t blocksize = page->mapping->host->i_sb->s_blocksize; 4224 4225 start += ALIGN(offset, blocksize); 4226 if (start > end) 4227 return 0; 4228 4229 lock_extent_bits(tree, start, end, &cached_state); 4230 wait_on_page_writeback(page); 4231 clear_extent_bit(tree, start, end, 4232 EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC | 4233 EXTENT_DO_ACCOUNTING, 4234 1, 1, &cached_state, GFP_NOFS); 4235 return 0; 4236 } 4237 4238 /* 4239 * a helper for releasepage, this tests for areas of the page that 4240 * are locked or under IO and drops the related state bits if it is safe 4241 * to drop the page. 4242 */ 4243 static int try_release_extent_state(struct extent_map_tree *map, 4244 struct extent_io_tree *tree, 4245 struct page *page, gfp_t mask) 4246 { 4247 u64 start = page_offset(page); 4248 u64 end = start + PAGE_SIZE - 1; 4249 int ret = 1; 4250 4251 if (test_range_bit(tree, start, end, 4252 EXTENT_IOBITS, 0, NULL)) 4253 ret = 0; 4254 else { 4255 /* 4256 * at this point we can safely clear everything except the 4257 * locked bit and the nodatasum bit 4258 */ 4259 ret = clear_extent_bit(tree, start, end, 4260 ~(EXTENT_LOCKED | EXTENT_NODATASUM), 4261 0, 0, NULL, mask); 4262 4263 /* if clear_extent_bit failed for enomem reasons, 4264 * we can't allow the release to continue. 4265 */ 4266 if (ret < 0) 4267 ret = 0; 4268 else 4269 ret = 1; 4270 } 4271 return ret; 4272 } 4273 4274 /* 4275 * a helper for releasepage. As long as there are no locked extents 4276 * in the range corresponding to the page, both state records and extent 4277 * map records are removed 4278 */ 4279 int try_release_extent_mapping(struct extent_map_tree *map, 4280 struct extent_io_tree *tree, struct page *page, 4281 gfp_t mask) 4282 { 4283 struct extent_map *em; 4284 u64 start = page_offset(page); 4285 u64 end = start + PAGE_SIZE - 1; 4286 4287 if (gfpflags_allow_blocking(mask) && 4288 page->mapping->host->i_size > SZ_16M) { 4289 u64 len; 4290 while (start <= end) { 4291 len = end - start + 1; 4292 write_lock(&map->lock); 4293 em = lookup_extent_mapping(map, start, len); 4294 if (!em) { 4295 write_unlock(&map->lock); 4296 break; 4297 } 4298 if (test_bit(EXTENT_FLAG_PINNED, &em->flags) || 4299 em->start != start) { 4300 write_unlock(&map->lock); 4301 free_extent_map(em); 4302 break; 4303 } 4304 if (!test_range_bit(tree, em->start, 4305 extent_map_end(em) - 1, 4306 EXTENT_LOCKED | EXTENT_WRITEBACK, 4307 0, NULL)) { 4308 remove_extent_mapping(map, em); 4309 /* once for the rb tree */ 4310 free_extent_map(em); 4311 } 4312 start = extent_map_end(em); 4313 write_unlock(&map->lock); 4314 4315 /* once for us */ 4316 free_extent_map(em); 4317 } 4318 } 4319 return try_release_extent_state(map, tree, page, mask); 4320 } 4321 4322 /* 4323 * helper function for fiemap, which doesn't want to see any holes. 4324 * This maps until we find something past 'last' 4325 */ 4326 static struct extent_map *get_extent_skip_holes(struct inode *inode, 4327 u64 offset, 4328 u64 last, 4329 get_extent_t *get_extent) 4330 { 4331 u64 sectorsize = btrfs_inode_sectorsize(inode); 4332 struct extent_map *em; 4333 u64 len; 4334 4335 if (offset >= last) 4336 return NULL; 4337 4338 while (1) { 4339 len = last - offset; 4340 if (len == 0) 4341 break; 4342 len = ALIGN(len, sectorsize); 4343 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0); 4344 if (IS_ERR_OR_NULL(em)) 4345 return em; 4346 4347 /* if this isn't a hole return it */ 4348 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) && 4349 em->block_start != EXTENT_MAP_HOLE) { 4350 return em; 4351 } 4352 4353 /* this is a hole, advance to the next extent */ 4354 offset = extent_map_end(em); 4355 free_extent_map(em); 4356 if (offset >= last) 4357 break; 4358 } 4359 return NULL; 4360 } 4361 4362 /* 4363 * To cache previous fiemap extent 4364 * 4365 * Will be used for merging fiemap extent 4366 */ 4367 struct fiemap_cache { 4368 u64 offset; 4369 u64 phys; 4370 u64 len; 4371 u32 flags; 4372 bool cached; 4373 }; 4374 4375 /* 4376 * Helper to submit fiemap extent. 4377 * 4378 * Will try to merge current fiemap extent specified by @offset, @phys, 4379 * @len and @flags with cached one. 4380 * And only when we fails to merge, cached one will be submitted as 4381 * fiemap extent. 4382 * 4383 * Return value is the same as fiemap_fill_next_extent(). 4384 */ 4385 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo, 4386 struct fiemap_cache *cache, 4387 u64 offset, u64 phys, u64 len, u32 flags) 4388 { 4389 int ret = 0; 4390 4391 if (!cache->cached) 4392 goto assign; 4393 4394 /* 4395 * Sanity check, extent_fiemap() should have ensured that new 4396 * fiemap extent won't overlap with cahced one. 4397 * Not recoverable. 4398 * 4399 * NOTE: Physical address can overlap, due to compression 4400 */ 4401 if (cache->offset + cache->len > offset) { 4402 WARN_ON(1); 4403 return -EINVAL; 4404 } 4405 4406 /* 4407 * Only merges fiemap extents if 4408 * 1) Their logical addresses are continuous 4409 * 4410 * 2) Their physical addresses are continuous 4411 * So truly compressed (physical size smaller than logical size) 4412 * extents won't get merged with each other 4413 * 4414 * 3) Share same flags except FIEMAP_EXTENT_LAST 4415 * So regular extent won't get merged with prealloc extent 4416 */ 4417 if (cache->offset + cache->len == offset && 4418 cache->phys + cache->len == phys && 4419 (cache->flags & ~FIEMAP_EXTENT_LAST) == 4420 (flags & ~FIEMAP_EXTENT_LAST)) { 4421 cache->len += len; 4422 cache->flags |= flags; 4423 goto try_submit_last; 4424 } 4425 4426 /* Not mergeable, need to submit cached one */ 4427 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys, 4428 cache->len, cache->flags); 4429 cache->cached = false; 4430 if (ret) 4431 return ret; 4432 assign: 4433 cache->cached = true; 4434 cache->offset = offset; 4435 cache->phys = phys; 4436 cache->len = len; 4437 cache->flags = flags; 4438 try_submit_last: 4439 if (cache->flags & FIEMAP_EXTENT_LAST) { 4440 ret = fiemap_fill_next_extent(fieinfo, cache->offset, 4441 cache->phys, cache->len, cache->flags); 4442 cache->cached = false; 4443 } 4444 return ret; 4445 } 4446 4447 /* 4448 * Emit last fiemap cache 4449 * 4450 * The last fiemap cache may still be cached in the following case: 4451 * 0 4k 8k 4452 * |<- Fiemap range ->| 4453 * |<------------ First extent ----------->| 4454 * 4455 * In this case, the first extent range will be cached but not emitted. 4456 * So we must emit it before ending extent_fiemap(). 4457 */ 4458 static int emit_last_fiemap_cache(struct btrfs_fs_info *fs_info, 4459 struct fiemap_extent_info *fieinfo, 4460 struct fiemap_cache *cache) 4461 { 4462 int ret; 4463 4464 if (!cache->cached) 4465 return 0; 4466 4467 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys, 4468 cache->len, cache->flags); 4469 cache->cached = false; 4470 if (ret > 0) 4471 ret = 0; 4472 return ret; 4473 } 4474 4475 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo, 4476 __u64 start, __u64 len, get_extent_t *get_extent) 4477 { 4478 int ret = 0; 4479 u64 off = start; 4480 u64 max = start + len; 4481 u32 flags = 0; 4482 u32 found_type; 4483 u64 last; 4484 u64 last_for_get_extent = 0; 4485 u64 disko = 0; 4486 u64 isize = i_size_read(inode); 4487 struct btrfs_key found_key; 4488 struct extent_map *em = NULL; 4489 struct extent_state *cached_state = NULL; 4490 struct btrfs_path *path; 4491 struct btrfs_root *root = BTRFS_I(inode)->root; 4492 struct fiemap_cache cache = { 0 }; 4493 int end = 0; 4494 u64 em_start = 0; 4495 u64 em_len = 0; 4496 u64 em_end = 0; 4497 4498 if (len == 0) 4499 return -EINVAL; 4500 4501 path = btrfs_alloc_path(); 4502 if (!path) 4503 return -ENOMEM; 4504 path->leave_spinning = 1; 4505 4506 start = round_down(start, btrfs_inode_sectorsize(inode)); 4507 len = round_up(max, btrfs_inode_sectorsize(inode)) - start; 4508 4509 /* 4510 * lookup the last file extent. We're not using i_size here 4511 * because there might be preallocation past i_size 4512 */ 4513 ret = btrfs_lookup_file_extent(NULL, root, path, 4514 btrfs_ino(BTRFS_I(inode)), -1, 0); 4515 if (ret < 0) { 4516 btrfs_free_path(path); 4517 return ret; 4518 } else { 4519 WARN_ON(!ret); 4520 if (ret == 1) 4521 ret = 0; 4522 } 4523 4524 path->slots[0]--; 4525 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]); 4526 found_type = found_key.type; 4527 4528 /* No extents, but there might be delalloc bits */ 4529 if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) || 4530 found_type != BTRFS_EXTENT_DATA_KEY) { 4531 /* have to trust i_size as the end */ 4532 last = (u64)-1; 4533 last_for_get_extent = isize; 4534 } else { 4535 /* 4536 * remember the start of the last extent. There are a 4537 * bunch of different factors that go into the length of the 4538 * extent, so its much less complex to remember where it started 4539 */ 4540 last = found_key.offset; 4541 last_for_get_extent = last + 1; 4542 } 4543 btrfs_release_path(path); 4544 4545 /* 4546 * we might have some extents allocated but more delalloc past those 4547 * extents. so, we trust isize unless the start of the last extent is 4548 * beyond isize 4549 */ 4550 if (last < isize) { 4551 last = (u64)-1; 4552 last_for_get_extent = isize; 4553 } 4554 4555 lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 4556 &cached_state); 4557 4558 em = get_extent_skip_holes(inode, start, last_for_get_extent, 4559 get_extent); 4560 if (!em) 4561 goto out; 4562 if (IS_ERR(em)) { 4563 ret = PTR_ERR(em); 4564 goto out; 4565 } 4566 4567 while (!end) { 4568 u64 offset_in_extent = 0; 4569 4570 /* break if the extent we found is outside the range */ 4571 if (em->start >= max || extent_map_end(em) < off) 4572 break; 4573 4574 /* 4575 * get_extent may return an extent that starts before our 4576 * requested range. We have to make sure the ranges 4577 * we return to fiemap always move forward and don't 4578 * overlap, so adjust the offsets here 4579 */ 4580 em_start = max(em->start, off); 4581 4582 /* 4583 * record the offset from the start of the extent 4584 * for adjusting the disk offset below. Only do this if the 4585 * extent isn't compressed since our in ram offset may be past 4586 * what we have actually allocated on disk. 4587 */ 4588 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) 4589 offset_in_extent = em_start - em->start; 4590 em_end = extent_map_end(em); 4591 em_len = em_end - em_start; 4592 disko = 0; 4593 flags = 0; 4594 4595 /* 4596 * bump off for our next call to get_extent 4597 */ 4598 off = extent_map_end(em); 4599 if (off >= max) 4600 end = 1; 4601 4602 if (em->block_start == EXTENT_MAP_LAST_BYTE) { 4603 end = 1; 4604 flags |= FIEMAP_EXTENT_LAST; 4605 } else if (em->block_start == EXTENT_MAP_INLINE) { 4606 flags |= (FIEMAP_EXTENT_DATA_INLINE | 4607 FIEMAP_EXTENT_NOT_ALIGNED); 4608 } else if (em->block_start == EXTENT_MAP_DELALLOC) { 4609 flags |= (FIEMAP_EXTENT_DELALLOC | 4610 FIEMAP_EXTENT_UNKNOWN); 4611 } else if (fieinfo->fi_extents_max) { 4612 u64 bytenr = em->block_start - 4613 (em->start - em->orig_start); 4614 4615 disko = em->block_start + offset_in_extent; 4616 4617 /* 4618 * As btrfs supports shared space, this information 4619 * can be exported to userspace tools via 4620 * flag FIEMAP_EXTENT_SHARED. If fi_extents_max == 0 4621 * then we're just getting a count and we can skip the 4622 * lookup stuff. 4623 */ 4624 ret = btrfs_check_shared(root, 4625 btrfs_ino(BTRFS_I(inode)), 4626 bytenr); 4627 if (ret < 0) 4628 goto out_free; 4629 if (ret) 4630 flags |= FIEMAP_EXTENT_SHARED; 4631 ret = 0; 4632 } 4633 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) 4634 flags |= FIEMAP_EXTENT_ENCODED; 4635 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) 4636 flags |= FIEMAP_EXTENT_UNWRITTEN; 4637 4638 free_extent_map(em); 4639 em = NULL; 4640 if ((em_start >= last) || em_len == (u64)-1 || 4641 (last == (u64)-1 && isize <= em_end)) { 4642 flags |= FIEMAP_EXTENT_LAST; 4643 end = 1; 4644 } 4645 4646 /* now scan forward to see if this is really the last extent. */ 4647 em = get_extent_skip_holes(inode, off, last_for_get_extent, 4648 get_extent); 4649 if (IS_ERR(em)) { 4650 ret = PTR_ERR(em); 4651 goto out; 4652 } 4653 if (!em) { 4654 flags |= FIEMAP_EXTENT_LAST; 4655 end = 1; 4656 } 4657 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko, 4658 em_len, flags); 4659 if (ret) { 4660 if (ret == 1) 4661 ret = 0; 4662 goto out_free; 4663 } 4664 } 4665 out_free: 4666 if (!ret) 4667 ret = emit_last_fiemap_cache(root->fs_info, fieinfo, &cache); 4668 free_extent_map(em); 4669 out: 4670 btrfs_free_path(path); 4671 unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1, 4672 &cached_state, GFP_NOFS); 4673 return ret; 4674 } 4675 4676 static void __free_extent_buffer(struct extent_buffer *eb) 4677 { 4678 btrfs_leak_debug_del(&eb->leak_list); 4679 kmem_cache_free(extent_buffer_cache, eb); 4680 } 4681 4682 int extent_buffer_under_io(struct extent_buffer *eb) 4683 { 4684 return (atomic_read(&eb->io_pages) || 4685 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) || 4686 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)); 4687 } 4688 4689 /* 4690 * Helper for releasing extent buffer page. 4691 */ 4692 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb) 4693 { 4694 unsigned long index; 4695 struct page *page; 4696 int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags); 4697 4698 BUG_ON(extent_buffer_under_io(eb)); 4699 4700 index = num_extent_pages(eb->start, eb->len); 4701 if (index == 0) 4702 return; 4703 4704 do { 4705 index--; 4706 page = eb->pages[index]; 4707 if (!page) 4708 continue; 4709 if (mapped) 4710 spin_lock(&page->mapping->private_lock); 4711 /* 4712 * We do this since we'll remove the pages after we've 4713 * removed the eb from the radix tree, so we could race 4714 * and have this page now attached to the new eb. So 4715 * only clear page_private if it's still connected to 4716 * this eb. 4717 */ 4718 if (PagePrivate(page) && 4719 page->private == (unsigned long)eb) { 4720 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)); 4721 BUG_ON(PageDirty(page)); 4722 BUG_ON(PageWriteback(page)); 4723 /* 4724 * We need to make sure we haven't be attached 4725 * to a new eb. 4726 */ 4727 ClearPagePrivate(page); 4728 set_page_private(page, 0); 4729 /* One for the page private */ 4730 put_page(page); 4731 } 4732 4733 if (mapped) 4734 spin_unlock(&page->mapping->private_lock); 4735 4736 /* One for when we allocated the page */ 4737 put_page(page); 4738 } while (index != 0); 4739 } 4740 4741 /* 4742 * Helper for releasing the extent buffer. 4743 */ 4744 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb) 4745 { 4746 btrfs_release_extent_buffer_page(eb); 4747 __free_extent_buffer(eb); 4748 } 4749 4750 static struct extent_buffer * 4751 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start, 4752 unsigned long len) 4753 { 4754 struct extent_buffer *eb = NULL; 4755 4756 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL); 4757 eb->start = start; 4758 eb->len = len; 4759 eb->fs_info = fs_info; 4760 eb->bflags = 0; 4761 rwlock_init(&eb->lock); 4762 atomic_set(&eb->write_locks, 0); 4763 atomic_set(&eb->read_locks, 0); 4764 atomic_set(&eb->blocking_readers, 0); 4765 atomic_set(&eb->blocking_writers, 0); 4766 atomic_set(&eb->spinning_readers, 0); 4767 atomic_set(&eb->spinning_writers, 0); 4768 eb->lock_nested = 0; 4769 init_waitqueue_head(&eb->write_lock_wq); 4770 init_waitqueue_head(&eb->read_lock_wq); 4771 4772 btrfs_leak_debug_add(&eb->leak_list, &buffers); 4773 4774 spin_lock_init(&eb->refs_lock); 4775 atomic_set(&eb->refs, 1); 4776 atomic_set(&eb->io_pages, 0); 4777 4778 /* 4779 * Sanity checks, currently the maximum is 64k covered by 16x 4k pages 4780 */ 4781 BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE 4782 > MAX_INLINE_EXTENT_BUFFER_SIZE); 4783 BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE); 4784 4785 return eb; 4786 } 4787 4788 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src) 4789 { 4790 unsigned long i; 4791 struct page *p; 4792 struct extent_buffer *new; 4793 unsigned long num_pages = num_extent_pages(src->start, src->len); 4794 4795 new = __alloc_extent_buffer(src->fs_info, src->start, src->len); 4796 if (new == NULL) 4797 return NULL; 4798 4799 for (i = 0; i < num_pages; i++) { 4800 p = alloc_page(GFP_NOFS); 4801 if (!p) { 4802 btrfs_release_extent_buffer(new); 4803 return NULL; 4804 } 4805 attach_extent_buffer_page(new, p); 4806 WARN_ON(PageDirty(p)); 4807 SetPageUptodate(p); 4808 new->pages[i] = p; 4809 copy_page(page_address(p), page_address(src->pages[i])); 4810 } 4811 4812 set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags); 4813 set_bit(EXTENT_BUFFER_DUMMY, &new->bflags); 4814 4815 return new; 4816 } 4817 4818 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info, 4819 u64 start, unsigned long len) 4820 { 4821 struct extent_buffer *eb; 4822 unsigned long num_pages; 4823 unsigned long i; 4824 4825 num_pages = num_extent_pages(start, len); 4826 4827 eb = __alloc_extent_buffer(fs_info, start, len); 4828 if (!eb) 4829 return NULL; 4830 4831 for (i = 0; i < num_pages; i++) { 4832 eb->pages[i] = alloc_page(GFP_NOFS); 4833 if (!eb->pages[i]) 4834 goto err; 4835 } 4836 set_extent_buffer_uptodate(eb); 4837 btrfs_set_header_nritems(eb, 0); 4838 set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags); 4839 4840 return eb; 4841 err: 4842 for (; i > 0; i--) 4843 __free_page(eb->pages[i - 1]); 4844 __free_extent_buffer(eb); 4845 return NULL; 4846 } 4847 4848 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info, 4849 u64 start) 4850 { 4851 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize); 4852 } 4853 4854 static void check_buffer_tree_ref(struct extent_buffer *eb) 4855 { 4856 int refs; 4857 /* the ref bit is tricky. We have to make sure it is set 4858 * if we have the buffer dirty. Otherwise the 4859 * code to free a buffer can end up dropping a dirty 4860 * page 4861 * 4862 * Once the ref bit is set, it won't go away while the 4863 * buffer is dirty or in writeback, and it also won't 4864 * go away while we have the reference count on the 4865 * eb bumped. 4866 * 4867 * We can't just set the ref bit without bumping the 4868 * ref on the eb because free_extent_buffer might 4869 * see the ref bit and try to clear it. If this happens 4870 * free_extent_buffer might end up dropping our original 4871 * ref by mistake and freeing the page before we are able 4872 * to add one more ref. 4873 * 4874 * So bump the ref count first, then set the bit. If someone 4875 * beat us to it, drop the ref we added. 4876 */ 4877 refs = atomic_read(&eb->refs); 4878 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) 4879 return; 4880 4881 spin_lock(&eb->refs_lock); 4882 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) 4883 atomic_inc(&eb->refs); 4884 spin_unlock(&eb->refs_lock); 4885 } 4886 4887 static void mark_extent_buffer_accessed(struct extent_buffer *eb, 4888 struct page *accessed) 4889 { 4890 unsigned long num_pages, i; 4891 4892 check_buffer_tree_ref(eb); 4893 4894 num_pages = num_extent_pages(eb->start, eb->len); 4895 for (i = 0; i < num_pages; i++) { 4896 struct page *p = eb->pages[i]; 4897 4898 if (p != accessed) 4899 mark_page_accessed(p); 4900 } 4901 } 4902 4903 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info, 4904 u64 start) 4905 { 4906 struct extent_buffer *eb; 4907 4908 rcu_read_lock(); 4909 eb = radix_tree_lookup(&fs_info->buffer_radix, 4910 start >> PAGE_SHIFT); 4911 if (eb && atomic_inc_not_zero(&eb->refs)) { 4912 rcu_read_unlock(); 4913 /* 4914 * Lock our eb's refs_lock to avoid races with 4915 * free_extent_buffer. When we get our eb it might be flagged 4916 * with EXTENT_BUFFER_STALE and another task running 4917 * free_extent_buffer might have seen that flag set, 4918 * eb->refs == 2, that the buffer isn't under IO (dirty and 4919 * writeback flags not set) and it's still in the tree (flag 4920 * EXTENT_BUFFER_TREE_REF set), therefore being in the process 4921 * of decrementing the extent buffer's reference count twice. 4922 * So here we could race and increment the eb's reference count, 4923 * clear its stale flag, mark it as dirty and drop our reference 4924 * before the other task finishes executing free_extent_buffer, 4925 * which would later result in an attempt to free an extent 4926 * buffer that is dirty. 4927 */ 4928 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) { 4929 spin_lock(&eb->refs_lock); 4930 spin_unlock(&eb->refs_lock); 4931 } 4932 mark_extent_buffer_accessed(eb, NULL); 4933 return eb; 4934 } 4935 rcu_read_unlock(); 4936 4937 return NULL; 4938 } 4939 4940 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS 4941 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info, 4942 u64 start) 4943 { 4944 struct extent_buffer *eb, *exists = NULL; 4945 int ret; 4946 4947 eb = find_extent_buffer(fs_info, start); 4948 if (eb) 4949 return eb; 4950 eb = alloc_dummy_extent_buffer(fs_info, start); 4951 if (!eb) 4952 return NULL; 4953 eb->fs_info = fs_info; 4954 again: 4955 ret = radix_tree_preload(GFP_NOFS); 4956 if (ret) 4957 goto free_eb; 4958 spin_lock(&fs_info->buffer_lock); 4959 ret = radix_tree_insert(&fs_info->buffer_radix, 4960 start >> PAGE_SHIFT, eb); 4961 spin_unlock(&fs_info->buffer_lock); 4962 radix_tree_preload_end(); 4963 if (ret == -EEXIST) { 4964 exists = find_extent_buffer(fs_info, start); 4965 if (exists) 4966 goto free_eb; 4967 else 4968 goto again; 4969 } 4970 check_buffer_tree_ref(eb); 4971 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags); 4972 4973 /* 4974 * We will free dummy extent buffer's if they come into 4975 * free_extent_buffer with a ref count of 2, but if we are using this we 4976 * want the buffers to stay in memory until we're done with them, so 4977 * bump the ref count again. 4978 */ 4979 atomic_inc(&eb->refs); 4980 return eb; 4981 free_eb: 4982 btrfs_release_extent_buffer(eb); 4983 return exists; 4984 } 4985 #endif 4986 4987 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info, 4988 u64 start) 4989 { 4990 unsigned long len = fs_info->nodesize; 4991 unsigned long num_pages = num_extent_pages(start, len); 4992 unsigned long i; 4993 unsigned long index = start >> PAGE_SHIFT; 4994 struct extent_buffer *eb; 4995 struct extent_buffer *exists = NULL; 4996 struct page *p; 4997 struct address_space *mapping = fs_info->btree_inode->i_mapping; 4998 int uptodate = 1; 4999 int ret; 5000 5001 if (!IS_ALIGNED(start, fs_info->sectorsize)) { 5002 btrfs_err(fs_info, "bad tree block start %llu", start); 5003 return ERR_PTR(-EINVAL); 5004 } 5005 5006 eb = find_extent_buffer(fs_info, start); 5007 if (eb) 5008 return eb; 5009 5010 eb = __alloc_extent_buffer(fs_info, start, len); 5011 if (!eb) 5012 return ERR_PTR(-ENOMEM); 5013 5014 for (i = 0; i < num_pages; i++, index++) { 5015 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL); 5016 if (!p) { 5017 exists = ERR_PTR(-ENOMEM); 5018 goto free_eb; 5019 } 5020 5021 spin_lock(&mapping->private_lock); 5022 if (PagePrivate(p)) { 5023 /* 5024 * We could have already allocated an eb for this page 5025 * and attached one so lets see if we can get a ref on 5026 * the existing eb, and if we can we know it's good and 5027 * we can just return that one, else we know we can just 5028 * overwrite page->private. 5029 */ 5030 exists = (struct extent_buffer *)p->private; 5031 if (atomic_inc_not_zero(&exists->refs)) { 5032 spin_unlock(&mapping->private_lock); 5033 unlock_page(p); 5034 put_page(p); 5035 mark_extent_buffer_accessed(exists, p); 5036 goto free_eb; 5037 } 5038 exists = NULL; 5039 5040 /* 5041 * Do this so attach doesn't complain and we need to 5042 * drop the ref the old guy had. 5043 */ 5044 ClearPagePrivate(p); 5045 WARN_ON(PageDirty(p)); 5046 put_page(p); 5047 } 5048 attach_extent_buffer_page(eb, p); 5049 spin_unlock(&mapping->private_lock); 5050 WARN_ON(PageDirty(p)); 5051 eb->pages[i] = p; 5052 if (!PageUptodate(p)) 5053 uptodate = 0; 5054 5055 /* 5056 * see below about how we avoid a nasty race with release page 5057 * and why we unlock later 5058 */ 5059 } 5060 if (uptodate) 5061 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags); 5062 again: 5063 ret = radix_tree_preload(GFP_NOFS); 5064 if (ret) { 5065 exists = ERR_PTR(ret); 5066 goto free_eb; 5067 } 5068 5069 spin_lock(&fs_info->buffer_lock); 5070 ret = radix_tree_insert(&fs_info->buffer_radix, 5071 start >> PAGE_SHIFT, eb); 5072 spin_unlock(&fs_info->buffer_lock); 5073 radix_tree_preload_end(); 5074 if (ret == -EEXIST) { 5075 exists = find_extent_buffer(fs_info, start); 5076 if (exists) 5077 goto free_eb; 5078 else 5079 goto again; 5080 } 5081 /* add one reference for the tree */ 5082 check_buffer_tree_ref(eb); 5083 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags); 5084 5085 /* 5086 * there is a race where release page may have 5087 * tried to find this extent buffer in the radix 5088 * but failed. It will tell the VM it is safe to 5089 * reclaim the, and it will clear the page private bit. 5090 * We must make sure to set the page private bit properly 5091 * after the extent buffer is in the radix tree so 5092 * it doesn't get lost 5093 */ 5094 SetPageChecked(eb->pages[0]); 5095 for (i = 1; i < num_pages; i++) { 5096 p = eb->pages[i]; 5097 ClearPageChecked(p); 5098 unlock_page(p); 5099 } 5100 unlock_page(eb->pages[0]); 5101 return eb; 5102 5103 free_eb: 5104 WARN_ON(!atomic_dec_and_test(&eb->refs)); 5105 for (i = 0; i < num_pages; i++) { 5106 if (eb->pages[i]) 5107 unlock_page(eb->pages[i]); 5108 } 5109 5110 btrfs_release_extent_buffer(eb); 5111 return exists; 5112 } 5113 5114 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head) 5115 { 5116 struct extent_buffer *eb = 5117 container_of(head, struct extent_buffer, rcu_head); 5118 5119 __free_extent_buffer(eb); 5120 } 5121 5122 /* Expects to have eb->eb_lock already held */ 5123 static int release_extent_buffer(struct extent_buffer *eb) 5124 { 5125 WARN_ON(atomic_read(&eb->refs) == 0); 5126 if (atomic_dec_and_test(&eb->refs)) { 5127 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) { 5128 struct btrfs_fs_info *fs_info = eb->fs_info; 5129 5130 spin_unlock(&eb->refs_lock); 5131 5132 spin_lock(&fs_info->buffer_lock); 5133 radix_tree_delete(&fs_info->buffer_radix, 5134 eb->start >> PAGE_SHIFT); 5135 spin_unlock(&fs_info->buffer_lock); 5136 } else { 5137 spin_unlock(&eb->refs_lock); 5138 } 5139 5140 /* Should be safe to release our pages at this point */ 5141 btrfs_release_extent_buffer_page(eb); 5142 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS 5143 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) { 5144 __free_extent_buffer(eb); 5145 return 1; 5146 } 5147 #endif 5148 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu); 5149 return 1; 5150 } 5151 spin_unlock(&eb->refs_lock); 5152 5153 return 0; 5154 } 5155 5156 void free_extent_buffer(struct extent_buffer *eb) 5157 { 5158 int refs; 5159 int old; 5160 if (!eb) 5161 return; 5162 5163 while (1) { 5164 refs = atomic_read(&eb->refs); 5165 if (refs <= 3) 5166 break; 5167 old = atomic_cmpxchg(&eb->refs, refs, refs - 1); 5168 if (old == refs) 5169 return; 5170 } 5171 5172 spin_lock(&eb->refs_lock); 5173 if (atomic_read(&eb->refs) == 2 && 5174 test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags)) 5175 atomic_dec(&eb->refs); 5176 5177 if (atomic_read(&eb->refs) == 2 && 5178 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) && 5179 !extent_buffer_under_io(eb) && 5180 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) 5181 atomic_dec(&eb->refs); 5182 5183 /* 5184 * I know this is terrible, but it's temporary until we stop tracking 5185 * the uptodate bits and such for the extent buffers. 5186 */ 5187 release_extent_buffer(eb); 5188 } 5189 5190 void free_extent_buffer_stale(struct extent_buffer *eb) 5191 { 5192 if (!eb) 5193 return; 5194 5195 spin_lock(&eb->refs_lock); 5196 set_bit(EXTENT_BUFFER_STALE, &eb->bflags); 5197 5198 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) && 5199 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) 5200 atomic_dec(&eb->refs); 5201 release_extent_buffer(eb); 5202 } 5203 5204 void clear_extent_buffer_dirty(struct extent_buffer *eb) 5205 { 5206 unsigned long i; 5207 unsigned long num_pages; 5208 struct page *page; 5209 5210 num_pages = num_extent_pages(eb->start, eb->len); 5211 5212 for (i = 0; i < num_pages; i++) { 5213 page = eb->pages[i]; 5214 if (!PageDirty(page)) 5215 continue; 5216 5217 lock_page(page); 5218 WARN_ON(!PagePrivate(page)); 5219 5220 clear_page_dirty_for_io(page); 5221 spin_lock_irq(&page->mapping->tree_lock); 5222 if (!PageDirty(page)) { 5223 radix_tree_tag_clear(&page->mapping->page_tree, 5224 page_index(page), 5225 PAGECACHE_TAG_DIRTY); 5226 } 5227 spin_unlock_irq(&page->mapping->tree_lock); 5228 ClearPageError(page); 5229 unlock_page(page); 5230 } 5231 WARN_ON(atomic_read(&eb->refs) == 0); 5232 } 5233 5234 int set_extent_buffer_dirty(struct extent_buffer *eb) 5235 { 5236 unsigned long i; 5237 unsigned long num_pages; 5238 int was_dirty = 0; 5239 5240 check_buffer_tree_ref(eb); 5241 5242 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags); 5243 5244 num_pages = num_extent_pages(eb->start, eb->len); 5245 WARN_ON(atomic_read(&eb->refs) == 0); 5246 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)); 5247 5248 for (i = 0; i < num_pages; i++) 5249 set_page_dirty(eb->pages[i]); 5250 return was_dirty; 5251 } 5252 5253 void clear_extent_buffer_uptodate(struct extent_buffer *eb) 5254 { 5255 unsigned long i; 5256 struct page *page; 5257 unsigned long num_pages; 5258 5259 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags); 5260 num_pages = num_extent_pages(eb->start, eb->len); 5261 for (i = 0; i < num_pages; i++) { 5262 page = eb->pages[i]; 5263 if (page) 5264 ClearPageUptodate(page); 5265 } 5266 } 5267 5268 void set_extent_buffer_uptodate(struct extent_buffer *eb) 5269 { 5270 unsigned long i; 5271 struct page *page; 5272 unsigned long num_pages; 5273 5274 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags); 5275 num_pages = num_extent_pages(eb->start, eb->len); 5276 for (i = 0; i < num_pages; i++) { 5277 page = eb->pages[i]; 5278 SetPageUptodate(page); 5279 } 5280 } 5281 5282 int extent_buffer_uptodate(struct extent_buffer *eb) 5283 { 5284 return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags); 5285 } 5286 5287 int read_extent_buffer_pages(struct extent_io_tree *tree, 5288 struct extent_buffer *eb, int wait, 5289 get_extent_t *get_extent, int mirror_num) 5290 { 5291 unsigned long i; 5292 struct page *page; 5293 int err; 5294 int ret = 0; 5295 int locked_pages = 0; 5296 int all_uptodate = 1; 5297 unsigned long num_pages; 5298 unsigned long num_reads = 0; 5299 struct bio *bio = NULL; 5300 unsigned long bio_flags = 0; 5301 5302 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags)) 5303 return 0; 5304 5305 num_pages = num_extent_pages(eb->start, eb->len); 5306 for (i = 0; i < num_pages; i++) { 5307 page = eb->pages[i]; 5308 if (wait == WAIT_NONE) { 5309 if (!trylock_page(page)) 5310 goto unlock_exit; 5311 } else { 5312 lock_page(page); 5313 } 5314 locked_pages++; 5315 } 5316 /* 5317 * We need to firstly lock all pages to make sure that 5318 * the uptodate bit of our pages won't be affected by 5319 * clear_extent_buffer_uptodate(). 5320 */ 5321 for (i = 0; i < num_pages; i++) { 5322 page = eb->pages[i]; 5323 if (!PageUptodate(page)) { 5324 num_reads++; 5325 all_uptodate = 0; 5326 } 5327 } 5328 5329 if (all_uptodate) { 5330 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags); 5331 goto unlock_exit; 5332 } 5333 5334 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags); 5335 eb->read_mirror = 0; 5336 atomic_set(&eb->io_pages, num_reads); 5337 for (i = 0; i < num_pages; i++) { 5338 page = eb->pages[i]; 5339 5340 if (!PageUptodate(page)) { 5341 if (ret) { 5342 atomic_dec(&eb->io_pages); 5343 unlock_page(page); 5344 continue; 5345 } 5346 5347 ClearPageError(page); 5348 err = __extent_read_full_page(tree, page, 5349 get_extent, &bio, 5350 mirror_num, &bio_flags, 5351 REQ_META); 5352 if (err) { 5353 ret = err; 5354 /* 5355 * We use &bio in above __extent_read_full_page, 5356 * so we ensure that if it returns error, the 5357 * current page fails to add itself to bio and 5358 * it's been unlocked. 5359 * 5360 * We must dec io_pages by ourselves. 5361 */ 5362 atomic_dec(&eb->io_pages); 5363 } 5364 } else { 5365 unlock_page(page); 5366 } 5367 } 5368 5369 if (bio) { 5370 err = submit_one_bio(bio, mirror_num, bio_flags); 5371 if (err) 5372 return err; 5373 } 5374 5375 if (ret || wait != WAIT_COMPLETE) 5376 return ret; 5377 5378 for (i = 0; i < num_pages; i++) { 5379 page = eb->pages[i]; 5380 wait_on_page_locked(page); 5381 if (!PageUptodate(page)) 5382 ret = -EIO; 5383 } 5384 5385 return ret; 5386 5387 unlock_exit: 5388 while (locked_pages > 0) { 5389 locked_pages--; 5390 page = eb->pages[locked_pages]; 5391 unlock_page(page); 5392 } 5393 return ret; 5394 } 5395 5396 void read_extent_buffer(const struct extent_buffer *eb, void *dstv, 5397 unsigned long start, unsigned long len) 5398 { 5399 size_t cur; 5400 size_t offset; 5401 struct page *page; 5402 char *kaddr; 5403 char *dst = (char *)dstv; 5404 size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1); 5405 unsigned long i = (start_offset + start) >> PAGE_SHIFT; 5406 5407 if (start + len > eb->len) { 5408 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n", 5409 eb->start, eb->len, start, len); 5410 memset(dst, 0, len); 5411 return; 5412 } 5413 5414 offset = (start_offset + start) & (PAGE_SIZE - 1); 5415 5416 while (len > 0) { 5417 page = eb->pages[i]; 5418 5419 cur = min(len, (PAGE_SIZE - offset)); 5420 kaddr = page_address(page); 5421 memcpy(dst, kaddr + offset, cur); 5422 5423 dst += cur; 5424 len -= cur; 5425 offset = 0; 5426 i++; 5427 } 5428 } 5429 5430 int read_extent_buffer_to_user(const struct extent_buffer *eb, 5431 void __user *dstv, 5432 unsigned long start, unsigned long len) 5433 { 5434 size_t cur; 5435 size_t offset; 5436 struct page *page; 5437 char *kaddr; 5438 char __user *dst = (char __user *)dstv; 5439 size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1); 5440 unsigned long i = (start_offset + start) >> PAGE_SHIFT; 5441 int ret = 0; 5442 5443 WARN_ON(start > eb->len); 5444 WARN_ON(start + len > eb->start + eb->len); 5445 5446 offset = (start_offset + start) & (PAGE_SIZE - 1); 5447 5448 while (len > 0) { 5449 page = eb->pages[i]; 5450 5451 cur = min(len, (PAGE_SIZE - offset)); 5452 kaddr = page_address(page); 5453 if (copy_to_user(dst, kaddr + offset, cur)) { 5454 ret = -EFAULT; 5455 break; 5456 } 5457 5458 dst += cur; 5459 len -= cur; 5460 offset = 0; 5461 i++; 5462 } 5463 5464 return ret; 5465 } 5466 5467 /* 5468 * return 0 if the item is found within a page. 5469 * return 1 if the item spans two pages. 5470 * return -EINVAL otherwise. 5471 */ 5472 int map_private_extent_buffer(const struct extent_buffer *eb, 5473 unsigned long start, unsigned long min_len, 5474 char **map, unsigned long *map_start, 5475 unsigned long *map_len) 5476 { 5477 size_t offset = start & (PAGE_SIZE - 1); 5478 char *kaddr; 5479 struct page *p; 5480 size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1); 5481 unsigned long i = (start_offset + start) >> PAGE_SHIFT; 5482 unsigned long end_i = (start_offset + start + min_len - 1) >> 5483 PAGE_SHIFT; 5484 5485 if (start + min_len > eb->len) { 5486 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n", 5487 eb->start, eb->len, start, min_len); 5488 return -EINVAL; 5489 } 5490 5491 if (i != end_i) 5492 return 1; 5493 5494 if (i == 0) { 5495 offset = start_offset; 5496 *map_start = 0; 5497 } else { 5498 offset = 0; 5499 *map_start = ((u64)i << PAGE_SHIFT) - start_offset; 5500 } 5501 5502 p = eb->pages[i]; 5503 kaddr = page_address(p); 5504 *map = kaddr + offset; 5505 *map_len = PAGE_SIZE - offset; 5506 return 0; 5507 } 5508 5509 int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv, 5510 unsigned long start, unsigned long len) 5511 { 5512 size_t cur; 5513 size_t offset; 5514 struct page *page; 5515 char *kaddr; 5516 char *ptr = (char *)ptrv; 5517 size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1); 5518 unsigned long i = (start_offset + start) >> PAGE_SHIFT; 5519 int ret = 0; 5520 5521 WARN_ON(start > eb->len); 5522 WARN_ON(start + len > eb->start + eb->len); 5523 5524 offset = (start_offset + start) & (PAGE_SIZE - 1); 5525 5526 while (len > 0) { 5527 page = eb->pages[i]; 5528 5529 cur = min(len, (PAGE_SIZE - offset)); 5530 5531 kaddr = page_address(page); 5532 ret = memcmp(ptr, kaddr + offset, cur); 5533 if (ret) 5534 break; 5535 5536 ptr += cur; 5537 len -= cur; 5538 offset = 0; 5539 i++; 5540 } 5541 return ret; 5542 } 5543 5544 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb, 5545 const void *srcv) 5546 { 5547 char *kaddr; 5548 5549 WARN_ON(!PageUptodate(eb->pages[0])); 5550 kaddr = page_address(eb->pages[0]); 5551 memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv, 5552 BTRFS_FSID_SIZE); 5553 } 5554 5555 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv) 5556 { 5557 char *kaddr; 5558 5559 WARN_ON(!PageUptodate(eb->pages[0])); 5560 kaddr = page_address(eb->pages[0]); 5561 memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv, 5562 BTRFS_FSID_SIZE); 5563 } 5564 5565 void write_extent_buffer(struct extent_buffer *eb, const void *srcv, 5566 unsigned long start, unsigned long len) 5567 { 5568 size_t cur; 5569 size_t offset; 5570 struct page *page; 5571 char *kaddr; 5572 char *src = (char *)srcv; 5573 size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1); 5574 unsigned long i = (start_offset + start) >> PAGE_SHIFT; 5575 5576 WARN_ON(start > eb->len); 5577 WARN_ON(start + len > eb->start + eb->len); 5578 5579 offset = (start_offset + start) & (PAGE_SIZE - 1); 5580 5581 while (len > 0) { 5582 page = eb->pages[i]; 5583 WARN_ON(!PageUptodate(page)); 5584 5585 cur = min(len, PAGE_SIZE - offset); 5586 kaddr = page_address(page); 5587 memcpy(kaddr + offset, src, cur); 5588 5589 src += cur; 5590 len -= cur; 5591 offset = 0; 5592 i++; 5593 } 5594 } 5595 5596 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start, 5597 unsigned long len) 5598 { 5599 size_t cur; 5600 size_t offset; 5601 struct page *page; 5602 char *kaddr; 5603 size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1); 5604 unsigned long i = (start_offset + start) >> PAGE_SHIFT; 5605 5606 WARN_ON(start > eb->len); 5607 WARN_ON(start + len > eb->start + eb->len); 5608 5609 offset = (start_offset + start) & (PAGE_SIZE - 1); 5610 5611 while (len > 0) { 5612 page = eb->pages[i]; 5613 WARN_ON(!PageUptodate(page)); 5614 5615 cur = min(len, PAGE_SIZE - offset); 5616 kaddr = page_address(page); 5617 memset(kaddr + offset, 0, cur); 5618 5619 len -= cur; 5620 offset = 0; 5621 i++; 5622 } 5623 } 5624 5625 void copy_extent_buffer_full(struct extent_buffer *dst, 5626 struct extent_buffer *src) 5627 { 5628 int i; 5629 unsigned num_pages; 5630 5631 ASSERT(dst->len == src->len); 5632 5633 num_pages = num_extent_pages(dst->start, dst->len); 5634 for (i = 0; i < num_pages; i++) 5635 copy_page(page_address(dst->pages[i]), 5636 page_address(src->pages[i])); 5637 } 5638 5639 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src, 5640 unsigned long dst_offset, unsigned long src_offset, 5641 unsigned long len) 5642 { 5643 u64 dst_len = dst->len; 5644 size_t cur; 5645 size_t offset; 5646 struct page *page; 5647 char *kaddr; 5648 size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1); 5649 unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT; 5650 5651 WARN_ON(src->len != dst_len); 5652 5653 offset = (start_offset + dst_offset) & 5654 (PAGE_SIZE - 1); 5655 5656 while (len > 0) { 5657 page = dst->pages[i]; 5658 WARN_ON(!PageUptodate(page)); 5659 5660 cur = min(len, (unsigned long)(PAGE_SIZE - offset)); 5661 5662 kaddr = page_address(page); 5663 read_extent_buffer(src, kaddr + offset, src_offset, cur); 5664 5665 src_offset += cur; 5666 len -= cur; 5667 offset = 0; 5668 i++; 5669 } 5670 } 5671 5672 void le_bitmap_set(u8 *map, unsigned int start, int len) 5673 { 5674 u8 *p = map + BIT_BYTE(start); 5675 const unsigned int size = start + len; 5676 int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE); 5677 u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start); 5678 5679 while (len - bits_to_set >= 0) { 5680 *p |= mask_to_set; 5681 len -= bits_to_set; 5682 bits_to_set = BITS_PER_BYTE; 5683 mask_to_set = ~0; 5684 p++; 5685 } 5686 if (len) { 5687 mask_to_set &= BITMAP_LAST_BYTE_MASK(size); 5688 *p |= mask_to_set; 5689 } 5690 } 5691 5692 void le_bitmap_clear(u8 *map, unsigned int start, int len) 5693 { 5694 u8 *p = map + BIT_BYTE(start); 5695 const unsigned int size = start + len; 5696 int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE); 5697 u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start); 5698 5699 while (len - bits_to_clear >= 0) { 5700 *p &= ~mask_to_clear; 5701 len -= bits_to_clear; 5702 bits_to_clear = BITS_PER_BYTE; 5703 mask_to_clear = ~0; 5704 p++; 5705 } 5706 if (len) { 5707 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size); 5708 *p &= ~mask_to_clear; 5709 } 5710 } 5711 5712 /* 5713 * eb_bitmap_offset() - calculate the page and offset of the byte containing the 5714 * given bit number 5715 * @eb: the extent buffer 5716 * @start: offset of the bitmap item in the extent buffer 5717 * @nr: bit number 5718 * @page_index: return index of the page in the extent buffer that contains the 5719 * given bit number 5720 * @page_offset: return offset into the page given by page_index 5721 * 5722 * This helper hides the ugliness of finding the byte in an extent buffer which 5723 * contains a given bit. 5724 */ 5725 static inline void eb_bitmap_offset(struct extent_buffer *eb, 5726 unsigned long start, unsigned long nr, 5727 unsigned long *page_index, 5728 size_t *page_offset) 5729 { 5730 size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1); 5731 size_t byte_offset = BIT_BYTE(nr); 5732 size_t offset; 5733 5734 /* 5735 * The byte we want is the offset of the extent buffer + the offset of 5736 * the bitmap item in the extent buffer + the offset of the byte in the 5737 * bitmap item. 5738 */ 5739 offset = start_offset + start + byte_offset; 5740 5741 *page_index = offset >> PAGE_SHIFT; 5742 *page_offset = offset & (PAGE_SIZE - 1); 5743 } 5744 5745 /** 5746 * extent_buffer_test_bit - determine whether a bit in a bitmap item is set 5747 * @eb: the extent buffer 5748 * @start: offset of the bitmap item in the extent buffer 5749 * @nr: bit number to test 5750 */ 5751 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start, 5752 unsigned long nr) 5753 { 5754 u8 *kaddr; 5755 struct page *page; 5756 unsigned long i; 5757 size_t offset; 5758 5759 eb_bitmap_offset(eb, start, nr, &i, &offset); 5760 page = eb->pages[i]; 5761 WARN_ON(!PageUptodate(page)); 5762 kaddr = page_address(page); 5763 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1))); 5764 } 5765 5766 /** 5767 * extent_buffer_bitmap_set - set an area of a bitmap 5768 * @eb: the extent buffer 5769 * @start: offset of the bitmap item in the extent buffer 5770 * @pos: bit number of the first bit 5771 * @len: number of bits to set 5772 */ 5773 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start, 5774 unsigned long pos, unsigned long len) 5775 { 5776 u8 *kaddr; 5777 struct page *page; 5778 unsigned long i; 5779 size_t offset; 5780 const unsigned int size = pos + len; 5781 int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE); 5782 u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos); 5783 5784 eb_bitmap_offset(eb, start, pos, &i, &offset); 5785 page = eb->pages[i]; 5786 WARN_ON(!PageUptodate(page)); 5787 kaddr = page_address(page); 5788 5789 while (len >= bits_to_set) { 5790 kaddr[offset] |= mask_to_set; 5791 len -= bits_to_set; 5792 bits_to_set = BITS_PER_BYTE; 5793 mask_to_set = ~0; 5794 if (++offset >= PAGE_SIZE && len > 0) { 5795 offset = 0; 5796 page = eb->pages[++i]; 5797 WARN_ON(!PageUptodate(page)); 5798 kaddr = page_address(page); 5799 } 5800 } 5801 if (len) { 5802 mask_to_set &= BITMAP_LAST_BYTE_MASK(size); 5803 kaddr[offset] |= mask_to_set; 5804 } 5805 } 5806 5807 5808 /** 5809 * extent_buffer_bitmap_clear - clear an area of a bitmap 5810 * @eb: the extent buffer 5811 * @start: offset of the bitmap item in the extent buffer 5812 * @pos: bit number of the first bit 5813 * @len: number of bits to clear 5814 */ 5815 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start, 5816 unsigned long pos, unsigned long len) 5817 { 5818 u8 *kaddr; 5819 struct page *page; 5820 unsigned long i; 5821 size_t offset; 5822 const unsigned int size = pos + len; 5823 int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE); 5824 u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos); 5825 5826 eb_bitmap_offset(eb, start, pos, &i, &offset); 5827 page = eb->pages[i]; 5828 WARN_ON(!PageUptodate(page)); 5829 kaddr = page_address(page); 5830 5831 while (len >= bits_to_clear) { 5832 kaddr[offset] &= ~mask_to_clear; 5833 len -= bits_to_clear; 5834 bits_to_clear = BITS_PER_BYTE; 5835 mask_to_clear = ~0; 5836 if (++offset >= PAGE_SIZE && len > 0) { 5837 offset = 0; 5838 page = eb->pages[++i]; 5839 WARN_ON(!PageUptodate(page)); 5840 kaddr = page_address(page); 5841 } 5842 } 5843 if (len) { 5844 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size); 5845 kaddr[offset] &= ~mask_to_clear; 5846 } 5847 } 5848 5849 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len) 5850 { 5851 unsigned long distance = (src > dst) ? src - dst : dst - src; 5852 return distance < len; 5853 } 5854 5855 static void copy_pages(struct page *dst_page, struct page *src_page, 5856 unsigned long dst_off, unsigned long src_off, 5857 unsigned long len) 5858 { 5859 char *dst_kaddr = page_address(dst_page); 5860 char *src_kaddr; 5861 int must_memmove = 0; 5862 5863 if (dst_page != src_page) { 5864 src_kaddr = page_address(src_page); 5865 } else { 5866 src_kaddr = dst_kaddr; 5867 if (areas_overlap(src_off, dst_off, len)) 5868 must_memmove = 1; 5869 } 5870 5871 if (must_memmove) 5872 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len); 5873 else 5874 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len); 5875 } 5876 5877 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset, 5878 unsigned long src_offset, unsigned long len) 5879 { 5880 struct btrfs_fs_info *fs_info = dst->fs_info; 5881 size_t cur; 5882 size_t dst_off_in_page; 5883 size_t src_off_in_page; 5884 size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1); 5885 unsigned long dst_i; 5886 unsigned long src_i; 5887 5888 if (src_offset + len > dst->len) { 5889 btrfs_err(fs_info, 5890 "memmove bogus src_offset %lu move len %lu dst len %lu", 5891 src_offset, len, dst->len); 5892 BUG_ON(1); 5893 } 5894 if (dst_offset + len > dst->len) { 5895 btrfs_err(fs_info, 5896 "memmove bogus dst_offset %lu move len %lu dst len %lu", 5897 dst_offset, len, dst->len); 5898 BUG_ON(1); 5899 } 5900 5901 while (len > 0) { 5902 dst_off_in_page = (start_offset + dst_offset) & 5903 (PAGE_SIZE - 1); 5904 src_off_in_page = (start_offset + src_offset) & 5905 (PAGE_SIZE - 1); 5906 5907 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT; 5908 src_i = (start_offset + src_offset) >> PAGE_SHIFT; 5909 5910 cur = min(len, (unsigned long)(PAGE_SIZE - 5911 src_off_in_page)); 5912 cur = min_t(unsigned long, cur, 5913 (unsigned long)(PAGE_SIZE - dst_off_in_page)); 5914 5915 copy_pages(dst->pages[dst_i], dst->pages[src_i], 5916 dst_off_in_page, src_off_in_page, cur); 5917 5918 src_offset += cur; 5919 dst_offset += cur; 5920 len -= cur; 5921 } 5922 } 5923 5924 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset, 5925 unsigned long src_offset, unsigned long len) 5926 { 5927 struct btrfs_fs_info *fs_info = dst->fs_info; 5928 size_t cur; 5929 size_t dst_off_in_page; 5930 size_t src_off_in_page; 5931 unsigned long dst_end = dst_offset + len - 1; 5932 unsigned long src_end = src_offset + len - 1; 5933 size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1); 5934 unsigned long dst_i; 5935 unsigned long src_i; 5936 5937 if (src_offset + len > dst->len) { 5938 btrfs_err(fs_info, 5939 "memmove bogus src_offset %lu move len %lu len %lu", 5940 src_offset, len, dst->len); 5941 BUG_ON(1); 5942 } 5943 if (dst_offset + len > dst->len) { 5944 btrfs_err(fs_info, 5945 "memmove bogus dst_offset %lu move len %lu len %lu", 5946 dst_offset, len, dst->len); 5947 BUG_ON(1); 5948 } 5949 if (dst_offset < src_offset) { 5950 memcpy_extent_buffer(dst, dst_offset, src_offset, len); 5951 return; 5952 } 5953 while (len > 0) { 5954 dst_i = (start_offset + dst_end) >> PAGE_SHIFT; 5955 src_i = (start_offset + src_end) >> PAGE_SHIFT; 5956 5957 dst_off_in_page = (start_offset + dst_end) & 5958 (PAGE_SIZE - 1); 5959 src_off_in_page = (start_offset + src_end) & 5960 (PAGE_SIZE - 1); 5961 5962 cur = min_t(unsigned long, len, src_off_in_page + 1); 5963 cur = min(cur, dst_off_in_page + 1); 5964 copy_pages(dst->pages[dst_i], dst->pages[src_i], 5965 dst_off_in_page - cur + 1, 5966 src_off_in_page - cur + 1, cur); 5967 5968 dst_end -= cur; 5969 src_end -= cur; 5970 len -= cur; 5971 } 5972 } 5973 5974 int try_release_extent_buffer(struct page *page) 5975 { 5976 struct extent_buffer *eb; 5977 5978 /* 5979 * We need to make sure nobody is attaching this page to an eb right 5980 * now. 5981 */ 5982 spin_lock(&page->mapping->private_lock); 5983 if (!PagePrivate(page)) { 5984 spin_unlock(&page->mapping->private_lock); 5985 return 1; 5986 } 5987 5988 eb = (struct extent_buffer *)page->private; 5989 BUG_ON(!eb); 5990 5991 /* 5992 * This is a little awful but should be ok, we need to make sure that 5993 * the eb doesn't disappear out from under us while we're looking at 5994 * this page. 5995 */ 5996 spin_lock(&eb->refs_lock); 5997 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) { 5998 spin_unlock(&eb->refs_lock); 5999 spin_unlock(&page->mapping->private_lock); 6000 return 0; 6001 } 6002 spin_unlock(&page->mapping->private_lock); 6003 6004 /* 6005 * If tree ref isn't set then we know the ref on this eb is a real ref, 6006 * so just return, this page will likely be freed soon anyway. 6007 */ 6008 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) { 6009 spin_unlock(&eb->refs_lock); 6010 return 0; 6011 } 6012 6013 return release_extent_buffer(eb); 6014 } 6015